commit b2ab3f6079
Author: Gerald Combs <gerald@wireshark.org>
Date:   Wed Nov 18 11:20:07 2026 -0800

    Build: 4.4.20

commit 8aeda3d4b6
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 18 18:27:23 2026 +1000

    WiMAX: Don't register the generic CRC expert info name twice

    The 4.4 backport of "WiMAX: Fix UCD loop, CID limit and several
    decoding errors" registered the new "CRC missing - the frame is too
    short" expert info as wmx.genericCrc.missing, the name of the existing
    "CRC is not included" note. The second registration replaced the first
    in the name lookup, so filtering on wmx.genericCrc.missing no longer
    found the note. Use wmx.genericCrc.missing_mal as master, 5.0 and 4.6
    do.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit d580c7ab70
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 18 14:14:19 2026 +1000

    PROFINET: Fix recursion and several bounds problems in PN-RT/SXP/MRP/DCP

    PN-RT: a reassembled FRAG PDU is handed back to the ethertype table and
    can be PN-RT again; its inner FRAG PDU shares the reassembly ID of the
    same frame, so the same data was dissected again until the layer limit
    was hit. Do not reassemble again while dissecting a reassembled PN-RT
    frame (nesting counter via p_get_proto_depth()). The reassembly key
    shifted the frame number by 2 but the fragment ID has 4 bits, so keys
    of different frames collided; shift by 4. The per-ID start table is now
    only updated on the first pass and the key is kept per frame with
    p_add_proto_data().

    SXP: the per-conversation service type cache was updated on every pass,
    so the Info column of a middle fragment differed between one pass and
    two passes. Update the cache only on the first pass and store the
    resolved service type per PDU for later passes.

    PROFINET: pn_find_aruuid_frame_setup() returned the last list entry
    when nothing matched; return NULL (all callers check for NULL).

    RSI: the unsigned test "length - offset - ... > 0" was always true;
    use signed remaining length. Register the reassembly table in
    proto_register instead of an init routine (RSI and SXP).

    MRP: test the remaining length of the tvb that offset belongs to,
    dissect unknown TLVs from that tvb, add the LinkInfo bitmask to the
    TLV tree, and do not wrap the uint8 length on short TLVs.

    PTCP: an unknown TLV advanced the offset twice and was added to the
    wrong tree.

    DCP: stop the option loop for odd block lengths, reject a too short
    Configuration Domain block, and check the PDU is at least 10 bytes
    before reading the length in the heuristic.

    4.4: Hand-ported. This branch has no SXP dissector, no RSI security
    dissector and no MRP InLink dissector, so those hunks are dropped.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit bba73eced3954556392deb35a3dd179f970efcd9)

commit fd6d45fcfa
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 18 14:12:01 2026 +1000

    PROFINET: Fix crashes, CPU blow-ups and per-pass state in PN-IO

    The Read/Write request and response header blocks, and the RTC1 object
    extraction, compared the stored device and controller MAC addresses
    with memcmp() over six bytes of pinfo->dl_src/dl_dst without checking
    the address type. On a link without Ethernet addresses (for example raw
    IP carrying UDP port 34962) the data pointer is NULL and the second
    packet crashed. The addresses are now only stored and used when both
    are Ethernet addresses, and compared with addresses_equal().

    The PROFIenergy float64 measurement values (Set_Meter and
    Query_Measurement) were dissected with dissect_dcerpc_uint64() into an
    FT_DOUBLE field, which is a dissector bug report (an abort with
    WIRESHARK_ABORT_ON_DISSECTOR_BUG) and reads an uninitialized flag. They
    are now added with proto_tree_add_item().

    With a GSD path configured, resolve_pa_profile_submodule_name() wrote up
    to MAX_NAMELENGTH bytes into a module name pointer that had been
    redirected to the exact-size GSD text, a heap overflow that also
    corrupted the GSD cache. The GSD text is now copied into the
    MAX_NAMELENGTH buffer.

    dissect_block() ignored the offset returned by the block body
    dissectors and advanced by the declared body length, but several
    dissectors (PDevData, IandM5, AssetManagementInfo, the CIM blocks) read
    past their declared body, so the enclosing loop dissected the same bytes
    again. A chain of N nested blocks cost 2^N or Fibonacci(N) block
    dissections. Every block body dissector now gets a tvb subset that ends
    at the declared body end (offsets stay absolute), so nesting is linear.
    The IandM5 and AssetManagementInfo entry loops also stop at the end of
    their data.

    IOCRBlockReq searched the station's IO data object and IOCS lists
    linearly for every entry (and allocated an object even for duplicates),
    quadratic in the number of entries. The lists now have a hash index by
    slot/subslot, and an object is only allocated when it is new.

    The AR list was a GList searched and appended to linearly, quadratic in
    the number of ARs. It is now a wmem_map keyed by the ARUUID.

    dissect_PNIO_C_SDU_RTC1() walked the whole object list for every object
    of every RTC1 frame. It now finds the object by frame offset in a hash
    index that is extended lazily with the entries added to the list. Each
    object count iteration looks up one IO data object and one IOCS object
    instead of scanning the lists, which gives the same result for
    well-formed configurations.

    The AR records were modified on every pass, and the "changed from"
    expert infos and the ARUUID/MAC information shown for a frame read the
    state left by all frames, so -2 and the GUI showed different PN-IO data
    than the first pass. The records are now only updated while the frame
    has not been visited. pnio_ar_info() keeps a per-frame copy of the AR
    state from the first pass and shows it on later passes, the "changed
    from" warnings are remembered per frame and shown again, and an AR is
    not found by frames before the one that created it. The conversation
    filter data now refers to the per-frame copy and is set instead of
    being prepended to file scope data on every pass.

    The RTC1 dissector no longer rewrites the PROFIsafe toggle bit and
    status byte state, and no longer allocates a new module name string for
    every object of every frame, after the first pass.

    hf_pn_io_gating_cycle (PDSyncData v1.0/v1.2 and the CIMNetConf blocks)
    was never registered, so those blocks ended in an "Unregistered hf"
    dissector bug report (an abort with WIRESHARK_ABORT_ON_DISSECTOR_BUG).
    Register it. Also do not call g_dir_close() on a GSD directory that
    could not be opened.

    Small fixes: reduce the RS_TimeStatus extension length by the optional
    time stamp so it is not decoded twice, guard the vendor specific fiber
    optic data length against underflow, fix the TxPortGroup array size
    precedence, and use int for two offsets that were 16 bit.

    4.4: Hand-ported, on top of the partial backport 0abfe269f0 that already
    gave the block body dissectors a tvb subset and bounded the IandM5 and
    AssetManagementInfo loops; this adds the rest. This branch has no
    conversation_address_list, GSD cache, ARSXP or CIM blocks, so the MAC
    address, GSD cache and those block hunks are dropped. The module name
    overflow is fixed where the Expected Submodule data is copied into the
    IO data object, and the unregistered hf_pn_io_gating_cycle does not
    exist here.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5c3a303860394fb649b940de9b13d37725453f95)

commit 6b640bc115
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    WiMAX: Make map decoding independent of earlier frames and fix a loop

    A compressed DL-MAP or Reduced AAS DL-MAP inside a PDU burst was always
    decoded from the start of the burst, not from the current offset, and
    the IE loop was bounded by the burst length, not by the map length. A
    burst made of many 4-byte map headers decoded the whole map again for
    every header, which is quadratic in the burst size and ran into the
    tree item limits. Decode from a subset at the current offset and stop
    the IE loop at the CRC, with an expert info if IEs are left over. The
    unsigned "length -= 15" that wrapped for short maps is clamped.

    Whether a frame is downlink was decided from bs_address, which is set
    by the first FCH burst in the file, so the decision for a frame
    dissected before it changed on the second pass. Decide it once per
    frame and keep it with p_add_proto_data().

    The compact DL-/UL-MAP state (harq_mode, cid_type, max_logical_bands,
    band AMC type and symbol counts) and N_layer and RCID_Type were never
    reset, so a map was decoded with the state left by earlier frames and,
    with a second pass, by later ones. Reset them at the start of each
    HARQ-MAP, DL-MAP, compressed DL-MAP, Reduced AAS DL-MAP and UL-MAP.
    cqich_id_size is deliberately left alone, it carries the UCD setting.

    Generate the CRC tables once at registration instead of on each PDU
    burst, so that a HARQ-MAP or MAC header reached without a burst is not
    checked against an empty table. The plugin.c generator does not match
    the handoff functions of the AAS feedback, ARQ, DREG and PMC messages
    to a protocol, so they were never called; call them from
    proto_reg_handoff_wimax().

    Fix small errors: the TLV length check in the m2m dissector compared
    the TLV start and reported every TLV as malformed; the HARQ_DL_MAP_IE
    length is 8 bits, not 2; the IR CTC/CC Dedicated_DL_Control_IE and
    RCID_IE results were added to the bit offset in the wrong unit; the
    appended UL-MAP of a Reduced AAS DL-MAP was decoded from bit 0; the
    invalid nested TLV items in the CS parameter decoder used a negative
    length; the Short-HMAC digest length subtracted an absolute offset; the
    DL-MAP CRC-16 field was FT_UINT32, so the checksum read 4 bytes and the
    end of every Reduced AAS DL-MAP was malformed.

    4.4: wimax_compact_dlmap_ie_decoder.h has no extern declarations of
    the state variables on this branch, so only the declaration of
    wimax_reset_compact_map_state() is added.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit df1131d982ac06d0cc19dad6715a459aaa9897b5)

commit d9bf3ed5e4
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    opcua: Fix nesting state leaks and per-pass encryption state

    The nesting counters (the per-frame Variant/DataValue counter and
    pinfo->dissection_depth) were not restored when an exception ended a
    message. tcp_dissect_pdus() goes on with the next UA message in the
    frame, so the leaked levels added up over the messages of one segment
    until increment_dissection_depth() hit its DISSECTOR_ASSERT. Run each
    message from a wrapper that resets the counter and restores both
    values, also on an exception.

    The encryption mode kept in the conversation was rewritten on every
    pass, so a frame before the OpenSecureChannel was dissected with the
    state left behind by the last message, and "tshark -2" differed from
    one pass. Only change the conversation on the first pass, and keep the
    mode and signature length each message started with in per-frame
    protocol data, keyed by the message number within the frame.

    Also: skip the 4 byte ArrayDimensions length of an oversized Variant
    instead of returning with the counter raised and the offset unchanged;
    give the X.509 decoder a tvb limited to the certificate ByteString;
    stop the signature (Sign) and padding plus signature (SignAndEncrypt)
    lengths from wrapping when the message is shorter, with a new expert
    info; print the NUL-terminated copy of a keylog DSB in the debug
    message, not the raw block; log "no keyset found" only when there is
    none.

    4.4: This branch has no AEAD decryption and no ws_debug; the
    footer length guard is also added in decrypt_opcua() before the
    padding byte is read, and the keylog DSB copy uses g_malloc0 instead
    of reading one byte past the block.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit fc0c9c07c9529bced1eafd10bb2ad16da2145d17)

commit 012a97ff1d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    usbdump: Bound the packet length and fix state and error handling

    The captured length from the frame header went to the buffer
    allocation unchecked, so a small (compressed) file could make the
    reader allocate up to 4 GiB. Reject lengths larger than
    WTAP_MAX_PACKET_SIZE_USBPCAP (128 MiB), the limit for USB
    encapsulations.

    Also print the converted version in the unsupported-version message and
    return WTAP_OPEN_ERROR, since an error is set. Do not let a random
    access read change the sequential multiframe tracking, and stop ignoring
    a failed read of the next multiframe size.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e7e95ec8aada056910023fb8bb5a1366f8d8ba6a)

commit efe75c9aba
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    WiMAX: Fix UCD loop, CID limit and several decoding errors

    The UCD Uplink Burst Profile sub-TLV loop did not check the sub-TLV
    length. A 4-byte length of 0xFFFFFFFA wrapped length + value offset to
    zero and the loop never advanced, hanging tshark. Stop on a length
    above MAX_TLV_LEN or below 1, as the DCD and REG loops do.

    The generic MAC header decoder reported a dissector bug when a 65th
    distinct CID carried a fragment, and its CID table was never reset
    between files. Move the tables to file scope, clear them in the
    defragment init routine, and show the fragment unreassembled with an
    expert info instead. The bs_address copy is now freed on reset.

    When the generic MAC header was reached directly, for example through
    Exported PDU, first_gmh stayed false and the reassembly ids of the
    second pass differed from the first. A direct call now counts as the
    first GMH. The new part of frag_num_array is cleared like
    cid_adj_array.

    REG-RSP decoded the whole rest of the message for every vendor
    specific, vendor id or MAC version TLV, which is quadratic. Decode only
    that TLV.

    Smaller fixes: CINR feedback in the type 2 header skipped a byte, the
    FPC station loop condition was inverted, ARQ-Discard and ARQ-Reset used
    offsets that did not account for the stripped type byte, the ARQ
    feedback IE item length was the whole message, the CID update and SA
    challenge sub-TLV loops compared an absolute offset with a length, the
    RNG-RSP sub-TLV check tested the wrong type, short HMAC tuple lengths
    and REG-RSP HMAC/CMAC offsets were wrong, UCD backoff exponents of 32
    or more shifted out of range and some UCD items had the wrong length,
    and the invalid sub-TLV items in DCD/UCD used a wrong offset and
    length.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit a29567aa46db4c1f2ade7579e06a8c8f83a922f8)

    [Backport: Needs 2f449f7335 (WiMAX: Fix some array bound overruns), picked before this commit. release-4.4 added the CINR mean and deviation as two-byte items at the same offset; took the master lines, which add each FT_UINT8 field as one byte. Kept the 4.4 abbreviation of ei_mac_crc_malformed.]

commit 8553ffec9f
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Jul 2 21:42:24 2025 -0400

    WiMAX: Fix some array bound overruns

    Coverity 1192590, 1192591, 1192592, 1477573, and 1477931.

    (cherry picked from commit 2f449f733525c7b78f1507668b5a662c2c55012d)

commit 4016ba46b5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    PROFINET: Fix memory safety problems in the CBA dissectors

    A DCOM OBJREF with OXID or OID 0 creates an interface without a
    parent. cba_ldev_find() and cba_pdev_find() dereferenced the parent
    unchecked and crashed on a later CBA call to that IPID. Both now check
    for a NULL parent.

    The physical device was stored in the same private_data pointer that
    cba_ldev_find() returns as a logical device, so a get_ACCO or Connect
    request on the IPID of a physical device wrote past the end of the
    smaller structure. Both structures now start with a type tag that the
    lookups verify.

    The array sizes of Connect, Connect2, Disconnect, SRT DisconnectCR and
    SRT Connect, and the type description sizes of Connect2 and SRT Connect,
    sized allocations (up to 34 GB) and loops. They are now limited to the
    bytes remaining in the packet divided by the minimum element size, with
    an expert info. Connection tracking, whose lists are searched linearly,
    was quadratic: each logical device list and each SRT frame now keeps at
    most 5000 connections, later ones are reported with an expert info and
    not tracked. The code that fills in connection results now accepts
    connections that are not tracked.

    ICBAAccoServerSRT::ConnectCR passed hfindex 0 with a NULL tree, which
    asserts in the registrar; pass -1. The PN-RT heuristic looked up frames
    with a destination address that is not an Ethernet address, which can be
    NULL; it now requires AT_ETHER.

    DiagConsConnections responses passed an uninitialized item pointer to
    proto_item_set_text() when dissected without a tree, because
    dissect_dcom_HRESULT_item() only sets it with a tree. That crashed on
    the first pass of tshark -2. Initialize it to NULL.

    Also loop on u32ArraySize >= 2 in dissect_PBAddressInfo() so an odd
    size no longer wraps, and set qc_reported to true after the first QC
    expert info, as intended.

    The cleanup routine freed only the list of physical devices; free the
    logical device, frame and connection lists hanging off them too.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit bc8ec2c6770914cb1954d9bf4671b5b6433d87d8)

    [Backport: release-4.4 val_to_str() has no allocator argument; kept the 4.4 call where the QC expert info is reported. packet-dcom-cba-acco.h on 4.4 lacks master's Doxygen comments; added the kind tags, connection counters and cba_free_all() declaration without them.]

commit 0abfe269f0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    PROFINET: Fix crashes, CPU blow-ups and per-pass state in PN-IO

    The Read/Write request and response header blocks, and the RTC1 object
    extraction, compared the stored device and controller MAC addresses
    with memcmp() over six bytes of pinfo->dl_src/dl_dst without checking
    the address type. On a link without Ethernet addresses (for example raw
    IP carrying UDP port 34962) the data pointer is NULL and the second
    packet crashed. The addresses are now only stored and used when both
    are Ethernet addresses, and compared with addresses_equal().

    The PROFIenergy float64 measurement values (Set_Meter and
    Query_Measurement) were dissected with dissect_dcerpc_uint64() into an
    FT_DOUBLE field, which is a dissector bug report (an abort with
    WIRESHARK_ABORT_ON_DISSECTOR_BUG) and reads an uninitialized flag. They
    are now added with proto_tree_add_item().

    With a GSD path configured, resolve_pa_profile_submodule_name() wrote up
    to MAX_NAMELENGTH bytes into a module name pointer that had been
    redirected to the exact-size GSD text, a heap overflow that also
    corrupted the GSD cache. The GSD text is now copied into the
    MAX_NAMELENGTH buffer.

    dissect_block() ignored the offset returned by the block body
    dissectors and advanced by the declared body length, but several
    dissectors (PDevData, IandM5, AssetManagementInfo, the CIM blocks) read
    past their declared body, so the enclosing loop dissected the same bytes
    again. A chain of N nested blocks cost 2^N or Fibonacci(N) block
    dissections. Every block body dissector now gets a tvb subset that ends
    at the declared body end (offsets stay absolute), so nesting is linear.
    The IandM5 and AssetManagementInfo entry loops also stop at the end of
    their data.

    IOCRBlockReq searched the station's IO data object and IOCS lists
    linearly for every entry (and allocated an object even for duplicates),
    quadratic in the number of entries. The lists now have a hash index by
    slot/subslot, and an object is only allocated when it is new.

    The AR list was a GList searched and appended to linearly, quadratic in
    the number of ARs. It is now a wmem_map keyed by the ARUUID.

    dissect_PNIO_C_SDU_RTC1() walked the whole object list for every object
    of every RTC1 frame. It now finds the object by frame offset in a hash
    index that is extended lazily with the entries added to the list. Each
    object count iteration looks up one IO data object and one IOCS object
    instead of scanning the lists, which gives the same result for
    well-formed configurations.

    The AR records were modified on every pass, and the "changed from"
    expert infos and the ARUUID/MAC information shown for a frame read the
    state left by all frames, so -2 and the GUI showed different PN-IO data
    than the first pass. The records are now only updated while the frame
    has not been visited. pnio_ar_info() keeps a per-frame copy of the AR
    state from the first pass and shows it on later passes, the "changed
    from" warnings are remembered per frame and shown again, and an AR is
    not found by frames before the one that created it. The conversation
    filter data now refers to the per-frame copy and is set instead of
    being prepended to file scope data on every pass.

    The RTC1 dissector no longer rewrites the PROFIsafe toggle bit and
    status byte state, and no longer allocates a new module name string for
    every object of every frame, after the first pass.

    hf_pn_io_gating_cycle (PDSyncData v1.0/v1.2 and the CIMNetConf blocks)
    was never registered, so those blocks ended in an "Unregistered hf"
    dissector bug report (an abort with WIRESHARK_ABORT_ON_DISSECTOR_BUG).
    Register it. Also do not call g_dir_close() on a GSD directory that
    could not be opened.

    Small fixes: reduce the RS_TimeStatus extension length by the optional
    time stamp so it is not decoded twice, guard the vendor specific fiber
    optic data length against underflow, fix the TxPortGroup array size
    precedence, and use int for two offsets that were 16 bit.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 5c3a303860394fb649b940de9b13d37725453f95)

    [Backport: release-4.4 PN-IO differs too much from master (no AR map, GSD module cache, object lists rework, or the Ethernet address compares, which were added in 4.6) to take the whole commit. Only the dissect_block() part was backported: every block body dissector gets a tvb subset ending at the declared body, and the IandM5 and AssetManagementInfo entry loops stop at the end of their data, so nested blocks no longer cost 2^N dissections. The other changes described above are not included.]

commit c148f5b876
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    IrDA: Fix an XID hints over-read and decode IAP results consistently

    dissect_xid() built the service hints text by overwriting its first two
    characters with " (". When hint bits were set but none of the named
    ones, the string was just ")" and the overwrite removed its terminating
    NUL, so reading it ran off the end of the stack buffer. Append the
    hints only when at least one is named.

    dissect_iap_result() honored the list length only when building a tree,
    so the LMP conversations created by the attribute value dissectors
    depended on whether a tree existed and frames decoded differently with
    "tshark -2". Read and honor the list length in the no-tree branch too.

    add_lmp_conversation() dereferenced a NULL pointer if an existing
    conversation had no IrLMP data; allocate the data in that case. Only
    look at the IrDA pseudo-header when the encapsulation is IrDA, as it
    holds unrelated data when reached through SLL, SIR or Exported PDU.

    Make attr_len unsigned so that large octet sequence lengths do not
    wrap, and actually append the octet count to the Info column.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 098de61478015b0acf7db0e6a217d0a2e7f5af91)

commit 1ca2879ae1
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    Gryphon: Fix an endless loop and a dissection depth leak

    cmd_usdt_register_non_legacy() looped forever on a USDT register block
    whose ID count is zero, because neither the offset nor the remaining
    length advanced. Skip the 20 byte block and stop if less than 20 bytes
    remain.

    The dissection depth was incremented in dissect_gryphon_message(),
    decode_command() and decode_response() and only decremented on normal
    return. An exception in a PDU skipped the decrements, and
    tcp_dissect_pdus() goes on with the next PDU of the frame, so a frame of
    many malformed PDUs ended in the depth assertion. Restore the depth
    after each PDU.

    Also use the field offset in cmd_bits_in()/cmd_bits_out() instead of the
    constant 1, make the NUL-terminated name fields FT_STRINGZ, avoid a
    wrap of the node name count in resp_ldf_get_node_names() when it is 0,
    and loop over events only while a whole 20 byte event remains.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 8b589bc58ce4ebd640062e01e1266a6f4c4056b6)

    [Backport: added #include <epan/exceptions.h>, which master already has, for the TRY block.]

commit 7401060b13
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    Falco Events: Harden against packet records and failed extraction

    dissect_falco_event() is registered as a dissector and can be reached
    from Exported PDU, whose record is REC_TYPE_PACKET. It then read
    rec_header.syscall_header, a union member that overlays the packet
    header, and passed its wth field to open_sinsp_capture(), which writes
    through it. Return 0 for records that are not REC_TYPE_SYSCALL. The
    event_param_data passed as dissector data is NULL when called without
    data; check it before use in dissect_sinsp_enriched().

    extract_syscall_conversation_fields() returned before it initialised
    its output when the protocol was disabled or the capture could not be
    opened, so the conversation filter callbacks and get_fd_stream_info()
    used an uninitialised container_id. Initialise the output first. On an
    extraction failure it called REPORT_DISSECTOR_BUG() from callbacks that
    run outside dissection, where nothing catches the exception; return
    false instead.

    Check the lineage number against N_PROC_LINEAGE_ENTRIES before indexing
    lineage_trees, and check hf_id_to_addr_id for NULL before use.

    Use PRId64 for the signed tid and return NULL, not false, from
    get_fd_stream_info().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit fc65ef052401ca3fabf59f9128cae5e33892f311)

    [Backport: Applied to plugins/epan/falco_bridge/packet-falco-bridge.c (dissect_falco_bridge()), the release-4.4 name of the Falco Events dissector. 4.4 already initialises the conversation fields first and has no protocol-enabled or sinsp_capture_open_failed() early returns, no get_fd_stream_info() and no evt/fd layer dissector calls; the NULL event_param_data check was added to the 4.4 SSC_FD parent item code, and the invalid lineage number now skips the field instead of calling ws_error().]

commit 5364272e1b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    iLBC codec: Bound the iLBC frame count and handle decode errors

    The iLBC size query accepted any payload whose length was a multiple of
    38 or 50 bytes, but libilbc decodes only one to three frames per call
    and returns -1 otherwise. The -1 was stored in a size_t, so the decoder
    reported 2^64-2 bytes and the RTP player and sharkd read far past the
    output buffer.

    Return 0 for payloads of more than three frames (150 bytes) and treat a
    negative result from the library as no samples.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 24e3d327e00df32885ea3103dbc3c302a5242df8)

commit a6ed80f04e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    Codecs: Size the G.726 output for bits carried between packets

    For 3 and 5 bits per sample (G726-24/-40, AAL2-G726-24/-40) spandsp
    keeps the leftover bits of one call and uses them in the next, so a
    packet can decode one more sample than inputBytes * 8 / bits. The size
    query did not account for it and spandsp wrote two bytes past the
    allocated buffer.

    Round the sample count up, which covers the at most bits-1 carried bits.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 38ffbd7d0e6647fe7db40fbeffa7f19a9e8e89a8)

commit d654a97c81
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 16 15:07:07 2026 +1000

    AMR, AMR-WB: Fix the length check for octet-aligned payloads

    decode_many checked the payload size by adding one TOC byte per frame
    already processed, but the speech data starts after all of the TOC
    entries. For frame i the check was short by frames-1-i bytes, so the
    memcpy of the frame read up to 31 (AMR) or 60 (AMR-WB) bytes past the
    payload.

    Count CMR and all TOC entries up front and add only the speech size per
    frame.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 8274108df1faaaab81013db722c50c19453e44d6)

    [Backport: plugins/codecs/amrwb does not exist on release-4.4, so only the AMR-NB part applies; master-only mode 9-14 handling around the hunk not taken.]

commit 6158077523
Author: Gerald Combs <gerald@wireshark.org>
Date:   Tue Nov 17 13:01:01 2026 -0800

    GitLab CI: Switch the Windows Arm64 package to Qt 6.8.3

commit aa41ed091d
Author: John Thacker <johnthacker@gmail.com>
Date:   Tue Nov 17 09:35:26 2026 -0500

    MSYS2 setup: Install flex

    flex was removed from their base-devel package, so now we have to
    select it by name.

    https://github.com/msys2/MSYS2-packages/commit/e4545bfa4361cee00ef066667405c4c7fc82fc9e

    (cherry picked from commit 9bd5abc57fb1f7eccf6b7ad2889e34715827ee3a)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 9cafb979b8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 14 06:38:28 2026 +1000

    RBM: Bound object nesting and fix NULL strings and struct names

    Nested objects were bounded only by increment_dissection_depth(), which
    ends in a DISSECTOR_ASSERT at the limit. Nesting is now limited to 64
    levels per packet (kept with p_get_proto_depth()) and deeper objects get
    an expert info and consume the rest of the data.

    After a nested "," object with an unsupported version the type and
    value were NULL and were printed with "%s". The defaults are kept in
    that case, and a NULL hash key or value is printed as "Unknown".

    The struct name length integer was assumed to be one byte; use the
    length returned by get_rbm_integer(), so names of 123 bytes or more
    decode correctly.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 968e78807db1b688166783dfdbda76bceb56d3c8)

commit 77d434c57a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:43 2026 +1000

    VNC: Fix Tight assert and several decoding errors

    A Tight rectangle whose compression control nibble is 0xA to 0xF only
    raised an expert info and left bytes_needed at -1, so the following
    DISSECTOR_ASSERT(bytes_needed != -1) fired on a malformed packet.
    Set bytes_needed to 0 after the expert info.

    vnc_client_cut_text_extended() called abs() on the length, which is
    undefined for INT_MIN. Negate in unsigned arithmetic instead. The
    original-format text was also added with length message_length - 4
    while the offset advanced by message_length; use message_length.

    vnc_server_cut_text() returned the offset as "bytes needed", which made
    the caller ask for more data after every complete message. Return 0. It
    also did not skip the 3 padding bytes after the message type, so the
    length was read from the padding and the text was decoded as the next
    message; skip them.

    SetColourMapEntries read the colour count from absolute offset 4;
    read it relative to the message start. It also asked for 3 bytes more
    than the message holds, because the bytes already consumed were counted
    again. The width * height * bpp
    products in the raw, ZRLE raw and rich cursor encodings were computed in
    signed int and could overflow; compute them in unsigned.

    On 4.4 the extended client cut text code does not exist, so only the
    other fixes are ported.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 3299f95c56d503ee1b1396f23f0950b250487171)

commit 4f522e4dfb
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 06:57:16 2026 +1000

    Z39.50: Bound RPN, element data type and GRS-1 nesting

    RPN queries (RPNStructure/rpnRpnOp), structured element data types
    (ElementDataType/structured) and GRS-1 subtrees (ElementData/subtree)
    can nest without limit. The generated code bounds these cycles only
    with increment_dissection_depth_by_n(), which asserts at the tree
    depth limit, so a search request of about 1.5 KB with 250 nested RPN
    operators on the default port ended in a dissector bug, and aborted
    under WIRESHARK_ABORT_ON_DISSECTOR_BUG. Count the nesting per packet,
    stop at 32 levels with an expert info, and leave the rest undissected.

    The MARC leader expert infos printed an invalid octet with %c, which
    puts a raw byte that may not be valid UTF-8 into the label (the fuzzer
    aborts on that). Print it in hex.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit d4ffa5ff0b982379025539e6557e1104e0639d7d)

commit 6761eba238
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    swIPe: Bound nesting of swIPe and IPv6 headers

    swIPe passes its payload to IPv6, and the IPv6 next header can be swIPe
    again. A packet with about 250 such nested headers hit the layer limit
    assertion in call_dissector_work(). Count the nesting with
    p_get_proto_depth() and stop with an expert info after 4 levels.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit d129054529564540ea8fa67a2bf34faf58ca7ab8)

commit cd777637a2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    Sparkplug: Fix leak of the split topic on early returns

    The vector returned by g_strsplit() was not freed on the three paths
    that return when the group ID, message type or edge node ID is missing,
    so every MQTT PUBLISH with a short spBv1.0 topic leaked it. Free it
    before each return.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit b190b4c1287dcca16f28b636c78ff0ea30a979b9)

commit b12f6fd5dd
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:02 2026 +1000

    SIP: Fix Session-ID leak and bound message/sip nesting

    A message/sip body was passed back to the SIP dissector without any
    limit, so nested messages ended in the generic layer assertion. Allow
    at most four nested levels, kept in a per-packet depth with
    p_get_proto_depth(), and add an expert info for deeper bodies.

    dissect_sip_session_id_header() read the logme parameter name five
    bytes past the start of the parameter even if the line ended earlier.
    The read could throw and leak both GByteArrays. Limit the read to the
    end of the line.

    The test for the '=' in a Session-ID parameter was inverted and used
    the wrong length, so ;remote=<uuid> was never decoded. Search the rest
    of the line and take the branch when the '=' is found.

    Remove a redundant assignment of a bool to comma_offset in the
    History-Info dissection. Allocate the generated trace ID string in the
    packet pool when the frame was already visited, so that redissecting
    does not grow the file scope.

    On 4.4 the History-Info redundant assignment does not exist, and the
    Session-ID '=' test already had the right sense, so only its length is
    fixed.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit fd1e757240b0c6a5e17c48c022c3d62f9b5e1f37)

commit b676649959
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:01 2026 +1000

    RTnet: Bound nesting of tunneled RTmac headers

    With RTmac version 2 and the tunnel flag set, the next protocol is
    looked up in the ethertype table using the packet's own type field.
    Type 0x9021 is RTmac itself, so every 4-byte header called
    dissect_rtmac() again on the rest of the frame. A small frame ran into
    the generic layer-count assertion in call_dissector(). The version 1
    default case had the same loop.

    Count the nesting per packet with p_get_proto_depth() and, beyond four
    levels, hand the rest to the data dissector instead of recursing.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e333e6228fdac2e4bae776cb5ae888fc8e5a64df)

commit 410db6e2d5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:00 2026 +1000

    RDP: Bound channel PDU and ServerData allocations, fix leaks

    Every ServerData SC_NET and SC_MCS_MSGCHANNEL created new dissector
    handles with create_dissector_handle() on every pass, once per channel
    id, and never freed them. A capture with many ServerData PDUs listing
    thousands of channels used hundreds of megabytes, doubled with -2 and
    grew with each reload. The handles are now created once in the handoff
    routine, registered in the "t124.sd" table only on the first pass, and
    at most RDP_MAX_CHANNELS channel ids are registered per ServerData.

    The reassembly buffer of a virtual channel PDU was preallocated with
    the length announced by the first chunk, up to 4 GiB per frame in file
    scope. The array now grows with the data that is actually present.
    pendingLen is clamped instead of underflowing when a chunk is larger
    than the announced length.

    The negotiate request copied the server address into file scope on
    every pass; it is now done on the first pass only. The flags peeked in
    the licensing stage no longer leak into the virtual channel PDU path,
    where a channel length with bit 0x08 set was shown as encrypted.

    The tvb of a reassembled virtual channel PDU was created with
    tvb_new_real_data() on the first pass and kept in file scope, where it
    was never freed; every reassembled PDU leaked it, together with the
    subset tvbs that the channel dissectors created from it. Keep the
    reassembled data and its lengths instead, and build a child tvb of the
    packet on every pass.

    A multitransport link created from the MultiTransport request left its
    UDP conversation pointer uninitialized, and the lookup from a UDP
    conversation compared against it. Zero the links when allocating them.

    On 4.4 there is no channel PDU reassembly or ServerData chunk code, so
    only the zero-initialised transport links, flags initialisation,
    once-only t124.sd registration with reusable handles (bounded by
    RDP_MAX_CHANNELS) and first-pass-only server address update are ported.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 686feab83ecf3f9943a57c1e33e22b87615d96fb)

commit 32bf93511e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 14:33:51 2026 +1000

    NMEA0183: Fix FT_CHAR handling and a few length errors

    The ETL number of shafts and XTR direction fields are FT_CHAR but were
    added through proto_tree_add_item_ret_string(), which reports a
    dissector bug for every such sentence. Add them as items instead.

    All 122 FT_CHAR fields were registered with BASE_NONE. A one character
    field holding a non-printable octet such as 0xbd then reached the
    "Invalid base" dissector bug when the label was formatted, outside of
    any exception handler, so printing the tree aborted (tshark -V). Use
    BASE_HEX, which shows such a character as '\xbd'.

    An empty or missing FT_CHAR (or numeric) field was added with a length
    of 0 from the packet, which throws and loses the rest of the sentence.
    Add a zero length item of the field type instead, and show only one
    byte, with an expert info, for an FT_CHAR field longer than one
    character.

    When '*' directly follows the sentence id the data length wrapped and
    the checksum was decoded as data; treat it as empty data. Check the
    binary file descriptor length before computing the length of the status
    field and add an expert info if it is too short. Use the length of the
    message, not its end offset, for the item of the second and later
    UdPbC sentences. Make the AAM units of radius field FT_CHAR with the unit names so that
    it shows the letter.

    The latitude and longitude direction fields were added with
    proto_tree_add_item_ret_string() and a NULL scope, which leaked the
    string; use pinfo->pool. decode_time() and decode_date() split the
    field at byte offsets, which cut a UTF-8 sequence when the field held a
    non-ASCII byte (shown as U+FFFD); reformat only fields of plain digits,
    and keep the whole first character of a unit label.

    On 4.4 only the direction field leak and the empty-data length wrap
    apply; the rest of the code touched upstream does not exist here.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c4c8168e01bbdcbaecc7d14b4222def2d2dd3e91)

commit 3cea0adbc1
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    X2AP: Bound nesting of X2AP-Message

    An X2APMessageTransfer IE carries an X2AP-Message octet string that is
    dissected as a whole X2AP-PDU, which can again be an X2APMessageTransfer.
    About 250 levels fit in one SCTP chunk and end in a depth assertion
    (an abort with WIRESHARK_ABORT_ON_DISSECTOR_BUG) or a tree depth
    exception. Allow at most 2 nested X2AP-Message levels per packet and
    report further levels with a new expert info, x2ap.nested_too_deep.

    A local PrivateIE-ID did not clear direct_ref_present, so a later
    private IE after a global one was decoded with the previous OID. Clear
    the OID state at the start of PrivateIE-ID.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 13a85feb9bbafeb78af378f3a74049eb5ef99299)

commit d42bbd49cb
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    UHD: Print the message id as hex instead of %c

    The 32-bit message id from the packet was formatted with %c in the
    protocol item label and in the Info column fallback text. An id whose
    low byte is 0x80 or higher puts invalid UTF-8 into the label, which
    aborts a debug build under the fuzzer's UTF-8 log domain, and other ids
    were shown as an unrelated character.

    Show the id through its value string, with the id in hex when it is not
    known.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit b070f37a6c62310adb176faf5171e250c3ef97cc)

commit 93c22885e4
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:01 2026 +1000

    TNS: Fix a Data Descriptor assert and several bad reads and loops

    A Data Descriptor PDU with a header length below 8 but a non-zero data
    length passed tcp_dissect_pdus() and then hit a DISSECTOR_ASSERT_CMPINT
    on the header length. Report it with an expert info and stop instead.

    tns_format_vector() added 8 to its position for the optional header
    without checking it against the image length. The following length
    check then compared against a negative remainder converted to unsigned,
    so it passed and the loop read up to 145 bytes beyond the buffer. Return
    NULL when the position is past the end.

    A TTI_RXD row with columns that consume no bytes (NULL columns without
    a data length, or "same as previous row" columns) cost one iteration
    per column for each one-byte row, which is quadratic in the packet
    size. Count such columns per message, stop at 100000 and add an expert
    info; real result sets never reach this.

    The rendered text of OSON (JSON) values contains raw bytes from the
    packet and can be invalid UTF-8, which aborts the fuzzer. Sanitize it
    with get_utf_8_string() before it goes into the item label.

    Also check a chunk length against the remaining bytes in
    tns_chunks_len() so the offset cannot overflow, and compute the thin
    TTI_IOV bind count in unsigned arithmetic.

    On 4.4 only the Data Descriptor header length assert applies; the other
    hunks are for code that does not exist there.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit ad25f35d223b18df0e209c2a807ee38820b9a3a3)

commit 51e4ab591b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    Thread: Bound UDP Encapsulation recursion and fix TLV offsets

    A MeshCoP UDP Encapsulation TLV can lead through CoAP back into MeshCoP
    without limit, ending in the generic layer limit assertion. Limit the
    nesting to 8 levels per packet and add an expert info beyond that.

    Check that the UDP Encapsulation TLV is at least 4 bytes long before
    reading the ports, and use the TLV length for the child tvb length.

    Fix the double offset advance for IPv6 Address TLVs in the Thread
    Address and Thread BL dissectors (and a stray whole-TLV IPv6 item in
    BL), and advance past the value of the Domain Prefix, Commissioner
    Token, TRI Hostname and Commissioner PEN Signature/Token TLVs. The
    Energy List TLV advanced the offset per entry and then again by the
    whole TLV length, so the TLVs after it were misparsed.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 532e587f094e170119fba45b047a63293cfe6ada)

commit 004dd14ecc
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    STUN: Bound nesting through DATA and ChannelData, fix password algs

    The DATA attribute and ChannelData payloads go to the "stun" heuristic
    list, which contains STUN itself when the stun_turn heuristic is
    enabled. Nested DATA indications recursed until the layer limit
    DISSECTOR_ASSERT fired. Allow at most 4 levels of STUN inside STUN,
    count them with p_get_proto_depth() and add an expert info beyond that.

    The PASSWORD-ALGORITHM(S) length check was inverted: valid attributes
    with several algorithms were flagged as too short, and a parameter
    longer than the attribute was decoded into the following attributes.
    The unsigned remaining length then wrapped around. Fix the comparison,
    do not subtract more than what is left, and put the padding item after
    the parameter data instead of at its start.

    att_length_pad was a uint16_t, so a padded length of 0x10000 or more
    truncated to 0 and bypassed the check that the attribute stays inside
    the message. Make it unsigned.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 7f3720db469e0ffbe8dcbca4f8e7d01ae87dd80a)

commit 976d293198
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    MACsec: Restore pkt_len when the next dissector throws

    The persistent pinfo->fd->pkt_len is reduced by the SecTAG, ICV and
    FCS lengths before the Ethertype dissector runs and was restored only
    if no exception was thrown. A malformed payload left the frame with a
    wrong length for later dissection passes. Restore it in a FINALLY
    block.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5581a2c88a95b1249e0195af400dd05956fac666)

commit 2c6d525e2c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    PN532 HCI: Fix infinite loop on long payloads

    The preamble and postamble scans used a uint16_t as the byte index.
    With a payload longer than 65537 bytes that contains no 00 FF start
    code, the index wrapped from 65535 to 0 and the loop never ended,
    hanging the dissector on a crafted USB transfer. Use an unsigned
    index and read the frame length with proto_tree_add_item_ret_uint().

    The data checksum of extended information frames was computed but 0
    was passed to proto_tree_add_checksum(), so it was never verified.
    Pass the computed value.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 1502c31234dba313582190aaa1af56c2ee9038f6)

commit efaf42d0db
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:00 2026 +1000

    QUIC: Fix an assertion on coalesced packets and bound frame lengths

    When a coalesced Initial packet had a Length too short for the 16-byte
    header protection sample, quic_decrypt_header() threw an exception on
    the first pass, so no packet info was allocated for the next coalesced
    packet. On later passes header deprotection is skipped, the first packet
    returned normally, and DISSECTOR_ASSERT(quic_packet) failed. Fail
    cleanly, with the same result on every pass, when the sample is not
    available, and report an expert info and stop instead of asserting when
    the per-packet info of a coalesced packet is missing.

    Replace the DISSECTOR_ASSERT_NOT_REACHED() in the STREAM and CRYPTO
    reassembly with an expert info and return, and reject STREAM frames
    whose offset plus length exceeds UINT32_MAX, as the reassembly uses
    32-bit sequence numbers.

    Frame lengths from 64-bit varints (CRYPTO, NEW_TOKEN, STREAM, DATAGRAM,
    CONNECTION_CLOSE reason) were truncated to 32 bits or int. Report an
    expert info and stop dissecting the frames if the length exceeds the
    remaining data.

    Clear the server connection ID data with a memset of the right type
    when a Retry is processed, instead of one sized for quic_cid_t.

    Also alias ALPN "roq-*" to "roq" for DATAGRAM frames and 0-RTT, as
    "h3-*" already is; RoQ datagrams of a "roq-09" connection were not
    dissected.

    On 4.4 the roq aliasing, datagram and server_cids memset hunks and the
    STREAM reassembly assert do not apply and are dropped.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 155c41014f49709ea61708e0491f99163402b48d)

commit a776d88bfe
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:43 2026 +1000

    WCP: Fix history window wrap and the second pass dissector bug

    A packet whose decompression failed on the first pass has no stored
    data. On the second pass (tshark -2, or any redissection in the GUI)
    the checks run against the final window state, can pass, and then
    REPORT_DISSECTOR_BUG("Can't find uncompressed data") fired, aborting
    under WIRESHARK_ABORT_ON_DISSECTOR_BUG. Make the checks that depend on
    the window state only on the first pass, record a failure with its
    expert info, and show the same expert info at the same point on later
    passes, so that tshark -2 gives the same result as one pass.

    The history window wrap test compared the pointer before incrementing
    it, so the write pointer could reach the end of the buffer and the next
    byte was written to buffer[MAX_WIN_BUF_LEN], one past the array. Wrap
    after incrementing, in decompressed_entry() and for literal bytes. A
    packet that ends exactly at the end of the window now leaves buf_cur at
    the start of the window instead of getting stuck.

    wcp_save_data() had an unreachable wrap branch and underflowed the
    length for packets shorter than 3 bytes. Reject short packets and
    lengths larger than the window, and copy in two pieces when the data
    wraps.

    On release-4.4 the offsets are int and proto_tree_add_expert_remaining()
    does not exist, so the equivalent calls are used.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 0c57be2d84d46357a540708c4afb6adafafdf2d2)

commit a49ef81c2d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    UCP: Fix a dissector bug in 23R and some length handling

    add_23R() dissected the IVR5x field, which is an FT_STRING, with
    UcpHandleByte(), which uses proto_tree_add_uint(). This reported a
    dissector bug (an abort when WIRESHARK_ABORT_ON_DISSECTOR_BUG is set)
    for a packet found by the default heuristic when a tree is built. Use
    UcpHandleString() as add_23O() does.

    tvb_ensure_bytes_exist() was passed the absolute end offset as a
    length; pass the length from the current offset. In check_ucp() the
    loop bound "length - 2" wrapped when the ETX was at offset 1; compare
    as "offset + 2 < length".

    On release-4.4 the tvb_ensure_bytes_exist() calls already use the right
    length, so only the check_ucp() loop bound and the 23R IVR5x handling
    are ported.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 245ca520f90fc3bb6a3e022b60db3e255c422b87)

commit c35dddb67c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 06:57:17 2026 +1000

    ZigBee Security: Fix uninitialised PAN and bound the NWK key ring

    For an APS relay downstream message without an 802.15.4 layer (ZigBee
    Direct over Bluetooth) addr16.pan was never set but was hashed in the
    short address lookup, giving a nondeterministic source address. Zero
    the structure and use the PAN of the NWK hints in that case.

    The NWK key ring only compared a new key against the head of the list,
    so each Transport-Key packet with a distinct key grew the ring and
    every secured frame tried all keys (4000 keys x 4000 frames took 7
    seconds). Compare against the whole ring and keep at most 64 keys per
    PAN; keys beyond that are ignored with a new expert info, which is
    remembered per frame so that later passes report it too.

    On release-4.4 only zbee_sec_add_key_to_keyring() exists (no _panid
    variant), and the proto data uses proto_get_id_by_filter_name() as the
    rest of the file does.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c241e4bf6946cd81a25ef45aea25b10f8493df9e)

commit 8551418a13
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    Telnet: Fix IAC search, NULL suboption tvb and length underflows

    find_unescaped_iac() never reported "not found" since abbe56cd76:
    tvb_find_uint8_length() leaves the offset at the end of the data, so
    dissect_telnet() always called telnet_command() past the end and every
    packet without an IAC was marked Malformed. It also made the TN3270 and
    TN5250 "one more segment" path unreachable. Return a bool with an
    output offset, and treat an IAC in the last byte as a command instead of
    throwing while looking at the next byte.

    unescape_and_tvbuffify_telnet_option() returns NULL for suboptions of
    10240 bytes or more. telnet_sub_option() passed that NULL tvb to the
    suboption dissector or to proto_tree_add_item(), which asserts. Add an
    expert info and return instead.

    Fix length handling in suboption dissectors: AUTHENTICATION SEND with an
    odd length (loop now runs while at least 2 bytes remain), the
    authentication data length underflow when fewer than 2 bytes are left,
    and the TN3270E CONNECT, IS, REASON and FUNCTIONS requests, which used
    a length that extended one byte past the suboption or wrapped the
    remaining length when it was fully consumed.

    On release-4.4 find_unescaped_iac() still returns an int with -1 for not
    found, so only the IAC-in-the-last-byte case is fixed there, and the
    tvb_find_uint8() length is adjusted instead of tvb_find_uint8_length().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit eff4b8be43a81cf5959ead014d18b4c5ac873bd7)

commit ae5bae5455
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SSH: Fix aborts and over-reads in key exchange and key setup

    A stale mac_id (hmac-sha2-256 negotiated in an earlier key exchange)
    made ssh_derive_symmetric_keys() overwrite the 64 bytes needed by
    chacha20 with 32. ssh_decryption_setup_cipher() then copied 64 bytes
    from a 32-byte key. Only raise "need" to 32, and copy the chacha20
    key only if at least 64 bytes were derived.

    ssh_kex_shared_secret() asserted on a missing public or private key
    and passed a zero or one modulus from a DH group exchange to
    gcry_mpi_powm(), which aborts the process with a divide by zero. Return
    NULL for a missing key and for a modulus <= 1; the caller already
    handles NULL.

    ssh_set_mac_length() accepted any -<bits> suffix, so a MAC length
    larger than DIGEST_MAX_SIZE (48 bytes) led to an over-read of calc_mac.
    Reject MAC lengths above DIGEST_MAX_SIZE.

    ssh_read_e_pq() and ssh_read_f_pq() returned false on an invalid
    length, but the callers test for < 0, so the expert info was never
    shown. Return -1. Also set the offset after the protocol banner to
    the next line start instead of adding it to the offset.

    On release-4.4 the banner offset and PQ key readers (ssh_read_e_pq/f_pq)
    are not applicable, and only the chacha20 key length check is ported to
    ssh_decryption_setup_cipher().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e2064591c19a34592fc07a2626a9e851dad422d3)

commit ec1f7e8c1a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:02 2026 +1000

    SCYLLA: Fix an endless loop and an uninitialized response frame

    dissect_scylla_negotiation_pdu() loops while the unsigned remaining
    length is non-zero and subtracts a feature length taken from the packet
    without checking it. A feature length such as 0xfffffff8 for
    CONNECTION_ID moved the offset back by exactly the header size on each
    pass while the length stayed unchanged, so tshark looped forever.

    Loop only while a full 8 byte feature header remains and stop with a
    new malformed expert info when the feature length exceeds the remaining
    negotiation data.

    The request and response map entry was allocated with wmem_new() and
    response_frame_num was never set for a request whose response does
    not arrive, so later passes read an uninitialized value and showed a
    random "Response frame" link. Allocate it with wmem_new0().

    On release-4.4 only the wmem_new0() change is applicable; the
    negotiation feature loop does not exist there.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit a691cee366fa0071d122e37a4d95e2de793c5671)

commit 35b1ad66f1
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 17 16:54:10 2026 +1000

    RESP: Fix an out-of-bounds read in the JSON detection

    When a string is inside a request array and looks like JSON, the
    dissector indexed the string with the length declared in the packet and
    built the JSON tvbuff with that length, but the string was read with the
    number of bytes actually captured. A snaplen-truncated frame declaring a
    huge length read far past the buffer and crashed; a verbatim string read
    two bytes past it even in a complete packet. Use the length actually
    read.

    A declared string length is now clamped to INT_MAX - 4, so adding the
    CRLF length can no longer overflow, and the 64-bit value is no longer
    truncated to int.

    The verbatim string value skipped the encoding and the colon but was one
    byte too long, including the CR. Use captured - 4, and require at least
    4 bytes before reading the encoding.

    Bulk errors used the bulk string fields; use the registered bulk error
    fields.

    On release-4.4 only the clamp of the declared bulk string length is
    applicable; the JSON detection and RESP3 code do not exist there.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 228f3d334163e098a5da7745d1c0e1f20caf2c11)

commit 2c4aa837ce
Author: Guy Harris <gharris@sonic.net>
Date:   Tue Nov 17 07:21:25 2026 +0000

    sll: Read the SLL2 packet type as one byte

    (cherry picked from commit 0adda0f6fd2e934c9b22aa9a6d7d87c9643e2d47)

    6ed9fba9 sll: Read the SLL2 packet type as one byte

    Co-authored-by: Brandon Chiu <38291053-chiub@users.noreply.gitlab.com>

commit 64f21a3a5e
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Nov 16 21:41:32 2026 -0500

    ERF: NULL the PAD block after unreferencing it.

    So that the caller wtap_read() doesn't try to unreference it
    a second time. We could only unreference it only at the end
    of the loop, but we would still have to NULL it in case
    erf_read_header failed before ever creating the block.

    Fixup 055f2a8f5f06d93a42c27d101f414807cb466fc8

    (cherry picked from commit 85de43a1de4b1990517d400ef1f9b770d49bfdcd)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit e9669aec39
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Nov 16 22:12:15 2026 -0500

    GQUIC: Encoding is not a bool

    The Wireshark encoding parmeter is an unsigned, not a bool. However,
    to be extra careful and avoid changing a struct type in a exported header
    (even though the only function using that type does not have public
    visibility), map true and false to ENC_LITTLE_ENDIAN and ENC_BIG_ENDIAN,
    respectively, unlike on master and the unrelease 5.0, where we can make
    the correct fix.

    This fixes dissection of the ENC_LITTLE_ENDIAN case and a MSVC C4305 in
    the latest versions of MSVC.

    Ping #21761

    (cherry picked from commit bd9881c46e801dc7d217e7cce74415be736a15e0)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 650bb1d43a
Author: Gerald Combs <gerald@wireshark.org>
Date:   Mon Nov 16 08:57:55 2026 -0800

    packaging: Update our macOS dSYM .dmg compression algorithm [skip ci]

    Switch to ULFO; it appears that dmgbuild no longer supports UDBZ, which
    we probably shouldn't have been using anyway.

commit 6a6f0b4d58
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Oct 12 23:07:16 2026 -0400

    BLF: Check the APP_TEXT length against the object length

    The text length in an APP_TEXT object was used to allocate a buffer
    (up to 4 GiB) before checking that the text fits in the object. This
    displays bytes from adjacent containers and is not a security issue,
    but does indicate a malformed file. Reject the object if the text is
    longer than the rest of the object.

    Thanks to AISLE Security and Simon Huang for separately reporting this
    issue.

    Fix #21770

    (cherry picked from commit c32a552ebd47a610e47a6dc43b5e620aefe8b778)

commit 2c890bef68
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Nov 15 07:03:04 2026 -0500

    Qt: Ensure that the packet list items are vertically center aligned

    A QItemView by default falls back to Qt::AlignLeft | Qt::AlignVCenter
    if Qt::TextAlignmentMode returns an invalid QVariant(). If we return
    only a horizontal text alignment flag, it does *not* use the default
    vertical alignment of Qt::AlignVCenter, but aligns with the top edge.
    Note that Qt::AlignCenter is a two-dimensional flag, hence setting it
    also aligned the text vertically in the center.

    Ping #21735

    https://doc.qt.io/qt-6/qt.html#AlignmentFlag-enum

    (backported from commit ab5ad1acb47632f2f17294cacec1c9ee3c281143,
    and slightly changed to work on Qt 5.)

commit ecf428062a
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Nov 15 09:07:14 2026 -0500

    charsets: Fix Windows-1252 0x8E and 0x9E

    The cp1252 table, added in commit 03c5da8d89 ("Add Windows code page
    1252."), mapped 0x8E/0x9E to U+0172/U+0173 (LATIN CAPITAL/SMALL LETTER
    U WITH OGONEK, Lithuanian letters) instead of U+017D/U+017E (LATIN
    CAPITAL/SMALL LETTER Z WITH CARON, Czech/Slovak letters) from the
    Microsoft mapping the table documents as its source:

        https://www.unicode.org/Public/MAPPINGS/VENDORS/MICSFT/WINDOWS/CP1252.TXT

        0x8E    0x017D    #LATIN CAPITAL LETTER Z WITH CARON
        0x9E    0x017E    #LATIN SMALL LETTER Z WITH CARON

    The table is used for display only, through tvb_get_string_enc() and
    tvb_get_stringz_enc() with ENC_WINDOWS_1252: the MySQL dissector decodes
    result-set and query text with it (MySQL's "latin1" is CP1252) and the
    TNEF dissector falls back to it.  After this fix those strings show
    Z with caron where they previously showed the Unicode REPLACEMENT
    CHARACTER; nothing else changes.

    Assisted-by: Claude
    Signed-off-by: 3j27 <3j27@proton.me>

    (cherry picked from commit ecc5ebe55e1b5f8e6f0777f4d91d5c2486145687)

    Co-authored-by: 3j27 <3j27@proton.me>

commit 48d7500a5a
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Nov 15 10:16:34 2026 +0000

    [Automatic update for 2026-11-15]

    Update manuf, services, enterprise numbers, translations, and other items.

commit 5a3724571e
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Nov 14 23:02:13 2026 -0500

    ERF: Support writing with CLI tools

    Since editcap and tshark started adding the IDBs of the input file
    with wtap_dump_add_idb() (d0270415a9), writing an ERF file from the
    command line has failed on the first record with "Adding IDBs isn't
    supported by this file type", even with tshark -2. Writing from the
    GUI works because it passes all the IDBs when opening the file.

    wtap_dump_add_idb() adds the IDB to wdh->interface_data, from which
    the ERF writer takes the interface information when it builds the
    metadata record, so adding a stub that returns true is sufficent to
    work as well as it did before the above commit; slightly better even
    for IDBs that appear mid-file.

    A proper fix would involve doing work in erf_add_idb to properly
    give interfaces sources and names, and pulling some of the code
    out from where it is scattered. That's probably too much to backport.

    Assisted-by: Claude Opus 5.5 verified the solution

    (cherry picked from commit 284ca34b210ef82ac2fd86e595d1c51514c5b36b)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 969cd15428
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 14 06:38:28 2026 +1000

    ELF: Fix assertions, an endless loop and quadratic memory use

    value_guard() asserted on any file offset or size above INT_MAX. These
    values come from the file, so throw ReportedBoundsError instead.

    An eh_frame_hdr whose binary search table encoding has length zero (for
    example DW_EH_PE_omit) never advanced the offset in the table loop, and
    the 32-bit entry counter wrapped for counts of 2^32 or more. Return when
    the entry length is not positive and use a 64-bit counter.

    Symbol and section names were copied from the string table without a
    length limit for every symbol and section, so many entries pointing at
    one very long string used quadratic memory. Fetch names through a helper
    that reads at most 256 bytes.

    Also read a 2 byte FDE count as 16 bits, compute the trailing blackhole
    size from offset plus size, and attach the invalid data encoding and
    program and section header entry size expert infos to their own items.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit ca86e4571b9d4c0d5e6e838bf6ffdaf3f081e16f)
    [Backport: 4.4 reads the header fields with explicit byte order tests
    and lacks the invalid class, data encoding and entry size expert infos;
    applied the value_guard(), name length cap, eh_frame_hdr loop and
    blackhole size changes only. The 2 byte FDE count was already read as
    16 bits.]

commit bbb62cbf7f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 14 06:38:28 2026 +1000

    BLF: Bound log container handling

    A LOG_CONTAINER whose object length was smaller than its headers made
    the payload length negative. The sub-tvb then ran to the end of the
    file, the loop bound wrapped, and the function returned a length that
    made the outer scan dissect the following container again. Chained
    containers doubled the work each time (exponential memory and CPU).
    Report an expert info and stop when the object length is too short,
    and use the computed payload length for the sub-tvb, the payload item
    and the loop bound; the latter now uses the length of the decompressed
    data instead of the compressed length.

    Compressed containers found inside a container are no longer
    decompressed (a double zlib stream amplified about a million times).
    This adds the blf.nested_container expert info.

    Also fix the header date: the month is 1-based for mktime(), and the
    weekday name and printed fields now use the raw values instead of the
    normalized struct tm.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 683a9baeabab14b85cc1b2611f87d3b575541372)
    [Backport: 4.4 has no frame_number argument in the BLF object
    functions; added only the in_container argument. 4.4 shows the header
    date with proto_tree_add_bytes_format() from the raw values, so the
    header date change is not needed.]

commit 8f9988f626
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 14 14:17:12 2026 +1000

    BER: Check the field before adding an empty bit string

    dissect_ber_constrained_bitstring() checks hf_id before adding the item
    for a bit string with content, but not for an empty one (a length of 1
    with a zero pad octet). A caller that passes -1 because it adds the
    bits itself, as LI_XER does for the NCI, then hit the "Unregistered hf"
    assertion on a packet with an empty bit string. Add the item only for
    a registered field, as the other path does.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit f823e3d28f1bc4c6a9eab6d609662487ac11c851)

commit 6168ad0434
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 14 06:38:28 2026 +1000

    RFC 7468: Allocate the base64 decode buffer once

    The decode buffer was grown with wmem_realloc() on the packet pool for
    every base64 line. On that pool realloc allocates and copies and free is
    a no-op, so memory use was quadratic in the number of lines (100000
    lines exhausted memory). The decoded data is never larger than 3/4 of
    the remaining input plus a few bytes for the decoder state, so the
    buffer is now allocated once, before the loop.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit fa45c16a96d27298c076424c00becd1cb1832a41)

commit bd6499309b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 14 06:38:28 2026 +1000

    MP4: Fix an assertion in ctts boxes of version 1

    The version 1 ctts sample offset field is FT_INT32, but it was added
    with proto_tree_add_item_ret_uint(), which reports a dissector bug.
    Add the item normally and read the signed value with tvb_get_ntohl().

    Also reset the movie header timescale at the start of each dissection so
    a stale value from an earlier file is not used, name the ctts sample
    offset fields "Sample offset" instead of "Sample count", and size the
    elst entry subtree by the real entry length.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 9da48d647360d397ce92b04c643210bf0a5727f5)

commit 01ffde38b4
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 14 06:38:28 2026 +1000

    JPEG: Follow each EXIF extension IFD pointer type only once

    Every ExifIFDPointer, GPS or Interoperability pointer entry in an IFD
    started a full walk of the IFD it pointed to. Pointers can repeat or
    point back at IFD0, so a 64 KB APP1 segment cost about 29 million IFD
    entries and several segments added up. Each of the three extension
    types is now followed at most once per TIFF structure (tracked in a
    bitmask passed down the IFD chain); further pointers get an expert info.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit dd4a9f5c777584440717acd403de5fb5a0bbd3d2)

commit 20c36cfd23
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 14 06:38:28 2026 +1000

    file-file: Ignore calls without file data

    dissect_file_record() assumes its data argument is a file_data_t. The
    dissector is also registered by name, so Exported PDU calls it with
    NULL data, and colorizing the frame then dereferenced NULL. Return 0
    when no file data is passed.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c313519be58b10b90549fc2d755392924841b05a)

commit b0cdbde150
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:45 2026 +1000

    XMPP: Bound element nesting instead of asserting

    xmpp_unknown_items() used DISSECTOR_ASSERT(level < ETT_UNKNOWN_LEN),
    so about 20 nested unknown elements in a stanza aborted the dissector
    with a packet-triggered assertion. Add an expert info
    (xmpp.too_deep) and stop descending at that level instead.

    xmpp_xml_frame_to_element_t() recursed once per nested XML tag, bounded
    only by the assertion in increment_dissection_depth(), so a few
    hundred nested elements triggered that assertion. Record the nesting
    depth in each xmpp_element_t and stop converting children below
    XMPP_MAX_ELEMENT_DEPTH (64), with the same expert info.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 23ffb89f8f2114960e9ec7029b07294b010e3e95)

commit bf49d3af40
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:45 2026 +1000

    X509SAT: Bound Criteria recursion, fix GUID offset

    Criteria recurses through "not", "and" and "or" and was bounded only by
    the dissection depth counter, which ends in a DISSECTOR_ASSERT. A
    Guide attribute with a few hundred nested "not" values aborted. Stop
    16 levels before the tree depth limit with an expert info.

    The GUID dissector did not advance the offset past the GUID, so it
    returned the wrong length. Add the missing "offset += len".

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    [4.4: pinfo->dissection_depth is an int here; cast it for the comparison.]

    (cherry picked from commit 068ee845238aa427fffc61401549b4770cfde750)

commit a55b9b504c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    WMIO: Bound object nesting

    A chain of objects, each referencing the next through a parent class
    method signature, recursed once per object. The only bound was
    increment_dissection_depth(), which asserts when reached, so a packet
    with about 250 objects caused a dissector bug failure.

    Keep a per-packet nesting counter with p_get_proto_depth() and stop
    with a new expert info (wmio.too_deep) at 32 nested object blocks.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 3119b62066ba4e5e739e51c8048e30d29ae758c7)

commit 885421f9a6
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    WireGuard: Do not process a response without a decrypted initiator

    A Handshake Response linked to an Initiation whose encrypted static
    field failed to decrypt reached DISSECTOR_ASSERT(hs->initiator_skey).
    Only process the response when the initiator static key is known.

    Also limit the key log line parsing (leading spaces and the '='
    search) to the current line instead of the end of the buffer.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit d1f9c84cd9f356f39cebe0c569c363665481dc74)

commit 2ff0e8534e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 03:52:53 2026 +1000

    follow: Free the fragment lists when resetting a stream

    follow_reset_stream() frees the records held in the fragment lists but
    not the list nodes. A follow stream that ends with data still waiting
    for a gap to be filled leaks the nodes.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 0eb23ca00703f39706a360c18dedcaad7e25f235)

commit ca3ce4180c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 03:52:53 2026 +1000

    UDPCP: Allocate the first destination address in file scope

    The conversation data is file scoped, but its address was copied with
    copy_address(), which allocates with g_malloc() and is never freed.
    Copy it into file scope.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 05b537fee61880035853416d0872439d0971fe56)

commit ba91792e87
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    USB-I1D3: Do not assert on a response that was already matched

    If a response threw an exception after it was linked to its request,
    the previous packet was not updated, and the next response found the
    same transaction with its response already set and hit a
    DISSECTOR_ASSERT(). Treat such a response as unmatched instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 2cb2bf9d3407cb3349487e609a1fda5d42ce01f2)

commit 48c7f784b6
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    UDPCP: Do not abort on a missing conversation

    On the first pass the dissector looked up the conversation with
    find_conversation() and passed the result to
    conversation_get_proto_data() without checking it. When no UDP dissector
    had created a conversation, for example when the dissector is reached
    through Exported PDU, the lookup returned NULL and REPORT_DISSECTOR_BUG
    aborted. Use find_or_create_conversation() on the first pass, and on
    later passes stop dissecting the conversation state if no conversation
    is found.

    Also add the missing closing parenthesis to the "No ACK seen" and "No SN
    frame seen" expert info texts.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 3d427eb7e66a863015d1e8480a52bb92ba9aa2f1)

commit 80ac05cc37
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:36:49 2026 +1000

    epan: Stop at the protocol layer depth limit without asserting

    call_dissector_work() asserted that the number of protocol layers is
    below gui.max_tree_depth. Encapsulations that can contain themselves
    can nest without limit in a crafted packet: 600 IPv4-in-IPv4 headers
    on raw IP, or TZSP carrying raw IP to the TZSP UDP port again (about
    125 levels in one 4 KB frame), reached the assertion with default
    settings, and the test suite and the fuzzer abort on it with
    WIRESHARK_ABORT_ON_DISSECTOR_BUG. That is bad packet data, not a
    dissector bug. Throw a DissectorError with a message instead, as the
    tree depth limit in proto.c does.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c4f22a9871ce093cb56039682e2b2822578462e9)

commit 144143d559
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 06:57:17 2026 +1000

    ZigBee APS: Bound recursion through nested tunnel commands

    An APS Tunnel command hands the rest of the frame back to the APS
    dissector, which can be another Tunnel command. Each level costs only
    11 bytes, so a large frame nests hundreds of levels until
    call_dissector_work() hits its layers assertion and aborts.

    Count the nesting per packet with p_get_proto_depth() and stop at 4
    levels with a new expert info (zbee_aps.tunnel_depth) instead of
    dissecting the inner frame.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 930996be5b08f59117a6ca7a9cc572502b4cb222)

commit 5cc4c157dc
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    Teredo: Bound nesting of IPv6 inside Teredo

    IPv6 carried in UDP port 3544 carrying Teredo carrying IPv6 and so on
    was only stopped by the generic layer limit assertion. Limit the
    nesting to 8 levels per packet and add an expert info beyond that.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit fc2613e19de398f44b3418acb06e5d87200853fd)

commit 24d16e8a4d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    TEAP: Bound nesting of PAC-Info and nested TLVs

    PAC-Info attributes can contain PAC-Info again, and the NAK, Request
    Action, EAP-Payload and Intermediate-Result TLVs re-enter the TEAP
    dissector. Each level costs only a few bytes, so a small packet reached
    the dissection depth assertion with the default settings.

    Keep a per-packet nesting counter with p_get_proto_depth(), allow at
    most 16 levels, and report anything deeper with a new expert info
    (teap.too_deep) instead of dissecting it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c65dba7a7ccf0b8e490fe3db7fa5b932def4c2fa)

commit fa19d3625c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    TCAP: Bound parameter nesting and the SRT chain length

    dissect_tcap_param() recursed on nested constructors without a limit
    other than increment_dissection_depth(), which asserts at the limit.
    A packet with a few hundred nested constructors in a component that no
    subdissector handles aborts.  Pass a nesting level and stop after 16
    levels with the new expert info tcap.param_too_deep.

    The service response time matching walked the whole chain of begin
    transactions with the same transaction ID for every packet, which is
    quadratic when a capture reuses one OTID with more than the lost-message
    timeout between Begins.  Stop at 1000 entries in a chain, add the new
    expert info tcap.srt_chain_too_long and skip SRT matching for those
    packets.

    Also use unsigned for the length of the originating and destination
    transaction IDs, which was truncated to 8 bits.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 2cc977b633324cd346eab77a4bebd52ed6ccb19d)

commit 478ec8469d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    SV: Do not add a string value to the FT_BYTES gmidData field

    When the gmIdentity of a GmidData is not 8 bytes long, the dissector
    added the string "????" to hf_sv_gmidData, which is an FT_BYTES field.
    This fails the field type check in proto.c and aborts on any packet with
    ethertype 0x88BA that has such a gmIdentity.  Add the bytes of the
    element with proto_tree_add_item() instead; the malformed expert info is
    already added.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 8694407aa3e814a4ac3b93ec7cdf24495c5c7c7e)

commit 08d266f7b8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SUA: Bound parameter nesting and reject short parameters

    Source/Destination Address, Routing Key, Registration Result and
    Address Range parameters dissect their contents as parameters again
    with no bound, so a packet could nest them until the dissection depth
    assertion fired. Limit the nesting to 8 levels per packet with an
    expert info.

    Stop with an expert info on a parameter length smaller than the 4-byte
    parameter header, which made the derived value lengths wrap.

    Reset the routing indicator for each message so a Global Title or Point
    Code outside an address does not use the previous packet's value.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 2d2870abb0096f88e656cfe21e72e2b2d89276fd)

commit 5cbdbac26a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SPNEGO: Bound nesting of SPNEGO tokens

    A NegTokenInit can list SPNEGO itself as a mechanism and carry another
    SPNEGO token as its mechToken, which is then passed back to the SPNEGO
    dissector.  An InnerContextToken handled by spnego-wrap can loop the
    same way.  A packet of a few kilobytes nests several hundred levels and
    ends in the dissection depth assertion in packet.c.

    Keep a per-packet nesting counter with p_get_proto_depth() in
    dissect_spnego() and dissect_spnego_wrap().  Beyond 8 levels, add the
    new expert info spnego.too_deep and stop.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c7527ebe20a1b0a25faee7ae2c9e329ccf89f0b2)

commit b6b85406b1
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SOCKS: Do not re-enter the UDP dissector for the payload

    A SOCKS5 UDP datagram whose DST.PORT equals the relay port made the
    UDP conversation lookup find the same SOCKS conversation again, once
    per 10-byte header, until the layer depth assertion fired. Guard the
    payload dissection with a per-packet flag (p_get_proto_depth()) so the
    SOCKS UDP dissector is not entered recursively.

    Also take the destination port of the UDP header in both the tree and
    no-tree cases, into a local variable instead of the conversation, so
    the output no longer depends on whether a tree is built, and read the
    SOCKS4 destination IPv4 address at offset 4 instead of at the version
    byte.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit dac70d53df58c336d561fd12e4bd96242c5d58c7)

commit 30340d04e2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SNMP: Fix NULL enterprise lookup and per-PDU offsets

    snmp_lookup_specific_trap() passed a NULL enterprise OID to strcmp()
    when a Trap-PDU had a generic trap of 6 but no OID in the enterprise
    field, and an entry exists in the enterprise specific trap UAT.  Return
    NULL in that case.

    Look up the USM user (and so compute the localized keys, which hashes
    1 MB of password data for each new engine ID) only for messages that are
    authenticated or encrypted.

    SNMP over TCP with several PDUs in one segment decoded only the first:
    message_length and the returned length were relative to the start of the
    tvb instead of the start of the PDU.  Make them relative to the PDU.

    Fix the key length accounting for OID index components in VarBind
    parsing, which subtracted one byte too many and could wrap, and clear
    the 64 byte password buffer so that an empty password does not hash
    uninitialized stack.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c837e2553e549653249b274e2becb67009f432ac)

commit 3d84b9440d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SNDCP: Do not pass a NULL tvb to IP after reassembly

    process_reassembled_data() returns NULL when the reassembled data
    belongs to another layer, for example an N-PDU that carries GSMTAP and
    SNDCP with the same N-PDU number, seen on a second pass. The dissector
    then called the IP dissector with a NULL tvb, which fails an assertion
    in tvbuff.c. Only hand the tvb on when it is not NULL.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 4ec5b55800321bdf0389eb28a7aef45bd14f1ad0)

commit 9750e6ec64
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    TECMP: Restore pkt_len when the Ethernet dissector throws

    dissect_ethernet_payload() overwrites the persistent pinfo->fd->pkt_len
    while the Ethernet dissector runs and restored it afterwards only if no
    exception was thrown. After a short or malformed Ethernet entry the
    frame kept the reduced length, so later dissection passes saw a wrong
    frame.len and a bogus "Frame length is less than captured length"
    error. Restore the value in a FINALLY block.

    The FR POC State and 10BASE-T1S control messages named the interface
    with an interface_id that was never set, so it was always 0. Read the
    value from the packet. The CAN Replay Fill Level message advanced the
    offset one byte too far after the queue loop, which made the returned
    length and the TECMP item one byte too long. Zero-initialise
    tecmp_info in dissect_data() so subdissectors never see an
    uninitialised data_flags.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 996b1e3c0439242820774b2a2b1a4cc17444a3ad)

commit 7f9696cbdc
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 18:51:02 2026 +1000

    TCP: Use the MPTCP tap data in the MPTCP endpoint table

    The MPTCP conversation table registered the TCP endpoint callback,
    which reads the tap data as a struct tcpheader. The MPTCP tap passes
    a struct tcp_analysis, so "tshark -z endpoints,mptcp" and the MPTCP
    tab of the Endpoints dialog read the wrong structure and crashed on
    any capture with MPTCP. Add an MPTCP endpoint callback that takes the
    addresses and ports from the MPTCP meta flow, as the conversation
    callback does.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit f1f8bddf0bfdef71bf33985382885687cf06900f)

commit 6a8f346417
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:02 2026 +1000

    SMB: Bound export object memory for sparse files

    insert_chunk() sized the export object buffer to the file offset plus
    length of each chunk, up to 4 GiB, and zeroed all of it. Two small
    packets writing at offset 0 and at 0xFFFFFF00 of the same FID made
    tshark allocate and touch about 4 GiB. A chunk that would grow the
    buffer by more than 64 MiB beyond the data gathered so far is now
    treated like an allocation failure, so the file is marked out of memory.

    Also stop Write AndX from wrapping the 16-bit data offset when it skips
    the pipe write length, and clamp the NT_TRANS_CREATE extended attribute
    length to the data that is left, so that 0xFFFFFFFF is no longer taken
    as -1 (to end of tvb).

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 2f32c965e8b0641c4509983e1bc3018c3e687ef1)

commit 3136efa53b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:02 2026 +1000

    SIMULCRYPT: Bound ECM_group nesting and the ECM datagram length

    An EIS ECM_group parameter is dissected by calling
    dissect_simulcrypt_data() with the EIS interface type, which can
    contain another ECM_group, and so on. A 64 KB message nests about
    16000 levels deep and the recursion was only stopped by the proto
    tree depth limit, with a dissector bug exception. Count the nesting
    per packet with p_get_proto_depth() with a limit of 4, add
    increment_dissection_depth() around the nested call, and report
    "ECM_group nested too deeply" with a new expert info beyond the limit.

    The ECM datagram was passed on with a tvb that extends to the end of
    the message instead of the parameter length. Use
    tvb_new_subset_length() with the parameter length.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 34bf9dbc46fbe3b325aecc935a3770d99144f9cc)

commit 93cb149928
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:01 2026 +1000

    SCTP: Bound the fragments stored per message

    Reassembly of unordered DATA chunks put every fragment of a stream
    into one message, because the stream sequence number is forced to 0.
    Each new fragment costs two walks of the fragment list, so a capture
    with many unordered middle fragments took quadratic time. Store at
    most 1024 fragments per message; later ones are not reassembled and
    get an expert info and a Info column note.

    The SACK chunk made the cumulative TSN ack relative to the first TSN
    for display and then passed that value to sctp_ack_block(), which
    subtracts the first TSN again, so SACKs never marked DATA chunks as
    acknowledged with relative TSNs on (the default). Pass the raw
    cumulative TSN to sctp_ack_block() and use the relative value only for
    display, as the NR-SACK code already does.

    The tap was queued in a FINALLY clause even when the chunks threw an
    exception, and the taps read the chunk tvbs outside of any TRY block.
    A chunk of length 0 made tshark -z sctp,stat abort with an unhandled
    exception. Queue the tap only when the packet was dissected completely.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 83577f306268cea942c1e909db94c91ac3f629ab)

commit 43924c3a83
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:01 2026 +1000

    SAP Router: Stop routed messages from dissecting themselves

    A route whose last hop names the SAP Router port (3298 or 3299) made
    the Routed message branch look the destination port up in the sapni.port
    table, which found SAP Router again. It ran on the same data at offset
    0 in the same conversation, reached the same branch and recursed until
    the layer limit failed a DISSECTOR_ASSERT.

    Do not hand the payload to the sub-dissector when the destination port
    is in the SAP Router port range, report an expert info instead, and wrap
    the nested call in increment_dissection_depth() and
    decrement_dissection_depth().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit ed22c4624bd20c8cb77728269381ee0e1c998d97)

commit 66c24a5989
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 22:31:19 2026 +1000

    RNSAP: Do not add a protocol layer per IE, and bound IE nesting

    Every protocol IE and extension was dispatched through its dissector
    table with dissector_try_uint(), which adds an "rnsap" entry to the
    frame's protocol layers each time. A PDU with a few hundred IEs (or
    nested extension containers) reached the layer limit and failed the
    assertion in call_dissector_work() ("saved_layers_len <
    prefs.gui_max_tree_depth"), a dissector bug triggered by the packet;
    with WIRESHARK_ABORT_ON_DISSECTOR_BUG set, tshark aborted.

    Dispatch the IEs without adding a layer, as RANAP, S1AP and X2AP do,
    and since that no longer limits recursion, bound the nesting of IEs and
    extensions to 32 per packet with an expert info.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    [4.4: dissector_try_uint_with_data() is called dissector_try_uint_new() here.]

    (cherry picked from commit 7b0661d4f72cbff74e0b6a1fdc42ecb23b502b7e)

commit c6014f127b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    RSVD: Fix a dissector bug on every SRB status response

    The SRB status response branch added the SrbStatus field with
    proto_tree_add_bits_item(), but the field has a 0x7f bitmask. The bits
    API reports a dissector bug for a field with a bitmask, so every
    RSVD_TUNNEL_SRB_STATUS response raised an exception when a tree was
    built (and aborts with WIRESHARK_ABORT_ON_DISSECTOR_BUG). Use
    proto_tree_add_item(), which applies the mask.

    The length passed to the sub-dissectors was truncated to 16 bits and
    could go negative for tunnelled SCSI data over 32 KB. Use int.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6ae439a1a72d5cb8a8092e0b788873ef562f0b1b)

commit ac1721722b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    RSVP: Bound the Juniper TLV loop, drop the unused request hash

    The JUNIPER_PROPERTIES object loop ran the packet's 16-bit TLV count
    without regard to the object length, so each object re-walked all the
    bytes after it, which is quadratic in the packet size (about 1.3 s per
    64 KB datagram). A TLV length below 2 also moved the offset backwards.
    Stop at the end of the object and report a TLV length below 2, or one
    that runs past the object, with an expert info.

    The request hash was keyed on the conversation index only, so all keys
    between one address pair shared a bucket and every packet walked the
    whole chain. Nothing ever read the hash value, so remove the hash, its
    key and value structures and the conversation lookup that only fed it.

    Small decoding fixes: compare the upstream label with a 32-bit value;
    print the P2MP IPv6 sender with tvb_ip6_to_str() and read the Sub-Group
    ID at the right offset; make the 3GPP2 Profile ID a 16-bit field and
    read Max_Latency, Max_Loss_Rate and Delay_Var_Sensitive with the widths
    they are advanced by; reject a 3GPP2 QoS list length below 2 instead of
    letting it wrap; skip the data of unknown 3GPP2 information elements;
    advance GENERALIZED_UNI sub-objects by their 16-bit length and read the
    egress label direction at the right offset.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 890608bba54499201e9f8efa4de199bab5681362)

commit 596b7bcb92
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Nov 13 21:58:02 2026 -0500

    PKCS12: Don't use a stale or NULL OID for EncryptedData

    The EncryptedData footer passed the file-scope object_identifier_id
    to dissector_change_string(), dissector_reset_string() and
    PBE_decrypt_data(). That pointer is allocated from the packet pool and
    is only set by a SafeBag, CertBag, CRLBag, SecretBag or PKCS12Attribute
    earlier in the dissection, but EncryptedPrivateKeyInfo is also
    registered for pkcs-9-at-encryptedPrivateKeyInfo and can be reached
    without one. The OID was then NULL, crashing in g_str_hash(), or
    pointed into the pool of an earlier packet.

    Remember the bag OID in per-packet proto data and skip the temporary
    dissector swap and decryption when there is none.

    This code was removed from later releases when the PKCS #8 structures
    moved to the AKP dissector (3b74ede300).

    Assisted-by: Claude

commit 2a6b7ec514
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 18:54:24 2026 +1000

    nstrace (mep): Use AT_IPv4 with tvb_address_to_str()

    Eight calls passed FT_IPv4 as the address type to
    tvb_address_to_str(), which selects some other (variable length)
    address type and ends in ws_assert_not_reached(), aborting on every
    such packet. Use AT_IPv4.

    Also fix the length of two items, which was computed as
    start_offset - offset (negative) instead of offset - start_offset.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 57527797d8e6a756dc3897eb08b72c94b48f4b5d)
    [Backport: 4.4 val_to_str() has no wmem scope argument; context
    conflict only.]

commit c9b7bb17d3
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:45 2026 +1000

    XML: Fix a NULL dereference for a top-level CipherValue

    after_untag() looked up the encrypted element as the grandparent of a
    CipherValue element. For a CipherValue directly under the root frame
    the grandparent is NULL and xml_get_tag() dereferenced it, crashing on
    a single UDP 3702 packet containing <CipherValue>QUJD</CipherValue>.

    Check for the missing grandparent and skip the KeyInfo lookup.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 32edeb48f3921b3419d5990fbb7e5ca2711572e9)

commit dcac35e3c4
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:45 2026 +1000

    XMCP: Use unsigned offsets when walking attributes

    The attribute offset in dissect_xmcp_message() and
    decode_xmcp_attr_value() was a uint16_t, but a message length of up to
    65535 plus the 20 byte header can reach 65555. The offset wrapped to a
    small value, the bounds check passed again and the attribute loop never
    ended, using memory without limit when no tree was built. Use unsigned
    offsets, including the one in the FLAGS loop.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6e51f58c7327661c32fa6d8e95f0b500d8503c9e)

commit a756a6faec
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    X509IF: Bound recursion and DN string building

    Refinement, ContextCombination and AttributeCombination recurse through
    "not", "and" and "or" and were bounded only by the dissection depth
    counter, which ends in a DISSECTOR_ASSERT. A certificate extension with
    OID 2.5.18.6 and a few hundred nested "not" values aborted. Stop 16
    levels before the tree depth limit with an expert info.

    Building the DN string copied the whole DN for every RDN and
    registered the frame end routine for every RDN, RDNSequence and AVA,
    which made memory and CPU quadratic in the number of RDNs. Stop
    prepending once the DN string is 1024 bytes long and register the
    frame end routine once per frame.

    Also reset ava_hf_index at the end of the frame, so that an exception
    inside an AVA does not leave a stale format for later frames.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit cbff2d264e422a64ca0bac0cea53386d681d1bde)
    [Backport: pinfo->dissection_depth is an int on this branch, so the new
    depth checks compare it with (int)prefs.gui_max_tree_depth, as packet.c
    does; packet-x509if.c regenerated.]

commit 83dca4f755
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    X11: Fix modifier map OOB read and keysym assert

    A truncated GetModifierMapping reply or SetModifierMapping request
    left the conversation with a large keycodes_per_modifier and the old,
    smaller modifiermap arrays, so the next KeyPress read past the arrays.
    listOfKeycode() now builds the arrays locally and commits them only
    when the whole list was read, and the callers store
    keycodes_per_modifier only after that.

    A GetKeyboardMapping reply with a huge reply length stored the length
    in a signed int and tripped DISSECTOR_ASSERT(keycode_count >= 0). Use
    an unsigned length, clamp the keycode count to 256 and report invalid
    counts with an expert info instead of an assertion.

    A ChangeKeyboardMapping with keysyms_per_keycode 0 created an empty
    keysym array whose first element keycode2keysymString() read
    uninitialized. Reject it in listOfKeysyms() and return "<Unknown>" for
    an empty array.

    Also fix small decoding bugs: the arc width, height and angles were
    all shown with the y value, arc and color item lists were 12 bytes per
    item but sized as 8, consistentWithOrder() and the opcode check used
    absolute offsets instead of offsets relative to the request, and the
    keysym selection for the Lock case used a stale value instead of
    syms[1].

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6d804dcf3ce569e8b9c3d1589d356f101c78941b)

commit 28c10b2ebc
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    WTP: Refuse nested concatenated PDUs

    A concatenated PDU (first byte 0) was dissected by recursing into
    dissect_wtp_common() for every sub-PDU, and a sub-PDU could itself be a
    concatenated PDU. Each level needs only 2-3 bytes, so one packet could
    nest about 10900 levels deep, which is a stack exhaustion risk when no
    tree is built. The specification does not allow a concatenated PDU
    inside another one, so a nested one is now reported with a new expert
    info (wtp.nested_concat) and not dissected. Recursion depth is thus at
    most one level.

    Also pass ENC_NA instead of the packet byte as the encoding of the
    continue flag item.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit cb00d7b716ae09d7ae21661038a8588771fbb7d9)

commit d921adde3d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    WSMP: Use an unsigned offset in the legacy element loop

    The offset was a uint16_t, so in a frame longer than 64 KB without a
    terminating element it wrapped to 0 before reaching the end of the
    tvb, and the loop never ended. Use an unsigned offset so the tvb
    bounds exception ends the loop.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 083b94e1a6d85e0af8dcc6eaf15326823fae5d17)

commit c6d4451aff
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    WoWW: Fix CPU blow-ups on unterminated strings and packed GUIDs

    A C string without a terminating NUL in the next 32 KB made
    get_null_terminated_string_length() return 0, so add_cstring() did not
    advance the cursor. Loops driven by a packet supplied count then ran up
    to 2^32 times, each scanning 32 KB, and the AddonInfo loop in
    CMSG_AUTH_SESSION did the same over the decompressed data. Find the NUL
    with tvb_find_uint8() and, when there is none, consume the rest of the
    tvb so the item ends at the boundary or throws.

    add_packed_guid() and add_update_mask() created a child tvb and a data
    source for every GUID and every update mask. add_new_data_source()
    appends to a list, so a frame with many packed GUIDs cost quadratic
    time and one allocation per GUID. Add the GUID as a uint64 value over
    the bytes it occupies, and the update mask index as a uint value, with
    no extra data source.

    The packet size in add_header_to_tree() was a uint16_t and wrapped for
    large size fields; use unsigned.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit b58049b154f8e2639680b32d7392630ab456e8c0)
    [Backport: 4.4 allocates from wmem_packet_scope() instead of
    pinfo->pool, so add_update_mask() no longer uses pinfo and marks it
    _U_; get_null_terminated_string_length() uses tvb_find_uint8() because
    4.4 does not have tvb_find_uint8_remaining(). 4.4 also lacks e32674357c
    (packet-woww.c: Avoid use of uninitialized memory), so
    add_packed_guid() still used wmem_alloc(); this change removes that
    buffer and also fixes that uninitialized read.]

commit b793bf10c8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    WLCCP: Bound TLV nesting and stop asserting on sibling count

    dissect_wlccp_tlvs() recursed with "_depth++", which passes the
    parent's current value and increments the parent's counter once per
    child. The counter therefore counted siblings, not nesting: a container
    with 101 child TLVs hit DISSECTOR_ASSERT(_depth < 100), while nested
    containers always got depth 0 and were not bounded at all, so about
    700 levels of 8 byte TLVs reached the tree depth limit.

    Pass _depth + 1 to the child and stop with a new expert info
    (wlccp.tlv_too_deep) at a nesting level of 32. The function now takes
    pinfo. The "offset > old_offset" asserts become loop exits.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5585326710ebe5ceb13599ebee6514753a65cde7)

commit 4abc016918
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:44 2026 +1000

    SMB: Bound conditional ACE nesting and SRA value counts

    A conditional ACE could nest COMPOSITE tokens about 13000 levels deep
    (five bytes each), reaching the tree depth limit and a dissector bug
    assertion. Pass a nesting level down and stop with an expert info at
    32 levels.

    The SYSTEM_RESOURCE_ATTRIBUTE ACE did not check its value count or
    value offsets against the ACE size, so many offsets pointing at one
    long string cost O(N*S) time and memory. Cap the value count by what
    the ACE can hold (size / 4) and skip value offsets beyond the ACE,
    both with expert infos.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 7be2c644eb32216a93ef2819dd944248317ff367)
    [Backport: in 4.4 dissect_nt_ace_system_resource_attribute() and
    dissect_nt_ace_system_resource_attribute_value() do not take pinfo;
    added a pinfo parameter to the former (passed from dissect_nt_v2_ace())
    for the new expert infos and left the value function unchanged.]

commit e359a11361
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:43 2026 +1000

    Wi-Fi NAN: Bound nesting of vendor specific attributes

    The NAN Vendor Specific attribute passed the rest of the frame, not its
    own length, to the 802.11 vendor IE dissector, which reaches the NAN
    beacon dissector again. The outer loop then dissected the same bytes
    once more, so k back-to-back attributes took 2^k - 1 dissections and a
    single beacon exhausted CPU and memory. Nested attributes also
    recursed until the stack overflowed.

    Pass only attr_len bytes to the nested dissector, and count the nesting
    per packet with p_get_proto_depth(). Beyond 8 levels an expert info is
    added and the nested IE is skipped. The same guard is used for the
    Element Container attribute.

    Also fix the Extended Mesh attribute's mesh ID length, which mixed the
    attribute length with a frame offset; it is now attr_len - 13.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit b40d9f3dd79ce284530958af614dee673191f251)
    [Backport: 4.4 calls dissector_try_uint_new() instead of
    dissector_try_uint_with_data() and does not include epan/tfs.h or
    wsutil/array.h; kept those, added only epan/proto_data.h, and replaced
    the vendor specific IE tvb created with tvb_new_subset_length(..., -1).]

commit 0e1c6460fa
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:43 2026 +1000

    WebSocket: Cap decompressed permessage-deflate output

    A permessage-deflate message was inflated with only a 2 GiB limit and
    the output was kept in file scope for later passes. Deflate expands
    about 1000:1, so a small capture of compressed frames exhausted memory
    with default settings. Stored output is now capped at 10 MiB per
    message. The rest is still inflated and discarded so that the sliding
    window stays in sync for later messages, and a new expert info reports
    the truncation.

    Restore pinfo->fragmented when a continuation fragment returns early,
    otherwise a following complete data frame in the same segment was not
    dissected.

    Clamp the 64-bit extended payload length instead of truncating it to 32
    bits, so that a large length is no longer decoded as a small one and the
    frame length cannot wrap.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit d18c887fcb7bb49aededf62b7bb270d84018ab53)
    [Backport: 4.4 declares websocket_uncompress() with separate zlib-ng
    and zlib prototypes instead of zlib_streamp; added the truncated
    parameter to both.]

commit 2afd545da9
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 13 13:32:43 2026 +1000

    WASSP: Bound sub-TLV recursion and avoid a conversation assert

    dissect_wassp_sub_tlv() recursed once per nested block TLV with no
    limit (about 16000 levels, stopped only by a 16 bit offset wrapping).
    Pass a depth and stop at 16 levels with a new expert info. Real
    messages nest only a few levels. Also make org_offset an int so it
    does not truncate offsets above 65535 in reassembled messages.

    A fragmented message reached through Exported PDU has no addresses or
    ports, so find_conversation_pinfo() returned NULL and
    DISSECTOR_ASSERT(conv) failed. Use find_or_create_conversation().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 9ab310f634967bd136691472af67b8e8b9644546)

commit 352f44d748
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    USBIP: Initialize the fallback transaction

    When a return PDU had no matching command, the fallback transaction was
    allocated without clearing it, so the endpoint and direction taken from
    it were uninitialized and chose the USB endpoint and conversation. Use
    wmem_new0().

    Also set ret_frame of a victim transaction in the unlink response only
    on the first pass, and attach the "not implemented" expert infos to the
    USBIP protocol item, which was NULL or an unrelated generated item.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 920bf1c9c4b06c23b7493393d9865329449a493f)

commit 5484468a96
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    USB HID: Bound Collection nesting, report sizes and usage expansion

    A report descriptor made of Collection items nested more deeply than
    the tree depth limit reached an assertion in the recursive item
    dissector. Nesting is now limited to 32 levels; beyond that an expert
    info is added and the rest of the descriptor is skipped.

    A Report Size above 32 bits, or a report_size * report_count above
    65535 * 8 bits (computed in 64 bits), is now rejected when the
    descriptor is parsed. Sizes above 64 bits made the bit item functions
    report a dissector bug, and the 32-bit product of an array field could
    wrap and let the per-element loop run about four billion times.

    Usage ranges were limited per field only, so a short descriptor could
    allocate hundreds of megabytes of file-scope memory. The expanded
    usages of one descriptor are now limited to 131072 in total.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit baa7f9374fd373e6b11e12c04eea23ce146d94d3)

commit 6f15757f54
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    USB: Fix FreeBSD empty composite assert and SET_INTERFACE slowdown

    A FreeBSD control transfer with two or more zero-length data frames
    counted each of them, but tvb_composite_append() drops empty members, so
    the composite had no members and tvb_composite_finalize() failed its
    assertion. Count only frames that contribute data.

    Every INTERFACE descriptor appended an entry to the interface's
    alternate settings array, and every SET_INTERFACE request scanned the
    whole array, which made dissection quadratic in the file size. Keep one
    entry per bAlternateSetting, which bounds the array at 256 entries.

    A string descriptor request with wLength below 2 made the string length
    wrap to 254 or 255 and the packet was reported as malformed. Use a
    length of 0 in that case.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 69af274413ceb65563480234f5683bbe358dab06)
    [Backport: 4.4 does not assemble the data frames of a FreeBSD control transfer into a composite tvbuff, so the empty composite fix has nothing to apply to and is dropped; the alternate settings and string descriptor changes apply as they are]

commit bad7438cdd
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    UMTS RLC: Bound RB id, duplicate history and reassembly channels

    The radio bearer id from the packet indexed the RRC ciphering
    tables (31 entries) unchecked in is_ciphered_according_to_rrc() and
    rlc_decipher(), a heap read past the end with RLC over UDP. Ids of
    31 or more are now treated as not ciphered and skipped.

    rlc_is_duplicate() kept every PDU of a channel in a list searched
    linearly for each PDU, so the cost grew quadratically (40000 frames
    took 32 s). The history is now capped at 1024 entries per channel
    (the SN window is 1024 for AM, so real traffic does not reach it),
    oldest first, and the list is only modified on the first pass.

    Each new channel allocated a 32 KiB reassembly table. The number of
    channels with such a table is now capped at 1024; further channels
    are not reassembled.

    get_reassembled_data() looked up with an uninitialised channel when
    the channel assignment failed; rlc_frag_assign() now returns the
    result and the caller gives up.

    The incomplete-sequence check now handles SN wrap-around, and the
    fragment list print no longer wraps for a zero-length fragment.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 92a4af027f158e9de3e1df82c25273edbc529237)

commit 72fcb0abc9
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    UMTS MAC: Reject HS-DSCH C/T 15 and harden MAC-is reassembly

    An HS-DSCH PDU with the C/T field set to 15 gives logical channel
    id 16, which indexes lchId_type_table and lchId_rlc_map (16 entries)
    one past the end. Report the reserved value with an expert info and
    discard the PDU, as RACH and FACH already do.

    MAC-is reassembly asserted on a duplicate head, middle or tail
    segment for the same TSN and on an impossible SS/SDU combination;
    both now add an expert info and discard the segment. find_head and
    find_tail looped forever when every TSN slot held a middle segment;
    they now stop after MAX_TSN steps. The TSN is read with
    tvb_get_bits16() so a 14 bit TSN no longer fails an assertion, and
    the fragment data is copied before the fragment is linked so a
    throwing copy cannot leave uninitialised data behind.

    The C-RNTI was added to the tree from a fixed bit offset 4 instead
    of the current bit offset.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 136f41990a7b48c1cdecf698ef76940a1e80fa51)
    [Backport: context conflict: 4.4 returns tvb_new_subset_length_caplen(..., -1) for a complete MAC-is SDU where master uses tvb_new_subset_length(); keep 4.4's call]

commit dc230cafe7
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    UMTS FP: Fix PCH assertion and paging indication bitmap over-read

    update_pch_conversation_info() asserted that the conversation channel
    is PCH. fp_info from FP Hint, DCT2000 or K12 need not match the
    conversation found by address and port, so a crafted Exported PDU file
    reached the assertion. Return when the conversation is not a PCH one or
    has no channel specific info.

    The relevant paging indication bitmap was stored with the length of the
    frame that carried it and shown with the length of a later frame, which
    could be longer (heap over-read) or zero (NULL data, assertion in
    tvb_get_ptr). Store the bitmap length with the bitmap, and show the
    bitmap only if it exists, using the stored length.

    The NBAP lchId tables have 16 entries but the fake logical channel id,
    which starts at 5 and grows for every new transport channel id, was
    used as an index without a bound. Use 0 for ids of 16 or more.

    The RACH new IE present flags were set whenever the division matched,
    not from the flag bit; take them from the bit. fp.hsdsch-interval and
    fp.pusch-duration hold values up to 2550 and are now FT_UINT16. The
    calculated HS-DSCH rate used integer division that gave 0 for
    intervals above 100; compute it in 64 bits and clamp to 32 bits.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 301d913f187cf3f3aea870cede96756a3500a7d7)

commit 083d2a3e26
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:02 2026 +1000

    UAVCAN-CAN: Fix a NULL dereference when a frame is dissected again

    In a multi-frame transfer, a frame without the start-of-transfer bit
    is skipped on the first pass when its transfer has not been seen yet.
    Such a frame gets no per-frame data. A later start frame with the same
    transfer id creates the transfer, so on the next pass (tshark -2, or
    the GUI redissecting) the earlier frame was accepted, took the visited
    branch, and dereferenced the missing per-frame data. Decide on later
    passes as the first pass did: accept a multi-frame frame only if it has
    per-frame data, and also return if that data is missing.

    A reassembled payload shorter than the two-byte CRC made the CRC
    offset and length wrap. Add an expert info and skip the CRC and DSDL
    dissection in that case.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 02f8fec251b529bff49e2aeb4843d3a14da91b11)

commit 7c4f7501f7
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:44:01 2026 +1000

    TR: Bound the old Linux detection and capture reads by the length

    check_for_old_linux_tvb() fetched at most 19 bytes but compared up to
    2*x bytes, reading beyond the pointer. Stop when 2*x exceeds the number
    of bytes available.

    capture_tr() called check_for_old_linux() unconditionally and read
    bytes in the frame without checking the length, including pd[0x20..0x27]
    and the byte after the 14-byte header. It also used 0xe0 instead of
    0x0e as the offset of the LLC header after the RIF, which read about 255
    bytes out. Run the Linux workaround only when fix_linux_botches is set,
    bound every read by len and fix the offset.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 5319ff29ec36d3acc880f8f7387b593053565bd9)
    [Backport: context conflict: 4.4 reads first2_sr with pntoh16() where master uses pntohu16(); keep pntoh16() with the corrected offset]

commit 0698cf0c24
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 14:36:49 2026 +1000

    reassemble: Find the next fragment number from the first gap

    fragment_add_seq_next() and the other REASSEMBLE_FLAGS_NO_FRAG_NUMBER
    users walked the whole fragment list for every new fragment to find
    the next fragment number, so a long sequence was quadratic. A USB-MIDI
    SysEx message that never ends, 3 bytes per fragment, took 25 seconds
    for a 165 KB capture. Start the walk at first_gap, the last fragment
    before the first gap: when the fragment numbers are generated there is
    no gap, so it is the last fragment, and the result does not change
    otherwise.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 7bd8c2870c3cf2a6ed16a150bb1ad52007c20eef)
    [Backport: comment-only conflict: 4.4 does not have the XXX comment about first_gap that master replaced]

commit b5fe478d36
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Nov 13 06:51:56 2026 -0500

    GNSS: ensure GAL nominal page type for data word handoff

    The GAL data word handoff needs the GAL even and odd pages within the
    same UBX-RXM-SFRBX message to be of nominal page type (not alert type).
    Ensure that this is the case (also fixing a related clang analyzer
    finding).

    (cherry picked from commit ddae4452a976fa1e844e8337dbf1bd85462fe16b)

    Co-authored-by: Timo Warns <timow+git@DiningPhilosopher.DE>

commit 64ec145eb7
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 06:57:17 2026 +1000

    ZVT: Bound recursion through TLV subsequences

    dissect_zvt_tlv_subseq() calls dissect_zvt_tlv_seq() again for the
    subsequence tags, and nothing limited the depth. Two bytes per level
    nest about 32000 levels in one APDU, which overflows the stack without
    a tree and hits the tree depth limit with one. Count the nesting per
    packet with p_get_proto_depth()/p_set_proto_depth(), stop at 8 levels
    and add a new expert info (zvt.tlv.too_deep).

    When dissect_zvt_tlv_len() fails, data_len kept the value of the
    previous object and was used to advance the offset. Stop the loop
    instead.

    The time and date fields copied fixed digit positions out of the BCD
    string even when a stop digit ended it early, copying uninitialised
    bytes. Show the short string as it is in that case.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 3b82a94e88371329bf03272fde24bbdcb5f7c1df)
    [Backport: context conflicts: 4.4 does not include epan/tfs.h and its APDU length expert info text differs; keep both.]

commit d2da36e546
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 06:57:17 2026 +1000

    ZigBee ZCL SE: Initialize Valid To time and fix several field decodings

    In the Key Establishment suite 2 certificate, valid_to_time was only
    set when the Valid To field was not 0xFFFFFFFF. For 0xFFFFFFFF
    ("does not expire") uninitialized stack memory was stored as the value
    of the zbee_zcl_se.ke.cert.valid.to field and showed up in -T json,
    -T fields and display filters. Initialize it to zero.

    The Change Debt label length was a uint8_t, so a length octet of 0xFF
    wrapped the label length to 0 and the following fields were decoded
    from the label bytes. Make it unsigned.

    The Get Scheduled Events Minimum Issuer Event ID was an FT_UINT8 field
    added with the remaining length. Make it FT_UINT32 with a length of 4.

    Read the Proposed Tenancy Change Control bitmask and the Request New
    Password duration in little endian, as everywhere else in ZCL.

    Decode the Get Profile Response End Time and the Display Message Start
    Time with ENC_TIME_ZBEE_ZCL so they use the ZigBee epoch. The value 0
    is still shown as "Now": the End Time field already had the now_strings
    table, and it is added to the Start Time field.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 7d5105d81ea4278a30d874e90de255838f43c467)
    [Backport: 4.4 shows the Get Profile Response End Time and the Display Message Start Time as FT_UINT32 with custom ZigBee time formatting, which already uses the ZigBee epoch, so the ENC_TIME_ZBEE_ZCL and field changes are dropped.]

commit 870aa8e904
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 06:57:17 2026 +1000

    ZigBee ZCL: Bound ARRAY/SET/BAG nesting and fix structured write offsets

    An ARRAY, SET or BAG attribute value can have an element type that is
    itself an ARRAY, SET or BAG, and each level costs only three bytes.
    The dissection recursed without a bound until the proto tree depth
    limit ended it with a "Dissector bug" error. Limit the nesting to 16
    levels with a per-packet counter (p_get_proto_depth()) and add an
    expert info, zbee_zcl.too_deep, when the limit is reached.

    Write Attributes Structured and its response read the selector
    indicator and the indexes at offset 0 instead of the current offset,
    so they showed the frame control byte and misaligned what followed.
    Use *offset.

    The Command Identifiers subtree of the Discover Commands Received and
    Generated responses had the negative length *offset - tvb_len. Use
    tvb_len - *offset and drop the meaningless loop bound. The minimum and
    maximum reporting intervals and the timeout period of the reporting
    configuration records were added to the parent tree instead of the
    record subtree; add them to the subtree.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit fcd897ecaba86c41a45e430c9e6a3b55f84af8f3)
    [Backport: context conflicts: 4.4 reads the selector indicator with tvb_get_uint8() after adding the item, so keep that and add the item at *offset; keep 4.4's else branch for the timeout period, adding it to the record subtree; 4.4 has no Command Identifiers subtree in the Discover Commands responses, so that hunk is dropped. dissect_zcl_attr_data() and the array and set helpers have no pinfo on 4.4 (master added it in 50f737a9b676a56b929566bdad061a6caabc679f, which changes a public function and so the ABI), so count the nesting in a file-scope variable, decremented through CLEANUP_PUSH() so that an exception does not leave it raised, and add the expert info with a NULL pinfo.]

commit b360edbaef
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 06:57:17 2026 +1000

    ZigBee TLV: Fix status map lookup and TLV length handling

    The Network Status Map TLV indexed the joined status string table
    with a 3 bit value, but the table has only 5 entries and a terminator,
    so values 5 to 7 printed a NULL string or read past the end of the
    array. Use val_to_str_const() with a "Reserved" default.

    The total TLV length and start offset of the ZDP and ZigBee Direct
    local TLV dissectors were uint8_t, so they wrapped for large offsets
    and lengths and the extra-bytes check was wrong; use unsigned. The
    extra-bytes item now covers the bytes not yet dissected instead of a
    fixed length. Use 1U when shifting by up to 31 bits to avoid undefined
    behaviour, and check the NWK hints for NULL before using them.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit e4cc19e29e3b8781c14454d36f469e5737beee27)
    [Backport: context conflict: 4.4 looks up the NWK protocol with proto_get_id_by_filter_name() and has a different case layout for the relay commands; keep those and add the NULL checks of the NWK hints.]

commit 8e095b83c5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 06:57:17 2026 +1000

    ZigBee NWK GP: Bound the key ring and fix length handling

    Every Commissioning and Commissioning Reply frame carrying a key
    prepended a record to the global key ring on every pass, with no
    duplicate check. Each fully encrypted frame tries every key, so N such
    frames and M encrypted frames cost N x M AES-CCM* decryptions per pass
    (117 s for 16000 each), and the ring grew again on each redissection.
    Use one helper at all four sites that adds a key only on the first
    pass, skips a key already anywhere in the ring and stops adding when
    the ring holds 256 keys, with a new expert info
    (zbee_nwk_gp.keyring_full), remembered per frame so that later passes
    report it too.

    The payload length was a uint8_t, so a 256 byte payload was reported
    as missing and longer ones were truncated. Make it unsigned and check
    that the payload exists before subtracting. The Read Attributes loop
    used a uint8_t counter that wraps for a list length of 255 and read
    one identifier too many for odd lengths; use an unsigned counter and
    require a full identifier. The heuristic now checks that the first
    byte exists before reading it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit e276f23421ce65e17ccb377752889eeb9c2731ab)
    [Backport: context conflict: 4.4 tested the uint8_t payload length for <= 0 and added the expert info with proto_tree_add_expert(); take master's check of the reported length.]

commit ed8019ca00
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 06:57:17 2026 +1000

    ZigBee Direct: Bound the session key ring and fix key checks

    Every Dump Info notification carrying a session key added a key to
    the key ring, and every later encrypted packet tried every key twice,
    so a capture with N keys and M encrypted packets cost N x M x 2 AES-CCM
    operations (8000 x 8000 took 87 seconds). Skip a key that is already in
    the ring and keep at most 32 keys reported over the air; further keys
    are ignored with an expert info. Keys from the UAT are not counted.

    The ignore_late_keys test for the per-PAN NWK keyrings was inverted and
    used only the keys that appeared after the packet; keep the keys at or
    before it. A UAT row was accepted when any one of its three fields
    parsed; require all three. Use the size of the zvd_ieee field when
    copying the ZVD address.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit b45d8ee04d24a39b1bfaa968ccd8b0ff42687c3c)
    [Backport: context conflict: on 4.4 the key record's frame number and label are set before the duplicate check; set them only when the key is added, as on master, so the label is not leaked for a skipped key.]

commit 107bda3e77
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 12 07:27:31 2026 +1000

    YAMI: Bound nested parameters and check the message header size

    Nested parameters recurse with only increment_dissection_depth(),
    which asserts at the tree depth limit. At 12 bytes per level, a
    parameter nested 600 deep (7.2 KB, reachable through an Exported PDU
    naming "yami") ended in a dissector bug and aborted under
    WIRESHARK_ABORT_ON_DISSECTOR_BUG. Stop at 32 levels with an expert
    info.

    A message header size below 4 moved the data message parse back into
    the frame header. Add an expert info and skip the message, and compute
    the frame size in 64 bits so a large payload size does not overflow.

    A binary value or binary array element was added with the pointer to
    its data but the length of its length field, data and padding, so
    proto_tree_add_bytes_format_value() copied up to 7 bytes more than
    tvb_get_ptr() had checked; a binary value at the end of a frame read
    past the buffer. Add the item with the length of the data. The items
    of a string array now cover the string, not only its length field.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 8c4d06d08c3c91a5a4fd9732d1db5bf6ae3f8500)
    [Backport: context conflict: 4.4 read the header and payload sizes with tvb_get_letohl() before adding the items; use proto_tree_add_item_ret_int() as on master (both fields are FT_INT32).]

commit 9f9617eb3f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    TN5250: Bound nested Restore Screen and fix several offsets

    A Restore Screen command dissected the rest of the buffer as further
    outbound commands, so a chain of ESC 0x12 pairs recursed once per two
    bytes until the tree depth limit made the packet a dissector bug. Keep
    a per-packet nesting count and stop with a new expert info
    (tn5250.restore_screen.nested) when more than 4 Restore Screen commands
    are nested.

    Fix offset arithmetic. The remaining length in WSSF Image/Fax Control,
    Image/Fax Download, True Transparency Write and WSF Pass-Through was
    computed as length - (start + offset) instead of length - (offset -
    start), which dropped valid data or moved the offset backwards; only
    add the item and advance when something remains. Erase to Address did
    not step over the row and column, Video/Audio Controls advanced 2
    bytes after a 4 byte field, Restore Partial Screen advanced 1 byte
    after a 2 byte length and did not advance past its data, Write Text
    Screen did not advance past the line image, and dissect_wcc always
    returned 2 even when an 0xFF escape byte was skipped.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit f712a4c9f4baf92c4acfd262f550c4c0f46bf94f)
    [Backport: include conflict only; only epan/proto_data.h is added.]

commit 67fc15ad64
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    TLS: Fix handshake reassembly asserts and CPU blow-ups

    A ClientHello followed by a handshake fragment in the same record made
    the reassembly key (which contained the packet direction, recomputed
    after ssl_set_server() changed the server address) disagree with the
    reassembly id, so the next fragment hit DISSECTOR_ASSERT(fh). The
    reassembly ids are unique, so the table now compares the id only.

    The CCS frame numbers were updated on every pass. A later pass could
    therefore treat a partial handshake message as plaintext that the first
    pass had considered encrypted, and fail DISSECTOR_ASSERT(frag_info).
    Update them on the first pass only, and show the data instead of
    asserting if a later pass still finds no fragment.

    The connection ID session list got one entry per connection_id
    extension, and it is walked for every DTLS record. Add a session only
    once (flag in SslSession).

    The JA4 cipher and extension lists were built with
    wmem_list_insert_sorted(), which is quadratic. Append to arrays and
    sort once.

    Also bound the supported_groups and ech_outer_extensions loops by the
    vector end, limit ECHConfig contents to the ECHConfig length, loop over
    the CertificateURL list by byte length, use ENC_NA for the cert types
    item, and set alpn_name and the ALPN app handle on the first pass only.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 19312eb820d2f748c8279d526cac1c88b1ec6443)
    [Backport: this branch has no ECH outer extensions dissector, so that
    loop bound is not taken, and SslSession only gains cid_listed.]

commit 1b2c8342ae
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    Tibia: Fix heap over-read parsing world list IP addresses

    ipv4tonl() skipped one character after each octet without checking for
    the end of the string, so an empty or short IP string in a world list
    entry made it read past the terminating NUL of the string buffer.
    Stop and return 0 when an octet cannot be parsed or is not followed by
    the expected '.'.

    Also avoid printing a NULL error string when the RSA decryption fails
    without setting one.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 7ddfd4103ab786672a54897c69def2f55722aef4)

commit 87198cce82
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    Thrift: Stop at the nesting limit in the binary protocol

    When the nesting limit (default 25) was reached, the generic
    TBinaryProtocol dissect_thrift_binary_linear() and
    dissect_thrift_binary_struct() returned THRIFT_SUBDISSECTOR_ERROR
    without advancing the offset. Their callers only treat
    THRIFT_REQUEST_REASSEMBLY as an error, so a list at the limit with
    elements of type list, set, map or struct and a count of up to
    0x7fffffff looped on the same offset, adding an expert info on each
    pass. A single small packet could use gigabytes of memory and a lot of
    CPU time. Return THRIFT_REQUEST_REASSEMBLY like the compact protocol
    functions do, so the callers stop.

    For the old, non-strict binary message header the message type was read
    at offset + length field + name length, although the offset was already
    past the length and the name. Read it at the offset.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e85b7b8eb6192af6492085c428b9ae77ac8f7070)

commit 2c03eb89c2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    TFTP: Fix Export Objects invalid free and reachable assertion

    In Export Objects, a later pass could keep using the file-scope wmem
    buffer left over from the first pass and hand it to the tap, which
    releases it with g_free(). That is an invalid free and corrupts the
    heap. Always keep the payload buffer in file scope and always give the
    tap a copy.

    A second transfer in one conversation that is larger than the first
    tripped DISSECTOR_ASSERT_CMPINT() on the payload offset. Check the
    bounds instead and drop the inconsistent object.

    Also make the heuristic dissector fall back to the conversation if
    none was stored on an earlier pass.

    The filename and payload copies were made in the dissector and freed
    by the Export Objects tap callback, so they leaked whenever the queued
    tap data was not delivered, as in the first pass of tshark -2. Pass
    pointers to the dissector's memory and make the copies in the tap
    callback.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5610e666dbaafd3d3eee76bb70e67c2287471f4d)

commit dc56b863c9
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    TDS: Fix an endless loop in ALL_HEADERS and several decoding errors

    dissect_tds_all_headers() added the 32-bit header length from the
    packet to the unsigned offset without a check. A length such as
    0xFFFFFFF8 wrapped the offset back to an earlier header, so the loop
    never ended and burned CPU on a single SQL batch packet. Reject a
    header length that runs past the ALL_HEADERS length, and a total
    length smaller than its own 4 byte field.

    Small decoding fixes:

    The decision whether a packet goes through in-order reassembly was
    taken from a flag that is set while dissecting, so a second pass
    could choose differently and show a single-buffer packet as Data.
    Store the decision per frame with p_add_proto_data() and set the flag
    only on the first pass.

    The TDS 4/5 login packet now changes the conversation encodings only
    on the first pass, and shows the int4 option with the int4 value.
    The capability length clamp was reversed and wrapped, and the
    capability loop did not count the length field itself. The param and
    row format token lengths are 2 bytes and now use the int2 encoding.
    A language token with length 0 no longer underflows. The nanoseconds
    of TIMEN, DATETIME2 and DATETIMEOFFSET were always 0 because of a cast
    placed before the multiplication, and TIMEN lengths above 5 are
    rejected. smalldatetime used a length of 8 for a 4 byte value, and
    the UDT assembly qualified name length no longer wraps in 16 bits. The
    TDS 4 and 5 protocol preferences left the date encodings of the
    conversation uninitialized.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 527bdf9f20cf712145f273d08b6dcfb3e42e4019)
    [Backport: the colmetadata assembly qualified name length is read with
    tvb_get_uint16() on this branch; resolved to master's version. The offsets
    are int here, so the remaining length is cast for the header length
    check.]

commit 37f6e9c8eb
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    TCPCL: Bound segments per transfer, fix session extension table

    The segment and ack lists of a transfer are searched linearly for every
    XFER_SEGMENT and XFER_ACK, and fragment_add_seq_next() walks the
    fragment list of the transfer on every segment. The transfer ID comes
    from the packet, so a capture with many small segments of one transfer
    makes the dissection quadratic (40000 segments of 18 bytes take 93
    seconds). Track at most 5000 segments and 5000 acks per transfer; a
    further segment is not analyzed or reassembled and gets an expert info.
    The sequence bookkeeping now also runs when the sequence analysis
    preference is off, so that the limit applies to reassembly alone.

    The SESS_INIT session extension items were looked up in the transfer
    extension table, so a session extension of type 1 was shown as the
    transfer extension "Transfer Length". Use the session extension table.

    The TCPCLv3 SHUTDOWN message advanced the offset by 1 after the 2 byte
    delay field; advance by 2.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit cfcf58c410e4ed20548062cbbfdd6cfca42cf7db)

commit b11c7148c6
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    TCP: Fix NULL dereferences and asserts in option and analysis code

    The TCP option dissectors in the "tcp.option" table assume they are
    called from tcp_dissect_options() with a struct tcpheader and an
    existing TCP conversation. An Exported PDU record naming that table
    passes NULL data and may have no conversation, so the SACK-permitted,
    MSS, TFO, MPTCP and SNACK dissectors dereferenced a NULL pointer, and
    the MSS, window scale, SACK, SCPS and SNACK dissectors looked up the
    proto data of a NULL conversation. Return early when the header is
    missing and tolerate a missing conversation.

    The SCPS extended capabilities loop used an 8-bit offset that wrapped
    for options longer than 255 bytes, so it never ended. Use wider types.

    Replace the DISSECTOR_ASSERT in the MPTCP mapping lookup, which fails
    when a DSS mapping starts inside the segment, with the missing-mapping
    expert info. Skip checksum verification, instead of asserting, when the
    address type is not IPv4, IPv6 or ILNP. Cap the number of duplicate
    frames listed in the tree, which was quadratic in the size of the
    group.

    The SNACK option read the ACK number at offset 8 of the 6-byte option,
    which threw on every valid option; take it from the TCP header. Pass
    true as more_frags to fragment_add_out_of_order() instead of a
    sequence number.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 9123fd80a1c2d21aa59673e3edb9dcf7ace88833)
    [Backport: the duplicate frame list cap is not taken, since the TCP
    duplicate packet detection it bounds is not on this branch.
    find_conversation_strat() takes three arguments here, and the MSS option
    dissector does not look up the conversation, so it only gets the NULL
    header check. There is no ILNP address type, so the checksum is verified
    only for IPv4 and IPv6, the cases the address switch handles.]

commit 024bb4e054
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:29 2026 +1000

    sysdig-event: Bound the parameter count and fix Info column lengths

    The V2 and V2 large header length dissectors looped over the 32-bit
    nparams value of the record, up to four billion iterations, even when
    there was no tree. Clamp the count to the number of length fields that
    fit in the tvb.

    The Info column code always read 2-byte parameter lengths, which is
    wrong for V2 large records that use 4-byte lengths. Also add the missing
    break after the FT_UINT64 case.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 54f2354802736954ddc87a34b08c91a2314b2a9c)

commit 456b7e5cb2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    STANAG4607: Stop dwell target loop that consumes no bytes

    A dwell segment carries a 16 bit target report count. When the
    existence mask has none of the target report bits set,
    dissect_target() adds nothing and does not advance the offset, but the
    loop still ran count times, up to 65535 per 48 byte dwell. A 64 KB
    packet holds about 1350 such dwells, so a 2.6 MB capture took about 27
    seconds. Leave the loop when a target report consumes no bytes.

    prt_ba16_none() compared a 16 bit value against 65536, which is always
    true, so the Nominal Sensor Cross Range Standard Deviation always showed
    "No Statement". Compare against 0xffff like prt_none16().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 9739db6a58cea293ad2d8b5715a0ae633b8f3e21)

commit ddfd1bad24
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SPICE: Reject wrapping message lengths

    A server message size close to 4G wrapped the PDU length (size plus
    header) to a small value in the data state. The message was then
    dissected as unknown, the offset did not advance and the loop over the
    PDUs in the frame never ended, hanging tshark at 100% CPU. The same
    wrap was possible in the four SASL loops, which were only stopped by an
    exception. Reject lengths that would wrap with a new expert info
    (spice.invalid_length), and also stop the data loop if a PDU made no
    progress.

    Display the SASL data message length big endian, as it is read, and
    write the server link capabilities offset as the sum of its parts
    instead of relying on the link header size being equal to them.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit efeeb70943092d986e1fe2fab8d1b9af0974c407)

commit f1c80e078b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SPDY: Keep decompressed headers per frame

    Every decompressed header block was appended to a file-scope list that
    was searched linearly, keyed only by stream ID and frame type, for every
    header frame. With many frames on different streams this took quadratic
    time (60000 frames took 46 s). The search also skipped the first list
    entry, so on a second pass the first header frame was decompressed again
    with an advanced zlib stream and failed, and blocks were shared between
    conversations that reuse stream IDs.

    Keep the list of decompressed blocks per frame with p_add_proto_data()
    instead, so a lookup only walks the blocks of the frame being dissected
    and later passes find what the first pass stored.

    Show the GOAWAY status with the GOAWAY status names and remove a stray
    closing parenthesis from its text.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 8b285b699627fb7b7905bc783112480e7fcf2ab8)
    [Backport: context conflicts only. spdy_data_frame_t is still defined next
    to the removed header_info_list on this branch and is kept, only the
    epan/proto_data.h include is added, and val_to_str() takes no allocator.]

commit f2dae2682f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SOMEIP-SD: Fix endless loop in the configuration option parser

    The configuration string was walked with a uint8_t position, so for a
    string longer than 255 bytes the position wrapped to 0 and the loop
    never ended, which hung tshark with a CPU spin (reachable through
    Exported PDU without any preference). Use an int position and read the
    element lengths from the tvb instead of the decoded string, which also
    keeps the element boundaries in step with the tvb offsets for non-ASCII
    bytes.

    Add an expert info instead of continuing for an option whose length
    leaves no room for the configuration string, and remove a dead second
    truncation check on the entries array.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 7357837e42adbfbbfa09553e96b4b0bbf86c1ab6)

commit 998314e3d9
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SolarEdge: Fix an out of bounds read and device data recursion

    The OPTIMIZER2 device record fetched 6 bytes with tvb_get_ptr() but then
    read the 7th byte, so a record ending at the end of the packet read one
    byte past the buffer. Fetch 7 bytes, so a short packet throws instead.

    dissect_solaredge_devicedata() called itself once per device record, and
    the dissection depth was released before the call, so a packet of up to
    64 KB could recurse about 8000 levels. Iterate over the records in the
    caller instead.

    Also fix the end of the device list, which compared an absolute offset
    with a relative length and stopped 20 bytes early, advance the offset
    past the POST_DATA payload, and use a length of 4 instead of an end
    offset for the invalid length expert item.

    solaredge_set_key() opened the system cipher handle but closed it only
    when the system key preference was valid, so every SET_KEY message
    leaked a handle with the default (empty) key. Check the key and the
    message length before opening the cipher, and always close it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit caeecf4de116faa513c2215712ce5781cc86767a)
    [Backport: the over-read, recursion, device list end, POST_DATA offset and
    expert item length hunks only. The SET_KEY code is still inline in
    dissect_solaredge_recursive() on this branch (solaredge_set_key() was
    split out on master later), so the cipher handle leak fix is not taken.]

commit 14473dc62b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 12:20:28 2026 +1000

    SNDCP-XID: Fix offset wrap infinite loop and two small bugs

    offset and the parameter length were uint16_t, with the length
    truncated from tvb_reported_length(). A 128 KB Exported PDU made the
    offset wrap to 0 and the parameter loop never ended. Use unsigned
    values and compare without the subtraction.

    Also advance past the full version parameter (2 + length) instead of a
    fixed 3 bytes, and show the compression entity length at the offset it
    was read from.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit f60a7627a83bef37a61bea7014a081f7f139db1e)

commit cb99473337
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Nov 11 10:08:12 2026 -0500

    PROFINET: Key the per-AR station info by AR UUID in its own map

    The PNIO dissector stored and looked up the per-AR stationInfo in the
    conversation's proto_data using the AR UUID (its first 32-bit word, or a
    fixed "fake" AR UUID) as the key. conversation_add_proto_data() /
    conversation_get_proto_data() index that table by registered protocol
    id, so using an AR UUID there misuses the API and puts attacker-controlled
    packet data into the protocol-id keyspace.

    When the AR UUID word equals another dissector's runtime protocol id on
    the same conversation, the two dissectors exchange structures of
    different types: that dissector retrieves the PNIO stationInfo as its own
    type, and PNIO retrieves its structure as a stationInfo. With WCP, for
    example, a crafted AR UUID equal to proto_wcp makes wcp_uncompress()
    treat an 80-byte stationInfo as a wcp_circuit_data_t and follow an
    uninitialized buffer pointer, a null-pointer store that crashes TShark
    (packet-wcp.c) from packet data with default settings.

    Keep the per-AR stationInfo in a per-conversation wmem_map keyed by AR
    UUID, stored under a dedicated registered protocol id
    (proto_pn_io_station_info), via new pn_find_station_info() /
    pn_add_station_info() helpers. The AR UUID is now only a map key, never a
    conversation proto_data key, so it can no longer collide with a protocol
    id. The per-conversation AR-UUID-to-stationInfo mapping is otherwise
    unchanged.

    Fixes #21745

    Thanks to Tomer Fichman for the report and PoC.

    Assisted-by: Claude
    (backported from commit a0ab51c632e661349203f60700a818f9414b5038)
    [Backport: applied by hand. This branch's packet-dcerpc-pn-io.c has
    diverged enough that the cherry-pick produced a large spurious conflict,
    and 4.4 does not have the current_fake_aruuid / extract-without-AR path,
    so only the current_aruuid call sites exist (seven in
    packet-dcerpc-pn-io.c, one in packet-pn-rtc-one.c). Added the same
    helpers, proto_pn_io_station_info registration and declarations as the
    master commit.]

commit 966bc9a651
Author: Evan Huus <eapache@gmail.com>
Date:   Tue Nov 10 21:32:34 2026 -0500

    Dissectors: Complete seven alphabetical-review fixes

    Preserve MAC-LTE HARQ state if copying a truncated PDU throws, reject
    LBMC property ranges that overlap earlier strings, and initialize ORAN
    beam arrays before replacing the scalar fallback.

    Keep MQTT and CITP redissection state per PDU instead of per frame.
    Separate DNP3 reassembly tables by link conversation without truncating
    link addresses, and give each HCI USB message a distinct reassembly ID
    so completed child decodes survive redissection.

    Generate small synthetic regression inputs within the existing protocol
    test classes and check one- and two-pass decoding. Correct the HCI USB
    test to require the retained child decodes.

    Assisted-by: Codex
    (backported from commit 00d85421dff9aba2bc4691e475cea4cfffd8267c)
    [Backport: only the MAC-LTE HARQ use after free and LBMC message
    property overlap fixes, which are identical to master; the ORAN beam
    array code is not on this branch, the MQTT, CITP, DNP3 and HCI USB
    changes are redissection correctness fixes, and the tests are not
    included.]

commit 78813a0084
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:02 2026 +1000

    SML: Fix endless loops from length underflow and backwards offset

    get_length() subtracted the number of length bytes from the decoded
    value without checking that the value was at least that large. A length
    byte of 0x00, or a multi-byte length with a small value, wrapped to a
    huge unsigned number. Loops over list entries then advanced by zero
    bytes up to 4G times, hanging tshark. Clamp the result to 0, so every
    field advances by at least its length bytes.

    In AttentionRes, a large attentionNo length moved the offset backwards
    into the previous message, so the message loop in dissect_sml_file()
    re-parsed it forever. Stop dissecting the message when the length is
    larger than the remaining data.

    Also pass &value_List_Entry to proto_tree_add_subtree_format() so that
    proto_item_set_end() on the entry actually does something.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit ae0165bea3de73aa627a5c13c5dd9ce46def623e)

commit dfa5a6406e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:02 2026 +1000

    SMB2: Bound decompression and the redirect IP loop, free async tables

    Compression Pattern_V1 and the LZ payloads of a chained
    COMPRESSION_TRANSFORM were only limited to 16 MB per payload, so a
    44-byte PDU could allocate 32 MB, and every PDU of a TCP segment adds
    up in the same frame. Keep a running total of decompressed bytes per
    frame and refuse any payload that would take the total over 16 MB
    (MAX_UNCOMPRESSED_SIZE), for the chained, Pattern_V1, uncompressed and
    non-chained LZ paths.

    The asyncs and asy_cancels hash tables of a conversation were never
    destroyed when the conversation went away, which leaked on every
    capture with a STATUS_PENDING response. Destroy them in
    smb2_conv_destroy().

    An OriginalMessageSize of 0xFFFFFFFF in a transform header wrapped the
    offset, so the next message was dissected backwards. Clamp the advance
    to the data present. The create context NextEntryOffset was stored in
    a uint16_t and truncated; use uint32_t.

    The share redirect error context loops over an IP address count taken
    from the packet. For an unknown address family
    dissect_windows_sockaddr_storage() returns the offset minus one, so the
    offset moved back and forth and the loop ran up to 2^32 times; without
    a tree nothing stopped it and memory grew without bound. Stop with an
    expert info when an address does not move the offset forward.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit a9f4db7b6c0ce685767206d6079f8118b007800e)
    [Backport: kept this branch's const uint8_t *tag (tvb_get_string_enc()
    returns uint8_t * here) and took master's uint32_t chain_offset;; this
    branch's smb2_conv_info_t has no asyncs or asy_cancels tables, so their
    destruction in smb2_conv_destroy() was dropped; otherwise identical.]

commit 4346061bff
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:02 2026 +1000

    SigComp: Fix UDVM memory and loop bounds

    With the sigcomp.decomp.msg preference on, the state/execution trace
    data source was created over the UDVM memory with a length of state
    address plus state length, which can reach 0x1FFFE bytes while the
    memory is 65536 bytes; showing the bytes read past the heap buffer.
    Fail the state access when the state does not lie within the UDVM
    memory.

    The reference operand decoder read buff[65536] for a 0xC0 operand at
    the last address. Wrap the address instead.

    STATE-ACCESS read the partial identifier from UDVM memory only up to
    the end of memory and then used uninitialised stack bytes. Wrap the
    address like the rest of the UDVM does.

    COPY-OFFSET was charged one cycle for a loop of up to 65535 steps,
    which lets a small message burn far more than the cycle budget.
    Charge the offset to the cycle count; the existing check before the
    next instruction then fails the decompression.

    The decompressed message is passed to SIP, which hands data starting
    with 0xF8 back to SigComp, so bytecode that outputs its own message
    recursed until the layer limit assertion. Limit the SigComp to SIP
    nesting to 4 levels, with an expert info, and count the dissection
    depth.

    Smaller fixes: the SHA-1 instruction wrote only every other digest
    byte (n was incremented twice) and its copy loop used the wrong start
    and length for the split at byte_copy_right and rejected a copy ending
    exactly at the end of memory (now like the CRC loop); literal operands
    masked the 10nnnnnn form with 0x1f instead of 0x3f and decoded the
    11000000 form from the wrong bytes; INPUT-HUFFMAN accepted bits_n up
    to 31, shifting a 16 bit value out of range, now limited to 16 as in
    RFC 3320.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 4ca0547ed3927c209ec82776b0fa35156e64303b)
    [Backport: this branch's val_to_str() takes no allocator, so the new
    diagnostic uses the three argument form; otherwise identical.]

commit 3f756eb028
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:02 2026 +1000

    SEL Protocol: Fix an endless loop in Read Response dissection

    A Data Format Response can announce a quantity of 0xFFFF values for a
    data item. dissect_fastmsg_readresp_frame() used 16-bit variables for
    the loop counter and the byte count, so the counter wrapped and
    "cnt <= quantity" was always true, while 2 * 0xFFFF wrapped so that the
    per-value length became zero and the offset never advanced. Without a
    tree this hung tshark at 100% CPU. Use a 32-bit counter and data size.

    The first pass of dissect_selfm() reused "offset" while it stored the
    Data Format Response and Device Description items, so the following
    function code checks read the wrong byte and the tree was built from a
    wrong offset, showing malformed packets on a one-pass run. Use a
    separate offset for those loops and drop the first-pass call to the
    read response dissector, which the tree path runs on every pass anyway.

    Shifting the 32-bit unsolicited SER status word by 32 or more is
    undefined; report those elements as not asserted. Make the raw 16-bit
    analog value field FT_INT16 so negative values are shown as negative.

    Data item, data region and analog channel names are copied from the
    packet into fixed buffers and shown in item labels, which must be valid
    UTF-8. Replace octets that are not ASCII.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 54668b1059684ced253efbef3aa46c212519d5e4)

commit 9635fa57e6
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:01 2026 +1000

    SCCP: Fix an infinite loop in XUDT/LUDT optional parameter scanning

    The variable and optional pointers in dissect_xudt_common() and
    dissect_sccp_message() were uint16_t. When the tvb is longer than 65535
    bytes (reachable through wtap encapsulation or Exported PDU), the loop
    that skips optional parameters looking for SEGMENTATION or END wrapped
    the pointer back into the buffer before tvb_get_uint8() could throw, and
    never terminated. Use unsigned for the pointers so the scan runs off the
    end of the tvb and throws.

    Also make the length_offset argument of sccp_reassemble_fragments()
    unsigned so LUDT data offsets above 65535 are not truncated, clamp the
    parameter length used with the dt1_ignore_length preference to 16 bits
    instead of truncating, and in looks_like_valid_sccp() reject pointers
    equal to the length, since reading the length byte there would throw
    from a heuristic that must not.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 140343913656201bdc496e6b77429c8475296b44)
    [Backport: the conflict was in the dissect_xudt_common() parameter
    lines; took master's unsigned offset and pointer types; otherwise
    identical.]

commit eaa045767c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:01 2026 +1000

    S7COMM: Bound recursion through nested TIS replace jobs

    A TIS "Replace job" request carries a job subfunction taken from the
    packet. If it is Replace job again, s7comm_decode_ud_tis_jobs() and
    s7comm_decode_ud_tis_data() call each other once per 8 bytes of data,
    without any limit. COTP and S7COMM reassembly make arbitrarily large
    PDUs possible, so a crafted packet overflowed the stack. Count the
    nesting per packet with p_get_proto_depth() and stop with a new expert
    info beyond 8 levels, and wrap the nested call in
    increment_dissection_depth()/decrement_dissection_depth(). Remove the
    comments claiming the packet runs out before the stack.

    Also stop decoding a BSTACK OB entry when the remaining length goes
    negative, and do not compute a block statistics line count from
    td_size - 2 when td_size is less than 2.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit b60fd3331dfd3b3484abb2411073dd3db7f6be15)
    [Backport: this branch's s7comm_decode_ud_tis_subfunc(),
    s7comm_decode_ud_tis_data() and s7comm_decode_ud_tis_jobs() did not
    take pinfo, which the nesting limit needs; added the parameter to the
    three functions and their callers; otherwise identical.]

commit 1b50cb227c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:01 2026 +1000

    RTSP: Fix abort on dest_addr without closing quote

    In rtsp_create_conversation(), sscanf() of "\"%u.%u.%u.%u\"" returns 4
    even when the closing quote is missing, so strstr() returned NULL and
    g_strndup(tmp, tmp2 - tmp) asked for about 2^64 bytes, which makes GLib
    abort. Check the result of strstr() for the closing quote, and for the
    ':' in the two other address forms, and report a bad address instead.

    dest_addr and src_addr were both stored through the same ip4_addr
    variable, and set_address() only keeps a pointer, so the destination
    address silently became the source address when both were present. Use
    separate variables.

    A Content-Type header with an empty media type made the parameter end
    offset smaller than the value offset, so the length wrapped and the
    frame was shown as malformed. Use a length of zero in that case.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit ff24626c21266671de5032a1d2761d9fce714f82)
    [Backport: the conflicting declaration line also carried master's
    rtp_dyn_payload local, which this branch does not have; declared only
    dst_ip4 and src_ip4; otherwise identical.]

commit 3220351d6d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:01 2026 +1000

    RTPS: Bound TypeObject and typecode dissection, fix stack pointers

    A TypeObject received in discovery could make dissect_user_defined()
    run for hours: the members of a structure beyond the display limit are
    still walked, so a chain of structures costs N^D calls. The preference
    that enables user data dissection is bypassed because a custom type with
    the id of a builtin type (for example InstanceStateDataResponse) was
    found before the builtin one. Look up the builtin types first, and stop
    a sample after 50000 dissect_user_defined() calls with the existing
    user data dissection error.

    rtps_util_add_typecode() recursed through TK_SEQUENCE without counting
    the nesting, so a large reassembled PID_TYPECODE overflowed the stack.
    Count the nesting for sequences, and stop with a new expert info when
    the dissection depth gets close to the maximum tree depth, before the
    depth assertion fires.

    Members of a typecode struct or union could overlap, all pointing at
    the same nested typecode, which made the work and the memory grow
    exponentially. Stop when a member starts before the end of the nested
    typecode of the previous one. The indent string is now allocated from
    the packet pool instead of the epan scope, where it was never freed.

    dissection_info was allocated without zeroing, so an array with an empty
    bound sequence used an uninitialized bound. Use wmem_new0().

    The RTPS_ROOT_MESSAGE_KEY and RTPS_CURRENT_SUBMESSAGE_COL_DATA_KEY
    proto data pointed at locals of dissect_rtps() and
    dissect_rtps_submessages(), which RTPS-PROC used after they returned.
    Allocate both from the packet pool.

    Only fill the TypeObject maps on the first pass, so that redissection
    gives the same result. PID_EXTENDED is now compared for equality instead
    of as a mask.

    rtps_util_add_type_id_v2() recurses for the element type of plain
    sequence, array and map type identifiers without a bound. Without a
    tree a chain of nested sequence identifiers in a reassembled
    PID_TYPE_INFORMATION overflowed the stack. Count it with
    increment_dissection_depth() and stop with a new expert info before
    the depth limit, as the typecode code does.

    rtps_util_look_for_secure_tag() advanced by the submessage length
    without the 4 octet header and subtracted the length twice, so a zero
    length submessage after a protected body looped forever when PSK
    decryption was enabled. Skip the header and the body.

    The type mapping kept for a discovered endpoint was allocated with
    wmem_new() and only fields_visited and the GUID flags were cleared, so
    a DATA_BATCH of an endpoint whose discovery data had no type id looked
    up an uninitialized type id (valgrind). Allocate it with wmem_new0().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 81b49cc2abaf1757a6a31176b93219491d3cf878)
    [Backport: this branch's rtps_util_try_dissector() calls
    dissect_user_defined() directly on the packet tvb without master's TRY
    block or ei_rtps_user_data_dissection_error, so the visit limit is
    reset there and its exception ends the packet as malformed (comment
    adjusted); kept this branch's dissect_rtps_submessage_v1() call while
    storing the pool-allocated submessage_col_info pointer. This branch has
    no XTypes v2 type identifiers (rtps_util_add_type_id_v2()) or RTPS
    header decryption context, so those parts and ei_rtps_type_id_too_deep
    were not taken; dropped master's forward declarations of
    rtps_util_add_generic_guid_v2(),
    rtps_util_add_timestamp_sec_and_fraction() and
    rtps_util_add_seq_number(), which are defined before use here and whose
    signatures differ; cast pinfo->dissection_depth (an int here) to
    unsigned in the depth checks; the mutable and union member mapping
    functions do not take pinfo here, so those maps are still filled on
    every pass; otherwise identical.]

commit 5201798de0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 11 05:03:01 2026 +1000

    RTP: Fix RFC 2198 recursion, uninitialized info and A2DP payload length

    A RFC 2198 block whose payload type is again RFC 2198 (payload type 99
    by default, or any "red" dynamic type) was dispatched back to the RED
    dissector with only one byte consumed per level, until the generic
    layer limit assertion fired. Track the nesting with a per-packet depth
    counter and, for a block inside a block, show an expert info
    (rtp.rfc2198_nested) instead of recursing.

    dissect_rtp_rfc2198() filled its copy of struct _rtp_info only when
    called with data, but passed it to every block's payload dissector.
    With NULL data (Exported PDU) AMR or EVS read uninitialized pointers.
    Zero-initialize the copy and pass NULL on to the payload dissectors when
    there was no RTP info.

    For A2DP and VDP streams info_payload_len was decremented once per
    codec or SCMS-T header byte with no lower bound, so an RTP packet
    without payload wrapped it to 0xFFFFFFFF, which the RTP player then used
    as a copy length. Compute the length from the final payload offset, as
    the padding case already does, so it is clamped at zero.

    Also dissect the RTP header of PacketCable CCC after the 12 byte CCC
    header instead of from the start of the CCC header.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 24032a4fc1e7b3161137739637a12cca81e7d38a)
    [Backport: this branch's dissect_rtp_data() takes the block length
    twice (payload and data length); kept that and passed NULL when there
    was no RTP info as master does; otherwise identical.]

commit 0951187774
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:02 2026 +1000

    RTMPT: Dissect aggregate tags within their own length

    dissect_rtmpt_body_aggregate() dissected each FLV tag with the audio,
    video or command dissector on the whole message tvb, so the tag body was
    parsed to the end of the message and not just over the tag's DataSize
    bytes. A message made of many tags, each holding an AMF0 long string
    that runs to the end of the message, copied and converted the whole
    tail once per tag: memory and CPU grew with the square of the message
    size. Dissect each tag on a subset tvb of DataSize bytes. This also
    bounds the audio and video multitrack loops to their tag. The audio
    multitrack data item now uses the track length instead of the rest of
    the tvb.

    Also fix several decoding errors in the same file:

    The Chunk Stream ID was added with the length of the basic header and a
    6-bit mask, so IDs above 63 were shown wrongly. Show the full ID.

    AMF3 arrays did not skip the name of an associative element or the
    string and object references, so the name was decoded as the value;
    the subtree length now covers the length and name, and the duplicate
    type text read at the wrong offset is removed. AMF3 object, XML and
    byte array references and the XML string were not skipped either.

    In a video multitrack message with many codecs, the FourCC was read at
    the track ID offset. The AMF0 long string length item had length 2
    instead of 4.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 5da10873054190c0ffed704342f68980bf1ac76e)
    [Backport: kept this branch's uncast tvb_get_string_enc() assignment in
    the conflicting line and took master's iValueLength+iStringLength
    subtree length; otherwise identical.]

commit 5e388d8537
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:02 2026 +1000

    RPL: Bound the container recursion

    dissect_rpl_container() recursed into sub-vectors without a limit, so
    a crafted PDU nested thousands of levels deep. Limit the nesting to 8
    levels per packet with an expert info, and call
    increment_dissection_depth() around the recursive call.

    The default case passed the bytes after the header to the data
    dissector and then dissected them again as trailing data because the
    offset was not advanced. Return after the first call.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 5e77af44f5ccc990d72b6e9ddd0652df31b5d440)
    [Backport: this branch still builds the sub-container tvb with
    tvb_new_subset_length_caplen(length, reported_length) and declares
    length; kept that and added the nesting limit and depth calls around
    it; otherwise identical.]

commit 77448d53bf
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    RDT: Fix an endless loop on packets of 64 KiB or more

    The packet length was kept in a uint16_t in seven packet dissectors.
    When the length-included flag is clear, the length is the remaining
    data, which was truncated to 16 bits. With exactly 65536 bytes
    remaining the length became 0, the offset no longer advanced and
    dissect_rdt() looped forever, adding items. Use uint32_t for the
    length in all seven functions, and stop dissect_rdt() when a
    sub-dissector makes no progress.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 85b7f1b8902e7217319a7d6345d2d4db4a629d26)
    [Backport: this branch read the packet length with tvb_get_ntohs() plus
    proto_tree_add_item() in the seven conflicting places; took master's
    proto_tree_add_item_ret_uint() into the widened uint32_t
    packet_length;; cast tvb_reported_length_remaining() to uint32_t in the
    seven comparisons with the widened packet_length, as it returns int on
    this branch; otherwise identical.]

commit b449a72e66
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    QuakeWorld: Fix quadratic allocation in the command tokenizer

    COM_Parse() allocated a buffer for all the text remaining in the packet
    for every token, and Cmd_TokenizeString() kept tokenizing after the
    80 arguments it can store. A 64 kB connectionless packet of one
    character tokens allocated about 1 GB from the packet pool. Stop at
    MAX_ARGS tokens and allocate only the token, with wmem_strndup().

    Cmd_TokenizeString() also stored the pointer COM_Parse() returned, the
    rest of the text after the token, instead of the token, so the client
    to server commands were never recognized. COM_Parse() now returns the
    token through an out parameter and that is stored. An unterminated
    quoted token no longer moves the pointer beyond the terminating NUL.

    The length of the connect arguments item included one byte more than
    the string, past its end. Use len - start.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit f9beb9321fe030149f133c37a18790eced4cb2a9)
    [Backport: this branch's Cmd_TokenizeString() has no allocator
    parameter, so COM_Parse() is passed wmem_packet_scope(), which the old
    code allocated from; otherwise identical.]

commit f06111f30d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    QNET6: Bound the nesting of combined messages

    Every _IO_* message handler in the KIF MsgSend dissector calls
    dissect_qnet6_kif_msgsend_msg() again when the combine length has
    bit 0x8000 set. A message of only 4 bytes per level (_IO_CLOSE with
    combine_len 0x8004) is enough to nest tens of thousands of levels,
    overflowing the stack. A reassembled datagram of 20 kB crashes tshark
    with a 1 MB stack.

    Wrap the function: limit the nesting to 16 levels per packet with
    p_get_proto_depth()/p_set_proto_depth(), add an expert info beyond
    that, and call increment_dissection_depth()/
    decrement_dissection_depth() around the nested call, and replace the
    comment that relied on running out of packet first.

    The link extra item of _IO_CONNECT was added with the variable
    extra_len as its length, although the field is FT_UINT32; any length
    other than 1 to 4 raised a Malformed Packet. Use MIN(extra_len, 4).

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 3d9aa54bdafc0ee3b1db21020499a5c9be604e4e)
    [Backport: added only the expert.h and proto_data.h includes, as this
    branch's include list differs (no epan/iana-info.h); otherwise
    identical.]

commit 6246b60295
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    RELOAD: Fix an infinite loop and CPU blow-ups on malformed messages

    A message extension with a content length of 0xfffffff9 made the
    truncation check wrap to zero and the loop advance by zero, so the
    dissector never returned. Compare the length against the remaining
    extension bytes without overflow. The MessageContents length check is
    done in 64 bits as well, and before the extensions length is read.

    In StoredDataSpecifier the dictionary key loop always read from the
    start of the specifier instead of the key list, and the keys length was
    not checked against the specifier length, giving tens of millions of
    iterations per packet. Read from the current key and limit the keys
    length to the specifier's length field.

    dissect_diagnosticinfo() returned 2 + length instead of 4 + length and
    walked the length of the whole remaining list instead of its own, which
    was quadratic. Return 4 + length and bound the info by its own length.

    Also fix small errors: the hop counter in a DiagnosticResponse was not
    skipped, and the list subtree hung off the wrong item; the IPv6
    IpAddressPort item was 6 bytes long instead of 18; the ArrayRange item
    was 16 bytes instead of 8; the ICE candidate priority offset missed the
    candidate offset; the error code was read at the forwarding header
    length instead of the current offset (wrong after reassembly); the
    truncated FindKindData and SipRegistration data items were at the wrong
    offset; and a truncated NodeId list was added twice.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 1e80174127252f7da861bb537f0cf7f939a8bc7f)
    [Backport: this branch read the ICE candidate priority with
    tvb_get_ntohl() plus proto_tree_add_item() in the conflicting line;
    took master's proto_tree_add_item_ret_uint() at the corrected offset;
    otherwise identical.]

commit e3be3e4a4f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    RTCP: Bound transport-cc allocations and fix several decoding bugs

    The transport-cc feedback dissector allocated two zeroed arrays sized
    from the packet status count (up to 65535 entries) for every FCI,
    before validating anything. A 64 KB packet holding thousands of tiny
    FCIs, each failing the run length check and returning early, allocated
    about 1.3 GB from the packet pool. Allocate the two arrays once per
    RTCP packet with a fixed size of 65536 entries and only read back the
    entries that were filled in (delta_index), so no per-FCI zeroing is
    needed.

    Fix the XR IDMS block, which did not advance past the NTP receive
    timestamp (so the RTP timestamp overlapped it and the next block was
    decoded four bytes early) and which decoded the timestamp as raw
    seconds/nanoseconds instead of NTP.

    Fix the REMB loop, which assigned instead of accumulated the number of
    FCI entries read, so a second REMB block ran past the packet. Fix the
    MS dominant speaker history, which counted bytes in steps of two while
    each entry is four bytes. Fix an unsigned comparison that was always
    true for the trailing PSFB FCI, clamp the profile specific extension
    bytes consumed to the remaining length, and do not subtract an RTCP
    padding count that is larger than the packet.

    Use the floor revoke cause strings for floor revoke in the Info column,
    use the whole setup_method buffer when copying the setup method, and
    make the transport-cc receive delta padding field FT_BYTES so any
    padding length is valid.

    The TMMBR/TMMBN measured overhead (the low 9 bits of the last two
    octets of the FCI) was read from the last octet onwards, one octet past
    the FCI, so a packet ending with such an FCI was malformed.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 62ce1fd99059bf5f6c24f4de25360401dba454c2)
    [Backport: kept this branch's int offset and packet_len in
    dissect_rtcp_rtpfb_transport_cc_fci() and its caller while passing in
    the arrays allocated once per packet; dropped the padding count clamp,
    since this branch does not subtract the padding from the packet length;
    kept this branch's one-octet TMMBR measured overhead item, which does
    not read past the FCI; otherwise identical.]

commit d1cac33349
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    RTLS: Bound the nesting of compound messages

    An AR_COMPOUND_MESSAGE_REPORT can contain further compound message
    reports, each costing only 20 octets, and the dissector recursed
    without a limit. Limit the nesting to 8 levels per packet with
    p_get_proto_depth()/p_set_proto_depth(), report the rest with an expert
    info, and bracket the nested call with increment/decrement of the
    dissection depth.

    The NACK reserved flags mask was 0xFC on a 16 bit field; make it 0xFFFC
    so that it covers the remaining bits.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 20389c1f304007ccd460b23f9abeb5872d0eeaf7)

commit 4bd978b142
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:01 2026 +1000

    RTITCP: Fix a message length wrap that loops forever

    dissect_control_message() returned uint16_t, so a control message with
    a length of 0xfff0 and the 16 octet CRC header returned 0x10000 as 0.
    The caller then never advanced its offset and looped forever, creating
    file scope allocations on every iteration for request messages until
    memory ran out. Return and keep the length as unsigned, and stop the
    loop if a message makes no progress.

    Do not add the error code description when the attribute is shorter
    than 4 octets (the length wrapped to a huge value). Add the port
    attribute with length 4 and the connection type attribute with length
    1, instead of the attribute length, which raised an expert info for the
    fixed size fields.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 64b966e42c05a99a48378973ce1b85da63dc7031)

commit 0a8de984a5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:00 2026 +1000

    EGFX: Fix endless loop on bad pduLength and partial ZGFX output

    A pduLength that points past the end of the data (for example
    0xFFFFFFF0) made nextOffset wrap back to the start, so the dissector
    looped forever when no tree was being built. nextOffset is now
    unsigned and a pduLength larger than the remaining data ends the
    dissection with an expert info.

    rdp8_decompress() (epan/tvbuff_rdp.c) allocated the multipart output
    from the size in the packet without zeroing it and returned it even
    when the segments filled only part of it, exposing uninitialized heap
    data. It now fails unless the segments produce exactly the announced
    size, and the announced size is bounded by segment_count times the
    64 KiB maximum output of one segment, which also limits the allocation
    that was up to 4 GiB.

    rdp8_decompress() returned the input tvb itself for a descriptor other
    than the two ZGFX ones. EGFX stored the returned tvb in a file-scope
    tree and used it again on later passes, after the packet's tvb had been
    freed (use after free with tshark -2 or in the GUI). The decompressed
    tvbs were also never freed. rdp8_decompress() now returns NULL for an
    unknown descriptor, the result is chained to the packet's tvb, and EGFX
    keeps the decompressed bytes (already file scope) for later passes and
    makes a child tvb from them on each pass.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 81dcd586e595a645450f68be3575ec5d186fa5b7)
    [Backport: this branch's tvbuff_rdp.h has no doc comment for
    rdp8_decompress(), so master's comment update was dropped;; cast
    tvb_reported_length_remaining() to uint32_t in the new pduLength
    comparison, as it returns int on this branch; otherwise identical.]

commit 278f58f63b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:00 2026 +1000

    DRDYNVC: Fix uninitialized PDU info, NULL channel and huge allocation

    The per-PDU info structure was allocated without zeroing and inserted
    in the PDU tree before it was filled. A middle fragment of a
    reassembly, or a payload larger than the pending length, left the tvb
    pointer uninitialized and it was later used as a tvbuff (wild pointer
    dereference). Both allocations are now zeroed and the entry is only
    inserted once the PDU is accepted.

    A Create Request seen before the UDP transport is linked to the main
    conversation creates no channel on the first pass; on later passes the
    channel stays NULL and was dereferenced. It is now checked.

    The reassembly buffer was preallocated with the announced PDU length,
    up to 4 GiB; it now grows with the data. A first chunk larger than the
    announced length no longer underflows pendingLen and is treated as a
    single packet. When ZGFX decompression fails the raw data is used
    instead of a NULL tvb.

    Reassembled and decompressed payloads were kept as tvbs across passes
    and never freed; keep their (file scope) data instead, chain the
    decompressed tvb to the packet's tvb, and make a child tvb on each pass.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit c4b92f84b454d99af6e4d1670282b63c64b9c627)
    [Backport: this branch has no ZGFX decompression in DRDYNVC and still
    keeps a tvb in the PDU info, so the patch was rewritten for it: zero
    both PDU info allocations and insert the Data PDU info only once it is
    accepted, let the reassembly buffer grow instead of preallocating the
    announced length, and treat a first chunk larger than the announced
    length as a single packet. This branch's Create Request does not look
    up createConfirmFrameId, so no NULL channel check is needed there. The
    decompression fallback and the tvb lifetime changes were not taken.]

commit 4ae113fdd2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:26:00 2026 +1000

    RPCoRDMA: Do not use stale InfiniBand/iWARP info and fix a tvb leak

    The file-scope pointers to the InfiniBand and iWARP info structures
    point into the callers' stack frames and were never cleared. When
    dissect_rpcrdma() was reached through the registered "rpcordma"
    handle (Decode As, Exported PDU), it used the pointers left from a
    previous packet, a stack use-after-return. Register a wrapper for the
    handle that clears them, and clear them after both heuristic
    dissectors return.

    add_iwarp_padding() never freed the tvb it created for the padding
    bytes, leaking one for each unaligned last iWARP write or read
    response. Make it a child of the current tvb.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit d21240270783c37c88024c81a0f559433eea56ad)

commit 3a26f371b2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 20:25:59 2026 +1000

    RPC: Fix type confusion on reused XIDs and an SRT tap overflow

    A call whose XID was already in the conversation's table reused the
    existing call info even when program, version or procedure differed.
    The dissector was chosen from the new call but received the old
    call's private data, which another program may have filled with an
    integer (portmap GETPORT stores the protocol number there). An NFS
    ACCESS reply then dereferenced it as a pointer and crashed. The same
    happened through the portmap CALLIT path. Only reuse the entry when
    program, version and procedure match, otherwise treat the packet as a
    new call. The table is updated only on the first pass; later passes
    use a packet-scoped copy. For indirect (CALLIT) calls a mismatching
    entry is never replaced, so later passes match replies the same way
    as the first one.

    The SRT tap compared the procedure with a signed cast, so a procedure
    of 0x80000000 or more became a negative index into the procedure
    table. Compare unsigned.

    Also: fix the mask in dissect_rpc_authgss_context (0xffffffc masked
    the top bits), set the GSS context create and destroy frames only on
    the first pass, and make the length in find_and_dissect_rpc_fragment
    signed so a "need more data" result is not corrupted.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 571c325c8f01afc67ac63ebebb5fc818fd1a80c9)
    [Backport: the find_and_dissect_rpc_fragment() hunks conflicted with
    master's earlier unsigned offset rework, not on this branch; this
    branch's len is already int, so kept its code there; otherwise
    identical.]

commit 8cbfb413dc
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 18:54:24 2026 +1000

    nordic_ble: Do not mix proto data types under one key

    dissect_ble_delta_time() stored an integer and dissect_ble_timestamp()
    stored a pointer to a packet_times_t under the same proto data key.
    A frame containing PDUs of both kinds, dissected a second time, read
    the integer back as a pointer and crashed. Use separate keys for the
    two, derived from the current layer number so that several PDUs in
    one frame do not overwrite each other, and handle a missing entry
    when the frame is dissected again.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 41785dbef8a71f9fc67952e0b630f555c0ac93b3)

commit 3f44b1e69a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 18:54:24 2026 +1000

    nvme-tcp: Keep H2CData requests out of the keyed SGL request tree

    The H2CData requests were stored in n_q_ctx.data_requests under the key
    [frame][cid], the same tree the NVMe dissector uses for keyed SGL
    descriptors under [addr_lo][addr_hi][rkey][size]. A command with a
    keyed SGL descriptor whose address matched an earlier H2CData frame and
    command ID made the dissector treat a command context as a subtree,
    crashing in wmem_alloc(). Give NVMe/TCP its own tree for H2CData
    requests (h2c_requests in struct nvme_tcp_q_ctx) and leave
    n_q_ctx.data_requests NULL, which dissect_nvme_cmd_sgl() already
    checks.

    Also add the C2HTermReq Fatal Error Status item to the C2HTermReq
    subtree instead of the PDU tree.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 3970e8936ae8d4955bec774f0f108ff5da07e6b7)
    [Backport: the conflicting Termination Request hunk only had context
    from master's 33196a3e3d (the short PLEN check), which is not on this
    branch; kept this branch's code there and applied only the tree ->
    c2htermreq_tree change; otherwise identical.]

commit 4b26df0aeb
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 18:54:24 2026 +1000

    NVMe: Use one key shape for data_responses and fix small decode bugs

    nvme_add_data_tr_pkt() and nvme_lookup_data_tr_pkt() built the key as
    [frame] when the rkey (the command ID for NVMe/TCP) was 0 and as
    [cid][frame] otherwise, in the same wmem tree. A C2HData PDU with
    command ID 0 in frame N stored a command context at the first level
    for key N, and a later PDU with command ID N then descended into it
    as if it were a subtree, crashing in wmem_alloc(). Always use the two
    level key [rkey][frame]. NVMe/RDMA passes an rkey of 0 and so keeps a
    consistent key shape.

    Also fix three small problems. The Self-test log page decoder never
    incremented the result index, so every entry was labelled "latest 0".
    add_ctrl_rab() and add_ctrl_mdts() shifted by a raw byte value, which
    is undefined for values of 64 or more; they now print "overflow". The
    Get Log Page NUMD byte count was computed with the +1 done in 32 bits;
    do it in 64 bits.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit b2780a041d7c68b3e6267ab83f01f76b652c0bf8)
    [Backport: this branch does not keep the Get Log Page NUMD in the
    command context (master's numd store comes from an earlier master
    change that is not on this branch), so only master's ((uint64_t)val +
    1) * 4 byte count was taken in the conflicting hunk; otherwise
    identical.]

commit 8e325922bf
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MsgPack: Fix endless loop and unbounded recursion

    An unsupported type byte (0xC1) did not advance the offset. Inside an
    array or map with a 32-bit element count this looped up to 4 billion
    times, adding an expert info and a column string each time, until memory
    ran out. Consume the type byte after reporting it.

    Nested arrays and maps recursed without any limit, one byte per level,
    and overflowed the stack. Limit the nesting to 32 levels, tracked per
    packet with p_get_proto_depth(), and report "Nesting too deep" with a
    new expert info msgpack.too_deep.

    Also fix the "float 64" typo and give the negative fixint field its own
    filter name, msgpack.integer.negfixint, instead of registering
    msgpack.integer.fixint twice.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit a904409dcd501c984bd64c747626dc936413f65f)
    [Backport: this branch only decodes fixmap and fixarray (at most 15
    elements), so an unsupported type cannot loop billions of times here,
    but nested arrays and maps still recurse one byte per level. The patch
    was rewritten for this branch's older code: the 32 level nesting limit
    with msgpack.too_deep, and consuming the unsupported type byte. The
    float 64 and negfixint field changes do not apply.]

commit be154ec754
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 19:03:59 2026 +1000

    MCPE: Bound Batch nesting and drop an assertion

    A Batch packet can contain another Batch, and each level decompresses
    its payload and recurses. Only the tree depth limit stopped this, and
    only when a tree was built. Allow at most 16 nested Batches per packet,
    stop with an expert info (mcpe.batch.too_deep) beyond that, and wrap
    the nested dissector call with increment_dissection_depth() and
    decrement_dissection_depth().

    Replace DISSECTOR_ASSERT(message_id == 0xFE) in the heuristic
    dissector with a plain check, and use uint8_t for the packet id in
    test_mcpe_heur() so that ids above 0x7f are not sign-extended.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 8136315aa86a726973a7499bf73e3867f7192f72)

commit 2cda55bb4b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 12:37:10 2026 +1000

    IPMI: Bound nesting, FRU Unicode length and request list

    Nested Send Message / Get Message commands called back into
    dissect_ipmi_cmd() without any limit, so a packet with many nested
    levels overflowed the stack when ipmi.dissect_bus_commands is enabled.
    The uint8_t level counters also wrapped at 256. Stop at 32 levels (or
    when the layer count approaches the tree depth limit) with a new expert
    info, and wrap the command handler call in
    increment/decrement_dissection_depth().

    For FRU Unicode strings with an odd byte length, the character buffer
    was sized at 3 * len + 1 but parse_unicode() wrote 6 characters for
    each 2-byte pair, including a trailing partial pair, overflowing the
    heap buffer and reading past the string. Size it as (len / 2) * 6 so
    only whole pairs are decoded.

    The per-conversation list of unmatched requests grew without bound and
    was scanned linearly for every response, which is quadratic for
    captures whose requests share a timestamp. Keep at most 1024 entries
    and drop the oldest.

    Also fix two small decoding errors: the offset after the full 6-bit
    ASCII triplets was advanced by len / 4 instead of (len / 4) * 3, and
    the Type/Length Type item was given the shifted type instead of the
    type/length byte, so it always showed 0 with its 0xc0 mask.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 71f337897627ce3dcace6aa999fefd9e0290d76a)
    [Backport: this branch still has the fprintf() debug output under
    #ifdef DEBUG after the request list append; kept it and added master's
    request list bound before it; otherwise identical.]

commit e18d7274bc
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 4 16:22:14 2026 +1000

    extrememesh: Limit MCH nesting and fix the encapsulated Ethernet copy

    dissect_extrememesh_mch() recursed once per 20-byte MCH header with no
    limit, so a large frame could nest thousands of levels. Count the
    recursion with increment_dissection_depth() and stop with an expert
    info before the dissection depth limit is reached. Add the expert
    field extrememesh.mch.too_deep and register an expert module.

    dissect_extrememesh_eth_noaddr() passed the size of the addresses to
    tvb_memcpy() as the source offset instead of using it for the
    destination, so the copy always ran past the end of the packet and
    every packet with an encapsulated Ethernet frame without addresses (and
    the next protocols that use it) was shown as malformed.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 64664c2221009c78c0bd79463ec12690e7e83d14)
    [Backport: pinfo->dissection_depth is an int on this branch; cast it to
    unsigned in the comparison with prefs.gui_max_tree_depth to avoid a
    sign-compare warning; otherwise identical.]

commit 7a3957741c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 18:54:24 2026 +1000

    NTP: Fix endless loop and length wrap in NTS extensions

    dissect_ntp_ext() returned an offset that moved backwards when the
    extension length was below 8 or not a multiple of 4, so the loop over
    the decrypted NTS extensions in dissect_nts_ext() never ended. Return
    the end of the tvbuff after reporting the bad length.

    Reject an NTS ciphertext shorter than the AEAD tag with an expert info
    before decrypting, since ciphertext_len - tag_len wrapped around.

    In the control message dissector, loop over READSTAT items while at
    least four bytes remain, using a separate counter so that the
    authenticator offset still uses the real data length, and keep
    data_offset in an int so that it does not wrap above 65523.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 6bdb3903a5e0312e009c85ad2a31fa85a05cd10b)
    [Backport: only the control message part applies. This branch has no
    NTS decryption, and its dissect_ntp_ext() already returns the end of
    the tvbuff for an extension length below 8 or not a multiple of 4, so
    the NTS extension loop cannot move backwards here.]

commit 63c4f894e3
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Oct 31 05:55:57 2026 +1000

    ORAN FH CUS: Do not copy every section into each U-plane frame

    Do not dereference the eCPRI message type when called without one,
    as from an exported PDU capture.

    On the first pass, every U-plane frame took a copy of all the
    sections the C-plane had announced, each holding two arrays of 273
    beam IDs. A capture announcing all 4096 sectionIds and then sending
    small U-plane frames used about 5 MB per frame. Copy only the
    sections that the frame refers to, and do not match a C-plane entry
    that was never used. Allocate the tap data from the packet scope.

    Reject sample bit widths above 16 from the preferences instead of
    asserting in tvb_get_bits32().

    Fix the dissection of valid packets:
    - SE 1 read one TRX index per configured antenna, past the entries
      given by activeBeamspaceCoefficientMask
    - SE 11 with SE 6 made bundles for unset rbgMask bits when numBundPrb
      is smaller than rbgSize, and an all-zero bundle was added at its bit
      offset
    - SE 31 did not consume the mcsIndex byte
    - ST 4 sleep commands skipped two bytes after the symbol mask
    - measurement report 11 counted the layers from bytes times 4
    - uncompressed values took the largest positive value as negative
      and divided by zero for a width of 1
    - the PAPR item was added at the section offset twice
    - the previous frame in the stream was shown as 0 on the second pass
      and never on the first; show it with the sequence number error

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit ff29e5db3ad0caec0a8c51d0b94fd726df816ae9)

    Backport: Omit the power-display offset fix because that feature is absent
    on release-5.0.

    Assisted-by: Codex

    (cherry picked from commit 60d8cf126f35670a41dfc554dc587457bb555d16)

    Backport: Omit absent per-section state, newer extensions and compression features; retain applicable bounds, callback-data and tap-lifetime fixes.

    (cherry picked from commit 5af1e2fc5fc72b08d9eaff19b8b7513f6994d7fa)

    Backport: Only the data-pointer and sample-width guards apply. Section state, taps, sequence tracking, SINR and ST4 decoding are absent; the ext6 loop already uses the correct bound.

commit 306b52a569
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 4 13:47:44 2026 +1000

    SMB: Fix NT Create AndX, two-pass state and quadratic FID lookups

    The NT Create AndX response with an extended word count (42 or more)
    assigned the return value of dissect_smb_access_mask_bits() to the
    offset. That function returns 0 for the generated "Granted Access
    Rights" item, so the offset was reset to 0, the byte count was read
    from the SMB marker and the body was shown as extra bytes or, with
    an AndX chain, as malformed. Do not assign the return value.

    Matches found through ct->primaries (secondary Transaction requests and
    the final Transaction response) were not added to ct->matched, so on
    later passes (tshark -2, the GUI) the secondaries showed "Continuation
    of: <unknown>" and the response lost its "Response to" and time fields
    and its Trans2 sub-command. Add them to ct->matched on the first pass.

    pdata->multi_cmds was set during the first pass from the number of
    commands seen so far and was therefore different on later passes,
    which changed the Info column with -2. Set it for every pass from
    the position of the SMB in the frame, in one place, for all readers.

    The TID table used wmem_tree_lookup32_le() on the TID and replaced the
    entry when a TID was reused, so an unmapped TID showed another share
    and later passes showed the later share for earlier frames. Key the
    table on {tid, frame} and check the stored TID. The FID list was kept
    sorted with g_slist_insert_sorted() and walked linearly on every lookup,
    which is quadratic in the number of Create responses. Use a wmem_tree
    keyed on {tid, fid, opened_in} with lookup32_array_le() and a check of
    the stored tid and fid, export the lookup as smb_fid_tree_lookup()
    and use it in MS-WSP, which walked the list too, and remove the
    list. Drop "|| !si->sip" from the FID lookup, which made the result
    of later passes differ from the first pass.

    Export Objects could write uninitialized heap memory into an exported
    file: after a chunk too far into the file made the buffer be dropped,
    the next chunk allocated a new buffer but did not copy into it, and the
    parts of a file between chunks were never initialized. Do not start a
    new buffer once the content was dropped, and zero the buffer when it is
    allocated or grown.

    Export Objects looped over g_slist_nth_data() in find_incoming_file(),
    insert_chunk() and smb_eo_cleanup(), which is cubic with many files.
    Iterate the list nodes directly.

    Smaller fixes: the "Requested Dialects" subtree length was always 0;
    Find Close Data Length was read big endian; the Read response gave
    dissect_file_data_maybe_dcerpc() a FID of 0 because the request info was
    looked up afterwards; the Delete request allocated file scope memory on
    every dissection; Write AndX with the pipe write length lacked a byte
    count check and wrapped bc; the Mac Query FS Info backup time did not
    advance the offset; the Mac file and directory count labels were wrong;
    the "bytes until total data count" subtraction could wrap; the short
    name in the 4_3_4_6 info levels was padded and not length checked; and
    the FIND_FIRST2 response loop did not stop when no progress was made.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 610c8d81e3ecaf4e62501cc38f27b314c3ccd671)

    (cherry picked from commit 5325a7d66b2360d3ca690747b96b1d5577c5dfe2)

    Backport: Backport FID/TID indexing and export fixes; omit absent granted-access and newer multi-command tracking code.

    Assisted-by: Codex

    (cherry picked from commit f3852ce542a51b0d9e300d57736b1c953a961aad)

    Backport: Use older string helpers and preserve 4.4 command handling while replacing linear FID lookups.

commit 8d291dd887
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 08:49:37 2026 +1000

    FTP: Bound the working directory and fix several smaller problems

    The conversation's working directory was deep-copied into the packet
    data of every request and response, and a CWD could grow the path
    without limit. A capture with many long CWD commands therefore used
    memory proportional to packets times path length (gigabytes for a
    1 MB capture). Limit the stored path to 4096 bytes, keep the old path
    and add an expert info when a CWD, CDUP or PWD would exceed it, and
    share one immutable string between the packets instead of copying it.
    Adding a path to an empty working directory no longer modifies a
    string that earlier packets and data conversations already point to.

    The expert info for a bad EPRT or 229 argument started at
    "offset - linelen - 1", which wraps around as an unsigned value and
    throws an exception. Use the start of the line.

    The invalid PWD response expert info was added only in the first pass.
    Check the line in every pass and update the conversation only in the
    first one.

    The Export Objects tap copied the reported length of the data from the
    tvbuff, which throws on a truncated capture. Use the captured length.

    The data stream duration and bit rate were computed with 32 bit
    arithmetic that could overflow and divided by zero for a zero or
    negative gap. Use 64 bit values and show them only for a positive gap.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit d41b724587d9c84fdcc556bcf51f03fdcd5d636c)

    (cherry picked from commit 8283f7eb5c0d0c7325a0765846f44563003ed0b4)

    Backport: Thread packet_info through older path helpers and use captured export lengths.

    Assisted-by: Codex

commit 3f0a75d284
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Nov 4 17:22:29 2026 -0500

    EAP: Fix a zero-width AT_RES assert, the TEAP recursion and lengths

    An EAP-AKA AT_RES attribute with a RES Length of 0 was passed to
    proto_tree_add_bits_item() with a width of 0 bits, which is reported
    as a dissector bug. Flag the length as bad and skip the RES instead.

    EAP-TEAP outer TLVs and the tunnelled TLS data were handed to the
    TEAP and TLS dissectors without counting the dissection depth, so an
    EAP-Payload TLV that holds another EAP-TEAP message recursed until the
    layer limit was hit. Count the depth around both calls.

    Fix some lengths as well:
    - EAP-SAKE compared an absolute offset with a remaining length, so the
      last attributes were not shown. Pass the end offset instead.
    - The MS-Authentication-TLV loop made the same mix-up, so only the
      first TLV was shown. Track the end offset, check the value length
      against what is left, and move to the next TLV by its length.
    - The LEAP name length was truncated to 8 bits and could wrap; use an
      unsigned and clamp it at 0.
    - The EAP-MD5 extra data length was computed before Value-Size was
      clamped, so it wrapped when Value-Size was too big.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 3930757815ca4df6472d62f09c8ee663a5b3ab70)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit b6a12b0dca44568420e63ab17a45716140ab7124)

    Backport: Keep signed offsets and the release TEAP reassembly path, adding its recursion guard.

    Assisted-by: Codex

    (cherry picked from commit 89ee05c41228f890ca515bb29cf5298c57ffe986)

    Backport: AT_RES decoding is absent; apply TEAP recursion and length fixes using older SAKE signatures.

commit d30029ef07
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Nov 7 21:56:17 2026 -0500

    Diameter: Bound grouped AVP recursion and fix several decoding issues

    Grouped AVPs were dissected recursively without any bound, so a message
    with deeply nested grouped AVPs over the default TCP/SCTP port 3868
    overflowed the stack, also without a tree. Track the nesting with
    increment_dissection_depth()/decrement_dissection_depth().

    diam_sub_dis_t.item was only set by unsigned32_avp() and never reset,
    so an AVP subdissector that hides "the item created in
    packet-diameter.c" could hide the value of an earlier, unrelated AVP.
    Reset it before each AVP is dissected and after a grouped AVP, and set
    it in unsigned64_avp() as well.

    With the session_imsi preference enabled the Session-Id to IMSI map was
    updated and consulted on every pass, so "Associated IMSI" depended on
    the dissection order and tshark -2 differed from one pass. Update the
    map on the first pass only and keep the result per PDU with
    p_add_proto_data().

    Smaller fixes:
    - CC-Total-Octets printed a uint64_t with %d and claimed 16 bytes; use
      PRIu64 and return 8. Mark add_group_str() with G_GNUC_PRINTF and use
      %u for the unsigned Result-Code and Rating-Group values.
    - Framed-IPv6-Prefix masked the byte after the prefix, with the wrong
      bits, instead of the last partial prefix byte.
    - The dictionary "value=" processing instruction skipped only
      strlen("key=\"") characters, so the parsed value was always "=".
      Skip strlen("value=\"").
    - dissect_diameter_avps() did not default Subscription-Id-Type and
      User-Equipment-Info-Type to unknown, so they were treated as E.164
      and IMEISV.
    - 3GPP Path used the absolute comma offset as the item length, so the
      second and later entries were too long and the last one was lost.
    - 3GPP QoS transfer delay compared the whole octet instead of the
      6-bit value, giving nonsense delays.
    - SM-RP-UI did not restore the column writability or p2p_dir when GSM
      SMS threw an exception, and never restored p2p_dir at all; use
      TRY/FINALLY.

    The spare bits of DER-S6b-Flags (3GPP AVP 1523) were registered with a
    bitmask of 0 although they are in the bitmask array given to
    proto_tree_add_bitmask_with_flags(), so every S6b DER with that AVP
    failed DISSECTOR_ASSERT(hf->bitmask != 0) when dissected with a tree.
    Give them the mask 0xFFFFFFFE, as the other flag AVPs with one defined
    bit have.

    The map from command name to SRT row index was one global shared by
    every Diameter SRT table, and each table's init function replaced it.
    With two tables (two -z diameter,srt options or two Service Response
    Time dialogs) a row index created by one table was used by the other,
    which never created that row, and add_srt_table_data() wrote past the
    end of its procedures array. Look the command up in the rows of the
    table being updated instead.

    grouped_avp() cleared the grouped state (dis_gouped, group_avp_code and
    the Vendor-Id) when a nested grouped AVP ended, so the AVPs after it in
    the enclosing group were treated as top level: a Result-Code after the
    Granted-Service-Unit of a Multiple-Services-Credit-Control became
    Result-Code-Command-Level and was missing from the MSCC text. Restore
    the state of the enclosing group instead. Also undo the dissection
    depth (and the group state) when an exception leaves a grouped AVP, so
    that the next PDU of the same TCP segment does not start with a higher
    depth.

    In the "diameter_avps" dissector an EAP-Message AVP let EAP replace the
    "AVPs:" list in the Info column. Put the list back after the AVPs.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 48a21dcb7220557599f63c4e91b0c171f7e870bf)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit cf1ef40389e49ed5151146354b943dd99eb4374b)

    Backport: Keep the older grouped-AVP tree and IMSI helper APIs; omit absent group summary strings and adapt comma scanning.

    Assisted-by: Codex

    (cherry picked from commit 44a2ffcd2c9b7d8676991362df575d7377f8ba0c)

    Backport: IMSI association and libxml dictionary processing are absent; retain existing grouped-AVP guard and apply applicable field and statistics fixes.

commit 13b6d2028e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 19:03:59 2026 +1000

    Modbus: Fix quadratic request matching and short PDU handling

    Every Modbus query was prepended to a per-conversation list that was
    never trimmed, and every response walked that list looking for a
    matching function code, transaction id and unit id. A capture with many
    queries and many (unmatched) responses took O(queries x responses)
    time, 38 seconds for a 668 KB file. Replace the list with a wmem_map
    keyed on transaction id, unit id and function code that holds the most
    recent request. The matching result is unchanged: the most recent
    earlier request with the same key. The match is made on the first pass
    only and stored per frame as before.

    classify_mbrtu_packet() and dissect_mbrtu_pdu_common() kept the length
    in an uint8_t/uint16_t, which truncates for large PDUs. Use unsigned,
    and return early from the RTU dissector when fewer than 4 bytes are
    present, so that len - 2 and len - 3 cannot go negative. Return from
    dissect_mbtcp_pdu_common() after the invalid length expert info when
    the length field is below 2 instead of creating a zero-length subset.

    The modbus_conversation structure in packet-mbtcp.h changes from a
    list to a map.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5cde8a9cfac13bbc5ebb66593d5d4851406e84f6)

    (cherry picked from commit 15370477032b2cb00726f420689e2fab2e78c1cd)

    Backport: Place the short-PDU guard before older header reads and retain the older response-state fields.

    Assisted-by: Codex

commit f008e07bdf
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    PIM: Bound Register nesting and fix several decoding errors

    A Register message carries an IP packet, which can itself carry a PIM
    Register, so a packet could nest IP and PIM hundreds of levels deep
    until the generic layer limit assertion fired. Both the PIM v2 and v1
    Register dissection now allow at most 4 nested Register messages per
    packet, track the nesting with p_get_proto_depth(), wrap the nested
    call in increment/decrement_dissection_depth() and add an expert info
    (pim.nesting_too_deep) beyond the limit.

    The DF-Election subtype was read from the first byte of the message
    instead of the high nibble of the second one, so the Backoff and Pass
    bodies were never dissected. The Join Attribute RLOC address family was
    read at the address offset instead of the attribute offset, the RLOC
    item showed the outer address instead of the RLOC, and the IPv6 text
    used the IPv4 formatter.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit c604286642fffeca2ab16cba2458abf284df79d7)

    Assisted-by: Codex

    (cherry picked from commit 2a1cdbdbf860e3217f012d671e445ae4b109136e)

    Backport: Use the release address-family constants and headers.

    (cherry picked from commit 3920009745fbca5a53c4c03d0d8c11db9520be55)

    Backport: Use older address handling; omit unrelated expert fields absent on this branch.

commit 24f3f94145
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:32 2026 +1000

    PFCP: Bound grouped IE nesting and fix session tracking costs

    Grouped IEs called back into the IE dissector with no limit on the
    depth. A single UDP datagram with thousands of nested grouped IEs
    (3GPP or vendor/BBF) overflowed the stack when no tree was built.
    Stop at a nesting of 16 with a new expert info.

    dissect_pfcp_grouped_ie_wrapper() dereferenced its data argument, which
    is NULL when the vendor IE dissectors are called from Exported PDU.
    Return 0 in that case.

    The session tracking map was hashed on the SEID only and pfcp_fill_map()
    inserted the cross product of all SEIDs and addresses of a message, so
    a message with many F-SEID IEs took cubic time. Hash the address too and
    collect at most 64 SEIDs and 64 addresses per message, with an expert
    info when the limit is hit.

    Also fix several small bugs: the Remote GTP-U Peer network instance
    offset was advanced twice, pfcp_seid_exists() compared only 32 bits of
    the SEID, the F-SEID session lookup used a pointer as the key, the IE
    length parameter of dissect_pfcp_ies_common() was truncated to 16 bits,
    the BBF NAT port forward text always showed the first entry's address,
    and the SDF filter SMMII loop did not move to the end of each instance.

    The Nokia User Location Information IE called a "gtpv2.uli" dissector
    handle that was never registered, so every such IE hit the
    DISSECTOR_ASSERT in call_dissector() with default settings. Call the
    exported dissect_gtpv2_uli() instead. Also show the Source IP Address
    mask/prefix length as a number instead of reading 16 bytes as an IPv6
    address at the 1-byte field.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 8f301eebc13f5217f1bac14555c2907df37e0b80)

    Backport: Add both GTP headers needed for the direct ULI decoder call.

    Assisted-by: Codex

    (cherry picked from commit 0dd1ee22b951e5444c964113f3844dd0f8225dde)

    Backport: Use signed 32-bit offsets and lengths; omit absent SMMII decoding and retain the existing RTS flag check.

    (cherry picked from commit 93a1671f6eb36f8889cd2ab7d2ceddd8743908cc)

    Backport: Use existing GLib session maps with direct keys; omit absent Nokia ULI extensions.

commit 860d2e8cb5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    PROXY: Bound nested PROXY headers, end SSL sub-TLVs at the TLV end

    The payload after a PROXY v2 header can be another PROXY header, which
    re-enters the PROXY dissector through the TCP or UDP heuristic tables
    with no limit of its own. Count the nesting per packet and stop with an
    expert info beyond 4 levels.

    An SSL TLV dissected its sub-TLVs up to the end of the whole header
    instead of the end of the SSL TLV, so following TLVs were shown nested
    under it, and an SSL TLV shorter than 5 bytes was not checked. The
    sub-TLVs now end with the SSL TLV, which is skipped as a whole.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit af61759b2b28b83e2cb678af5099a61a945a02c9)

    Assisted-by: Codex

    (cherry picked from commit bcd9331acfd188899985901917b04a135da71e52)

    (cherry picked from commit 9994f01a977a831f92435cbc1a1d7a341fb2a1be)

    Backport: Use explicit integer encoding while applying SSL sub-TLV bounds.

commit b4a3748593
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 12:30:39 2026 +1000

    LBMR: Fix TNWG option loop and swapped LBTRDMA arguments

    dissect_lbmr_tnwg_opts() added each option length to a running total
    and then subtracted that total from the remaining length and added it
    to the offset on every pass, so offsets were wrong from the second
    option on. The context instance, address and domain handlers also
    return their own length byte, which can be 0, so a zero-length option
    never advanced the loop and hung the dissector. Use the length of each
    option, and add an expert info and stop when the length byte is shorter
    than the option header.

    The TNWG interest records had the same problem: a record length of 0
    kept the record loop at one offset for up to 65535 records, which ends
    in the too-many-items error. Stop with an expert info when a record is
    shorter than its header.

    In the TIR LBTRDMA case, lbtrdma_transport_add() and
    lbtrdma_transport_source_string() take (address, session_id, port) but
    were called with port and session_id swapped. Pass them in the right
    order.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 569ef5e1c7bcee4d2b626aeae5f5bba48391c1b9)

    (cherry picked from commit 278b47a3073334bdf576991c4c873af5f88858bc)

    Backport: Preserve the older contents helper signature.

    Assisted-by: Codex

commit f94536b08c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 19:55:21 2026 +1000

    IuUP: Fix an out-of-bounds read, an assertion and a loop

    The heuristic dissector passed a byte from the packet as the length of
    the two-octet array given to crc6_0X6F(), reading far beyond the stack
    array. It also read the header without checking that three bytes are
    available. Pass the constant length 2, which matches the header CRC
    computed by add_hdr_crc(), and reject packets shorter than 3 bytes.

    dissect_rfcis() used DISSECTOR_ASSERT(c < 64) on the number of RFCI
    records, which a packet can exceed. Report it with an expert info and
    stop. dissect_iuup_payload() did not advance when the RFCI had no
    subflows or all lengths were zero, looping forever; break out when no
    bits were consumed.

    The heuristic CRC10 check for data PDUs with CRC used offset 4 instead
    of 2 as add_payload_crc() does, reading two bytes past the end of the
    packet. Use offset 2.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 3de592af82382392e21cd284341f452a49ea9145)

    (cherry picked from commit e55e877081e6e7de9c6af89afa89cdf6a0103f4d)

    Backport: Adapt the RFCI decoder declaration to the older branch.

    Assisted-by: Codex

commit 82889e8ef8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 12:37:09 2026 +1000

    InfiniBand: Fix uninitialized read and multi-pass state in CM and RC send

    The connection context saved when a CM REQ is seen was allocated
    without being cleared. For an IPv4 REQ only four bytes of each GID were
    filled in, but a CM REP carried over IPv6 or InfiniBand addresses made
    attach_connection_to_pinfo() use sixteen bytes, so the remaining bytes
    were read uninitialized when creating the conversation. Allocate the
    context with g_new0() and record the address type and length of the REQ
    in it, then use those instead of the type of the REP packet.

    parse_PAYLOAD_reassemble_tvb() cleared pinfo->fd->visited and added the
    RC send fragment to the reassembly again on every dissection pass.
    Never write the visited flag, add fragments only on the first pass, use
    the existing reassembled-id lookup on later passes and set the
    conversation reassembly flags only on the first pass.

    create_conv_and_add_proto_data() overwrote pinfo->src with the LID
    address on the first pass only; use a local address. The Trap 259
    DataValid and SWLIDADDR fields were registered as FT_IPv6 but have a
    length of two bytes; make them FT_UINT16. Empty the CM context table
    when a capture file is closed, and add a missing break in
    dissect_general_info().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5cca374385ceb1b7575add61729747e907fb759e)

    (cherry picked from commit 9647fe85fedcc841d29248be515078891f8e5965)

    Backport: Keep the older conversation-data layout; reassembly enablement is already initialized only when allocating state.

    Assisted-by: Codex

    (cherry picked from commit ff515133ab618ae154f6f66f0b8d4d0a131e9c90)

    Backport: RC Send reassembly is absent; retain CM initialization, address state and capture cleanup.

commit 6b311f9436
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 12:37:10 2026 +1000

    IPPUSB: Initialize the PDU state and pass no data to HTTP

    pdu_store() allocated the multisegment PDU with wmem_new() so nxtpdu
    and the other fields were uninitialized and later tested; use
    wmem_new0().

    ippusb_last_pdu was never reset between capture files, and -1 was
    converted to 0xFFFFFFFF for the tree lookup, attaching stray fragments
    to an unrelated PDU. Look up only when it is not negative and reset it
    in a cleanup routine.

    The urb_info_t from the USB dissector was passed as data to the HTTP
    dissector, which treats it as a struct tcpinfo. Pass NULL instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 59a2bf93dc4f6193b2db5e016c1de8436ee1568d)

    (cherry picked from commit beeb6a149cada7a2788f7cf6393dbef037a7fc74)

    Backport: Keep older return-value and shutdown handling; add per-capture cleanup separately.

    Assisted-by: Codex

commit e86dc26b9d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:20 2026 +1000

    IEEE 802.15.4: Fix TAP padding leak and length underflows

    dissect_ieee802154_tap_tlvs() collected the TLV padding in a GByteArray
    using proto_tree_add_bytes_item(), which leaked the array when the item
    threw on a truncated packet. Add the item normally and compare against
    a static array of zeros with tvb_memeql(), so nothing is allocated.

    A TAP header length smaller than 4 made the TLV subset length wrap;
    reject it. A frame shorter than header plus MIC made a negative
    reported length when decryption failed; report an expert info and stop.

    The retry of the PAN ID dissector table with the destination PAN ID
    looked up the source PAN ID again; use the destination PAN ID.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit f4642b52c4c7c067f3534e4cfb9be25d2243e9da)

    (cherry picked from commit 3da804230e0be29b3714d9c284bc8866f5e52ae5)

    Backport: Preserve older dissector and padding-offset APIs.

    Assisted-by: Codex

commit d583c77f5d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:19 2026 +1000

    HTTP: Do not assert in streaming mode or on repeated CONNECT tunnels

    A Content-Length header in a message that is being dissected in
    streaming chunk mode (Transfer-Encoding: chunked with a streaming
    content type such as application/grpc-web, or pipelined messages that
    pick up the streaming state of another message) hit a
    DISSECTOR_ASSERT_HINT(). Add the "Content-Length and Transfer-Encoding
    set together" expert info and ignore the Content-Length instead, as
    Transfer-Encoding overrides it (RFC 9112 section 6.3).

    The assertions in dissect_http_on_stream() that the headers exist on a
    later pass and that all streaming state exists when a new streaming
    reassembly is started can be reached by pipelined messages too. Dissect
    the message without streaming mode when that state is missing instead
    of asserting.

    The http_seq statistics created a child node when the referer was the
    request URI itself, and walked the whole parent chain for every packet,
    which was quadratic for long referer chains. Do not create such a node
    and limit the walk to 64 levels.

    Two CONNECT requests to the same target, both answered, made the
    tunnelled data recurse through HTTP until the layer depth assertion in
    packet.c fired. Count the tunnel nesting with
    increment_dissection_depth() and pass the data to the data dissector
    with an expert info before the layer limit is reached.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6b443de041ada93792089c35e0749c6d7392a84f)

    (cherry picked from commit f3f82b21e54f65d938b916ac2ccdbb879e346ed1)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

    (cherry picked from commit 15bd31caf814c73950b22c7d46f25b9183b45b02)

    Backport: Initialize headers before body handling and retain the older tunnel lookup API.

commit c1690afd59
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 17:39:51 2026 +1000

    GTP: Fix NULL conversation dereference and uninitialized PDCP data

    An End User Address of PDP type 0x8d (IPv4v6) with length 18 added its
    dynamic IPv4 placeholder to the IPv6 field, a field type error that is
    reported as a dissector bug (and aborts with
    WIRESHARK_ABORT_ON_DISSECTOR_BUG). Use the IPv4 field, as the length 6
    case does.

    With session tracking and deinterlacing enabled, the conversation
    looked up with CONVERSATION_IP is NULL for IPv6 packets, and its
    conv_index was dereferenced. Use CONVERSATION_IPV6 for IPv6 and fall
    back to the non-deinterlaced map when no conversation is found.

    Zero-initialize the struct pdcp_nr_info passed to the PDCP NR
    dissector; only some members were set, so it read uninitialized stack.

    Return 3 + length from decode_gtp_data_req() instead of an absolute
    offset, swap the GSN address types of the 4 and 16 octet address cases,
    use the IE length for the MBMS Service Area subset, and keep the
    TEID-to-IMSI strings in file scope so they are not leaked by the
    epan-scope map being reset.

    The deinterlaced session lookup also leaked the list of map keys when it
    found a match. Free it on that return too.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6c17191aa4a6e36776904070722df4a73beb165c)

    (cherry picked from commit d10fb27cdab91a0faa47db8fc8a6a455dbdb4dea)

    Backport: Omit deinterlacing changes because that feature is absent; retain PDCP initialization.

    Assisted-by: Codex

    (cherry picked from commit 09a5fa37887b6da98e8574cfffae700d7e760241)

    Backport: Omit absent IMSI mapping; retain PDCP initialization and the corrected CDR return length.

commit 631ef981b7
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 17:39:51 2026 +1000

    GTPv2: Bound grouped IE nesting and fix NULL dereferences

    Grouped IEs call dissect_gtpv2_ie_common() recursively with no limit,
    so a datagram of nested grouped IEs exhausted the stack or hit the tree
    depth error. Limit the nesting with a per-packet depth counter, and
    stop with an expert info before the generic dissection depth limit.

    The IMSI and Cause decoders wrote to args when session tracking was on
    even though callers such as the JSON 3GPP dissector pass NULL. Check
    args first.

    find_conversation_strat() can return NULL (IPv6 or without
    ip.track_conv_id), and its result was dereferenced for the conversation
    index. Use index 0 in that case.

    gtpv2_match_response() also matched unrelated message types as
    responses because the default case fell through to the lookup. Return
    NULL there. Also check that at least 4 octets remain before subtracting
    4 from the remaining length.

    In the GSM key and triplets MM context the offset returned by the
    triplet decoder was dropped, so the fields after the triplets were
    decoded at the wrong offset, and the UMTS key, quadruplets and
    quintuplets MM context only decoded the Higher bitrates flag when no
    data was left. Use the returned offset and test for remaining data.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 4e1cf713827776319663763a4c3e09b51afc7cf9)

    Backport: Retain the separate Higher bitrates length read used on
    release-5.0.

    Assisted-by: Codex

    (cherry picked from commit e91efa44828be905fca9658e07c1bfa90c3f7616)

    Backport: Keep signed offsets and omit absent conversation-deinterlacing state.

    (cherry picked from commit f814c0ab145285ce7ab188db31d1407ad7aa2e83)

    Backport: Omit the IMSI session assignment added after 4.4; retain nesting and existing argument guards.

commit 145edcae28
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 08:49:38 2026 +1000

    GIOP: Fix unbounded work and two-pass differences

    The request list was a GList that was searched linearly for every
    request and reply, so a capture with many requests took quadratic time
    (40k requests and replies took about 100 s). Replace it with wmem_maps
    keyed by frame number and by request id, source address and port. The
    latest request still wins, as before.

    A frame with the ZIOP flag was inflated with no limit on the output, so
    a small packet could allocate hundreds of megabytes. Inflate with a
    local zlib loop that stops when the result would exceed the
    giop.max_message_size preference and reports an expert info.

    Reading the stringified IOR file (giop.ior_txt) called decode_IOR() with
    a NULL pinfo, which dereferenced pinfo->pool and crashed in the init
    routine. Use a dummy packet_info with a per-line scratch allocator
    (frame number 0, so expert infos are dropped) and catch exceptions.

    The repository id of a request was looked up in the object key hash on
    every pass, but the hash is filled from IORs seen later in the file, so
    a second pass could show different results than the first. Resolve it
    on the first pass and keep it. Do the same for the message type of the
    last fragment, and only update the request id to message type map on
    the first pass.

    The heuristic sub dissector probes were run without a TRY, so one probe
    that threw ended the search and marked the message malformed. Catch the
    exception and go on with the next sub dissector.

    Smaller fixes: tk_ulong and tk_ulonglong values did not get their text
    because the item was not assigned; the wchar length was an int8_t; the
    offset-progress checks in the struct, sequence, array and union loops
    did not update the old offset; sequence, array and value typecode
    parameters read fields inside an encapsulation with the outer byte
    order and boundary; the proto depth is reset at the start of each PDU in
    case an exception left it raised; the last-fragment override stored the
    message size in host order where the header holds wire order; format_text
    was given a length that included the terminating NUL; LocateReply
    exception and addressing disposition were added to the wrong tree and a
    LocateRequest length error was attached to the wrong item.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6b8a5b21215c9cb7af7061fd199fe44a6dbbd337)

    (cherry picked from commit 459703bd76ae85c7cd313bfdcdd7121af912f204)

    Backport: Retain older debug guards and decoder signatures.

    Assisted-by: Codex

commit 3eac92dd69
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 08:49:38 2026 +1000

    GeoNetworking: Fix IntX assertions, duplicate flag and lat/lon

    A security IntX field longer than four bytes made the dissector add a
    70-bit value to a 32-bit application ID field, and a 64-bit value to the
    32-bit local region field, so both ended in a bit width or type
    assertion. Fields longer than four bytes now only get the too-big
    expert info, and the PSID name is only shown for the application ID.

    The duplicate packet detection added its tree item, expert info and
    Info column text only on the first pass, so they were lost on a second
    pass such as with tshark -2. The result is now stored per packet with
    p_add_proto_data() on the first pass and shown on every pass.

    The latitude and longitude display functions multiplied the int32 value
    by 6, which overflowed, and used abs() on INT32_MIN. Compute them with
    64-bit values.

    IS_HT_KNOWN() lacked outer parentheses, so !IS_HT_KNOWN(x) only negated
    its first term and header types 0x43 and 0x33 were decoded misaligned.
    Add the parentheses, and accept the GeoAnycast sub-types (0x31, 0x32),
    which the check hidden by the missing parentheses rejected. The length
    counters of the certificate, signer, header field, recipient info and
    trailer loops are unsigned and were decremented by the size of each
    element, which wrapped when an element was longer than the declared
    length; stop the loop at zero instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 0d07cd498c99247eb47941782eae0b40ca3e27c2)

    (cherry picked from commit 05983fd3f783a789b17795b416545b71e93b496d)

    Backport: Bound IntX fields in the older conditional layout.

    Assisted-by: Codex

commit 00d5f0744c
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Nov 5 19:54:42 2026 -0500

    FMP: Fix crashes and several decoding errors

    dissect_fmp_devSerial() was called with a NULL packet_info for the
    serial disk identifier, which crashed in dissect_rpc_opaque_data().
    Pass pinfo.

    The FMP file handle is an FT_BYTES field but was dissected with
    dissect_rpc_string(), which asserts on the field type. Use
    dissect_rpc_data().

    The general string decoder checked the length word that follows the
    encoding mode instead of the encoding mode itself. The volume
    management info dissector did not advance past the volume type, so the
    server based list length and the following data were misread, and the
    hierarchical volume case dropped the offset it returned.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 103beb8f5905b83aa2fded66a4b8b76f60c2b7cb)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 9413fbf49afad833ed3e84e16d82a2edb8d9eceb)

    Backport: Use the older RPC decoder signatures.

    Assisted-by: Codex

commit eeda1ebd42
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 16:52:53 2026 +1000

    E2AP: Fix NULL tvb on TNL address and RAN function dissector choice

    An extended tnlAddress BIT STRING of length 0 gives a NULL value tvb,
    which was passed to tvb_captured_length() and failed an assertion.
    Check for a NULL tvb, show an IPv6 address only for a 16-byte address,
    show both addresses for a 20-byte IPv4 plus IPv6 address, and nothing
    for other lengths.

    When a RAN function name was stored, the initial service model
    dissector was taken from the available dissectors of an unrelated
    table entry, usually KPM, instead of those of the matched RAN function,
    so an RC indication was decoded as KPM and marked malformed. The
    choice was then corrected from the OID only when the frame had already
    been visited, so the decoding differed between one pass and two.
    Use the matched RAN function, and store the OID and update the mapping
    only on the first pass.

    Also set the top-level tree pointer after the protocol subtree is
    created, so the message name is again appended to the protocol item;
    keep track of the highest version when choosing a RAN function
    definition dissector; and read tnlPort at its bit offset, as it does
    not have to be octet-aligned.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6e1d7718598bb69b537e73fe4ea8894741636e6f)

    (cherry picked from commit 5690ef904ac8586f10008e5e6f0eb180bdc29aac)

    Backport: Set the release top_tree variable after tree creation; preserve the older private-data layout.

    Assisted-by: Codex

    (cherry picked from commit 9e606de793af74b54d68a171443254cfff5bf613)

    Backport: Retain the older first-available RAN decoder selection; the newer highest-version selection loop is absent. Regenerate ASN.1 output.

commit 137b2d814d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 2 11:26:16 2026 +1000

    DECT-NWK: Check identity bit widths, find each IE after the one before

    The fixed, network assigned and portable identity IEs passed their
    7-bit value length to proto_tree_add_bits_item() as a bit width, and a
    width of 0 or above the field size made it report a dissector bug.
    Treat such a width as malformed.

    The variable length IEs continued where their dissector stopped, not
    after the IE length, so an IE of an unexpected length was followed by
    IEs made up from its contents.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit ff3193289ad4245a28ba6ceeb6e8725dab796596)

    (cherry picked from commit 7fa24a9f261398dc5db7f0ee2bae8fd5c8d45d2c)

    Backport: Use bounded integer bit fields with explicit big-endian encoding on the older branch.

    Assisted-by: Codex

commit 0db952ccb3
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 02:53:56 2026 +1000

    CoAP: Do not assert without a TCP conversation, check OSCORE lengths

    A CoAP message whose port type is TCP but which was not dissected by
    TCP, for example in an Exported PDU naming "coap" with a TCP port type,
    has no conversation, and find_or_create_conversation_noaddrb() hit
    DISSECTOR_ASSERT(conv). Create the conversation instead.

    The OSCORE option did not check the Partial IV and the kid context
    against its own length. The Key ID length then wrapped, and the fields
    were read from the payload after the option. A request could also store
    a Partial IV of 6 or 7 octets, which a response without a Partial IV
    handed to OSCORE as the request's, and packet-oscore.c asserted that it
    is at most 5 octets. Stop at a field that does not fit in the option,
    and have OSCORE check the request Partial IV length too.

    The payload of a message over TCP took everything to the end of the
    segment, including the messages after it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e9e755c0696d89bf78c3be5812e0cf85879e307a)

    (cherry picked from commit 3c3df21d80ae5dc7f86fc648420db7c27987c3de)

    Backport: Keep the older payload decoder, which receives the explicit message end; retain conversation and OSCORE guards.

    Assisted-by: Codex

commit 2c803c7cff
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 10:18:28 2026 +1000

    BEEP: Loop over the messages in a frame, step over the previous payload

    Messages after the first one in a frame were dissected by a recursive
    call for each, so a frame of 500 or more messages (17 KB on TCP 10288)
    failed the DISSECTOR_ASSERT() in increment_dissection_depth(). Loop
    over the messages in dissect_beep() instead.

    When a frame continued the payload of the previous frame's message,
    the first pass created the per-frame data with the bytes left but did
    not step over them, so the frame was parsed from its first octet; the
    later passes did step over them, and tshark -2 gave different output.
    Create the per-frame data first and then always step over what is
    left, and add that payload to the BEEP tree instead of the top level.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 2a1d04b08772de4921fb9905db011d15ceae5cfe)

    (cherry picked from commit 92228e793eb4a67485901d0a3a2db2c5b083b052)

    Backport: Remove the older payload-preprocessing block now handled by the message loop.

    Assisted-by: Codex

    (cherry picked from commit 69dc79be029c8422c165e628dd1ea8ea35ab8817)

    Backport: Replace the older payload pre-processing with the bounded message loop.

commit fa6466c71e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 07:00:16 2026 +1000

    AVSP: Stop nested frames before the layer limit, step over samples

    Every AVSP subtype hands the rest of the frame to the ethertype
    dissector, which can carry AVSP again. A frame nesting a few hundred
    levels (about 3.5 KB) reached the layer limit in call_dissector_work()
    and failed its DISSECTOR_ASSERT(). Stop with an expert info when the
    frame is close to the limit.

    GREENT did not step over the sample data, so the header of a second
    sample was read from the first one's data. Also show "48bit UTC
    timestamp" for the GRE TAP UTC version and print the DzGRE IDs in
    decimal without a "0x" prefix.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 4c96cea92879e413bfd47a5182533e3b10e3b082)

    (cherry picked from commit 33f7a1e89795e4812bfcee1ee1c4693229eb3ff9)

    Backport: Retain the nesting bound; newer GRE telemetry subtypes are absent.

    Assisted-by: Codex

commit 0a26c9dfc4
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    pcap_pktdata: Bound nesting through PPI and PKTAP

    PPI and PKTAP hand the link-layer type found in the packet to this
    dissector, and LINKTYPE_PPI and the PKTAP type map back to the PPI and
    PKTAP dissectors, so a packet can nest them thousands of levels deep
    until the generic layer-depth assertion fires. Keep a per-packet nesting
    count with p_get_proto_depth()/p_set_proto_depth(), stop with an expert
    info beyond 4 levels, and wrap the nested call in
    increment_dissection_depth()/decrement_dissection_depth().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit c9a08f69d5e7cf976756a9e41ca77ddd2550c6a9)

    Assisted-by: Codex

    (cherry picked from commit c8f358f510ca05ca5aec486cc80f3dfb906dd505)

    Backport: Guard the existing encapsulation call and restore depth on exceptions.

    (cherry picked from commit 60f737a6be7723126bcbdb77efeee261e3b49bd2)

    Backport: Apply the fix to the older branch context.

commit 6dba05adf5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:16 2026 +1000

    DISP: Close the parenthesis of an unknown standard update strategy

    The Info column text for an unknown standard update strategy was
    "standard(N" without the closing parenthesis. Add it.

    disp.cnf gave the Time type (GeneralizedTime) the field type FT_STRING,
    but dissect_ber_GeneralizedTime() adds an FT_ABSOLUTE_TIME value, so a
    valid lastUpdate, beginTime, updateTime or update window failed the
    field type check in proto.c ("Dissector bug") with a tree; every
    updateShadow has a mandatory updateTime. Drop the TYPE_ATTR so the
    fields get the type asn2wrs uses for GeneralizedTime.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6fc075b8987779ba43e6baec700801628c686de9)

    (cherry picked from commit 55984419c1d31bfc3bbf84e1716eb810d7b12d7a)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

    Assisted-by: Codex

commit de0e8a39a6
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Oct 25 23:17:43 2026 +1000

    AIN: Dissect ReturnResult instead of asserting

    ANSI TCAP passes a ReturnResult to AIN with the operation code of the
    invoke it answers, which reached DISSECTOR_ASSERT_NOT_REACHED(). Dissect
    it with the operation code ANSI TCAP keeps.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit d978766097aea5707b811155dc77afa567e93038)

    (cherry picked from commit 1984a3b19f02074eb4054e09a3b7c7f7b0caabeb)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

    Assisted-by: Codex

commit 5679cee490
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Oct 27 17:44:58 2026 +1000

    ANSI MAP: Dissect the SMS bearer data only once

    The SMS bearer data was kept after it had been handed to the IS-637,
    IS-683 or IS-801 dissector, and every following TeleserviceIdentifier
    or ServiceIndicator parameter handed it over again. The BER SET decoder
    accepts repeated members, so a message with a few hundred of them added
    a layer each time until the layer limit in call_dissector_work() failed
    its DISSECTOR_ASSERT(). Forget the bearer data once it is dissected.

    A ReturnResult whose invoke ANSI MAP had not saved, for example because
    the invoke was malformed, was dissected with the operation code of the
    last packet dissected, so its decoding depended on the order packets
    were viewed in. Use the operation code TCAP matched instead. That is
    also used without a transaction ID, which was passed to snprintf() as
    a NULL string.

    Also show the matched operation for ReturnError and Reject in the Info
    column, and keep counting known operations in the statistics after an
    unknown one; a static flag stopped that for the rest of the session.

    Every MSLocation parameter was malformed, because the 1 or 2 octet
    resolution was added to an 8-bit field with length -1, and a 1-octet
    SMS-TeleserviceIdentifier threw, because it was read from the enclosing
    tvb.

    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    Assisted-by: Claude
    (cherry picked from commit 2e11416fba3014438eff60c8f50048c8a40299a0)

    (cherry picked from commit 3fcb594535515b5fb642ae5ccdadf45ea9775de9)

    Backport: Retain signed decoder signatures and regenerate the dissector.

    Assisted-by: Codex

    (cherry picked from commit 8e5638835e513c6c7fdcbe7b18982b79564587c2)

    Backport: Preserve the older value-string API and regenerate ASN.1 output.

commit 77c420e9cb
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 03:05:35 2026 +1000

    CredSSP: Do not hand a missing authInfo or pubKeyAuth to GSS-API

    When the [2] authInfo or [3] pubKeyAuth tag of a TSRequest held
    something other than an OCTET STRING, no tvb was returned and NULL was
    passed to the GSS-API wrap dissector, which hit the DISSECTOR_ASSERT
    that a tvb is given in proto_tree_add_item(). An Exported PDU naming
    "credssp" reaches it. Stop there, as the negoToken body already does.

    The Remote Guard package of a TSRemoteGuardCreds was kept from an
    earlier packet, so a later credential buffer could be decoded with the
    wrong package.

    The [1] credentials of a TSCredentials had the same problem: when they
    are not an OCTET STRING, NULL was passed to the password, smart card or
    Remote Guard credential dissector, which hit an assertion in the tvbuff
    code. The TSCredentials of an anonymous NTLMSSP exchange are decrypted
    without any keys, so this is reached with the default preferences.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 812b104e167df7121471d672bd23ad272c8fd75b)

    (cherry picked from commit 457fdefe601071fdd128f32d284ba1b723d82f43)

    Backport: Retain signed sentinel initializers and regenerate the dissector.

    Assisted-by: Codex

    (cherry picked from commit 13252e84f2d88f46d5a5e6bce749b761a3c2ef6a)

commit d84ad158c8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    PLDM: Fix a leak and 16-bit offset wrap

    The type map for GetPLDMCommands was created with wmem_map_new() in a
    NULL allocator scope for every request, so each request leaked a map,
    and the map held only the last request, rewritten on every pass. Create
    one autoreset map in proto_register_pldm(), insert into it only on the
    first pass, and remember the type looked up by a response with
    p_add_proto_data() so later passes decode it the same way.

    parse_fru_record_table(), dissect_FRU() and dissect_FWU() kept the
    offset in a uint16_t. For a PDU longer than 65535 bytes the offset
    wrapped and the FRU table loop reread the same bytes until the tree
    item limit stopped it. Use unsigned offsets.

    Compute the padding before the FRU table CRC as unsigned and only add
    the CRC when at least 4 bytes remain, and print a non-printable alpha
    version byte as '?'.

    dissect_pldm() did nothing without a tree, so on a first pass without
    one (tshark -2, the GUI) the request map stayed empty and responses lost
    their type. Dissect without a tree too.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit de72b06743f3da2ab48bdba7ea5523113730b766)

    Assisted-by: Codex

    (cherry picked from commit 285a22bcc8f64f14afbe3855fd915109b6fef951)

    Backport: Omit the FWU offset change because FWU support is absent.

    (cherry picked from commit ae142d8cffd0002ac5968bd4c9c2b6a863dc7c07)

    Backport: Apply the fix to the older branch context.

commit 9a45745ee9
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 12:57:20 2026 +1000

    NIS+: Fix a CPU blow-up in the callback data entries

    dissect_cb_entry() was not implemented and consumed no bytes, while
    dissect_rpc_array() calls it once for each of the up to 0xffffffff
    entries in the packet, so one NIS+ CB RECEIVE packet took about a minute
    of CPU time. The entries are nis_object, so dissect them with
    dissect_nisplus_object(), which consumes data and ends with a bounds
    exception, and remove dissect_cb_entry().

    Add the time item in dissect_nisplus_time() before advancing the
    offset, so it covers the time and not the four bytes after it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 7e5a7f7dd80ee9e111c919513ea5e1ab9c68a765)

    Assisted-by: Codex

    (cherry picked from commit 467de94eed3bc1b780ea882bfe41d8e031268530)

    Backport: Remove the older signed-offset version of the non-advancing callback.

    (cherry picked from commit 09a6f2bf8697231816dc18ff81a9548496317029)

    Backport: Apply the fix to the older branch context.

commit 2a2a4d9a8f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 12:57:20 2026 +1000

    NFS: Fix a CPU blow-up and bad lengths in the SCSI layout and snooping

    dissect_nfs4_vol_indices() added the same 4-byte item num_vols times
    without advancing the offset. With a count of 0xffffffff and in-bounds
    bytes no exception stopped the loop, so one GETDEVICEINFO reply took
    tens of seconds. Advance the offset by 4 for each index.

    With file name snooping enabled, a file handle length of 0x80000000 or
    more became negative and reached tvb_memdup(), which asserts. Reject a
    file handle length that is zero, negative or larger than 128 at the top
    of the three snooping helpers. Skip empty names when snooping READDIRPLUS
    entries, and make nfs_full_name_snoop() check name_len and the start of
    the buffer before reading the character before them, which was a one
    byte heap read before the buffer.

    Decode the SCSI SLICE and STRIPE volumes at the right offsets: start
    and length are 8 bytes and the stripe unit is 8 bytes (RFC 8154), so
    make those three fields FT_UINT64 and advance past them.

    Use the "unknown" permission tables in dissect_nfs4_acemask() for an
    object type that is neither a file nor a directory, so NULL is never
    passed to "%s".

    The snooping hash tables for READDIR parents and unmatched names never
    freed their keys, which leaked 8 bytes per call; free them with the
    table. Free the previous full name when a file handle is matched again
    instead of leaking it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 83ba457dadc3e682d3a13c2efb3e8fe417a95621)

    Assisted-by: Codex

    (cherry picked from commit 879528d78acc16f2b1754faa2c8e0f5f6f03e2ad)

    Backport: Keep the older integer-key snooping map and omit absent READDIR snooping; adapt the SCSI volume loop.

    (cherry picked from commit 0f8445b65f863a2b317b10a0c086f2cb563d6abb)

    Backport: Apply the fix to the older branch context.

commit 39b3e55a3a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 19:03:58 2026 +1000

    LTP: Fix assert on large client service IDs and header offsets

    A data segment with a client service ID that does not fit in 32 bits
    hit DISSECTOR_ASSERT(client_service_id < UINT32_MAX) and aborted the
    dissection. Such IDs now skip the dissector table lookup and fall
    through to the data dissector.

    ltp_check_reception_gap() computed next_fst - (prec_lst + 1) in
    uint64_t, which underflows when the claim does not start after the
    preceding one, producing a huge bogus gap. Return before subtracting
    when next_fst <= prec_lst + 1.

    The version and type of the segment header were added with a byte
    offset passed where a bit offset is expected; use 8*frame_offset and
    8*frame_offset+4.

    dissect_trailer_extn() advanced frame_offset instead of extn_offset and
    always returned 0, so the trailer length was lost. Accumulate in
    extn_offset, and use hf_ltp_trl_extn_len for the trailer length field
    instead of the header extension length field.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 4ed187d775c192487ad4932900ac2e80f0be0cbf)

    (cherry picked from commit 4d15064c071be52c7a452c1275d4bbde102ffcee)

    Backport: Omit absent client-service dispatch changes; retain the older SDNV API when fixing trailer offsets.

    Assisted-by: Codex

    (cherry picked from commit 01919f2fb0cb84ed649667afc8af037387aa0bd5)

    Backport: Apply the fix to the older branch context.

commit d13819323d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 19:55:21 2026 +1000

    ITS: Fix DSRC imports and bit offsets in ItsPduHeader

    CPM-PDU-Descriptionsv1.asn and ISO19321IVIv2.asn imported LaneID,
    NodeOffsetPointXY and Offset-B10..B16 from a module called DSRC, but
    the module in DSRC.asn is ETSI-ITS-DSRC. asn2wrs could not resolve the
    types and registered nine fields as FT_NONE. A CPMv1 OffsetPoint then
    decoded a CHOICE into an FT_NONE field and hit the field type assert,
    which any packet with such a CPM could trigger. Import from
    ETSI-ITS-DSRC, with its OID, like the other ITS modules do.

    The ItsPduHeader body added the byte count returned by the sub-dissector
    to a bit offset, and returned a byte count where a bit offset is
    expected. Shift both by 3.

    dsrc_velocity_fmt printed a signed int with %u; use %d.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 64248e2b0324a89a650c33238ed403fe51dca95f)

    (cherry picked from commit 47ef5ef1ff660cd72bca8e0d7ebb2be0a1e60274)

    Backport: Resolve the velocity format context and regenerate from release ASN.1 inputs.

    Assisted-by: Codex

    (cherry picked from commit 036c55af3c45e4c3e9ee1cea0ae34e0f9c8b160e)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

commit c1af5adcee
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 12:37:09 2026 +1000

    INAP: Fix a dissector bug assertion on the BCSM correlation ID

    The Digits dissector chooses the subtree type with a chain of hf_index
    comparisons and starts with ett = -1. CorrelationID is also used with
    hf_inap_bcsmEventCorrelationID (EventReportBCSM and
    RequestReportBCSMEvent), which was not in the chain, so
    proto_item_add_subtree() was called with -1 and hit its assertion on
    a packet. Add that field to the correlationID branch and leave any
    other unknown user of Digits as a plain OCTET STRING.

    Also report the errorCode, not the opcode, for an unknown
    returnErrorData, reset errorCode for every packet, return the number
    of captured bytes instead of a wrapped 8-bit PDU size, and advance the
    offset past an ExtensionField value without an OID while passing only
    the remaining bytes to the data dissector.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit fde8230d5a4445a4e3b2a59267e7b1c4da432311)

    (cherry picked from commit fff0b99588a18f0bc62ff328f6aed1c0599c20cd)

    Backport: Use the existing expert-info API and regenerate the dissector.

    Assisted-by: Codex

    (cherry picked from commit 9b738ab08b6aedd14f1c5188f9114fbb319488ac)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

commit 9a7486d0cd
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 4 16:22:14 2026 +1000

    fcswils: Fix nested zone set handling and small errors

    A ZONESET zone object was dissected recursively for its members, limited
    only by the dissection depth, which ends in a dissector assertion at
    the maximum tree depth for a few hundred nested objects. The format
    nests a zone set only once, which is also all get_zoneobj_len() accounts
    for, so a zone set inside a zone set now gets a malformed expert info
    and is not dissected further. This also keeps the object lengths
    consistent with what the dissector consumes.

    In the ESS vendor capability object the offset of the next object
    advanced by 12 instead of 8 for each vendor object, and the unsigned
    remaining length could wrap when it was not a multiple of 8.

    The opcode of an exchange was stored on every pass. Update it only on
    the first pass so that redissecting a response in any order uses the
    right request.

    The opcode of the request was kept once per exchange, so when an
    exchange was reused the first SW_ACC was decoded with the opcode of the
    last request on later passes (tshark -2).

    Save the opcode each reply was matched with per frame on the first pass
    with p_add_proto_data() and use it when the frame is visited.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 786306e2de982e6c3ae5eba5aaa65cd6fb1cad6a)

    (cherry picked from commit ba5300ce4015667ce8dccd87e2308233d35720c2)

    Backport: Retain the signed offset argument on the release branch.

    Assisted-by: Codex

    (cherry picked from commit 102848f7713acec60c7c6a12c81c01c708e46941)

    Backport: Apply the fix to the older branch context.

commit 184e0db51f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 4 16:22:14 2026 +1000

    ETW: Bound the formatted message size

    format_message() replaced every "%N" in the event message with the
    value of property N, with no limit on the size of the result. A record
    of a few hundred kilobytes could make the output several gigabytes and
    abort the process. The formatted output is now capped at 64 KiB; once
    the next substitution would exceed it, the rest of the message is
    appended unformatted.

    startLoc was a uint16_t and wrapped for messages longer than 65535
    characters, restarting later appends at the wrong place. Make it
    unsigned.

    Pass NULL instead of the frame's pseudo header to mbim.control, which
    casts its data to urb_info_t, and a zeroed struct tcpinfo to nbss,
    which expects one, rather than relying on the pseudo header happening
    to be all zeros.

    Do not add the PROV_TRAITS provider name when the traits size is less
    than 2, which gave a length of -2 or -1.

    The FILETIME to Unix time conversion used an unsigned subtraction, so
    timestamps before 1970 wrapped to a far future time. Use a signed value
    and normalize the nanoseconds so they are never negative.

    The source and destination addresses, the provider name and the message
    come from pinfo->pool, but were set with col_set_str(), which keeps the
    pointer and requires a string that outlives the packet. Use
    col_add_str() so the text is copied.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit b78ba46648b24181508a768cfa8f8c0e1ee65666)

    Backport: Resolve the provider-name encoding context against release-5.0.

    Assisted-by: Codex

    (cherry picked from commit 4323094614a220e2eb34295cc3d729c1832fed60)

    Backport: Backport MBIM callback-data and column string lifetime fixes; message formatting, traits, SMB dispatch and timestamp conversion are absent.

    (cherry picked from commit e260beebec3d1d66a442f56b3eaed3bd40ee633c)

    Backport: Apply the fix to the older branch context.

commit 8fc9aa3fe4
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Nov 7 21:58:57 2026 -0500

    DVB-CI: Fix CA PMT loop, two-pass state and length checks

    dissect_es() returned -1 for an elementary stream with a descriptor
    that is not a CA descriptor, and the CA PMT loop stored it in an
    unsigned variable. The "<= 0" check missed it and the offset moved
    back by one, so the same bytes were dissected forever when there was
    no tree. dissect_es() now returns 0 on error.

    An open session response created a new conversation on every pass. On
    the second pass each one was inserted behind later ones, which took
    quadratic time and doubled the conversations. Create it on the first
    pass only; later passes find it.

    The CI+ X.509 extensions were enabled around the certificate
    dissection without TRY/FINALLY, so a malformed certificate left them
    enabled, and certificates of other protocols were then decoded with
    the CI+ meaning of these OIDs. Disable them in a FINALLY block.

    The negotiated buffer size and the CAM's proposal were global state
    updated on every pass, so the "Payload too large" and buffer size
    checks differed between one and two passes. Keep the values that apply
    to a frame with p_add_proto_data(). The "Payload too large" expert
    info started at offset 2 with the length of the whole LPDU, threw an
    exception and stopped the reassembly; it now covers the LPDU.

    A hostname connection descriptor, or one of another type, whose length
    was too short and a CA descriptor shorter than 4 bytes gave a negative
    length, so the field ran to the end of the APDU and the offset moved
    back. They now get an expert info and the dissection stops.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 9c8b34ab3a8716f37fc543c0aacde92508d6629c)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 6604f29939a35a623e46d3b8e8f04df3f3921a19)

    Backport: Retain signed offsets and the existing explicit length-comparison casts.

    Assisted-by: Codex

    (cherry picked from commit eafb6e07ee648f4f53b8782d44bcd6176a3cb625)

    Backport: Apply the fix to the older branch context.

commit cbd716633a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 10:35:11 2026 +1000

    BVLC: Dissect the secured message of a Secure-BVLL, not the BVLL again

    For an unencrypted Secure-BVLL the dissector called itself on the same
    tvb from offset 0, which read the same Secure-BVLL header again. The
    only thing that changed per level was the 16-octet signature the
    security wrapper cut off the end, so a datagram of about 8 KB nested
    500 levels deep and failed the dissection depth assertion, on the
    default port. Dissect the data after the security wrapper instead, and
    do not decode a Secure-BVLL inside a Secure-BVLL.

    The security wrapper shortens the tvb to the secured data, and for an
    encrypted Secure-BVLL the dissector returned the shortened length, 0,
    so UDP tried BVLC again and the packet was malformed. Return the length
    from before the wrapper.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 7e03b3bae4954b2db3745e811548bcab1e14ce0c)

    (cherry picked from commit 66e6de8a64d426f9ae057ec6b69b9b6ce7ab8ef1)

    Backport: Use the signed return value of the older BACnet security wrapper API.

    Assisted-by: Codex

    (cherry picked from commit b7210294a4e17ed33d244c1e1f12f859a3593736)

    Backport: Apply the fix to the older branch context.

commit e1fa4e8918
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 10:16:49 2026 +1000

    B.A.T.M.A.N. Advanced: Stop nested payloads before the layer limit, decode TVLVs

    Unicast, broadcast and fragment packets hand their payload to the
    Ethernet dissector, and ethertype 0x4305 calls B.A.T.M.A.N. again. A
    frame nesting a couple of hundred levels (about 4.6 KB) reached the
    layer limit in call_dissector_work() and failed its DISSECTOR_ASSERT().
    Stop with an expert info when the frame is close to the limit.

    When f7461969ff made the offsets unsigned, the TVLV loop's "offset !=
    -1" became "offset != 0", which is false from the start, so no TVLV
    was decoded since. Drop the test, and keep the TVLV length in an
    unsigned so that a length field of 0xfffc or more cannot wrap it to
    less than the header and stop the loop advancing.

    Also read the OGM2 TVLV length from the OGM2 rather than from offset
    16 of the aggregate, read the ICMP v7 and v14 uid (and v14 reserved)
    from the right offsets, and restore pinfo->fragmented when a v15
    fragment number is not supported.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 8bfd58c7ee0c127670293bfef9a9c83368c3b534)

    (cherry picked from commit 98ab15783787e172d782f80a659f9e3f64f5e0d3)

    Backport: Retain signed offsets while removing the obsolete sentinel check.

    Assisted-by: Codex

    (cherry picked from commit 24f689666b4b56c84df48ea741c7ee186695eccd)

    Backport: Apply the fix to the older branch context.

commit 0880e3d743
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    PPI: Fix header length underflow and quadratic A-MPDU list

    The PPI header length was reduced by 8 and then by each field length
    without checks, so a length below 8 or a field longer than what is left
    underflowed, the loop read payload bytes as fields, and the payload was
    never dissected. Report an expert info and stop in both cases.

    Every frame of an A-MPDU listed all the fragments seen so far, so the
    number of tree items was quadratic in the number of MPDUs (about 60 MB
    of -V output for 2000 frames). List the segments only in the last frame
    of the aggregate.

    Also give the aggregate protocol item the remaining length (-1) instead
    of tot_len, which is always 0 at that point.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 18947b1020d8576441e3e657b02112b212ae6862)

    Assisted-by: Codex

    (cherry picked from commit 9ad710203d1876357a924356c47c388767fc8ce4)

    Backport: Apply the fix to the older branch context.

commit bf46a77d31
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    PANA: Bound Relayed-Message nesting, fix Vendor-Id offset

    The Relayed-Message AVP holds a whole PANA PDU and was dissected
    recursively with nothing but the packet length bounding the depth, about
    2700 levels. Count the nesting per packet and stop with an expert info
    beyond 4 levels. A nested PDU shorter than the 16 byte header, or whose
    msg_length is below 16, is now reported with an expert info instead of
    underflowing avp_length.

    The Vendor-Id of a vendor specific AVP was read at offset 8 of the
    buffer instead of offset + 8, so every AVP after the first used the
    first AVP's data to choose its type.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 4bfb67264097d069356a2985365afbaba19b876c)

    Assisted-by: Codex

    (cherry picked from commit df6522b8f00fdb415b4c7b6d995a10669a1d9ad9)

    Backport: Apply the fix to the older branch context.

commit ecd251e764
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    P1: Check for a missing MTAName before using it

    MTAName used the string tvbuff without a NULL check when building an
    address, which fails an assertion for an empty constructed string or a
    wrongly tagged MTAName, reachable from any X.509 Name. Check it in both
    branches.

    Also use the length of next_tvb, not of the parent tvbuff, for the
    unknown content type expert item, and make the content type integer in
    BuiltInContentType a plain local instead of a static.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit e63fe5b4ae5411be9732a462454241f0dbd3e152)

    Assisted-by: Codex

    (cherry picked from commit b2ace7eeaa66d5c3a5c512f6112b02ea970da5c7)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

commit afe5ee2e23
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    PKCS12: Fix stale pointers kept across packets and ber.oid handling

    The object identifier string is kept in a file scope pointer into the
    packet pool and was never reset, so a SafeBag, CertBag, CRLBag,
    SecretBag or PKCS12Attribute without its identifier passed the string
    of an earlier, freed packet to call_ber_oid_callback(), a heap
    use-after-free. Reset it at the start of each of these sequences.

    The salt tvbuff and iteration count used for decryption were likewise
    left over from an earlier packet when a PBE algorithm had no
    parameters, and generate_key_or_iv() used the freed tvbuff. Record per
    packet, with p_add_proto_data(), that parameters were parsed, and treat
    them as missing otherwise. The key derivation itself is unchanged.

    The PFX and AuthenticatedSafe dissectors changed the ber.oid entry for
    id-data and restored it only on normal return, so an exception left
    later CMS data being dissected as PKCS#12, and each call created a new
    dissector handle. Create the two handles once in handoff, and restore
    the previous handle in a cleanup function that also runs on exceptions,
    which also makes nesting restore correctly.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 8e6b0d8bbbd0a6bb9fb49d7d62bc14a57a2da9a1)

    Assisted-by: Codex

    (cherry picked from commit 87d4578f5914743a1cda3bc2affda835374f2aa4)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

commit 6378f08d69
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 12:57:20 2026 +1000

    nfapi: Fix HARQ TDD CPU loop and bound TLV recursion

    The HARQ indication Rel9 and Rel13 TDD value functions only raised an
    expert info for an invalid mode and then ran a loop of up to 255 or
    65535 iterations that reads nothing, so a 7-byte TLV could cost
    thousands of subtree operations. Return right after the expert info.

    dissect_tlv_list() recursed without limit through the TLV decoders
    (for example RX SR indication body to the SR PDU and back to the TLV
    list). Count the nesting per packet with p_get_proto_depth(), stop at
    8 levels with a new expert info (nfapi.too_deep), and use
    increment/decrement_dissection_depth() around the nested decode.

    nfapi_tag_vals_fn() passed a NULL tlv->name to "%s" for the gaps in
    the configuration tag table; print "Unknown" instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit cfbb5275787472b22b012be007715a7e603f32c3)

    Assisted-by: Codex

    (cherry picked from commit a5dac3499ddd27f6718991c1be2b305b91d20a6e)

    Backport: Apply the fix to the older branch context.

commit b53c000c88
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 12:57:20 2026 +1000

    NGAP: Fix NULL deref, nested message recursion and ranmode state

    set_stats_message_type() dereferenced stats_tap, which is only set in
    dissect_ngap(). A message PDU reached through the ngap.proc.* integer
    tables (for example from an exported PDU) crashed. Check for NULL.

    The NGAP-Message octet string is dissected by calling the NGAP
    dissector again, so RerouteNASRequest messages nested inside each other
    recursed until the layer assertion fired. Count the nesting per packet
    with p_get_proto_depth(), stop at 4 levels with a new expert info
    (ngap.too_deep), and use increment/decrement_dissection_depth().

    GlobalRANNodeID updated the conversation's RAN node type on every pass,
    so on a second pass an RRCContainer saw the node type of a later frame.
    Only update it on the first pass, and remember the value used for each
    frame with p_add_proto_data().

    The hidden 3gpp.tmsi field of FiveG-TMSI read the outer tvb instead of
    the octet string; use parameter_tvb. The NB-IoT TA key shifted a
    32-bit PLMN by 24 before widening it; cast to uint64_t first.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c55a99edc2a1bfadd9081f11710fdf48efc3c02c)

    (cherry picked from commit 730ee9c4c7042733ecaf2ee84c7170c2c727c6d8)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

    Assisted-by: Codex

commit 8e5165972b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MQTT-SN: Bound encapsulated message nesting, fix the node ID type

    An encapsulated message can contain another encapsulated message, and
    each level adds several tree levels, so deep nesting ended in the tree
    depth "Dissector bug" before the dissection depth limit. Allow at most
    16 nested encapsulated messages per packet and report deeper nesting
    with an expert info (mqttsn.nesting_too_deep).

    The Wireless Node ID in an encapsulated message has variable length,
    but the field was FT_UINT16, so an 8 byte ID raised "Trying to fetch an
    unsigned integer with length 8". Make it FT_BYTES.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 77ad6400a9702379a3d06acfbf84563a2fdb8807)

    (cherry picked from commit db4b586e43d700eba0277e7d00be6c6f3c05ceb4)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit e25e5bee62
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:01 2026 +1000

    MIME multipart: Bound nesting and fix gzip length after base64

    A multipart body part with a multipart content type re-entered
    dissect_multipart() with no limit, so a deeply nested body ended in the
    layer count assertion. Keep a per-packet nesting counter and stop with
    an expert info at 16 levels.

    When a part was base64 decoded and then gunzipped, the encoded length
    was used for the decoded data. Use the reported length of the data
    actually decompressed. Also drop an unreachable DISSECTOR_ASSERT in
    dissect_kerberos_encrypted_message().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 738fd25097651294f9800ca3d2838683c6326fc1)

    (cherry picked from commit ee62de03087c6d9d79ec0eb00093ed66b012102e)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit f77606b41e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 19:03:58 2026 +1000

    Logcat: Fix quadratic newline replacement and length handling

    With the one-line info column preference set, the message was scanned
    from its start for every newline and carriage return that was replaced,
    so a message made of many such characters took quadratic time. Replace
    the two g_utf8_strchr() loops with a single pass over the message.

    The message length was computed as length - tag_length - 1, which wraps
    when the tag is longer than the payload length field says, and the text
    subset used string_length - 1, which wraps for an empty message. Report
    an invalid payload length with an expert info and stop when the tag does
    not fit, compute the message length from the bytes already consumed, and
    only dissect the text lines when the message is not empty.

    The source address length was one byte more than the tag string, which
    includes its terminating NUL; use the string length.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e2259c23b36c7809ad01ee48a0cea701b09b881c)

    (cherry picked from commit 81b3b4d2590abd2b43b665182d941aee17d30742)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit da49a80ac2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 19:55:21 2026 +1000

    ISUP: Fix range and status assert and several field decoding errors

    The Range and status parameter computed the range as a uint8_t octet
    plus one, which wraps to 0 for an octet of 0xFF. The following
    proto_tree_add_uint_bits_format_value() call with 0 bits then reached
    REPORT_DISSECTOR_BUG. Compute the range as an unsigned value so the
    1..8 bit branch is never entered with a zero width.

    Reset the file-scope called number and cause value together with the
    calling number at the start of each message, so stale values from a
    previous packet are not reported to the tap.

    Fix the 2-octet CODEC list length in the BAT ASE dissection to combine
    the octets the same way the outer loop does, the 2-octet application
    context identifier, which read the same octet twice, and the calling
    geodetic location LPRI, which used the wrong mask and so did not match
    the field. Report an APM originating or destination address length of 1
    as malformed instead of computing a length of -1.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit c53d8de736192919d757d508d0696a475897f25f)

    Coverage captures and tests remain on master.

    Assisted-by: Codex

    (cherry picked from commit 842652587f709df0aad0cb5fd6103b5874db44c1)

    Backport: Apply the fix to the older branch context.

commit 64f7d1a6b0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:20 2026 +1000

    IEEE1905: Fix MIC check buffer overflow and small length bugs

    check_mic_tlv() allocated DECRYPT_MIC_CMDU_OCTETS + tlvs_len bytes but
    copied the 6 CMDU bytes, the 13 MIC TLV header bytes and all TLVs other
    than the MIC TLV into it, so a short MIC TLV overflowed the heap buffer.
    Size the buffer for everything written, check every copy against the
    remaining TLV data and the allocation, and use unsigned lengths. Reject
    a MIC TLV that is shorter than its fixed part plus the MIC with an
    expert info before the check is attempted.

    dissect_qos_management_descriptor() subtracted 14 from an unsigned
    16-bit length without checking it, and could loop forever on a
    zero-sized or oversized tagged field. Check the length and stop.

    Also fix the version extraction in check_mic_tlv(), where & bound
    tighter than the shift, and ieee1905_fragment_equal(), which compared the
    source address twice instead of source and destination, and the push
    button join notification TLV, which added its two MAC addresses with a
    length of 2 so that the TLV was always malformed.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit caa1144ca4c9a2d3d59e68c36a0d550715090eed)

    Backport: Backport the applicable length, loop and fragment-key fixes. The
    MIC checking and decryption feature is absent from release-5.0, so its
    changes are omitted.

    Assisted-by: Codex

    (cherry picked from commit 525183d21685a2d984cf4f48bc8365ea915b8564)

    Backport: Apply the fix to the older branch context.

commit 9a3c71f8bf
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:19 2026 +1000

    ID3v2: Do not assert on a short UFID frame

    A UFID frame whose size is smaller than its owner identifier hit
    DISSECTOR_ASSERT(size >= text_length). Report it with a malformed expert
    info and skip the frame instead.

    Also make the frame size field FT_UINT32; it is added with 4 bytes and
    was registered as FT_UINT8.

    The check of the high bits of the synchsafe tag size tested octets 7-10
    instead of 6-9, so a 10 octet tag threw an exception.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 52488eb99f6d3eec78874078a69d0491318cc02c)

    (cherry picked from commit 3c42d903baab8ad470bccaf308f823fa32affb15)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit ff430c0dec
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:19 2026 +1000

    ICMP: Bound error nesting and fix extension lengths

    An ICMP error message contains an IP datagram which can itself contain
    an ICMP error message, and so on, without any bound. Do not dissect the
    embedded datagram of an error message that is already inside an error
    message; show it as data with an expert info instead.

    Fix the unknown Mobile IP extension, which added length - 4 bytes
    although length is already the data length; the unknown AFI address
    sub-object of the interface information object, which used a negative
    length (offset - obj_end_offset); and the interface name sub-object,
    which accepted a length of 0 and then used length - 1. Report a length
    of 0 with an expert info.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit db32d5ef59cafde2ce2f73120354b1d2f220bd1c)

    (cherry picked from commit 5586acca441112d09b7bbfb1402b53b2740bed23)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit 8a2076d8e0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 17:39:52 2026 +1000

    H.223: Fix mux sublist size and avoid an assertion and endless loop

    mux_element_sublist_size() started at the second element of the list
    it was given, so a sublist consisting of a single element (for example
    one with untilClosingFlag, stored as a repeat count of 0) got size 0.
    That tripped DISSECTOR_ASSERT(length != 0), reachable with a crafted
    H.245 MultiplexEntrySend, and would have made the caller's loop run
    forever. The sum was also a uint32_t that could wrap with nested
    repeat counts.

    Start at the first element, sum in 64 bits and clamp, and return 0
    instead of asserting. The caller reports an expert info and stops when
    the sublist size is 0, and compares against the remaining length
    without overflow.

    Also make h223.mux.mpl an FT_UINT8 without the stale 0xf00f mask, hold
    the payload length in a uint32_t, allocate the mux table item only
    after the early returns in add_h223_mux_element() so that it is not
    leaked, and drop a duplicated include.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 87dabea9c3552ec4bdbec18c93f05cf234b39b47)

    (cherry picked from commit d3380b7ac59304a1265ed8b751a6beb5210ad463)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit e0a4b964f5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 16:52:54 2026 +1000

    EHDLC: Stop on a DataLen shorter than the header

    A TFP or P-GSL frame (SAPI 10, 11 or 12) with a DataLen of 0 never
    advanced the offset, because the "len == 0" guard at the end of the
    loop was skipped by the "continue". The same header was dissected
    again and again, each time calling the subdissector, until the layer
    depth DISSECTOR_ASSERT in epan/packet.c fired. With a DataLen of 0 or
    1, "len - 2" also passed a wrapped length to the subset tvb, and the
    same underflow could happen with "len - header_length" for the other
    frames.

    Add an expert info and stop dissecting when DataLen is shorter than
    the address and length fields, and again when it is shorter than the
    header including the control field. This replaces the old "len == 0"
    guard.

    dissect_ehdlc_xid() parsed IEs up to the end of the tvb, not the end
    of the frame, so it took any following frames as XID IEs. Bound the
    loop with the frame length.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 968ee1580b68d43bcb1d7ddf17732e462f764022)

    (cherry picked from commit c0c2c7d4b4306bc919c3db37a7dad3e79d577c15)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit 22c04aaa93
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 02:52:56 2026 +1000

    CMS: Stop nested content before the layer limit, keep digests apart

    The content of a ContentInfo, and the eContent of an
    EncapsulatedContentInfo, are dissected through the BER OID table, which
    can lead back to ContentInfo. Each level adds a layer, so a message
    nested about 600 times deep hit the DISSECTOR_ASSERT on the number of
    layers in call_dissector_work(). Stop with an expert info before the
    limit, as ASAM CMP does.

    The content that a messageDigest attribute is checked against was kept
    for the whole packet. A detached SignedData after another ContentInfo
    in the same packet had its digest checked against the content of the
    earlier message, and could show it as correct. Give each ContentInfo
    its own content.

    id-aes256-CCM was registered under 2.16.840.1.101.3.4.1.44, the OID of
    id-aes256-CFB. It is 2.16.840.1.101.3.4.1.47.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 43bdf2898eda033cc35a925c9e413a3baf40c8cc)

    (cherry picked from commit 0716b283bd5bb1a364a9afb549785fbd1a0c55d9)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

    Assisted-by: Codex

commit afacb4c25e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 10:21:05 2026 +1000

    BT LE: Do not use an uninitialized advertiser address, fix fields

    An extended advertising PDU without AdvA and without a known connection
    never set the source address buffer, but it was still passed to the
    AD dissector. A BIGInfo AD copied it into the file-scope broadcast ISO
    state, from where it was shown as the Central Address of the BIS PDUs
    and used as their source address. Zero the buffers.

    The extended advertising Host Advertising Data fragments were shown,
    and the reassembled data dissected, only when a packet was visited
    again, so a single pass decoded the fragments as complete Advertising
    Data. Do it on every pass; the reassembly state is already kept per
    frame.

    Connected ISO PDUs looked up the ACL connection tree and read the
    connection info as a connected ISO one; look up the tree LL_CIS_IND
    fills instead.

    Also count the CRC octets of PDUs longer than 253 octets in more than
    eight bits, keep the ACAD length from wrapping when the extended header
    is too short for its fields, read the LL_PERIODIC_SYNC_IND Sleep Clock
    Accuracy from the PDU instead of from the channel map offset, give the
    "Anonymous" address its full length and flag an overlapping control
    procedure on the first pass as well.

    LL_PHY_UPDATE_IND without a PHY change has no Instant, but the
    uninitialized instant variable was still stored with the control
    procedure and compared on later packets, so whether the procedure
    completed depended on stack contents. Start it at zero.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 0179d29d1b34b5738784441e01981873fe7f25ef)

    (cherry picked from commit c4b0af23ede57dea31e885f8a40b1139812d0955)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit 5ba0a65f80
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 10:17:19 2026 +1000

    BT HSP: Do not copy the rest of the packet for each command

    dissect_at_command() copied and upper-cased everything from the start
    of the command to the end of the (reassembled) data, for every command,
    so a 64 KB message of ';' commands allocated about 2 GB and took seconds
    per pass. Make the upper-cased copy once per packet and point into it.

    Also find the end of a line with a non-ASCII octet in it, and leave
    the delimiter out of an unknown command name.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 11c9f5e638d6abfb830919d46d92186719a73b26)

    (cherry picked from commit 3bc98e9c3b1b6078421f86c5f56d8880fd924db2)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit d56cf1d8ec
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 07:00:16 2026 +1000

    AYIYA: Stop nested tunnels before the layer limit

    AYIYA forwards its payload to the ip.proto table, which can carry
    AYIYA again over UDP. A packet nesting a few hundred levels reached
    the layer limit in call_dissector_work() and failed its
    DISSECTOR_ASSERT(). Stop with an expert info when the packet is close
    to the limit.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit b4f37794cd4bd77157f512ed7927c67b81dfc3a6)

    (cherry picked from commit 28d82c7f708782cff152b4d89bace292153edee5)

    Backport: Apply the fix to the older branch context.

    Assisted-by: Codex

commit ac6e2b2244
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Oct 27 19:30:50 2026 +1000

    TCAP: Stop handing components over before the layer limit

    Each component of an ITU TCAP message is handed to the subdissector of
    its application context or SSN, and each hand-off adds a protocol layer
    to the frame. A message with about 500 components reached the layer
    limit in call_dissector_work() and failed its DISSECTOR_ASSERT(),
    losing the rest of the frame. Stop handing components over when the
    frame is close to the limit, and report it with an expert info instead.

    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    Assisted-by: Claude
    (cherry picked from commit 371529607702311c9d1d96beddb708894294f921)

    (cherry picked from commit 8212158a770fb58da98e7d6001f0f7f20664600c)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

    Assisted-by: Codex

commit 56b4fb2d2f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Oct 27 19:30:49 2026 +1000

    ANSI TCAP: Stop handing components over before the layer limit

    Each component of a package was handed to ANSI MAP, AIN or the national
    opcode dissector, and each hand-off added a protocol layer to the
    frame. A package with about 500 components reached the layer limit in
    call_dissector_work() and failed its DISSECTOR_ASSERT(), losing the
    rest of the frame. Stop handing components over when the frame is
    close to the limit, and report it with an expert info instead.

    A result was matched with the last invoke saved for its transaction up
    to and including its own frame. In the first pass only the invokes
    before it in the frame were saved, in later passes all of them, so a
    package with an invoke after the result decoded differently with
    tshark -2. Match only invokes of earlier frames.

    Also keep the whole length of a transaction ID longer than 255 octets;
    it was truncated to 8 bits.

    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    Assisted-by: Claude
    (cherry picked from commit 275b50dfe8dbae2d369fa6974c7059326d1e6f46)

    (cherry picked from commit 1f9572d06ee6106eb1ff0b8065f5e6da8b330edf)

    Backport: Apply the fix to the older branch context, regenerating ASN.1 output.

    Assisted-by: Codex

commit 77526a1a07
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 17:33:58 2026 +1000

    SCCP: Check that the destination is a point code before using it

    dissect_sccp() checked only that the source address was an SS7 point
    code and then read the destination address data as an mtp3_addr_pc_t
    too. A carrier that sets only the source, such as PPCAP with a Source
    Address record and no Destination Address record, left the destination
    empty, and SCCP dereferenced its NULL data. A crafted Exported PDU file
    naming PPCAP crashed tshark with default settings.

    Check the destination type before using it, and do the same where the
    SCCP users table takes the point codes from both addresses.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 2b5b8f861ccf57e0444f489a27f7fbee1453e40f)

    Assisted-by: Codex

    (cherry picked from commit 596f3ea30766217f0461ddb6b13ccc8c7e9e7799)

commit b97b388d61
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 17:11:09 2026 +1000

    P7: Register the extended-registrations item as an OID

    The extended-registrations item of RegistrationTypes is decoded with
    dissect_ber_object_identifier_str(), but its field was generated as
    FT_NONE. A Register-MS argument or error that carries
    extended-registrations therefore hit DISSECTOR_ASSERT_NOT_REACHED in
    the BER OID helper with default settings.

    Give the field type FT_OID in p7.cnf so the decoder and the field
    agree, and the OID is shown.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 926bdc2126f8d2d4bfca7c90d00ee0b376167d28)

    Assisted-by: Codex

    (cherry picked from commit 3ffaea2cd3ff114ebbdcb8be5e33f730aeb8386f)

commit effd8e2cd2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    Portmap: Bound the nesting of CALLIT and RMTCALL calls

    A CALLIT/RMTCALL call or reply can carry another portmap CALLIT, to any
    depth, and only the generic layer limit assertion stopped it. Keep a
    per-packet nesting counter (limit 4), stop with an expert info beyond
    it, and bracket the indirect call with increment_dissection_depth() and
    decrement_dissection_depth().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 0da03349a7e9be6cd99c47602e5c9c4a62a76013)

    Assisted-by: Codex

    (cherry picked from commit b6cf2e479dabe831e501df8d5051ac5824fa4cc5)

commit d300abcf5b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    PPCAP: Use NUL terminated copies for string addresses

    The Node Id source and destination addresses were created with
    set_address_tvb(AT_STRINGZ) directly on packet bytes. A zero length
    gave NULL address data, which g_strlcpy() rejects, and a string without
    a NUL made the address code read past the field. Use a NUL terminated
    copy in the packet pool and skip the address when the length is 0.

    Create AT_IPv6 addresses with their real length of 16 and only when the
    field has at least 16 bytes, instead of 6.

    Reset the static SSN at the start of each packet so the TCAP SSN lookup
    no longer depends on earlier packets.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 757add1fb0a5062402b71f7105beb478678f03c1)

    Assisted-by: Codex

    (cherry picked from commit 2e2b4563f9d1809014d86e168234b1f5770a1d9a)

commit 3222818243
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    P_MUL: Do not assert on a relative message ID seen on a later pass

    message_id_offset used 0 both for "not set yet" and for "set to 0".
    When the first PDU had message ID 0 and a later one a different ID, the
    offset changed between the first and the second pass, the first PDU then
    got a bogus relative ID, register_p_mul_id() found no entry for it and
    DISSECTOR_ASSERT(pkg_data) fired.

    Track whether the offset is set with a separate flag, reset by the init
    routine and set only once (both for the message ID and the Ack loop),
    and return NULL instead of asserting when no saved data exists; the
    callers already handle NULL.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit a5d5f1de66487f12e649d09edfcd412b72f8cff2)

    Assisted-by: Codex

    (cherry picked from commit 460db738dc57a660af5893ed0d57b61c9931170a)

commit d6272d7b19
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    PKTC: Initialize the time of the revocation field

    dissect_pktc_mtafqdn_krbsafeuserdata() set only the seconds of an
    nstime_t before showing it, so the nanoseconds were stack garbage.
    Initialize the structure with NSTIME_INIT_ZERO.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit df1583cdb0e898b2d866c8de8b896080e7e16b50)

    Assisted-by: Codex

    (cherry picked from commit 0f85f6087272c1dc814522b4fc99b72c0342eac1)

commit 6ccc860b37
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:33 2026 +1000

    OSC: Fix dissection depth on error returns, initialise timetag

    dissect_osc_bundle() incremented the dissection depth for each bundle
    element but returned without decrementing it on the error paths, which
    leaves the counter too high for the rest of the packet. Decrement on all
    return paths.

    The nstime_t passed to proto_tree_add_time_format_value() for the
    Immediate timetag was uninitialised, in both the bundle and the message
    code. Initialise it to zero.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit db8db3aecf6037b674d7b907dcee7f26016904c3)

    Assisted-by: Codex

    (cherry picked from commit 230ec650ef021565ec1f2f3a3166589673bacc97)

commit f3577c4189
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:32 2026 +1000

    PER: Use the 64-bit API for wide integer fields in dissect_per_integer

    dissect_per_integer() accepted any FT_INT* or FT_UINT* field but always
    added it with proto_tree_add_int() or proto_tree_add_uint(), which only
    take fields of up to 32 bits. An unconstrained INTEGER mapped to a 40 to
    64 bit field (TETRA network-time is one) made the tree code report a
    dissector bug, and abort with WIRESHARK_ABORT_ON_DISSECTOR_BUG set.
    Use proto_tree_add_int64() and proto_tree_add_uint64() for those field
    types.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 4397b4a76f25403774c793bbcd92b53d74aa048f)

    Assisted-by: Codex

    (cherry picked from commit 0bcdaa592dedfebd79eddc8f3bc30b363c38bf62)

commit 34499e9f02
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 10 15:16:32 2026 +1000

    PPP: Bound nesting of PPP payloads and fix small decoding errors

    LCP Protocol-Reject, PPPMux sub-frames and Multilink single-fragment
    reassembly each hand packet-controlled data back to the PPP protocol
    table or to dissect_ppp() with no limit, so a crafted packet nests until
    the generic layer-depth assertion fires. Add a per-packet nesting
    counter (limit 8) kept with p_get_proto_depth(), with an expert info and
    a plain data display beyond it, and bracket the nested calls with
    increment/decrement_dissection_depth().

    PPPMux kept the last sub-frame PID in a static variable, so decoding a
    frame depended on the packets dissected before it. Make it a per-packet
    local.

    Small fixes: the LCP Quality-Protocol data length was length + 4 instead
    of length - 4; the Prefix-Elision option never advanced its offset and
    read the prefix one byte too far; the BCP MAC L and M bit masks pointed
    at the wrong byte; capture_ppp_hdlc tested pd[0] instead of pd[offset];
    the PPPMux flags were added to the top-level tree instead of the header
    subtree; and the LCP variable-length option error always named the
    Vendor option and "6" bytes instead of the actual option and length.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 759c5d02890372ddf463a8f7e2bb1486524deee0)

    Assisted-by: Codex

    (cherry picked from commit 682bf87d878ae9dd4e9e1592436cdefca1f025f1)

commit 5b171edba1
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 12:57:20 2026 +1000

    NMF: Bound the record size and fix the record offsets

    dissect_nmf_record_size() returned the offset of the varint instead of
    the offset after it and did not check the record size against the data.
    A Sized Envelope with a huge size made the offset wrap and
    proto_item_set_end() hit an assertion. Return the offset after the
    varint, which also fixes the Via and Upgrade strings that started at
    the length bytes, and add an expert info (the record size must not
    exceed the remaining reported length).

    dissect_nmf_record() returns an absolute end offset, so the payload
    loop must assign it rather than add it. Do not let 4 + length wrap in
    nmf_get_pdu_len(). Only update the conversation state on the first
    pass.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit b6c176ef5de949293aa2548bac5b20e75bfd546c)

    Assisted-by: Codex

    (cherry picked from commit 607d2fdd245d1caf0f102eb95b01306ba15a6c1f)

commit daac3c19af
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 16:52:52 2026 +1000

    DVB-S2-TABLE: Stop on unknown TBTP2 format, fix SNMP proto id

    A TBTP2 with an assignment_format other than 0-3 and 10-12 walked up to
    256 frames of up to 65536 assignments each, without consuming any
    bytes for an assignment. A 1.3 KB packet ran about 16 million loop
    iterations, 0.35 s per packet without a tree, and with a tree it hit
    the "possible infinite loop" check. Add an expert info for an unknown
    assignment format and stop dissecting the table there, because the
    size of an assignment entry is not known.

    The Network Layer Info descriptor passed the literal 1 to
    dissect_snmp_pdu() as the protocol id, which is the "Wireshark
    Columns" protocol. The SNMP PDU was shown under a "Wireshark Columns"
    item and the Protocol column was set to "Columns". Look up the SNMP
    protocol id in the handoff and pass that.

    The Route id loop count field had the mask 0x01 of the Route id flag
    field before it, so at most 2 route ids were read. Remove the mask, as
    for the other loop count fields of the descriptor.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit e3363ed1640ae429da40dcd048803cad56d438b1)

    Backport: Retain the separate assignment-format read used on release-5.0.

    Assisted-by: Codex

    (cherry picked from commit fcdde2569ebd3bcaf77ac4ab134296530d9071ff)

commit 1ddeb8274d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 18:54:24 2026 +1000

    nstrace: Fix an infinite loop on short NetScaler 3.5 records

    add35records() advanced by the record length taken from the packet.
    A record length smaller than the 3 byte record header (for example
    0) with a next-record type that is again a known type never advanced
    the offset, so the dissector looped forever. Stop with a new expert
    info (nstrace.bad_reclen) when the record length is below 3.

    dissect_nstrace() read the NetScaler pseudo-header without checking
    that the record has a NetScaler encapsulation, but the "ns" handle can
    also be reached by name, for example from Exported PDU. Return 0
    unless the encapsulation is one of WTAP_ENCAP_NSTRACE_*.

    Do not modify the pseudo-header in place when adding the VM name
    lengths to the Ethernet offset; use a local variable, so that a
    second dissection of the frame gives the same result.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit a9dbcc99a242014929c93805176efd4d2975e486)
    [Backport: only the added includes (epan/expert.h, epan/tfs.h, wsutil/array.h) conflicted (context); the infinite-loop fix (stop when the record length is below the 3-byte header), the encapsulation check and the in-place-pseudo-header fix applied cleanly.]

commit ea42ce1b60
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 12:30:40 2026 +1000

    LLDP: Fix TLV length handling in several TLVs

    The Time To Live TLV returned a fixed length and ignored the TLV
    length, so a TLV with a length other than 2 desynchronised the
    dissection of the following TLVs. Return the TLV length instead.

    A management address with an address string length of 0 made the
    address length -1 and moved the offset backwards; report it with an
    expert info and skip the TLV.

    The ONOS TTL value (32 bits) was added with the length of the rest of
    the TLV; check that the length is 4. The fixed-size Cisco ACI fields
    were added with the TLV length and subtracted unsigned sizes from it, so
    short TLVs wrapped the remaining length; check that the TLV holds the
    fixed size and use that size, and only walk complete APIC list
    entries. The Port Mode field is 2 bytes and now uses FT_UINT16.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 583a8da35512035bd5765ada5447904c105be510)
    [Backport: applied as-is; the only conflicts were context (Cisco ACI nodeip fixed 4-byte read, ONOS TTL length==4 check). The mgmt-address length-0 offset-back fix and the other TLV length checks applied cleanly.]

commit 10d6864fef
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:16 2026 +1000

    DLT: Fix the payload length and parameter error returns

    The message length counts from the start of the DLT message, but the
    offset includes the 16-byte storage header of .dlt files, so the
    payload item of every message read from a file was 16 bytes too
    short, or got a negative length. Subtract the start of the message.

    dissect_dlt_verbose_parameter() returns an unsigned length but
    returned -1 on errors, which moved the offset one byte back and kept
    the argument loop going on overlapping bytes. Return 0 so that the
    loop stops, and add a "buffer too short" expert info where there was
    none. A string with an unsupported coding is skipped by its length
    instead of returning -1.

    UTF-8 string arguments were added to the tree with ENC_ASCII, so
    non-ASCII characters showed as replacement characters in
    dlt.data.string while the Info column, built from the decoded buffer,
    was right. Add the item with the encoding of the argument.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 7bc40b3f73b58ae33c76d139e8fc008289e72b73)
    [Backport: the payload-length fix and the verbose-parameter -1->0 return (the overlapping-bytes loop fix) apply cleanly; dropped the cosmetic UTF-8 string-encoding refinement, which depends on master's encoding->str_encoding rename not present on 4.4 (4.4 keeps 'encoding' and ENC_ASCII).]

commit 5247d6d5e3
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 18:54:24 2026 +1000

    OpenFlow: Fix endless loops and unbounded recursion in v4, v5 and v6

    The v5 and v6 property, instruction and match helpers handled a too
    short element length with "offset = length", where length is the
    length of the message. For a message nested in another one (the data of
    OFPT_ERROR, OFPT_REQUESTFORWARD, OFPT_BUNDLE_ADD_MESSAGE) the nested
    message was dissected on the outer tvb at an absolute offset with the
    relative inner length, so this moved the offset backwards and the
    callers, which loop on the absolute element end, parsed the same
    elements forever while adding expert infos. Drop these assignments, so
    the offset stays after the element header and only moves forward, and
    dissect nested messages (v4 error too) in a subset tvb of their own
    length, so that offsets and lengths are relative to the inner message.
    The top level dissectors also use a subset tvb of the header length,
    as the Exported PDU tvb can be longer than the message.

    Nested messages (error, requestforward, bundle_add) had no recursion
    limit, so a single PDU with thousands of nested headers overflowed the
    stack in v4, v5 and v6. Count the nesting per packet with
    p_get_proto_depth()/p_set_proto_depth(), stop with an expert info
    after 4 levels, and wrap the nested call in
    increment/decrement_dissection_depth().

    Also fix small bugs: the v5 bundle property did not advance the offset
    over type and length, the v5 empty match (length 4) was reported as too
    short, the v6 OXS experimenter value is bytes and not a 32 bit
    integer (and is bounded), and the experimenter queue stats property
    expert items in v5 and v6 used the message length instead of the
    property length.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 2815b00f70d2fd836876b5347e319c6e59cb2dd0)
    [Backport: kept 4.x epan/ipproto.h and added epan/proto_data.h for the per-packet nesting counter; the master-only epan/iana-info.h include is not needed on 4.6/4.4. Loop/recursion bodies applied cleanly.]

commit 7dc58f9d16
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 12:57:20 2026 +1000

    NNTP: Do not read past the line when matching STARTTLS

    The line pointer from tvb_get_ptr() is not NUL terminated, so
    g_ascii_strncasecmp() against "STARTTLS" or "382" read beyond the
    packet data for a line that is a shorter prefix of the keyword. Check
    the line length and use tvb_strncaseeql(). Update the STARTTLS
    conversation state only on the first pass.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 04b11565823678fe357dbe2eb79caf1d7b3b9ba8)
    [Backport: kept 4.x int offsets and tvb_find_line_end() (tvb_find_line_end_remaining() is absent on 4.6/4.4); applied the security core - dropped the tvb_get_ptr()/g_ascii_strncasecmp() over-read and used the bounds-checked linelen>=N + tvb_strncaseeql() as on master.]

commit b78a5c2425
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 18:54:24 2026 +1000

    OCP.1: Bound the PDU size to avoid an endless loop

    The 32-bit PDU size was used unchecked. A size such as 0xFFFFFFF5 made
    get_ocp1_message_len() (TCP) and dissect_ocp1_pdu()/dissect_ocp1() (UDP)
    advance the offset by a value that wraps back to 0, so the same PDUs
    were dissected forever. The size must now be at least the rest of the
    header, the sum of the sizes must not wrap, and in the UDP path the PDU
    must lie inside the remaining data. Otherwise the loop stops, and the
    UDP path adds an invalid length expert info. Every step now advances by
    at least the header size and stays in the tvb.

    Also use two case labels for Command and CommandRrq instead of
    (A || B), which evaluated to 1, and use the message offset for the
    OcaLibVolType id and the OcaTaskStatus error code, which were shown at
    the start of the parent structure.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 69466cec10292fe08dd692ce7bb2bc8dd4d1faf3)

commit d4a9aa5444
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 18:54:24 2026 +1000

    NR RRC: Bound nested RRCReconfiguration and fix a PDU field type

    RRCReconfiguration and RRCReconfigurationComplete can contain
    themselves through the nr-SCG, nr-SCG-r16, condRRCReconfig-r16,
    ltm-CandidateConfig-r18, ReferenceConfiguration-r18 and nr-SCG-Response
    containers (and through the LTE RRC cycle). The recursion was not
    bounded and a crafted packet overflowed the stack. Use the existing
    nr_rrc_enter_nesting() limit (8 levels per packet) in both messages and
    stop with an expert info beyond it.

    dissect_nr_rrc_rach_ConfigCommonIAB_r16_PDU() passed an FT_NONE field to
    a CHOICE, which fails the field type assertion when a tree is built
    (reachable through XnAP). Use the generated FT_UINT32 field
    hf_nr_rrc_rach_ConfigCommonIAB_r16 instead.

    AbsoluteTimeInfo-r16 read twelve characters of the BCD string, which is
    shorter when a nibble is 0xF. Format it only when it has twelve digits
    and show the string as is otherwise.

    Only update the PDCP and MAC bearer, SRB and security state while the
    frame is first dissected, so that redissection does not rotate the
    previous algorithm state or apply the settings again.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 01c44bdfdc5773796e59f8a9c75cfb77f353f595)

commit fea8da6945
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 12:57:19 2026 +1000

    NDMP: Fix a NULL tvb after incomplete reassembly

    When the last fragment of an NDMP record arrived but reassembly was
    not complete, process_reassembled_data() returned NULL and the
    dissector went on to use new_tvb, which crashed. This could be
    triggered by fragments from different connections that share a
    sequence number, because the reassembly table was keyed on addresses
    only. Return when the reassembly is incomplete, and register the table
    with addresses_ports_reassembly_table_functions so that fragments from
    different ports are kept apart.

    Also label the command subtree with val_to_str_const() on the message
    number instead of indexing msg_vals[] with an index into the
    differently ordered ndmp_commands[] table.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit f2aec6551ee9f32e191beda0a2c7c1e073c42ba6)

commit e46ae62003
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 12:57:19 2026 +1000

    Netlink: Bound message length and attribute nesting

    A netlink message length close to 2^32 made pkt_end_offset wrap, which
    was also a signed int, so the loop in dissect_netlink() moved backwards
    and dissected the same messages forever, allocating memory until it ran
    out. A message longer than the remaining data now gets an expert info
    (netlink.bad_length) and is clamped to the remaining length, so the loop
    always moves forward; the offsets are unsigned.

    Attribute callbacks could recurse without limit, each level costing only
    four bytes of packet, which overflowed the stack. Add a nesting counter
    and the packet_info to struct packet_netlink_data and stop in
    dissect_netlink_attributes_common() at 32 levels with an expert info
    (netlink.nesting_too_deep); each level also counts as a dissection
    level. This bounds net_dm (HW_ENTRIES/HW_ENTRY), ovs_flow (CLONE,
    SAMPLE, ENCAP, CHECK_PKT_LEN), ovs_packet and nl80211 (IFTYPE_EXT_CAPA)
    in one place. The structure is created only in dissect_netlink() and
    lives for one message, so an exception leaves no stale state.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from 942a6029807d18ab2df8d5fd8f1c631c3c78398d)

commit 256fac1f6f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MMS: Fix NULL string, overflow and short bit strings

    GetLogicalDeviceDirectory, GetDataSetDirectory and GetServerDirectory
    passed a NULL item id to a %s format when no matching Identifier had
    been seen. Use an empty string.

    Compute the number of seconds of a binary time in time_t to avoid
    a signed integer overflow for large day counts. Print the unsigned
    pre-Cinfo value with %02u. Check that the quality, check and reported
    optional fields bit strings are long enough before reading them.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 0b621d6f9576a2e44499a25f19655699907832b6)

commit 2932c700a4
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 12:37:09 2026 +1000

    IPDR: Check lengths before skipping over them

    The SAMIS type 1 dissector added a length read from the packet to the
    offset without checking it. A length such as 0xFFFFFFFF wrapped the
    offset backwards (and len + 4 wrapped for the SF channel set). Stop
    with an expert info if a length runs past the end of the data.

    Add the CMTS IPv6 address only when its length is 16.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6b4050d572ac571a63efada6f8af6f42d7a044ab)

commit 101570b3f3
Author: Gerald Combs <gerald@wireshark.org>
Date:   Tue Nov 10 10:39:58 2026 -0800

    tools: Update krb5 to 1.22.3

    (cherry picked from commit 55b5def4b5f19e91f56abfd9621f4ffbca938dce)

    Conflicts:
            cmake/modules/FetchArtifacts.cmake

commit dd20fdecb3
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 16:50:29 2026 -0500

    IMF: Bound the sio-label parsing to each item

    dissect_imf_siolabel() searched for '=' from every ';' separated item to
    the end of the whole header. A long header with many items took time
    quadratic in its length (30 seconds for 2 MB). Compute the item first
    and bound the '=' search and the value whitespace trimming to it, so the
    cost is linear. An item without '=' is reported as an unknown parameter,
    and quoted type/label values shorter than two bytes are no longer indexed
    with an underflowing length.

    Assisted-by: Claude

    (backported from commit 919d4f9bf7002868ab2557ae968ecafda1da1924)
    [Backport: adapted to 4.4's int offsets and the -1-returning tvb_find_uint8()/tvb_find_line_end() (4.4 lacks the bool/out-param tvb_find_uint8_length()/tvb_find_line_end_length()); tvb_skip_wsp()/tvb_skip_wsp_return() are used for the whitespace skipping as on master. The upstream commit's three unrelated small fixes are dropped; 4.4's mailbox '>' test is already correct.]

commit 862cc3ce44
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 16:41:43 2026 -0500

    SVCCTL: Keep EnumServicesStatusW names in their buffer

    The service and display names of an EnumServicesStatusW response are at
    offsets from the start of the services buffer that were not checked, and
    many entries could point at the same long string, which was scanned and
    added again for each: a 1 MB response took 30 seconds, and 8 GB with a
    tree. Dissect the entries and their names in a subset tvb of the buffer,
    so a bad offset throws, and stop when the names add up to more than the
    buffer. The buffer size was also added to the offset as an int unchecked.

    Assisted-by: Claude

    (backported from commit 8c2f1f9c034b156b83cd2d5d1e145dcf33e38351)
    [Backport: kept 4.4's older-pidl int offsets; nothing dropped.]

commit d752506abb
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 12:30:39 2026 +1000

    LBMC: Fix memory blow-ups, NULL address and fragment handling

    dissect_lbmc_message_data() allocated a 64 KiB
    lbmc_extopt_reassembled_data_t and cleared it for every LBMC message in
    a packet, so a frame of 4-byte messages took over 1 GB. The structure
    is now allocated once per call and only its small header fields are
    reset per message.

    The reassembly, stream and substream key builders copied four bytes
    from address.data, which is NULL for AT_NONE addresses (Exported PDU
    without IP tags). One helper now returns 0 unless the address has data
    of at least four bytes.

    dissect_msg_properties() trusted the key and value offsets, so all
    fields could point at one long string and multiply the work. The
    offsets must now lie after the field table and increase from field to
    field, otherwise an expert info is added and dissection stops. The
    string size lookups now include the base offset.

    lbmc_message_add_fragment() rejects overlapping fragments, which left
    uninitialized bytes in the reassembled message, and the reassembly
    buffer is now zeroed. The duplicate check comes before the length
    checks so that a second pass (tshark -2) accepts the fragments it has
    already stored. The reassembled tvb is now a child tvb of the LBMC tvb
    instead of a real-data tvb that was never freed.

    The UMQ index command item lengths used index_len before it was
    assigned, and the application header chain used a uint8_t length that
    could wrap; it is now an int and the loop stops when an element is
    longer than the remaining data.

    The UMQ index command release and reserve headers computed the index
    length without the base command header, unlike the responses, so every
    one of them was malformed. Subtract the base header too, and compute the
    length before the item that uses it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 5f6631e1cfcafaae29f5e6207b0f6cf8afe8c84e)
    [Backport: adapted to 4.4's unbased message-properties key/value offsets (used the fix's offset+key_offset / offset+value_offset form and dropped the then-unused actual_key_offset).]

commit 63218980aa
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MiWi: Fix buffer overflow, address map and payload offsets

    miwi_short_to_str() writes "Broadcast" (10 bytes) for 0xffff, but the
    destination and source PAN id and short address strings were kept in
    8-byte wmem_new(uint64_t) buffers. This overflows the heap buffer for
    any frame with a broadcast address. Use 16-byte local arrays at all four
    sites.

    The miwi_map hash tables were never created, so every lookup and
    address update ended in a GLib critical. Create them in an init routine
    and destroy them in a cleanup routine, like packet-ieee802154.c does for
    its map. With the tables in place, fix the code that was unreachable
    before: add the inferred extended source address with
    hf_miwi_ext_src_addr instead of the FT_UINT16 short address field, and
    update the address map only on the first pass.

    Also hand only the data after the header to the data dissector for
    unknown commands and for data frames with a bad FCS, instead of the
    whole frame from offset 0.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 6432789eceeae8c0b678fe0959a7fcb8a48373e3)
    [Backport: kept 4.4's if(tree) guard on the short source-address display, with the hf_miwi_ext_src_addr correction applied inside it; nothing dropped.]

commit b2f6abb95a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 12:37:10 2026 +1000

    IPSec: Fix several memory-safety problems with SA configuration

    After a libgcrypt error the cipher handle was closed but
    cipher_hd_created stayed true, so the next packet of the SA used a
    freed handle. Clear the flag after every gcry_cipher_close() in the
    error paths. TWOFISH-CBC also never set the block length, so the block
    size check was skipped and every packet with a bad length failed in
    gcry_cipher_decrypt(); set it to 16.

    The IPv6 SA address filter parser could write before the start of a
    stack buffer and past its end for groups longer than four hex digits.
    Bound every write and reject such groups. The IPv4 parser applied the
    netmask with sscanf() before the expanded address was NUL-terminated,
    which read uninitialized stack memory. Terminate it first, and in both
    parsers read only the one hex digit being masked.

    Also look up the algorithm names with val_to_str_const() since the
    value_string arrays are not indexed by value, update firstValidSN only
    on the first pass, and do the AH length shift on an unsigned value.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 4f1979d0194509ff3592b5be0301f5e61bb3b173)
    [Backport: 4.4 has no ESP Decryption Secrets Block import, so that
    part was dropped.]

commit 1a6bb1923c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 16:28:30 2026 -0500

    Diameter: Bound grouped AVP recursion

    Grouped AVPs were dissected recursively without any bound, so a message
    with deeply nested grouped AVPs over the default TCP/SCTP port 3868
    overflowed the stack, also without a tree. Track the nesting with
    increment_dissection_depth()/decrement_dissection_depth().

    grouped_avp() also cleared the grouped state (dis_gouped and the
    Vendor-Id) to zero when a nested grouped AVP ended, so the AVPs after it
    in the enclosing group were treated as top level. Save and restore the
    enclosing group's state instead, and undo the depth and the state when
    an exception leaves a grouped AVP, so that the next PDU of the same TCP
    segment does not start with a higher depth.

    Assisted-by: Claude

    (backported from commit 48a21dcb7220557599f63c4e91b0c171f7e870bf)
    [Backport: extracted only the grouped-AVP recursion bound and the
     enclosing-group state save/restore from the upstream commit.]

commit ceb5f1d27a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 8 16:21:57 2026 -0500

    SMB: Do not export uninitialized heap memory in Export Objects

    Export Objects could write uninitialized heap memory into an exported
    file: the per-file buffer was allocated with g_try_malloc() and the
    parts of a file between the chunks that were actually seen were never
    initialized, so they were written out as whatever the heap held.
    Likewise, once a chunk too far into the file had made the buffer be
    dropped (is_out_of_memory), a later chunk started a fresh buffer and
    reconstructed only part of the file.

    Allocate the buffer with g_try_malloc0(), zero the newly grown part
    when the buffer is enlarged, and do not start a new buffer once the
    content has been dropped.

    Assisted-by: Claude

    (backported from commit 610c8d81e3ecaf4e62501cc38f27b314c3ccd671)
    [Backport: extracted only the Export Objects memory-safety change from
     the upstream commit.]

commit 747420ba99
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 12:37:10 2026 +1000

    IPv6: Fix NULL dereference and bounds errors in extension headers

    The routing type sub-dissectors (type 0, 2, 3, 4, 5 and 6) cast their
    data argument to struct ws_rthdr and dereferenced it without a check.
    Exported PDU calls any dissector reached through a uint dissector table
    with NULL data, so a crafted Exported PDU packet crashed Wireshark.
    Return 0 when data is NULL.

    The CRH dissector read one SID past the end of the header after the last
    SID, so every CRH whose SIDs were all non-zero ended up malformed and
    the payload after it was never dissected. Fetch each SID at the top of
    the loop, bounded by the SID count and the remaining length.

    The Line-Identification option with an option length of 0 computed a
    length of 0xffffffff and stepped the offset back by one byte. Report it
    with an expert info and stop.

    opt_start was a uint8_t and truncated option offsets larger than 255,
    so the undecoded remainder of an option was dissected as new options.
    Make it an int.

    The hop-by-hop, destination options, routing and fragment header
    dissectors cast their data to ws_ip6 although they can be reached from
    IPv4. Use WS_IP6_PTR() so other data is treated as no IPv6 header.

    ipv6_opt_type_hdr() kept its position in the option table in a static
    pointer, so after the first option type not in the table every lookup
    returned "any header" and the header mismatch warnings appeared in the
    first pass only. Start each lookup at the beginning of the table.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit d65c2b7bd2f52f2d009c48ff16b35652d80e01cc)
    [Backport: only dissect_routing6_srh conflicted (NULL check added, offset
    made unsigned as upstream); the CRH SID-count check casts octets_per_sid
    to int to avoid a -Wsign-compare warning (it is 2 or 4).]

commit ee3664dcf8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 08:00:37 2026 +1000

    LZ77: Limit the decompressed size, do not read literals past the input

    Each match of Plain LZ77 can be up to 4 GB long, and only one match was
    limited to 16 MB, so a few octets of matches expanded to gigabytes: a
    746 octet MAPI RPC_HEADER_EXT took 2.1 GB and 24 seconds to decompress,
    and a full 64 KB field about 4 GB, for every such frame. SMB2, HTTP and
    NETLOGON use the same decompressor. Limit the whole output to 16 MB. A
    literal after the end of the compressed data was read from the octets
    after it; fail instead, as [MS-XCA] does not allow it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 9dae068965b3e92e84d7e63dc26023987a702cce)
    [Backport: 4.4 has the older byte-by-byte match-copy loop, so only the
    output-size limit was applied, not master's unrelated block-copy
    rewrite.]

commit 9b4aed4988
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:19 2026 +1000

    ICQ: Bound multi packet recursion and fix MSG_CONTACTS

    icqv5_srv_multi() recursed into dissect_icqv5Server() without any
    limit. Count the nesting with increment_dissection_depth() and stop
    with an expert info before the tree depth limit is reached.

    Restore the two "separator found" tests in MSG_CONTACTS that were
    inverted when tvb_find_uint8() was converted to the unsigned
    tvb_find_uint8_length(), and fix the swapped length and value of the
    meta user Result item.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit b71c89fda9bccd8997bab3c3140556e93cd99406)
    [Backport: 4.4 predates the tvb_find_uint8_length() conversion, so the
    MSG_CONTACTS separator fix does not apply and was dropped; only
    <epan/prefs.h> was added from the include hunk. The multi packet
    recursion bound and the meta user Result fix are kept; the depth check casts gui_max_tree_depth to int, as other 4.4 dissectors do, to avoid a sign-compare warning.]

commit 4f4f7863db
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 22:09:48 2026 +1000

    H.265: Fix palette initializer and bit width handling

    The SPS screen content coding extension used (1 << BitDepth) - 1 as the
    width of each palette predictor initializer instead of BitDepth. For
    8-bit video this asked for 255 bits and raised a dissector bug, and for
    a bit depth of 32 the width wrapped to 0 so the loop made no progress
    while running up to 3 * 2^32 iterations. Use BitDepthY/BitDepthC as the
    width, and stop with an expert info if bit_depth_*_minus8 is larger than
    8 or sps_num_palette_predictor_initializers_minus1 is 128 or more.

    The PPS screen content coding extension decoded
    pps_num_palette_predictor_initializers as a signed Exp-Golomb value into
    an unsigned field, which asserted, and mixed byte and bit offsets so the
    loop re-read the same bits. Decode it as ue(v), keep a single bit
    offset, and stop with an expert info if the count is above 128 or a bit
    depth entry is above 8.

    lt_ref_pic_poc_lsb_sps used log2_max_pic_order_cnt_lsb_minus4 + 4 as a
    bit width without a bound; reject values above 12 in the SPS.

    hrd_layer_set_idx is read with ue(v) but registered as FT_UINT8, which
    asserted. Make it FT_UINT32.

    Smaller fixes: vps_num_units_in_tick and vps_time_scale are 32-bit
    fields, vps_time_scale was added with a length of 1 bit; the sub layer
    profile_idc was read at a byte offset used as a bit offset; the sub
    layer profile compatibility flags were all read from the same bit; the
    FU header NAL unit type uses six bits, not five.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit ee78dc15450af59743ab0c5f79092cee18d74500)
    [Backport: the SCC extension signatures were made unsigned to match the
    fix, and their forward declarations updated to match. 4.4 has no
    h265_type_values_ext / h265_type_summary_values_ext, so the unrelated
    NAL unit type extended-value-string change was dropped and the type
    mask kept at 0x1f. The palette-width security fix, the bit-depth and
    initializer-count guards, the profile_tier bit offsets and the
    vps_time_scale width fix apply unchanged.]

commit 6eb38210ad
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:17 2026 +1000

    DVB-CI: Fix CA PMT length checks, two-pass state and CI+ cert leak

    A hostname connection descriptor, or one of another type, whose length
    was too short, and a CA descriptor shorter than 4 bytes, gave a negative
    length so the field ran to the end of the APDU and the offset moved
    back. They now get an expert info and the dissection stops.

    An open session response created a new conversation on every pass, which
    took quadratic time and doubled the conversations; create it on the
    first pass only. The negotiated buffer size and the CAM's proposal were
    global state updated on every pass, so the checks differed between one
    and two passes; keep the per-frame values with p_add_proto_data(). The
    CI+ X.509 extensions were enabled around the certificate dissection
    without TRY/FINALLY, so a malformed certificate left them enabled and
    other protocols' certificates were then decoded with the CI+ meaning of
    these OIDs; disable them in a FINALLY block.

    Assisted-by: Claude

    (backported from commit 9c8b34ab3a8716f37fc543c0aacde92508d6629c)
    [Backport: the upstream commit also fixes a CA PMT endless loop where
     dissect_es() returns the -1 error sentinel into an unsigned es_info_len,
     so "<= 0" misses it and the offset moves back. That is a master/5.0-only
     regression from 8638423a19 ("Use expert_remaning and unsigned offsets.",
     2026-01-05), which is not on release-4.4: here dissect_es() returns int,
     es_info_len/ca_desc_len are int and the "<= 0" checks still catch the -1,
     so there is no loop. dissect_es() and the CA PMT loop are therefore left
     unchanged; only the length checks, the two-pass state and the CI+
     certificate leak are backported.]

commit 71a3c59030
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:16 2026 +1000

    DO-IRP: Fix TCP length wrap, length checks and request tracking

    get_do_irp_message_len() added the 20-byte envelope to the 32-bit
    message length, so a length of 0xffffffec or more wrapped to a small
    value or 0. With 0, tcp_dissect_pdus() asked for one more segment over
    and over, so the rest of the stream was swallowed and dissection time
    grew quadratically with the number of segments. A length that wraps
    is now returned as a value below the envelope length, which
    tcp_dissect_pdus() reports as a bounds error.

    decode_generic_data(), decode_pk_data() and decode_string() read the
    length with tvb_get_int32() and returned len + 4 as an int, and the
    credential and signed-info items used cred_len + 4 and sig_len + 4.
    A length near 0xffffffff wrapped, so the item had the wrong length and
    the caller's offset moved by a few bytes or backwards. The lengths are
    now read with tvb_get_uint32() and checked against the remaining
    data; a length beyond the data is reported with the new expert info
    do-irp.length.invalid and dissection stops.

    The UDP reassembly stored the 32-bit request and sequence IDs in
    uint16_t variables, so fragments of different messages were reassembled
    together. They are now uint32_t.

    The response frame of a request was updated on every pass, so the
    "Response In" link could differ between passes. It is now set only on
    the first pass, to the first response.

    The body length of an unsupported response code is handled as
    unsigned, so a consumed length beyond the body length no longer gives
    a negative subset length.

    Every envelope cleared the Info column, so a TCP segment with several
    messages showed only the last one. The column is now cleared once per
    segment (and per UDP datagram), and the messages are separated by
    " | ".

    Over TCP, a segment shorter than the 20 octet envelope was rejected
    before tcp_dissect_pdus() could ask for more data, so a message whose
    first segment was that short was lost. Such a segment now only has its
    version checked, and tcp_dissect_pdus() reassembles the envelope.

    Over UDP, the expected fragment count was set after the first fragment
    had been added with more fragments to come, so a message sent as a
    single fragment with the TC flag set was never reassembled. A message
    that fits into one fragment is now added as the last fragment.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 44f8fdc52f871c6bac9781485fbaf06cfe98fe38)
    [Backport: two small context conflicts (a decode_string check_len call
    and the unhandled-bytes tvb_ensure_bytes_exist64 guard) resolved in
    favour of the fix; otherwise applies unchanged.]

commit 4b7a1aae4e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:01 2026 +1000

    NAS EPS: Bound Replayed NAS message container nesting, fix leaks

    The Replayed NAS message container IE of a Security mode complete
    message was dissected by calling the EMM message dissector again, and
    the inner message could itself carry another container. Five bytes per
    level were enough to recurse thousands of levels deep and overflow the
    stack. The specification allows a single level, so a container inside
    a container is now reported with an expert info and not dissected, and
    the nested call is wrapped in increment_dissection_depth() and
    decrement_dissection_depth().

    The EEA2 deciphering output buffer was a GByteArray handed to a tvb
    that never freed it, leaking the plaintext of every ciphered message on
    each pass. It is now allocated from pinfo->pool. The key GByteArray is
    also freed when the key preference is invalid or applied again.

    The Satellite operation parameters IE tested the SF waiting time flag
    twice; the estimated UL delivery time duration is now conditional on
    its own indicator.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 46db1c5f4d8ee608cdd541ab14cab10f6ef769ae)
    [Backport: 4.4 has no EEA2 deciphering support (no deciphering_eea2_msg,
    no decipher key preference) and no S&F satellite operation parameters IE,
    so only the Replayed NAS message container recursion bound and its
    expert info were taken; the cipher/key leak and esfudtpi fixes do not
    apply.]

commit ef454543e6
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Nov 8 14:36:40 2026 -0500

    openflow_v6: Use a subset tvb to fix an infinite loop

    Use a subset tvb for the action so that a ReportedBoundsError
    is automatically thrown if the length in the action header is
    not long enough for what is actually contained.

    In particular, this fixes the case in fuzzed data where the length
    indicated in the header is zero, and an unknown type is encountered.
    An exception will be thrown while dissecting the header instead of
    the offset moving backwards.

    Essentially the same issue as #17847 / CVE-2022-0585, but in a
    different part of the dissector.

    Also, update the URL to the current location.

    Fix #20932

    (cherry picked from commit 3fa8b2e13086d81270f27993512635854e71df9a)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit c07689076c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 10:19:17 2026 +1000

    BT L2CAP: Allocate the SDU as it arrives, fix several lengths

    Each I-frame with SAR=Start allocated a file-scope buffer of the SDU
    length it claims, up to 64 KB, whatever data followed. A capture of
    small Start frames took about 4.5 KB of memory per 16-octet frame and
    gigabytes of address space. Copy what the Start frame carries and grow
    the buffer as continuation frames arrive.

    With the 4-octet extended control field the payload length still
    subtracted a 2-octet one, so the FCS was read past the frame and valid
    frames were malformed.

    The Credit Based Connection Request and Response compared the offset
    with an end computed for a command at the start of the PDU, so the
    CIDs of a later command were not decoded. Configuration options now
    step over their option length, so an option longer than its known
    fields does not make the next option start in its padding.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 5f50fe9294163bd7f69f8a1c1dcb81ab45d20b5c)
    [Backport: 4.4 has no extended (4-octet) I-frame control field and no
    dissect_i_frame_control_field() helper, so its inline 2-octet control
    dissection was kept and control_length fixed at 2, which makes the
    length checks identical to 4.4's original -= 6 / -= 4. The SDU
    allocate-as-it-arrives fix and the credit-based and options length
    fixes apply unchanged.]

commit 3ca0819578
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 4 16:22:14 2026 +1000

    FCSP: Fix negotiate loop and subtree placement

    dissect_fcsp_auth_negotiate() added the 32-bit parameter length from
    the packet to the offset without checking it. A length such as
    0xFFFFFFFC wrapped the offset back to the length field, and the loop
    over up to 2^32 protocols reread it until the tree item limits ended
    the dissection with a dissector bug. Stop the list with a new
    malformed expert info when the length goes beyond the packet. The
    offset in the DH-CHAP challenge handler is now unsigned like the
    others.

    The per-opcode handlers were given the top level tree instead of the
    FC-SP tree, so their items were placed outside the FC-SP subtree.

    An unknown DH-CHAP parameter tag with a non-zero length did not advance
    the offset, so the next iteration decoded the value of the unknown
    parameter as a tag. Skip the value, and count the parameter header
    against the remaining length as the known tags do.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 9e0148066fc13ac2c41a1427502ce85c8cbee9ae)
    [Backport: in the DH-CHAP challenge handler only the offset was made
    unsigned; 4.4 keeps its uint16_t name/param length locals (master had
    widened them separately).]

commit 7c08b6260b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 19:03:59 2026 +1000

    MAC-LTE: Fix OOB read, per-RNTI memory use and a DRX loop

    Extended LCIDs (32..95, from the eLCID subheader) were used to index
    ue_mappings->mapping[], which has only 39 entries, in
    get_mac_lte_channel_priority(), lookup_rlc_channel_from_lcid() and
    get_mac_lte_channel_mode(), reading past the heap block. Check the
    index against array_length() first.

    The HARQ resend tracking allocated about 78 KB per RNTI (LastFrameData
    held a 2048 byte array), so a small capture with many RNTIs used
    gigabytes. Keep each frame's data as a file-scope allocation of the
    bytes actually seen (at most MAX_EXPECTED_PDU_LENGTH, as before) and
    free it when the slot is reused. Resend detection is unchanged.

    update_drx_info() looped forever when the subframe number was above 9
    (or the SFN above 1023), since the loop never reaches such a value.
    Return early in that case.

    Also fix the expected SI-RNTI value in the expert text (0xFFFF), use an
    unsigned shift for the SCell bitmap test, and mark the preamble item,
    not a stale item, as generated.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 47acd0904950c207984840a538e11952c82e43a7)
    [Backport: kept 4.4's "int length" field in LastFrameData (master had
    already widened it to unsigned); only the data array was changed to a
    file-scope pointer. The rest applied unchanged.]

commit 4c4536f702
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MobileIP: Bound nesting of NAT traversal tunnel data

    A NAT traversal Tunnel Data message hands its payload to the IP
    dissector. A packet that wraps IPv4/UDP port 434/MIP/IPv4/... many
    times nests without limit until the layer limit in packet.c ends in a
    DISSECTOR_ASSERT.

    Count the nesting per packet with p_get_proto_depth() and stop at 4
    levels with a new expert info (mip.nattt_too_deep). The nested call is
    also wrapped in increment_dissection_depth() and
    decrement_dissection_depth().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 7d581a31043262f1761626bc546f37e09d095b19)

commit 15f302a438
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:01 2026 +1000

    IPsec: Bound the nesting of AH headers

    AH with next header 51 (AH), or next header 135 (MIPv6) whose Fast
    Binding Acknowledgment carries payload protocol 51, nested itself
    without bound until the layer limit in packet.c failed a
    DISSECTOR_ASSERT. Count the AH nesting per packet with
    p_get_proto_depth() and stop with a new ah.nesting_too_deep expert info
    after 8 levels.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c0b156fcc23ec0721e01e4ade4e4958a174f28ea)

commit 931516eb38
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MIKEY: Fix SRTP policy parameter hf lookup and small field errors

    The SRTP policy parameter table hf_mikey_sp_param was declared with
    SP_MAX + 1 entries but only SP_MAX - 1 parameter types are registered.
    A parameter of type 13 (SP_MAX) passed the bounds check and was added
    with hf index 0, which is a dissector bug "Unregistered hf" (an abort
    with WIRESHARK_ABORT_ON_DISSECTOR_BUG). Size the array SP_MAX so the
    bounds check rejects the type and the generic parameter field is used.

    The PKE payload length used the mask 0x3ff instead of the 0x3fff the
    field is registered with. The Certificate payload length item was added
    at offset 1 instead of 2. dissect_mikey returned -1 from the payload
    loop; return the amount of captured data like the other early exits.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 917aeb460edd35ec0075e22bb729f6bcc3594a74)

commit f4fd2538e8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MoldUDP64: Fix infinite loop on truncated frames

    When fewer than two bytes of a message block were captured but the
    reported length still had room for one, the block dissector returned 0
    without advancing and the loop in the caller never ended. Stop the loop
    when a block makes no progress.

    The message length and block length were 16 bit values that wrapped for
    a length of 0xFFFF, making the block advance by one byte and creating
    tens of thousands of items. Use unsigned values, and an unsigned
    message count which no longer wraps.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 138e1cd41d4ac7b593227e3d27b88db3d744f948)

commit f86be82155
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:01 2026 +1000

    MySQL: Fix an infinite loop and a reassembly assertion

    A MariaDB COM_STMT_BULK_EXECUTE for a statement prepared with zero
    parameters never advanced the offset: the row loop only moves forward
    through the per-parameter inner loop, which runs zero times. Any
    trailing byte after the bulk flags then kept the dissector looping
    forever (a CPU hang, or with a tree an unbounded number of row items).
    Stop the row loop when a row consumed no bytes.

    When a decompressed PDU header was split across two compressed packets,
    dissect_mysql_decompressed_pdus() asked for the missing header bytes
    without setting desegment_offset, so a stale offset from an earlier PDU
    made the streaming reassembly code fail an assertion. Set the offset to
    the start of the partial header.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 12bb86ba1ff2f7a9bb3448886769439605648f9a)

commit 532b097e17
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MIH: Fix composite tvb leak and NAI-encoded MIHF ID lengths

    dissect_mih_tlv() built a composite tvb for a MIHF ID starting with a
    backslash, outside any TRY block. The length guard measured from the
    start of the tvb instead of the field, so the first
    tvb_new_subset_length() could throw before any member was appended.
    The composite was then never chained to a parent and leaked for every
    crafted packet. Require that the whole ID exists with
    tvb_bytes_exist(tvb, offset + 1, mihf_id_len) so the loop cannot throw.

    The accepted lengths were 12, 64 and 128, which do not match the NAI
    encoding of a MAC address, IPv4 and IPv6 address (two characters per
    octet: 12, 8 and 32). Use 12, 8 and 32.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit ee9b0b8b1643e220323efcfc8f4d1eda80f1716b)

commit e9f45a6df0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    META: Fix NULL handle, uninitialised ATM info and TLLI read

    An unregistered protocol id in the META header made
    dissector_get_uint_handle() return NULL, which was passed to
    call_dissector_with_data() and hit DISSECTOR_ASSERT(handle != NULL).
    Fall back to the data dissector.

    The atm_phdr passed to the ATM dissector for DXT ERF AAL5 was only
    partly set, so the ATM dissector read uninitialised stack. Zero it.

    Read the 4-byte TLLI item with tvb_get_letohl() instead of
    tvb_get_letohs(), make the META protocol item cover the 8-byte header
    plus hdrlen, and report an item offset that would overflow the 16-bit
    offset as malformed instead of wrapping.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit f8509a06618746e9d636dce39f1d6485a7517c7d)

commit ea06910100
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MS-DO: Fix assertion and size overflow

    do_get_direction_str() hit DISSECTOR_ASSERT_NOT_REACHED() when
    pinfo->match_uint was neither the source nor the destination port, as
    happens with Exported PDU. Return "Unknown direction" instead.

    The check "buffer_size < message_size + 4" overflowed for a message size
    of 0xFFFFFFFC or more, so a bogus message passed as complete. Compare
    without the addition and compute the missing byte count without
    overflow as well.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5f407628b6bdeb1f86028ac31e53f5ab9a0fb4d6)

commit da67ad805f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:01 2026 +1000

    MSNLB: Use an unsigned offset

    The offset was a uint16_t. In a frame larger than 64 KiB it wrapped
    around in the Extended Heartbeat loop, so the loop never reached the end
    of the tvb and added the same items forever until memory ran out. Use an
    unsigned offset.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e2754261688a5d6ffd87113364d1756074683c36)

commit afa040590a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 21:00:00 2026 +1000

    MongoDB: Reject a nested OP_COMPRESSED message

    A compressed message whose original opcode is OP_COMPRESSED made
    dissect_mongo_op_compressed() nest one level per 9 byte header until
    the recursion DISSECTOR_ASSERT fired. With zlib, each level also
    decompressed up to INT_MAX bytes, so two nested levels in a 787 byte
    packet used 400 MB. A compressed message cannot contain another one:
    add an expert info and stop at the first nested OP_COMPRESSED.

    Replace the recursion_depth DISSECTOR_ASSERT in dissect_opcode_types()
    with an expert info and return, since the depth is controlled by packet
    data.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6eb47529156581b65bbe237e6ded51e29b08a545)

commit 4d94dc5fe5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 19:03:59 2026 +1000

    NR-RRC: Bound recursion through nested NR RRC messages

    The MRDC transfer messages ULInformationTransferMRDC and
    DLInformationTransferMRDC carry a complete UL-DCCH-Message or
    DL-DCCH-Message as an octet string, which was dissected by calling the
    message dissector directly. A packet in which each level contains the
    same message type again needs only about four bytes per level, so a
    packet of a few kilobytes overflowed the stack. This is reachable by
    default through LTE RRC and NR RRC.

    Allow at most 8 nested NR RRC messages per packet and stop with an
    expert info (nr-rrc.nesting_too_deep) beyond that; count each nested
    message as a dissection level too. The same limit now also guards the
    existing segment reassembly recursion, which called
    increment_dissection_depth() without checking the limit and could end
    in an assertion.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 8d142292aea2213557d812ccabb61b1f12386a1a)

commit 8d9162ec90
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 19:03:59 2026 +1000

    MAC-NR: Fix use after free on an unknown framing tag

    dissect_mac_nr_context_fields() freed the mac_nr_info on an unknown
    context tag and then returned true. Both callers took that as success,
    stored the freed pointer with p_add_proto_data() and went on to read it
    in dissect_mac_nr(); mac-nr-framed also wrote into and freed it again on
    the next pass. That is a heap use after free and a double free.

    Return false without freeing instead. Each caller now frees the struct
    only if it allocated it in this call; mac-nr-framed never frees a struct
    it got back from p_get_proto_data(). Nothing is stored before the
    context parse succeeded.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 615a958984291a053dabcdda8b60a3b909eb77df)

commit 1bcc84df5f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 12:30:40 2026 +1000

    LLRP: Bound Impinj parameter nesting and fix some offsets

    dissect_llrp_impinj_parameter() restarted the nested parameter depth
    at 0 when it dissected the optional trailing custom parameter, so the
    "depth > 16" stop never applied to chains of nested Impinj custom
    parameters. Pass the depth through and continue with depth + 1.

    For an unknown vendor in a CUSTOM_PARAMETER the length was computed as
    len - 8, which can be negative or wrong when the length field is
    invalid. Use the bytes left up to the end of the parameter instead.

    Fix the offsets of GET_BLK_PERMALOCK (4 bytes after the access
    password, 1 after the memory bank), the AccessSpecID of
    CLIENT_REQ_RESPONSE (4 bytes) and the offset skipped after the
    reader ID of IDENT (2 length bytes plus the string).

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit df4e42422f4cf0ba60d9404ee92624bcf0ab0fa7)

commit 10da242959
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 12:30:39 2026 +1000

    LISP: Bound LCAF address recursion and fix a depth leak

    get_addr_str() recursed through nested Instance ID and AS Number LCAFs
    with only increment_dissection_depth(), so a chain of nested LCAFs ran
    into the dissection depth assertion. The recursion now goes through a
    static helper with a level counter and stops at 4 levels, returning a
    placeholder string.

    dissect_lcaf() returned from the unknown LCAF type case without calling
    decrement_dissection_depth(), leaking one level per such LCAF. It now
    skips the data and leaves through the common exit.

    dissect_lcaf_afi_list() returned -1 for an unknown AFI, which callers
    used as an offset. It now returns the current offset after the expert
    info. dissect_lcaf_nonce_loc() used tvb_get_ntohl() for the one byte
    reserved field and the three byte nonce, reading past them. Strings
    returned by get_addr_str() can be NULL for an unknown AFI; they are
    now replaced by "(unknown)" when appended to the tree item text.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 2e7044b3c76b54f42a4caf2128110de63328a804)

commit 721eaca9c0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 12:30:39 2026 +1000

    LDAP: Fix a NULL tvb, quadratic filter strings and small bugs

    An empty constructed OCTET STRING as an AttributeValue leaves no
    tvbuff, and the default and the custom attribute UAT paths then passed
    NULL to tvb_ascii_isprint() or proto_tree_add_item(), which hits a
    DISSECTOR_ASSERT. Return early when there is no tvb.

    The display string of a SubstringFilter was rebuilt with
    wmem_strdup_printf() for every substring item, so the packet pool grew
    quadratically (63 KB of items used 530 MB). Build it in a wmem_strbuf
    instead and stop appending once it holds ITEM_LABEL_LENGTH (240)
    characters; a single item contributes at most that many characters.

    The and, or, not and present filters copied the whole display string of
    their children at every nesting level, and the and/or accumulators were
    rebuilt per item. Truncate the strings that feed Filter_string to
    ITEM_LABEL_LENGTH before wrapping them, without cutting a UTF-8
    character in two, and stop growing the and/or accumulators past that
    length.

    Deeply nested filters reached the DISSECTOR_ASSERT of the dissection
    depth limit (about 200 nested not filters). Stop with an expert info
    16 levels before that limit.

    A Credentials value of zero length (04 00) made get_ber_identifier()
    throw; return early instead. DirSyncFlags read the BER length at the
    identifier offset instead of after the identifier; use the offsets
    returned by the BER helpers. dissect_NetLogon_PDU() kept an unsigned
    length in a uint16_t; use unsigned.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit f4864b55a501a23f6516d0169a28e0011856b0ab)

commit ec417b6401
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 12:30:39 2026 +1000

    LBMPDM: Fix crashes and hangs on malformed messages

    A DECIMAL field with an exponent of -64 or less made the divisor
    10^n wrap to zero, so the modulo that followed raised SIGFPE. The
    positive exponent path and the negation of INT64_MIN also overflowed
    signed arithmetic. Exponents with an absolute value above 18 are now
    shown raw, and the arithmetic uses unsigned magnitudes with an
    overflow check.

    An offset table entry with a huge field id (up to 0x7fffffff) made the
    dissector allocate and fill an array of that many entries, up to 8 GiB
    from a 32 byte packet. Ids above 65535 now get an expert info and the
    table is dropped. Offset table segments are also checked against the
    reported length of the packet before the arrays sized by the segment
    length are allocated. Offsets from the table that point outside the
    data segment are skipped.

    The segment length was converted to a signed int and could become zero
    or negative, so the segment loop never advanced. The loop now rejects
    segments shorter than the segment header or longer than the message
    remainder or the packet, with an expert info. A message length smaller
    than the message header, which wrapped the remaining length, is
    rejected too.

    An array field whose elements have zero size iterated up to 2^32
    times at the same offset. The loop now stops with an expert info.

    The length of a string-named field definition (uint32 + 4) could wrap
    to zero so the definition loop did not advance. It must now be at
    least the minimum definition size and fit in the segment; otherwise
    the loop stops with an expert info.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit f79a39894c9388878fb781de7f76172db11a3053)

commit e93bca08fa
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Nov 7 12:30:39 2026 +1000

    L2TP: Fix session list growth and Ericsson AVP length handling

    update_session() appended the session to tunnel->sessions on every
    call, even when an existing entry was found and merely updated, and it
    ran on every dissection pass. A capture with many control messages
    built a list with duplicate entries, making each find_session() walk
    longer and the total work quadratic. Append only newly created
    sessions, and call update_session() only on the first pass.

    The Ericsson TEI to SC Map AVP created its subset with avp_len instead
    of avp_len - 1 after consuming the map count byte, so the subset could
    extend past the AVP. Require avp_len >= 1 and use avp_len - 1.

    The Ericsson Transport Configuration AVP subtracted 3 from avp_len
    without checking it, which underflows for short AVPs. Report a bad AVP
    length with the existing expert info when avp_len < 3.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 130a178184ae727ced3d7d81c4d34b6048f79cd7)

commit 65e8a28437
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 19:55:21 2026 +1000

    KT: Check the key, value, name and log sizes against the packet

    The ksiz, vsiz, nsiz and size fields are 32-bit values taken from the
    packet and were passed unchecked as the length of proto_tree_add_item()
    and added to an int offset. A value of 0xFFFFFFFF became the length -1,
    which means "to the end of the data", and moved the offset back by one
    byte. With a large record count every record then added the rest of the
    packet to the tree twice, which needs memory quadratic in the packet
    size.

    Call tvb_ensure_bytes_exist64() before every item whose length comes
    from the packet, so that a size larger than the remaining data ends the
    dissection with an exception instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 8a70bc0d904ab932f5311bb27b24f4b6214b9a2f)

commit acd3265cc5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 19:55:21 2026 +1000

    KINK: Bound the payload chain recursion

    control_payload() calls the handler of each payload, and every handler
    calls control_payload() again for the next payload. Nothing limited the
    depth, so a packet with a long chain of tiny payloads overflowed the
    stack.

    Count each payload as a dissection level and stop with an expert info
    before the depth reaches a fixed limit or the tree depth preference.
    Remove the comment in proto_reg_handoff_kink() that claimed the UDP
    transport keeps the recursion bounded.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit d62b70019c7d24f9a37a7d9909ee582802b41549)

commit a7bc995d06
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 19:55:21 2026 +1000

    Kafka: Fix record loop, asserts and field types

    A record size varint of 2^32-5 was truncated to int, so the record
    ended where it started and the record batch loop dissected the same
    bytes up to 0xffffffff times, an endless loop with unbounded memory
    use. Reject a record size larger than the remaining data with an
    expert info, and stop the batch loop when a record does not advance
    the offset.

    The loop limits in decompress_lz4() and the xerial path of
    decompress_snappy() were DISSECTOR_ASSERTs reachable from packet
    data. Treat reaching the limit as a decompression failure instead.

    Fix the offset returned for a bad string length in
    dissect_kafka_regular_string() (offset + 2, not 2), and make the fetch
    session ID, fetch session epoch and forgotten topic partition fields
    FT_INT32, matching the 4 bytes they are added with.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 9c64cfd2756323b7d68b0653db1177320c8e11fc)

commit a2585bbf45
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 19:55:21 2026 +1000

    tvbuff: Do not assert on the zstd loop limit

    tvb_uncompress_zstd() used DISSECTOR_ASSERT_HINT when the loop
    iteration limit was reached, which crafted compressed data can
    trigger. Remove the assertion and treat input left over after the
    loop as a decompression failure: return NULL, freeing the buffers
    allocated so far.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit c84853f19e388555282d8f1602246aa411d44d9b)

commit 83664b49e1
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 19:55:21 2026 +1000

    ISO 15765: Do not assert on too many segments, clamp the length

    The per-sequence-number counter of segments was a uint8_t that wrapped
    after 255 consecutive frames with the same 4-bit sequence number, and
    DISSECTOR_ASSERT(frag_id_high != 0) then aborted the dissection. This
    is reachable from a crafted LIN diagnostic capture. Check the counter
    before incrementing it; at UINT8_MAX add a new expert info and flag the
    transfer as being in error instead.

    When Consecutive Frames are missing, last_byte_seen can pass the
    announced length by more than the size of the current frame, so len and
    bytes_used wrapped. Clamp the excess to the frame length.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit fe03ebeb51a0fe26a0662164580738c721372d37)

commit 46bf9f5cd7
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 14:02:35 2026 +1000

    RPCoRDMA: Do not pass a NULL tvbuff on the second pass

    On the second pass add_send_fragment() returned the result of
    get_reassembled_data() for every SEND fragment. When the reassembly did
    not complete, for example for a malformed sequence of SEND First,
    Middle and Last, that is NULL, and dissect_rpcrdma_ib_heur() and the
    iWARP heuristic passed it on to packet_is_rpcordma(), which ended in an
    assertion on a NULL tvbuff. The first pass dissected the fragment as it
    is in that case, so the two passes also gave different results. Return
    the fragment itself when there is no reassembled data, as the first
    pass does.

    Also check the length in packet_is_rpcordma() before reading the XID
    and the message type.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 7af96626c3cd4f6a63c5ebd6cb6dd4abf66acb57)

commit accaf55a55
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 12:37:09 2026 +1000

    iperf3: Do not mix TCP and UDP conversation data

    The TCP and the UDP dissector store differently sized structures under
    the same protocol key of a conversation found by address and port only.
    A packet reaching the UDP dissector through Exported PDU with the
    addresses and ports of a TCP iperf3 flow got the 12 byte TCP structure
    and used it as the UDP one, reading past the allocation and using the
    bytes as a map pointer. The reverse mixed the UDP structure into the TCP
    sequence state and could hit a DISSECTOR_ASSERT_NOT_REACHED.

    Return early unless pinfo->ptype is PT_TCP in the TCP dissector and
    PT_UDP in the UDP dissector.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5aee7fd428b81554b6ce3108b02ba6597d63113b)

commit c42f630cd0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:20 2026 +1000

    ISAKMP: Fix dissection depth and fragment decryption failure

    dissect_payloads() incremented the dissection depth a second time at the
    end of each payload instead of decrementing it. A packet with many
    payloads therefore exhausted the depth limit and hit the assertion in
    increment_dissection_depth(). Decrement it instead.

    dissect_ikev2_fragmentation() used the result of dissect_enc() without
    checking it. When the fragment cannot be decrypted dissect_enc() returns
    NULL, which led to an assertion in tvb_reported_length(). Return in that
    case, before pinfo->fragmented is changed, and also when the decrypted
    data is empty.

    When no keys are available, dissect_enc() added the encrypted data item
    at offset + 4 with the full length, running past the end of the payload.
    Use length - 4, and handle payloads shorter than the 4 byte IV.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 93d8cc5d32db35ab7dc04b1df1d0e3c9bb065c21)

commit abfd693d07
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:20 2026 +1000

    IEEE 802.1CB: Initialise the ethertype data

    dissect_ieee8021cb() left ethertype_data.trailer_id uninitialised,
    so stack garbage was used as a header field index by the ethertype
    dissector. Register a Trailer field, set trailer_id to it and
    zero-initialise the structure.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 51c5c5f876d7d745c7d4aab15bda5bb5de222940)

commit 75cdd74136
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:20 2026 +1000

    IDMP: Check the reassembled tvb and fix the PDU length setup

    process_reassembled_data() can return NULL, which was then passed to
    the PDU dissector. Keep the result in a separate next_tvb and return
    when it is NULL. Zero-initialise the session data structure, and pass
    the 6 byte header length that get_idmp_pdu_len() reads as fixed_len
    to tcp_dissect_pdus() instead of 0.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 7a9c40da0fc0925b54ab1fb04b8b656969322074)

commit 27f4f6ff72
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:19 2026 +1000

    IAX2: Do not assert on a repeated frame in desegment_iax()

    Several IAX2 PDUs in one frame (for example via PPPMux) made
    desegment_iax() register the same frame number twice in the fragment id
    table, which hit a DISSECTOR_ASSERT(). Add an expert info and dissect
    the PDU without desegmentation instead.

    Trunk packets counted the distinct call numbers with a linked list that
    was searched and appended to for every entry, which is quadratic for
    large trunk packets. Use a wmem_map instead.

    The subtree of a trunk call without timestamp was 2 bytes too long
    (rlen + 6, but the header is 4 bytes); use rlen + 4. Initialize fd_head
    in desegment_iax().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 2aa846604953b32f5e4e832039e859e26003348a)

commit 68ed6600ed
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Nov 6 06:45:19 2026 +1000

    HSMS: Stop on an unknown item format code and on deep lists

    dissect_secs_variable() looks up the value length of the item format
    code in a map. An unknown code returns 0, and the following
    length % 0 and length / 0 trap with a division by zero. Add an
    expert info for the unknown format code and return -1 before
    dividing.

    Each nested list adds two levels to the protocol tree, so deeply nested
    lists reached the tree depth limit before the dissection depth limit and
    showed as a dissector bug. Stop with an expert info before the tree
    depth limit is reached.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit ca7099dd9ead8f0f2b7415275ce0ecff1c92280c)

commit 196432367b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 22:09:48 2026 +1000

    HICP: Avoid quadratic parsing of the parameter list

    The parameter loop called g_strv_length() on every iteration, which is
    quadratic in the number of ';' separated fields and let a large packet
    burn seconds of CPU. Loop until the NULL terminator instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 01785c1bff159276469d38de244445592f24b495)

commit 1c6858c144
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 22:09:48 2026 +1000

    HDFS: Fix a global over-read comparing type names

    Several tvb_memeql() calls compared a packet supplied length of up to
    65535 bytes against short string literals, reading past the end of the
    literal. Add a helper that requires the length to equal the literal's
    length before comparing, and use it for all of them.

    Also make hdfs.authlen a FT_UINT32 read with ENC_BIG_ENDIAN, matching
    the four byte length it displays.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 4a4d82fdd6c8945d519b7d3fae1aa138c8ddd062)

commit 9885cf2b89
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 22:09:48 2026 +1000

    HCLNFSD: Fix a heap over-read decoding the authorize ident

    hclnfsd_decode_obscure() allocated exactly ident_len bytes without a
    terminating NUL, and the AUTHORIZE call then took the username two
    bytes in and ran strlen() on it and the password behind it, reading
    past the buffer for short idents.

    NUL-terminate the decoded buffer, only extract username and password
    when the ident is longer than two bytes, and keep the password pointer
    inside the decoded buffer.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit a6124a3a3dd5bcab190742c9021062c4a30ccabe)

commit b63b50086b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 22:09:48 2026 +1000

    H.460: Limit RAS nesting through H.460.17

    A RAS message can carry GenericData 17/1 holding another RAS message,
    without end. The recursion was only stopped by the generic dissection
    depth limit, which ends in a DISSECTOR_ASSERT when no tree is built.

    Count the nesting with p_get_proto_depth()/p_set_proto_depth() and stop
    at 8 levels with an expert info.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit eae99ccc5b28cf9c641ee4e750a293e6eadcfe4d)

commit 1919deac7f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 22:09:48 2026 +1000

    H.248 Annex C: Check for a NULL tvb in acodec

    dissect_ber_octet_string() returns a NULL tvb for an empty constructed
    OCTET STRING, which was passed to tvb_reported_length() and threw a
    dissector bug. Return in that case.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit d705683dca188552408cc0d6b2f21e849a737796)

commit d8fab4037d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 22:09:48 2026 +1000

    H.248.7: Decode the apv specific parameters as octet string

    The announcement variant specific parameters field is an FT_STRING but
    was decoded with h248_param_ber_integer, which asserts in the BER
    integer code for a string field. Use the octet string decoder, like the
    other FT_STRING fields of the package.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e42b93ced25549a704e55250a494c333d7b035c8)

commit d4bd87f3b9
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 22:09:48 2026 +1000

    H.248: Fix NULL tvb and uninitialized context handling

    An empty constructed OCTET STRING as WildcardField or as
    SigParameterName gives a NULL tvb that was passed to tvb_get_uint8() or
    tvb_reported_length(), which throws a dissector bug. Check for it.

    In gcp_ctx(), contexts were allocated with wmem_new() and their prev
    pointer was left uninitialized, which was read when the context
    information is enabled. Allocate them zeroed. A missing context no
    longer asserts; a temporary one is used instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 1c6ed81c6844f41028b69b4031ec29253063c6ca)

commit a7d4c716d8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 22:09:48 2026 +1000

    H.245: Reset the upcoming OLC state on exceptions

    The file-scope upcoming_olc and upcoming_channel pointers are set while
    an OpenLogicalChannel or OpenLogicalChannelAck is dissected. When an
    exception was thrown inside (a truncated PDU), they were left pointing
    into freed packet memory. A later UnicastAddress reached from H.460,
    H.225 or SDP without going through dissect_h245_h245 then wrote through
    the stale pointer, a use-after-free.

    Push a cleanup handler around the body of both functions that clears
    the two pointers when an exception unwinds.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e4e24bc45e89a0973572a96b98f15a55e5926204)

commit e1280a5b64
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 18:55:34 2026 +1000

    H.245: Reset the H.223 channel parameters for every OpenLogicalChannel

    The forward and reverse H.223 logical channel parameters are kept in
    file-scope pointers that were never reset. An OpenLogicalChannel
    without reverse parameters therefore inherited those of an earlier
    one, and its OpenLogicalChannelAck then failed the DISSECTOR_ASSERT
    that compares the reverse channel of the ack and the request. The
    pointers also outlived the file-scope memory they point to.

    Reset both pointers at the start of each OpenLogicalChannel and when
    the file is closed, and replace the assertion, whose two sides both
    come from packets, with a check before adding the reverse channel.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 0ba2d749b2cd121334b2f13736922dd2781f72ca)

commit f9ffb35cba
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 17:39:52 2026 +1000

    GVSP: Limit the work done for GenDC component and part offsets

    The GenDC container descriptor was decoded by following the component
    offset array and, for each component, its part offset array. Nothing
    prevented the offsets from repeating, so a packet of 64 KB made the
    dissector decode N * P part headers, which takes seconds and exhausts
    the tree item limit.

    Require the component offsets and the part offsets of one component to
    increase strictly, limit the total number of part headers to 1024 and
    stop with an expert info otherwise.

    The component and part header subtrees also added offset + 16 twice,
    so they started in the wrong place. Use the already adjusted offsets.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 685f377e8b9eafeb2869dae0d4a453b96db9d1c8)

commit 007fa5fedb
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 17:39:51 2026 +1000

    GSM CBSP: Fix a NULL dereference for unsupported coding schemes

    dissect_cbs_data() returns NULL for a data coding scheme it does not
    support, for example a reserved one. dissect_cbsp_content_ie() passed
    the result to tvb_captured_length() and the tree code without a check,
    which ends in an assertion failure when a tree is built. Only add the
    page content when there is an unpacked tvb.

    Also set len_len to 1 for TLV_TYPE_TLV attributes. No attribute uses
    this type today, but the length octet was otherwise not accounted for.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 0ea3de7dc209ada3374f29f1ec40f2434f95a5bd)

commit 6f28b7e3ce
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 17:39:50 2026 +1000

    GQUIC: Do not assert on CRYPTO frames in older versions

    A CRYPTO frame (0x08) in a packet whose version is not known or is
    below Q050 hit DISSECTOR_ASSERT(version_valid && version >= 50). That
    is reachable with one crafted UDP packet. Report an expert info and
    skip the rest of the packet instead.

    Also fix a few tag parsing errors: the CADR unknown address type
    advanced tag_offset by tag_len + 4 instead of tag_len - 4, a PUBS tag
    shorter than 2 bytes made the end-offset arithmetic wrap, and the
    number of tags in an SCFG tag was not limited to what can fit in the
    remaining data, so tag_number * 8 could wrap.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit d8b2d7763826b148f92b7b62291315f42503abf3)

commit f7704efa5d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 08:49:38 2026 +1000

    GELF: Fix a memory leak when a chunked message is truncated

    The chunked message ID was fetched into a GByteArray that was only
    freed after later calls, which can throw on a short packet. The array
    then leaked. Copy the eight ID bytes into a stack buffer instead.

    BUILD_MESSAGE_ID() shifted the message id bytes by 3, 2, 1 and 0 bits
    instead of 24, 16, 8 and 0, which made different chunked messages share
    a reassembly id. Build the 32 bit value with the right shifts.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit bb2e75cb0206d04fc5344354b8088afa1b214785)

commit cf3813aa06
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 08:49:38 2026 +1000

    GCSNA: Fix endless loop in the service reject message

    The service reject dissector kept its bit offset in a uint16_t. The
    offset wrapped at 8192 bits for large packets, moved backwards and made
    the field loops run until the tree guards stopped them. Use uint32_t
    for the offset, the field count and the cause.

    The number of reserved bits before the encapsulated PDU was computed
    from the low two bits of the offset instead of the low three, so it
    could add a wrong number of bits or a whole byte when the offset was
    already aligned. Compute it modulo 8 and skip the field when it is
    empty.

    The Service Reject default case added the unsupported-reject data with
    a bit offset used as a byte offset, so cause values the dissector does
    not know were always malformed. Convert the bit offset to a byte offset.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 593b90fa2500a55e70f1d401ffd71378669671ce)

commit c948343111
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 08:49:37 2026 +1000

    FTDI MPSSE: Fix assertions reachable from crafted packets

    A data shifting command with a 16-bit length of 0xffff in byte mode
    needs a response of 0x10000 bytes. The response length was passed as
    uint16_t, wrapped to 0 and ended in DISSECTOR_ASSERT(response_length >
    0). Pass it as uint32_t, and ignore a zero length instead of asserting.

    A preliminary command record completed by reassembly in a later frame
    was never inserted in the command tree for that frame, so the second
    pass found a different record and hit an assertion. Insert the pointer
    when the completing frame differs from the one that created the record.

    A stale preliminary record left from another MPSSE instance (for
    example after the device was re-enumerated) tripped the cmd and
    response_length assertions. Treat a mismatching record as no longer
    preliminary and start a new one. On later passes, skip a missing or
    mismatching record instead of asserting, and drop the assertion on the
    preliminary flag when advancing to the last record.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit cc3997eaa51e0634e5260285d8f462e9b61bfc81)

commit a86db8569c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 08:49:37 2026 +1000

    FTDI FT: Fix assertion in reassembly and repeated state updates

    The reassembly code asserted that the desegment offset set by the
    subdissector is larger than the bytes already reassembled. The MPSSE
    dissector can set a smaller offset when it scans past a bad command, so
    a crafted packet hit DISSECTOR_ASSERT. Treat such an offset as a
    continuation of the current reassembly instead.

    The control request records and the bitmode records were stored again
    on every pass over a frame. Only update them on the first pass.

    Frames that are only part of a reassembly showed the protocol "FTDI
    MPSSE" in the first pass, because the first pass hands them to the
    subdissector, and "FTDI FT" in the second pass, which does not. Set the
    protocol column to "FTDI FT" in the first pass when nothing in the frame
    was completed.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 308458538a60c221311f8ad050bff7044cc11f77)

commit bf9d881d61
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 4 16:22:14 2026 +1000

    f5ethtrailer: Fix noise asserts and short trailer handling

    The "f5ethtrailer.noise" dissector can be called by name, for example
    from an Exported PDU file, with a NULL data argument. The noise
    subdissectors then hit DISSECTOR_ASSERT(tdata != NULL) and the frame
    shows a dissector bug. Return 0 for missing data in the entry point
    and replace the three assertions with plain checks.

    The heuristic subtracted the offset of the first non-zero byte from
    a trailer length that was already measured from that offset. With
    the unsigned wrap the size test passed on short remainders, and the
    FCS probe read past the end of the tvb, so a frame that is not an
    F5 trailer showed as malformed. Test the remaining length instead,
    and compare the DPT length against what remains from the magic,
    which also covers the case where the trailer starts after an FCS.

    An RST cause length of 1 to 8 made the string and item lengths negative
    in the old medium trailers and in the DPT medium v4 noise. Reject such
    lengths (treat them as bad in v4). The v3 RST cause length was read with
    a signed accessor; use tvb_get_uint8. In the low noise v2/v3 VIP name,
    the length check now starts at the name, one byte after the length.

    dissect_dpt_trailer() returned a DPT length smaller than the header,
    or larger than the data, as is; use the tvb length in that case
    so the heuristic does not return after adding items. A '%' in the
    brief_inout_chars preference produced an invalid printf format and
    is now rejected in favour of the default characters.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 1a8b90d57e10401b1c2315352aaf17cb09ef5d8f)

commit 1e97e26e53
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 4 16:22:13 2026 +1000

    ETI: Use FT_INT64 for fixed point fields registered as FT_UINT64

    NettingCoefficient, TradeToQuoteRatio and TradeToRequestRatio were
    registered as FT_UINT64, but the field tables dissect them as
    ETI_FIXED_POINT, which always adds a signed 64-bit value. Every
    SRQSNegotiationRequesterNotification, SRQSNegotiationNotification and
    RiskLimitsRptGrp entry therefore raised a dissector bug (and an abort
    with WIRESHARK_ABORT_ON_DISSECTOR_BUG) and cut off the rest of the
    message. Register the three fields as FT_INT64.

    The file is generated by eti2wireshark.py, which is not in the tree; the
    generator needs the same change.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 193e6ac655c3ff476add2ce4309f64d89389477f)

commit 405c38e8f4
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 4 13:47:44 2026 +1000

    Ethertype: Limit the nesting of encapsulated payloads

    Ethernet in Ethernet (0x6558), E-Tag, EtherIP and similar encapsulations
    all re-enter dissect_ethertype(). Deeply nested packets ran into the
    DISSECTOR_ASSERT on the layer limit in epan/packet.c, a "Dissector bug"
    that aborts with WIRESHARK_ABORT_ON_DISSECTOR_BUG=1.

    Stop calling the next dissector when the packet has come within 8
    layers of the limit, and add an expert info
    (ethertype.nesting_too_deep) instead. Dissectors that check the depth
    themselves stop 16 layers before the limit, so they still report the
    nesting first.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 55fdac962d4dd39ba22beb22ab36942ed72d12e5)

commit 815b9889ad
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 16:52:54 2026 +1000

    ERF: Bound Anchor ID linking and check parent_section length

    Every record with an Anchor ID extension header walked the list of all
    earlier records with the same Host ID and Anchor ID, added a "Linked
    Frame" item for each and marked each one as depended upon. A file in
    which every record carries the same Anchor ID took quadratic time,
    and, because mark_frame_as_depended_upon() stores the links in the
    frame data, quadratic memory when the frames are kept (-2 or the GUI):
    20000 records (1.7 MB) used up 2 GB. Now a non-definition record links
    only to the anchor definition records, which are kept in a list of
    their own, and a definition record links to all records with the
    Anchor ID as before. At most 64 links are shown and marked per record;
    an expert info says when there are more.

    The Anchor ID hash table compared the IDs masked with the extension
    header type (ERF_EXT_HDR_TYPE_ANCHOR_ID, 18) instead of the Anchor ID
    mask, so two different IDs could compare equal. Use
    ERF_EHDR_ANCHOR_ID_MASK.

    A parent_section tag shorter than 4 bytes was not treated as
    truncated, and its Section ID was added with a length of -1 or less,
    which runs to the end of the record. Give parent_section an expected
    length of 4, so it is reported as a truncated tag like the others.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5f20ba9a3b0ac32160ef262111e74e6f180d0e94)

commit 49721adcc4
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 16:52:54 2026 +1000

    EPL: Fix quadratic mapping handling and profile parser crashes

    Every SDO ObjectMapping was appended to the conversation's mapping list
    on every pass and the whole list was then sorted again, and each
    mapping registered a subtree index of its own that was never freed. A
    capture with many mappings took quadratic time and leaked a subtree
    index per mapping and pass. Record mappings only on the first pass,
    insert them at their sorted position, limit a list to 4096 entries
    (with an expert info for those left out) and use a single registered
    subtree for all PDOs.

    The SDO duplication table was searched in full for every frame that
    starts or ends a sequence, which is quadratic in the number of entries,
    and it was updated again on every pass, so later passes could mark
    frames as duplicates that were not. Key the table by source and
    destination node with one small table of sequence numbers per pair,
    which is cleared in one go, update it only on the first pass and keep
    the duplicated frame number in the per-packet data.

    The profile parser dereferenced a NULL string when an EDS file has no
    Description in [FileInfo], crashing when the preferences are applied,
    passed an uninitialised GError to g_key_file_load_from_file() and
    leaked the error, rejected an XDD file when an XPath expression
    matched nothing, read the value of XML attributes without children,
    and let the fields of one subobject carry over to the next one. Check
    the description and the attribute values, initialise and free the
    error, treat a missing node set as empty and reset the subobject for
    each element.

    Also decode Real64 values with an FT_DOUBLE field instead of the
    FT_FLOAT one, don't let the remaining PDO payload length wrap around
    when mappings overlap, show variable length types such as strings over
    the whole payload instead of with length 0, start each entry of a
    Write Multiple Parameter by Index response at its own offset, show the
    subindex instead of the index in its subindex field and determine
    whether the entry has a subindex correctly, and read the 8-bit
    subindex of a mapping in a Read Multiple Parameter by Index response
    as one byte.

    Changing the device profile table again after it had two entries for
    the same device type freed memory twice: the second entry is a shallow
    copy kept in the scope of the first, and drop_profiles() destroyed that
    scope once per entry and read the freed profile. Mark the copies and
    free only the profiles that own their scope.

    Free the buffer a mapping from a profile is read from, and the node
    ID string of a profile table entry, which were leaked each time the
    preferences were applied.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 9c73369fb4065926af634ecb4d0e6392b7e96df2)

commit 444a0ea6f5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:17 2026 +1000

    DRDA: Limit the recursion of nested collections

    A DDM collection (TYPDEFOVR, PKGSNLST) can contain another collection,
    and dissect_drda_collection() called itself through the opcode table
    without a limit, so a packet with deeply nested collections overflowed
    the stack, also without a tree. Bound the recursion with
    increment_dissection_depth().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 90f06458d4ad5825e9df7547236a5b41a3441cdf)

commit fdefadd331
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:17 2026 +1000

    DSR: Limit Next Header recursion and fix the flow id length

    A DSR header whose Next Header is DSR again called the DSR dissector
    recursively through the ip.proto table without a limit, which ended in
    an assertion or a stack overflow. Bound the recursion with
    increment_dissection_depth().

    The 16-bit flow identifier of the flow state header was added with a
    length of 1.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit fc23753166e28434b728957dce391d8e755226a8)

commit 5f6965635e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:17 2026 +1000

    RBM: Decode Ruby Marshal integers like Ruby does

    get_rbm_integer() left its value and length outputs unset for the
    type byte 0xfa (-6), so an Array or Hash used an uninitialised element
    count and the offset moved by an uninitialised length. This is reached
    by default when a Ruby Marshal file is opened, and in DRb.

    The function also differed from r_long() in Ruby's marshal.c for the
    type bytes 4 (a 4-byte positive integer, decoded as -1) and -5 (0,
    decoded as a 5-byte integer with an undefined shift). Decode all type
    bytes as Ruby does, so every path sets both outputs, and build the
    multi-byte values in an unsigned integer.

    dissect_rbm_inline() left its type and value outputs unset when the
    header version was not supported, and DRb then used them.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e5381a5c60d6e665fdcd8d717c3b7405bc38be48)

commit 169eee4c46
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:17 2026 +1000

    Link 16: Dissect a word without state instead of a dissector bug

    The Link 16 dissector is registered by name, so it can be called
    without its message state, for example from an Exported PDU file.
    It then reported a dissector bug. Use a local state in that case and
    dissect the word on its own.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6402356896918c8fe520d4c76ee382df8225ba5a)

commit dc76ae8b31
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:17 2026 +1000

    x509ce: Create the CI+ extension handles only once

    x509ce_enable_ciplus() created four new dissector handles each time it
    was called, which DVB-CI does for every certificate, so memory grew
    with every certificate dissected. Create the handles once in the
    handoff function and reuse them.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 82a5123b447dd51118fd7a834100f2d320a8ee36)

commit d8d123083f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 2 11:26:16 2026 +1000

    ETSI DCP: Zero the corrected data, reject an RS k of 0

    The Reed-Solomon corrected data was not initialized: when correction
    failed the later blocks were not written, and the octets after the
    last block were never written at all, but all of them were dissected
    as the AF packet and shown as a data source. Allocate it zeroed.

    With an RS k of 0 the Reed-Solomon correction decoded the same block
    again and again without moving its output, so the reassembled AF
    packet was garbage. Reject it as the comment says.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e808c1047bf49582a652cfab30e693601a225654)

commit c1a38c212f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 2 11:46:49 2026 +1000

    DeviceNet: Do not assert without CAN information

    The dissector asserted that it was given the CAN information, but an
    Exported PDU that names it calls it without any, so such a packet was
    reported as a dissector bug and tshark aborted with
    WIRESHARK_ABORT_ON_DISSECTOR_BUG set. Reject it as CANopen does.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 651eda8914bd4aad2cdd4687756bd6050f84ee2a)

commit 18a2ecd24f
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 2 11:26:16 2026 +1000

    DECT-MITEL-RFP: Fix the DTMF key encoding

    The DTMF key, an FT_CHAR field, was added with ENC_ASCII_7BITS, a
    string encoding whose bits read as two variable-length integer
    encodings at once, so every MEDIA-DTMF message failed an assertion in
    tvb_get_varint() when a tree was built. Add it with ENC_NA.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 852c999869b97e0d99d1c922cc5f496a4981ceb3)

commit cf0ae2732e
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 2 11:26:16 2026 +1000

    DCOM: Limit nested activation properties

    The ActivationPropertiesIn and ActivationPropertiesOut marshalers are
    also the dissectors of activation properties with those CLSIDs, so an
    activation blob whose property is another ActivationPropertiesIn
    recursed on the same octets until the stack overflowed and tshark
    crashed when it dissected without a tree. Count each activation
    property in the dissection depth.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit b40355229e5c32b28ebe1ae9d9ff1fc2f8ff781d)

commit 55e7b106e8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 2 10:24:00 2026 +1000

    pidl: Skip NUL-terminated strings in the conformant run

    The Wireshark backend dissects a NUL-terminated string element with
    dissect_null_term_wstring() or dissect_null_term_string(), which do not
    know about the conformant run. As the target of a pointer the string
    was dissected, and its octets consumed, in the conformant run as well,
    which hit the DISSECTOR_ASSERT on the octets a conformant run may eat:
    for example any SVCCTL EnumDependentServicesW response or
    GetServiceDisplayNameA request with a string. Skip these strings in the
    conformant run, and regenerate the dissectors that have them (EVENTLOG,
    FRSRPC, MAPI, SVCCTL and WITNESS).

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit b59d73005a0087dd9dfff8cd0e6dc5ec8bdd56ae)

commit 48d51756e5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 04:43:42 2026 +1000

    CoSine: Reject a fraction of a second that is not in hundredths

    The fraction of a second in the header line of a control blade trace is
    in hundredths. It was scanned with %9d and multiplied by 10000000 in an
    int, which overflows (undefined behavior) from 215 on, and gave wrong
    or negative nanoseconds for anything above 99. Report a damaged file
    instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 73a137881bfbd96910a312640c2e6311570993a7)

commit 45cb6e8af5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 03:11:37 2026 +1000

    D-Bus: Require a variant signature to be one complete type

    A variant with an empty signature closed its own subtree but left the
    type reader thinking the variant was still open, so the enclosing array
    or variant was never closed and the reader went on to the terminating
    NUL of the signature, which hit DISSECTOR_ASSERT_NOT_REACHED(). A
    signature of more than one complete type, such as "ii", was accepted and
    misparsed too. Report both as an invalid variant signature.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 22fbc8fbf4c6c080837c8cffe48f96f6bcb6d49d)

commit 499476a8af
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:43:57 2026 +1000

    CLNP: Stop the nesting of error reports before the layer limit

    The discarded PDU of an error report is dissected as CLNP again, so
    error reports nested in each other recursed until the layer limit
    asserted, or with a tree the tree depth limit, as each level adds two
    tree levels. Stop at a quarter of the tree depth with an expert info.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit da17f8137b726a375def4f64710f0960daee1b3d)

commit 2478d54229
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:43:56 2026 +1000

    Cisco ttag: Stop the nesting before the layer limit

    The payload of a ttag can be a ttag again, which nested until the
    layer limit asserted. Stop before the limit with an expert info.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 5542a67cf95e6f76d109040056161eff09c825b8)

commit a0295f92dc
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:43:56 2026 +1000

    Cisco MetaData: Stop the nesting before the layer limit

    The payload of Cisco MetaData over Ethernet or GRE can be Cisco
    MetaData again, which nested eight octets per level until the layer
    limit asserted. Stop before the limit with an expert info.

    Also give the Cisco MetaData over GRE item the length of its eight
    octet header.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 90ac3b8d4ae56604a09da687dad4d0411eadf3eb)

commit 4a0d17fb1b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:43:56 2026 +1000

    CBOR: Stop the recursion before the tree depth limit

    The recursion limit for nested arrays, maps, tags and indefinite
    length strings was the maximum tree depth itself, but each nesting
    level adds one or two tree levels below those of the layers above, so
    deeply nested items reached the tree depth limit first, which is
    reported as a dissector bug. Stop at a quarter of it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 253ffb576a5ffd0ba7a5bcf28fa591ec5ac2ed1c)

commit 586067ad07
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:37:21 2026 +1000

    CDMA2K: Do not add fields of more than 64 or of 0 bits, fix offsets

    The padding at the end of an addressing or authentication header
    record, a TMSI zone with an address length over 12 and a TMSI code or
    BCMC flow discriminator of length 0 were added as bit fields of more
    than 64 or of 0 bits, which is a dissector bug. Skip padding of more
    than 64 bits and check the lengths.

    The bit offsets were kept in 16 bits, so in a message past 8 KiB they
    wrapped back to the start of the packet and the message loop started
    over until the tree item limit stopped it. Keep them in 32 bits.

    The supplemental channel loops of an active set record used the
    counter of the enclosing pilot loop, and the channel count started at
    65535 when it was not in the message, so they did not end. Give them
    their own counter and start the count at 0.

    Also read the rejected order type, the RSC mode indication and the
    fast call setup ORDQ at their own offsets, and widen the TMSI fields
    to the bits added.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 112396cd011eb80e44d15652e42684d4ece7277e)

commit 001dddad90
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 10:33:00 2026 +1000

    BT SDP: Do not search the whole record for every attribute

    For a DID or HDP service record, dissect_sdp_service_attribute()
    walked all attributes of the record with findUintAttribute() for every
    attribute it dissected, so a single attribute list was quadratic in the
    number of attributes: a 64 KB Service Attribute Response took about six
    seconds per pass. Search only for the attributes that use the result:
    the DID Vendor ID and Product ID and the HDP Supported Features.

    The continuation state was saved as a hex string, and only as many
    characters as the state has octets were copied into the reassembly
    key, so states differing in their second half were taken for the same
    state. Save the octets instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 065e4aaeadd0d7fb6777f5f3319fb9f99aef96ed)

commit 63ebe9c33c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 05:43:27 2026 +1000

    BSSGP: Reset the SI item parameters for each item, fix several IE lengths

    The RR message function and subtree of a NACC SI item were set only
    for a known message type, and kept from the previous item otherwise;
    an unknown type in the first item added a subtree with ett -1, which
    failed DISSECTOR_ASSERT() (NS over UDP, with a tree), and in a later
    item decoded it with the previous item's function. Reset them for each
    item and add the subtree only for a known type.

    Channel Needed and Velocity Data used the length returned by the RR
    and GSM A helpers as an offset, so the IEs after them were dropped.
    FLUSH-LL and FLUSH-LL-ACK decoded the new NSEI as an RR IE, and
    SGSN-INVOKE-TRACE the OMC Id as a common IE, which stopped the
    dissection. Transaction Id, MBMS Session Duration and MBMS Service
    Area Identity List did not step over their value.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 87e094fcc5939ded8ec6986e3534385a03c42dc2)

commit 9596475486
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 05:37:29 2026 +1000

    BMC: Do not dissect the Data Coding Scheme twice, fix the schedule loop

    BMC decoded the Data Coding Scheme and then passed the octets after it
    to the Cell Broadcast dissector, which decoded the first octet of the
    CB data as the Data Coding Scheme again. Pass the Data Coding Scheme
    to it instead.

    With a CBS schedule period of 255 the 8-bit loop counter never ended
    the loop, so every such schedule message was malformed.

    Cell Broadcast: dissect_cbs_data() returns NULL for a reserved coding
    group or character set, and tvb_reported_length() then failed its
    assertion. Stop after the expert info instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e89b54694f4259acba6824c994ef21eacbb33554)

commit fc1c8e6850
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 10:24:35 2026 +1000

    BIER: Stop nested MPLS payloads before the layer limit

    MPLS hands a payload whose first nibble is 5 to BIER, and BIER hands
    Next Protocol 1 and 2 back to MPLS. A frame of MPLS labels and BIER
    headers repeated a few thousand times (plain Ethernet, ethertype
    0x8847) reached the layer limit in call_dissector_work() and failed
    its DISSECTOR_ASSERT(). Stop with an expert info when the frame is
    close to the limit.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 6580ca526f7422fca1a7265cfc5a389fb7cb6db9)

commit a83ae43963
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 10:18:49 2026 +1000

    Bencode: Restore the dissection depth at the end of a list

    A list that ended with its "e" returned without the
    decrement_dissection_depth() that matches the one before the loop, so
    each complete list left the depth one higher. 600 empty lists inside a
    list (from bzr on TCP 4155, BitTorrent or rtpproxy) reached the
    maximum depth and failed the DISSECTOR_ASSERT() in
    increment_dissection_depth().

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 0657b93577843b8956f61410a7b0f647e3ba512b)

commit f4807cdcf5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Oct 28 12:24:24 2026 +1000

    LI5G: Do not assert on attribute type 0

    Attribute type 0 has no field, so a conditional attribute of type 0
    was added with field index 0 and failed the "Unregistered hf"
    assertion. Skip it like an unknown type.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 20e6bc5e2008e2cb7d9da8583a8b52cfea9b111f)

commit 5bb2a572be
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Oct 25 23:17:43 2026 +1000

    AIM: Terminate the chat message before giving up on a long one

    aim_get_message() returned before clearing the buffer when a message was
    longer than 999 bytes, and its callers then printed the uninitialized
    buffer into the Info column.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 985d5a80b487cd4c24706aab32f2ba4699188fe4)

commit b6d2f85d79
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Oct 27 17:37:41 2026 +1000

    ANSI IS-637: Do not assert on User Data longer than its text

    The reserved bit count in User Data was never bounded. A subparameter
    longer than its text gave a count over 64 bits, which hit the
    no_of_bits <= 64 assertion in tvb_get_bits64() for GSM 7-bit text and
    reported a dissector bug for the other encodings. A text longer than
    its subparameter wrapped the unsigned count instead and read the
    following subparameters as text. Keep the count signed and report Short
    Data or Extraneous Data instead, in User Data, Multiple Encoding User
    Data and Service Category Program Data.

    Also fix some decoding errors:
    - A Multiple Encoding or Service Category Program Data text that starts
      on an octet boundary was read from one octet too late.
    - The teleservice and transport message lengths were truncated to 8
      bits, so messages over 255 octets were cut short.
    - The Broadcast flag of the last transport message was kept, so a
      teleservice message reached another way decoded its User Data as
      CMAS depending on which packet was dissected before.
    - A CMAS type 0 record of length 0 threw an exception instead of
      reporting Short Data.
    - The reserved bits at the end of CMAS User Data were shown one octet
      past it, so a CMAS message that ends with User Data threw an
      exception.

    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    Assisted-by: Claude
    (cherry picked from commit 0d2f9272b74314baed46af2cc10e53c90931d08a)

commit dbcaa13998
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Oct 26 16:50:41 2026 +1000

    TPM 2.0: Do not assert on a response without a command

    A response is matched to the last command seen. A response with no
    command before it, for example in a capture started mid-way or in
    other traffic on port 2321, fails a DISSECTOR_ASSERT(). Add a note
    instead and stop after the response header, and forget the last
    command when a new file is opened.

    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    Assisted-by: Claude
    (cherry picked from commit 6a8bd04a449ef76b829e2677baea5907153dbc58)

commit 6633b68372
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Oct 26 09:17:28 2026 +1000

    AMQP: Fix a field type assertion and a type confusion in AMQP 1.0

    An AMQP 1.0 list item is dissected by the type its encoding octet
    names, into the field of that list position. A boolean (0x41, 0x42)
    or a timestamp (0x83) sent for a field of another type, such as the
    maxFrameSize of an open, fails the field type assertion of
    proto_tree_add_boolean() or proto_tree_add_time(). Add an expert info
    instead, as dissect_amqp_1_0_zero() does for an integer sent for a
    field that is not one.

    A Message Decoding entry passes its "Additional Data" string as the
    data argument of the dissector it names. Only protobuf takes a string
    there (the message type); json and data-text-lines, for example, read
    it as a media_content_info_t, past the end of the string, and crash
    tshark on the first message that matches the entry. Pass it to
    protobuf only.

    While here, fix small errors found with synthetic captures:
    - Keep the AMQP 0-9 delivery tracking in file scope. It was kept per
      packet, so a second pass lost the Publish-Number and "Message in"
      items, and "Ack in" was never shown.
    - Recognise the 0-9-1 ("AMQP" 0 0 9 1) and 0-8 ("AMQP" 1 1 8 0)
      protocol headers, which use the 0-9 framing.
    - Acknowledge every outstanding message for a multiple ack of
      delivery tag 0.
    - Show the elements of 0-9 field arrays, which were added to a NULL
      item.
    - Name Exchange.Unbind-Ok by its method id, 51.
    - Read 0-10 tx commands from the start of the command, not one octet
      in, so that they are not all shown as tx.select.
    - Step over the size of the 0-10 message.subscribe arguments map
      before its entry count.
    - Give the 0-10 types 0xd0-0xdf and 0xf0-0xff their sizes of 9 and 0
      octets in maps, not 5.
    - Show 1.0 message and correlation ids sent as ulong in 64-bit fields.
    - Add the "unknown list type" expert info of 1.0 lists to a tree item
      that exists.

    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    Assisted-by: Claude
    (cherry picked from commit acd5cc67e74cec1b2e391cbbfb5c41a723a9429e)

commit df555edac0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Oct 26 09:02:52 2026 +1000

    AMP: Stop at a TNVC item larger than the packet

    cbor_info() gives an array a total size of UINT64_MAX, and a byte or
    text string with a 4 or 8 octet length or a semantic tag can give one
    above INT32_MAX. A Report Set whose TNVC data items hold any of these
    fails the DISSECTOR_ASSERT() on the item size. Report the item as
    malformed CBOR and stop, as the other size checks of the function do.

    While here, fix small errors found with a synthetic capture: step over
    the reports array head by its own size, not that of the Rx names array
    head; do not count down from a TNVC count below 2 into a wrapped loop
    bound; do not add a template item of length -1 for an empty non-literal
    template; and show Table Set messages, which were added to a subtree
    that is never created, and name their opcode.

    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    Assisted-by: Claude
    (cherry picked from commit f725109284c909a8825756782d18a4eefed719b6)

commit 27b1eb294b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Oct 24 11:34:42 2026 +1000

    6LoWPAN: Don't decompress a first fragment shorter than an IPv6 header

    When the datagram size of a FRAG1 or RFRAG first fragment was less
    than the 40-byte IPv6 header, HC1 and IPHC computed the reported
    payload length as dgram_size - 40 in an unsigned, which wrapped, and
    built a reassembly tvb with a reported length smaller than its
    captured length. The dissector already added an expert info for
    this; stop there instead of going on.

    Assisted-by: Claude
    (cherry picked from commit da498fdb1ed135efaa882603ae452460c2f663a6)

commit 49a7aec305
Author: Anders Broman <a.broman58@gmail.com>
Date:   Tue Dec 30 22:16:43 2025 +0100

    expert: Add ...remaining functions highligting to the end of tvb.

    proto_tree_add_expert_remaining()
    proto_tree_add_expert_format_remaining()

    (cherry picked from commit 0ac5c9b5f1f214c303191619728da95b23f9855c)

commit c98850a6b5
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Nov 8 10:18:56 2026 +0000

    [Automatic update for 2026-11-08]

    Update manuf, services, enterprise numbers, translations, and other items.

    PCI IDs failed.

commit e0a0ee0abd
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Nov 7 12:07:05 2026 -0500

    Packaging: MSYS2: Install the zlib 1.4 DLL

    Upstream zlib 1.4 changed the DLL names on Windows again.

    MSYS2's zlib 1.4 package includes both the old zlib1.dll (which existing
    not yet rebuild packages need) as well as a new libz-1.dll, but links -lz
    to libz-1.dll instead of zlib1.dll, so Wireshark built with it needs
    libz-1.dll. TShark in the NSIS installer from the GitHub MSYS2 workflow
    failed to start on all branches without it. Install it.

    Keep zlib1.dll, and which GLib, Qt, libxml2, GnuTLS, minizip-ng, and
    other libraries built with earlier zlib versions use.

    https://github.com/madler/zlib/releases/tag/v1.4

    We will also have to see what vcpkg does, and also remember this when
    rebuilding our libssh package.

    (backported from commit a233ff958d5fa3eef9f4ffcdff82c1b712b3e05f)

commit 496192c15a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 17:39:50 2026 +1000

    GSM A BSSMAP: Bound Field Element recursion and fix small errors

    The Cell Load Information Group field element (0x0b) is dissected by
    be_fe_cell_load_info_group(), which calls be_field_element_dissect(),
    which can dispatch to the same element again. The length of a field
    element was not checked against the enclosing element, so a packet
    could nest it without limit and overflow the stack, or hit the maximum
    tree depth. Stop with an expert info when the nesting gets too deep,
    stop when the cell id used up the whole element, and clamp the length
    of each field element to what is left of the enclosing element.

    be_res_avail() read a 32 bit value into a 16 bit variable, used the same
    value for the full rate and half rate counts, and set the item length to
    the length of the whole element. Read each 2 octet count separately,
    make the fields FT_UINT16 and drop the proto_item_set_len() calls.

    be_speech_codec_lst() and be_speech_codec() set the length of the last
    item only once after the loop. Set it for each item in the loop.

    [pinfo->dissection_depth is an int on this branch, so the comparison
    with prefs.gui_max_tree_depth casts the latter to int, as packet.c
    does, to avoid a -Wsign-compare warning. The expert info table
    conflicted only because this branch lacks the unrelated spacing change
    to the extraneous data message.]

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit af57259207f52ca57ce85d2db63bdfc9074b7044)

commit 78497bf4e0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 17:39:50 2026 +1000

    GPRSCDR: Reset the extension OID for every ManagementExtension

    The file-scope obj_id points to a string allocated from the packet pool
    when the identifier is dissected. It was never reset, so a
    ManagementExtension without an identifier used the string left by an
    earlier extension or packet after it had been freed, and passed it to
    call_ber_oid_callback().

    Set obj_id to NULL at the start of each ManagementExtension.

    Also make MSTimeZone read the second octet from parameter_tvb instead
    of the enclosing tvb, and require at least two octets.

    [packet-gprscdr.c was regenerated from the .cnf; the generated
    functions on this branch take an int offset.]

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 8a7bf25f3325bf60ccca49b9f66cffe43f80b955)

commit ffd3bb20b5
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 2 10:03:47 2026 +1000

    DCOM: Limit nested OBJREFs, register custom marshalers for every packet

    A custom OBJREF is handed to the marshaler of its interface, and some
    marshalers, such as the one for IStorage, dissect an OBJREF again, so a
    request with nested custom OBJREFs recursed until the stack overflowed
    and tshark crashed. Count each custom OBJREF in the dissection depth.

    The custom marshalers of ISystemActivator and WMIO were registered only
    when an activation or IWbemServices call was dissected, so an OBJREF in
    an earlier frame was not decoded on the first pass but was on the
    second. Register them when a file is opened. A handler OBJREF stored the
    IPID of its standard OBJREF in the IID.

    A VARIANT of type VT_CY decoded its value with field index 0 and failed
    an assertion on any such VARIANT. Decode it without a field.

    [This branch passes the VT_CY value pointer without the uint64_t cast
    that master has; the fix's version of the call is used.]

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 44343b79b0741ee76e2a3e9e045e9e6ed38b66a0)

commit 97ff97ec4c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 17:39:51 2026 +1000

    GSM MAP: Do not treat BER OID callback data as tcap_private_t

    dissect_gsm_map() casts its data pointer to struct tcap_private_t.
    The handle was also registered as a BER OID dissector, and ACSE and
    PRES call those with their own private data (a SESSION_DATA_STRUCTURE).
    dissect_gsm_map_GSMMAPPDU() then read acv and oid through the wrong
    type and ran strrchr() on a garbage pointer, crashing on a crafted
    SES/PRES stream.

    Register the BER OIDs through a wrapper handle that passes NULL data.
    The TCAP and SCCP registrations are unchanged.

    The Nokia extension fields traceReference and traceType, which use types
    imported from MAP-OM-DataTypes, were registered as FT_NONE, so a Nokia
    private extension carrying them hit the DISSECTOR_ASSERT_NOT_REACHED in
    dissect_ber_integer(). Give the two types their field types in the .cnf.

    [The 4.4 branch does not register resetContext-v3, so that OID
    registration is not part of this backport. packet-gsm_map.c was
    regenerated from the template and .cnf.]

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 8bb4163b37038a8ca98181cbc874a97f3bfef8f2)

commit e162dc501d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Nov 5 08:49:37 2026 +1000

    ForCES: Fix cubic time in nested LFB select TLV loops

    The LFBselect TLV loop called dissect_operation_tlv(), which itself
    looped over Operation TLVs up to the end of the tvbuff, and each of
    those called dissect_path_data_tlv(), which again looped to the end of
    the tvbuff. A packet of a few KB of small TLVs therefore took O(N^3)
    time (6 KB took about 9 seconds).

    Make dissect_operation_tlv() handle exactly one Operation TLV, with a
    length check, and give dissect_path_data_tlv() a subset tvbuff
    limited to that TLV, so each level only walks its own TLV. The tree
    for valid packets is unchanged. Stop the LFBselect loop on a TLV
    length below the TLV header size.

    [The Operation TLV type is read with tvb_get_ntohs() on this branch
    rather than with proto_tree_add_item_ret_uint(); the fix's version of
    dissect_operation_tlv() is used as is.]

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 1e2fb3338a7832d8bcdb77b83a2e20dd294d715a)

commit db84bb01fd
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Nov 5 23:20:57 2026 -0500

    WINSPOOL: Check the security descriptor length

    The length of a sec_desc_buf is from the packet and was added to the
    offset unchecked, so a length near 2^32 moved the offset back and the
    array around it was decoded again from earlier octets. Check that the
    descriptor is in the data.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 3af32a1be62754fbf0bd7289f76d3eba2de44e42)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 48c5038d3a
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Nov 5 23:19:07 2026 -0500

    SRVSVC: Do not loop on a security descriptor length that wraps

    The length of a sec_desc_buf is from the packet and was added to the
    offset unchecked. In an array of them, such as NetShareInfo level 1501,
    a length of 0xfffffffc moved the offset by nothing, so an array of 2^32
    entries dissected the same octets until tshark was killed. Check that
    the descriptor is in the data.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit b5f7e7a0b6055676bff52dbd468600a51f9fafb0)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 93498e1eb9
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Nov 5 19:59:11 2026 -0500

    DNS: Clear all of the client subnet address (union initialization)

    The address of an EDNS Client Subnet option is copied into a union of a
    uint32_t and 16 octets, initialized with {0}. Strictly according to the
    C standard since C99 (and explictly since C23) that initializes only the
    first member. GCC 15 and later and Clang 18 and later* no longer clear the
    rest of a union with automatic storage, so the last 12 octets are
    uninitialized on some compilers, including the latest GCC. (There's
    a -fzero-init-padding-bits=unions flag but we don't seem to have
    anywhere else in the code that does this with a union specifically
    so it's probably best to fix it.)

    C23 has {} which we might use eventually.

    This was exposed by the test in e139bcf392431805e13a1f0f5bbf25b0c30cca89
    and also running the test on very recent GCC.

    * - Technically, the LLVM first step treated the rest of the bytes as
    uninitialized, but the backend would zero out the rest.

    Assisted-by: Claude Opus 5.5 helped run this down

    (cherry picked from commit 958b219effa4ae9f788e87807e85dfa1c36a4264)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit f213fe6b01
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 16:52:54 2026 +1000

    ESS: Reset the security category OID for each category

    object_identifier_id is a file-scope pointer that is set to a string
    in pinfo->pool by the SecurityCategory type and the tagName fields,
    and was never reset. A SecurityCategory without its [0] type still
    reaches the [1] value, which then looked up the OID string of an
    earlier packet after that packet's pool had been freed, a heap use
    after free. The ess_dissect_attribute() and
    ess_dissect_attribute_flags() helpers compared the same stale pointer,
    or NULL, with strcmp() when a tag set had no tagName and the "ESS
    Category Attributes" table had entries.

    Reset the pointer to NULL at the start of SecurityCategory,
    RestrictiveTag, EnumeratedTag, PermissiveTag and InformativeTag, and
    skip the attribute name lookup when no OID has been seen.
    call_ber_oid_callback() already handles a NULL OID.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit c36cf5cf95d94b709996a928e8f29aa847a32867)

commit 4d446ec04d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Nov 4 17:29:09 2026 +1000

    Dissectors: Handle a missing or unexpected data argument

    A crafted Exported PDU file can name any registered dissector
    (EXP_PDU_TAG_DISSECTOR_NAME) or heuristic dissector
    (EXP_PDU_TAG_HEUR_DISSECTOR_NAME), and Exported PDU then calls it with
    the data argument NULL (or, for a handle that also has a TCP heuristic,
    with a struct tcpinfo). Many dissectors assume that data is always the
    structure their usual parent passes. Calling every registered name and
    every heuristic this way showed 36 that dereference a NULL pointer, two
    that abort in ws_assert_not_reached() (XCP over TCP and UDP) and about
    100 that fail a DISSECTOR_ASSERT, all with default settings and only by
    opening a file.

    Return early (0 for a dissector, false for a heuristic) when the data a
    dissector needs is missing, instead of dereferencing it or asserting,
    and handle the few values derived from it that were asserted on (an
    unexpected port type in XCP, a missing conversation in H.223, iPerf3,
    ZMTP, MS Proxy and SOCKS, a short IE length in PFCP, an unexpected
    address type for the default TCP MSS). WebSocket uses its data as HTTP
    upgrade information only when HTTP, HTTP/2 or HTTP/3 is the caller.
    The behaviour with the data the usual parent passes is unchanged.

    [The 4.4 branch does not have the Vertexcom HomePlug AV vendor dissector,
    the mac80211_hwsim and Open vSwitch netlink dissectors, XCP, the IEEE
    802.15.4 dissector that takes the FCS type as data, or the TCP default
    MSS, so those changes are not part of this backport. PEAP keys the EAP
    tvb with tls_group.]

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 1e2c175c21c8ced022763ae7e8c4f2a133913fac)

commit 0aede29dff
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 16:52:53 2026 +1000

    EBHSCR: Fix FlexRay tvbs backed by stack arrays and a LIN length

    The FlexRay frame and symbol packets built the 2-byte FlexRay header
    in an array on the stack and made a tvb of it that outlived the
    function, so fields and filters such as flexray[0:2] read stack memory
    after it was gone. The symbol packet tvb was also never freed. Take
    the header from pinfo->pool and make the symbol tvb a child of the
    packet tvb, so it is freed with it.

    A LIN payload of 1 byte made the data length wrap to 0xFFFFFFFF;
    report it as too short with a new expert info instead. The NMEA string
    is now limited to the captured payload, not the header length field.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit ebe8aa9b89d9f86ebcca362868a855140e7709e7)

commit 7a58ef9609
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Nov 4 21:07:38 2026 -0500

    epan: Check for label overflow on FT_DOUBLE both before and after units

    80f16015e2af296fb5ab8f676979bfed9731171f introduced new float display
    types. Unfortunately I expect that some people may reach for BASE_DEC as
    the "standard" choice, but really for most purposes BASE_NONE is better.
    In particular, BASE_DEC can overflow the standard label width.

    The ability to append units was added to FT_FLOAT and FT_DOUBLE, but
    as sprintf and g_strlcpy return the size of the attempted result,
    truncation must be checked both before and after adding units.

    Fixup 764982a7d97ac47f22bbc6e214a398a60b05feb5

    Fix #21726

    Thanks to Khaled Suliman of AISLE Research for the report and POC.

    (backported from commit 6ffc6fdd4738d1814a87f07dd356a37ce2aef21a)

    Assisted-by: Claude

commit d2f322d11f
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Nov 5 08:46:24 2026 -0500

    SCTP: Don't re-use fragmented DATA chunks in another reassembly

    If a DATA chunks are transmitted in such a way that a completed
    series would appear be enveloped in other, i.e., two DATA chunks
    with B/E bits 10 with no end chunk with B/E bits 01 between them,
    and vice versa, that is strictly illegal according to RFC 9260
    section 6.9:

    https://www.rfc-editor.org/info/rfc9260/#section-6.9-8.3.1

    In particular, if the chunks are received out-of-order with respect
    to TSN, this can cause a fragment's data to already be used for the
    apparent inner message and freed before the outer message is completed.
    This could lead to the creating a TVB with uninitialized heap data,
    because the fragments with NULL data are not copied but their length
    is used in the total length.

    Since add_fragment() does not create a fragment with length 0, returning
    early, the data element is always non-NULL unless the fragment has
    already been used in another reassembly. In such a case, do not re-use
    the fragment to complete a second reassembly, but leave the other
    fragments un-reassembled.

    If SCTP were able to use the standard fragment reassembly API, much of
    this would be handled and with additional expert info and warning
    fields.

    Fix #21712

    Thank to Khaled Suliman for the report and PoC.

    (cherry picked from commit 6a560feb097ca6c07ddcf97049ad3b4de7e4a5f5)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit a6e9f8bf6f
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Nov 5 08:43:53 2026 -0500

    TEAP: Bound the recursion of nested TLVs

    Request-Action and Intermediate-Result TLVs, and the TLVs after an
    EAP-Payload, carry more TLVs, and dissect_teap() was called for them
    without counting the dissection depth, so nested TLVs recursed without
    a limit and could exhaust the stack. Count the depth around the calls,
    as the NAK TLV does.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 975972014e8d989f0bb6fdff2f94deb7bf4b8238)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit efbce2e76b
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:16 2026 +1000

    DNP3: Fix a loop on empty octet strings and two-pass reassembly

    An octet string object (g110/g111) with variation 0 and no prefix
    takes no space, so a 32-bit item count made the item loop run up to
    2^31 times without advancing; this took minutes of CPU per packet
    with the default ports. Reject such an object with an expert info.
    Also stop the item loop, with an expert info, when it does not
    advance for more items than a bit-packed object needs, and stop it
    after an unknown object has taken the rest of the message.

    The application layer reassembly was keyed by the IP addresses and
    the extended transport sequence number only, so messages from
    different DNP3 link addresses behind the same IP address could be
    mixed, and the per-message data was keyed by the 6-bit transport
    sequence number. Two messages in one frame with the same sequence
    number got the same reassembled message on later passes. Add the
    link addresses to the reassembly key and offset the reassembly id by
    them, as the table of reassembled messages is keyed by frame and id
    only, and key the per-message data by the index of the message in
    the frame.

    The size of the variable part of the file objects g70v4, g70v5 and
    g70v6 wrapped when the object size was smaller than the fixed fields,
    and the offset then moved back. Add an expert info and take no data.

    The 32-bit analog command events g43v1 and g43v3 did not step past
    the 32-bit value, so the item was 4 octets short and the g43v3
    timestamp was read over the value. The g70v3 file command read the
    permissions at the start of the object instead of at their own
    position. The endpoint tap built the destination endpoint from the
    link source address, so every source was shown as both transmitter
    and receiver and destinations never appeared. Signed device
    attribute values were shown unsigned in the item text (-2 as 254);
    they are now shown signed.

    [The 4.4 branch has neither the extended sequence number reassembly
    nor the DNP3 conversation and endpoint taps, so the reassembly key and
    endpoint tap changes are not part of this backport.]

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 3136e8b0280668a30cd078c6c383d0a1269fb9e4)

commit 7b581ec2f9
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:21:29 2026 +1000

    DOF: Fix TCP framing loop, SID learning cost and length handling

    The TCP and tunnel framing kept the start of each PDU in a 16-bit
    field, so a PDU that started 64 KiB or more into a frame was never
    found and tshark looped forever, allocating a new entry each time.
    Keep the offset in an int.

    dof.tunnel needs the packet reference of the TCP framing, so an
    Exported PDU that names it reached DISSECTOR_ASSERT(). Reject a call
    without one.

    Learning a SID walked every DOF packet seen so far, so a capture in
    which each packet carries a new SID took quadratic time. Index the
    packets by SID ID and only update those.

    A DNPv1 length shorter than the address fields wrapped around and the
    payload was taken to the end of the packet; it now gets an expert
    info. OAP Advertise passed ENC_BIG_ENDIAN as the offset of its command
    control. An OAP alias resolved to a binding from a later frame on
    later passes only. A SID longer than 255 octets was truncated when it
    was learned, and the Object ID text was built from the reported length
    of a captured buffer.

    In the security modes, lengths of domains, identities, groups and
    blocks were kept in 8 bits and truncated; a range that went backwards
    wrapped into a huge allocation. The CCM key table compared new keys
    with the cipher handle instead of the key, decryption always used the
    session cipher instead of the one for the key period, and ciphertext
    shorter than the MAC was not rejected. validate_session_key() leaked
    its HMAC handle. A preference OID nested too deep asserted, the OAP
    alias table was leaked on every file load, and an Object ID longer
    than 65535 octets was truncated.

    When a sender's provisional SID ID was merged into the SID ID of a
    known SID, the packets that had stored the provisional ID were
    updated, but the sender key still mapped to the provisional ID, so a
    later packet from that sender went back to it, without the SID. The
    sender keys are now updated as well.

    An OAP value list at the end of the PDU (a response without values, a
    Set or Invoke without parameters) threw a bounds error. An empty value
    list is now left out of the tree.

    The expert infos for reserved DNP and DPPv2 flag bits were added with
    a NULL item and did not show in the tree. They are now added to the
    flags item.

    An interface ID with length code 0 is 16 octets long in the tree, but
    its text showed no data ("[{00}:{}]"). The text now uses the same
    lengths as the tree, bounded by the stored buffer.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 1c1a917e0762a6f316ce8e6ea3095b7c05043476)

commit 8b1efebc83
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Nov 2 10:03:47 2026 +1000

    SPOOLSS: Do not loop on key buffers or job levels, convert each string once

    The key buffer of EnumPrinterKey responses and GetCorePrinterDrivers
    requests was dissected while the offset was before an end taken from
    the packet, and a string at the end of the data returned the offset
    unchanged, so tshark looped forever. Check the size and stop when a
    string does not advance.

    EnumForms, EnumJobs and EnumPrinterDrivers responses dissected their
    entries in the buffer even when there was none, for a NULL pointer or a
    size that is too big, and hit the assertion that a tvb is given. An
    EnumJobs response with an unknown level added an expert info for every
    job without advancing, so a count of 2^32 took minutes and gigabytes.

    Each string in a buffer converted all the rest of the buffer to find
    its end, so many relative strings took quadratic time and memory: a
    64 KB EnumPrinterDrivers response took 17 seconds and 5 GB. Find the
    terminating NUL, and convert only the string; this also gives the right
    length for text that is not ASCII. Value sizes are checked before they
    are added to the offset, the printer data values stay within their
    size, and a REG_DWORD of less than 4 octets is no longer read from
    before it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 594a507563f58bd1ae37f9cf4473ef8162c07ac4)

commit f5b7cf610d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:17 2026 +1000

    DOP: Reset the binding type for each PDU

    The binding type OID was kept in a file-scope pointer into
    pinfo->pool and never reset, so a later PDU without a binding type
    used a string freed with an earlier packet when it looked up the
    binding parameter dissector and wrote the Info column. Reset it at the
    start of each DOP PDU, and do not resolve a missing OID.

    The agreement fields call the binding parameter callback without an
    Info column text, which passed NULL to a "%s" format. Append the text
    only when there is one.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 53a2b9b2d721bad9e8c80929f8cc43ae2343c5ff)

commit 4470604636
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Tue Nov 3 08:50:16 2026 +1000

    DOCSIS: Fix MAC management TLV loop hang and fragment checks

    DCC-REQ, DCC-RSP, DCC-ACK, DCD, CM-STATUS, CM-CTRL-REQ and OCD walked
    their TLVs with a uint16_t offset. A reassembled multipart message, or
    a large frame, of more than 64 KiB made the offset wrap, so the loop
    never ended and kept adding tree items until memory ran out. Use an
    unsigned offset in these loops and in the RNG-RSP commanded power and
    OCD subcarrier assignment helpers.

    DBC-REQ and REG-RSP-MP used fragment sequence number 0 as sequence -1
    in the reassembly, and accepted numbers above the fragment count.
    Report an expert info and skip reassembly for such fragments.

    A message length below 6 made the payload length underflow. Report an
    expert info and stop after the header.

    Smaller fixes:
    - UCD OFDMA IR power control: add the two 1-byte fields with length 1
      instead of the TLV length;
    - DPD subcarrier assignment vector: the length includes the 2-byte
      start field, so do not read two bytes past the vector;
    - DPD and OCD subcarrier assignment lists: the length includes the
      type byte, so do not read an index past the list;
    - MDD diplexer band edge: continue with the next sub-TLV after a bad
      length instead of stopping, and take the override value with the
      sub-TLV length instead of always reading two bytes;
    - CWT: give the "Upstream Subcarrier Indices" subtree the sub-TLV
      length instead of the parent length;
    - CWT-RSP: show the status, not the op code, as the status in Info;
    - OPT-REQ: check the 16-bit measurement duration, not its first byte,
      against 1024.

    The BPKM attributes handed the whole message to the ASN.1 dissectors
    of public keys, certificates, signatures and OCSP responses, which
    could then read past the attribute; they now get a buffer of the
    attribute length, and an empty attribute gets an expert info. The OCSP
    response loop stops when a response does not advance. The key exchange
    share field ID and the CMTS designation CA serial number were added
    with the wrong length or offset.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit c79694d5aa1098ba208556f213099abd136fcf30)

commit 3715c2e319
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 08:00:37 2026 +1000

    WMIO: Dissect each heap entry once, do not loop on object qualifiers

    The input and output signatures of a method are objects in the method
    heap, and their classes can have methods of their own whose signatures
    point at the same objects. Every reference was followed again, so a
    chain of objects in a 2.5 KB IWbemServices response took 4^(n-1) object
    dissections and tshark never finished. Property lookups that share a
    PropertyInfo dissected it and its qualifiers again for each lookup.
    Remember the heap entries dissected in an object and show the later
    references without following them.

    The encoding length of an object qualifier is unsigned but was added as
    a signed value, so a negative length moved the offset back to the same
    qualifier and the qualifier set looped forever. Throw when the length
    is beyond the data.

    A Unicode encoded string was fetched as big endian for its label, a
    UINT8 qualifier value was read signed, the text of a UINT16 value
    replaced the label of its qualifier, and an unknown dictionary index
    showed the field index.

    Array qualifiers had the same problems: the byte size of an array of
    integers was its 32-bit count times the element size added to the
    offset unchecked, so a count near 2^32 moved the offset back to the
    same qualifier, and an array of objects added each signed object length
    and could run about 2^32 times over objects of length 0. Check the array
    size and each object length against the data.

    An Encoded-String with flags other than 0 and 1 passed a NULL string to
    "%s" for its label.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 9f424d89eb30bc72eb32950ff4c6afc87747c59d)

commit 8a1854e9f8
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 11:44:53 2026 +1000

    DCERPC: Spread policy handles over the hash, do not grow on repeated closes

    Policy handles were hashed on octets 4-7 only, so handles that share
    them all went to one bucket and a stream of such handles took quadratic
    time: 40000 LSA handles took 43 seconds.

    A close of a handle whose latest instance was already closed did not
    fall in any instance, so a new instance was allocated and inserted by
    walking the whole list. Repeated closes of one handle value grew the
    list by one entry per frame, quadratic in time: 20000 LSA Close calls
    took 17 seconds. Extend the closed instance to the repeated close
    instead, keeping the frame of the first close, so a retransmitted close
    also shows the handle it closes.

    The expert info for a SID longer than its 28-octet dom_sid28 field said
    "Association rejected".

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 439c1a3513eda455396e7c390c8768b76c440c7c)

commit ae64e486a1
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 08:00:36 2026 +1000

    DCERPC: Do not assert on NDR64 counts, dissect deferred pointers once

    An NDR64 array max_count, offset or actual_count is 64 bits from the
    packet, and dissect_ndr_ucvarray_core(), dissect_ndr_uvarray(),
    dissect_ndr_byte_array() and dissect_ndr_vstring() checked it with
    DISSECTOR_ASSERT(val <= UINT32_MAX). Any interface used with the NDR64
    transfer syntax hit it, for example after a bind found by the TCP
    heuristic. dissect_ndr_cvstring() silently truncated the count instead.
    No array that big fits in a packet: add an expert info and throw. A
    64-bit value for a 32-bit field outside NDR64 also asserted; show it
    truncated with the same expert info.

    Deferred pointers were kept in lists that were appended to by walking
    them, indexed with g_slist_nth_data(), and searched from the bottom of
    the stack, and when the pointers of a pointed-to object were done the
    level below was scanned again from its start. Long arrays of structures
    with pointers and linked lists took quadratic or cubic time: a WINREG
    QueryMultipleValues request with 4000 nested elements took about a
    minute. Keep each level with its tail and the last pointer handled, as
    a stack with the newest level first, so that each pointer is visited
    once. Each level is also a level deeper in the tree, so a long linked
    list such as a DRSUAPI object list hit the maximum tree depth as a
    dissector bug; stop with an expert info before it.

    The hash key of every pointer was allocated in file scope on every
    pass although the table is rebuilt for each call; use the packet pool.

    dissect_ndr_ucarray_core() truncated an NDR64 max_count to 32 bits;
    check it the same way. The last fragment of a fault was added to the
    reassembly as if more fragments followed, so fault reassembly never
    completed. "-z dcerpc,srt" with an interface that has no dissector
    dereferenced NULL; report it as an invalid argument.

    The elements of an NDR array were dissected with no limit on nesting,
    and an element can hold an array of its own type. A MAPI restriction is
    dissected as an AND restriction holding an array of restrictions, so a
    65 KB RopGetSearchCriteria response overflowed the stack and crashed
    tshark. Count each array in the dissection depth.

    A count of wide characters that fits in 32 bits could still overflow
    the octet length of the string; check it as well.

    The must_check_size member of dcerpc_info is no longer used; keep it
    so that the layout of the structure does not change.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 6ae68da7a474312c8586b71c7c61a891398e8e9a)

commit 663d66c653
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 08:00:37 2026 +1000

    FRSRPC: Do not loop on a chunk size that wraps, do not assert on names

    The chunks of a CommPktChunkCtr are dissected until the end of the data,
    and each chunk adds its 32-bit size to the offset. A size near 2^32
    wrapped the offset back to the start of the chunk, so the same chunk
    was dissected forever, at 100% CPU without a tree. Stop when a chunk
    does not move the offset forward.

    The file name of a change order command was dissected as a
    NUL-terminated string followed by DISSECTOR_ASSERT() that it is shorter
    than 261 octets, so any name of 130 or more characters without a NUL
    hit the assertion. The field is a uint16 array of 261 elements, 522
    octets, but only 261 were skipped, so the padding after it was taken
    from the name. Add it as a zero-padded UTF-16 string of 522 octets.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 75370996800698695556188ed267983aa6746bba)

commit 9d9e2f2a03
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 08:00:37 2026 +1000

    GPEF: Do not loop on an EFS key that does not advance

    The keys of an EFS blob are dissected for a 32-bit key count from the
    packet, and each key moves the offset by its own length field. A length
    of 0 dissected the same key again for every count, about 2^32 times
    without a tree. LSA hands EFS policy blobs (opnums 53 and 54) to it, and
    the "efsblob" handle is reachable from an Exported PDU. Throw when a key
    length is shorter than the key header or beyond the data. The key item
    started at offset 0 instead of at the key.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 4225fff3244af72f0bc589d5423872373224ee44)

commit fd6e44c8c3
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:43:55 2026 +1000

    CAST: Do not let the PDU length wrap

    get_cast_pdu_len() added the header length to the data length in 32
    bits, so a data length close to 4 GiB gave a PDU length of 0 or a few
    octets. A length of 0 made TCP add every later segment of the stream
    to the same PDU and dissect all of it again for each one, which takes
    quadratic time. Return a length shorter than the header instead, which
    is reported as malformed.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 231876a06fb7cdd8dc5bd9283f7f10d3daebbc11)

commit ef69398e43
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Oct 25 09:14:11 2026 +1000

    Aeron: Don't add fragments to a message that is already complete

    With fragment reassembly enabled, a message that was already complete
    still took the next contiguous fragment, and a retransmitted begin
    fragment was added to it again. Each fragment with the end flag then
    rebuilt the whole message in a new file scope buffer, so memory grew
    with the square of the number of fragments: a 1.3 MB capture used
    6.4 GB.

    Ignore a retransmitted begin fragment of a complete message, and keep
    other fragments that follow a complete message as orphans. Also build
    the frame info lookup key on the stack; the key values were leaked on
    every lookup.

    Assisted-by: Claude
    (backported from commit 31bf085794733ece0436365ad3e0bbcc2292fc9a)

commit 34ada27af9
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 10:27:21 2026 +1000

    Bitcoin: Do not wrap the PDU length, check varint lengths

    get_bitcoin_pdu_length() added the 24 octet header to the 32-bit
    payload length, so a payload length of 0xffffffe8 gave a PDU length
    of 0, which TCP takes as "one more segment": every later segment of
    the stream was added and the whole buffer dissected again, quadratic
    in the length of the stream (8000 segments took 11 s). The heuristic
    is enabled by default. Return the largest length instead.

    String, data, addrv2 address, witness component and merkleblock flag
    lengths are varints that were cut to 32 bits, so 0xffffffff became -1,
    "to the end of the tvb", and the offset stepped back one octet. Throw
    when such a length is longer than what is left.

    Also put the header length macro in parentheses, and pass ENC_NA
    instead of the display flag BASE_SHOW_UTF_8_PRINTABLE as the encoding
    of two FT_BYTES fields.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 23a21ae696567f4bec17f4db0bf6150d59971497)

commit b1206a3854
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Nov 2 09:15:47 2026 -0500

    COROSYNC/TOTEMNET: Forget the last key when the keys change

    The index of the key that decrypted the last packet was a static
    variable of the dissector and was kept when the list of keys was split
    again after a preference change. With a shorter list it pointed past
    the end of the new array, so the next packet read a key pointer from
    beyond it or passed NULL to strlen(). Keep the index at file scope and
    reset it when the list is split.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 68cf3e555aaab718c894e88b92eb349f20a7a5ec)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 6f1e7252a2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 03:03:21 2026 +1000

    Couchbase: Check 32-bit value lengths, limit the JSON results

    The 32-bit lengths of multipath values and results, DCP cache transfer
    items and xattrs were passed as int lengths and added to the offset. A
    length of 0xFFFFFFFF became -1, which means "to the end" and does not
    throw, and moved the offset back one octet. Each spec then added an item
    holding the rest of the PDU while moving forward only a few octets, so a
    60 KB frame used 1.5 GB with a tree and a 250 KB one 5.2 GB. The value
    length of a frame, the body length minus the other sections, wrapped the
    same way. Throw when a length does not fit in the packet.

    Every multipath result was handed to the JSON dissector, which adds a
    layer each time, so about 490 results hit the DISSECTOR_ASSERT on the
    number of layers in call_dissector_work(). Stop calling it before the
    limit.

    When the path of a subdoc value was longer than the value, the value
    length wrapped and the value took the rest of the PDU.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 1492169519551013720ed95e26b1acc375d837d2)

commit bc8cbf0f3e
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Nov 2 05:34:07 2026 -0500

    COSE: Do not crash without media information, do not leak a header key

    The application/cose media type handle read the media parameters from
    its data argument without checking it. AMQP 0-9 content bodies, OBEX and
    ID3v2 pictures call the media_type table with no media_content_info_t
    and a type taken from the packet, so an AMQP message with the content
    type application/cose crashed tshark.

    When the principal key of a header map was found, its value was read
    before the key was released. If the value was missing, the exception
    skipped the release and the GVariant leaked.

    A cose-type media parameter that is not in the subtype table gave a
    NULL handle, which was passed to call_dissector_only() and hit its
    DISSECTOR_ASSERT, for example for an HTTP body with an unknown
    cose-type. Show the invalid tag expert info instead. An unsigned
    content format was added to an FT_INT64 field with
    proto_tree_add_cbor_uint64(), which is a dissector bug; make the field
    FT_UINT64.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 67e25542ff9986844936f2d79b8a8156960bd471)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit d3736a687d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Nov 1 03:15:12 2026 +1000

    DCCP: Do not wrap MP option lengths, check addresses for the checksum

    The length of an MP-DCCP option was reduced by one for the option type
    without checking that there is one, so an option with no value wrapped
    the length to 255. An MP_CONFIRM option then read past its own end and
    recursed into the following bytes, one level per 3 octets, which
    overflowed the stack and crashed tshark. With a length that cannot
    wrap, every level is inside the option above, which is at most 255
    octets. Not every MP option also stepped past the option type, so the
    next option was read one octet early. A Change or Confirm option with
    no value wrapped the length of the feature option the same way.

    The checksum pseudo-header asserted that the addresses are IPv4 or
    IPv6, which they are not in an Exported PDU naming "dccp". Do not check
    the checksum without them.

    The base of the relative sequence numbers was set when the Request or
    Response was dissected, on every pass, so the Request and the packets
    before it showed absolute numbers on the first pass and relative ones
    on the second. Set it on the first pass, before the sequence number of
    the Request is shown, and use it only from that frame on.

    The acknowledgement number of a Reset, CloseReq, Close, Sync and SyncAck
    is 48 bits at offset 2 of the subheader, but was shown as 3 octets at
    offset 1 and the relative value was masked to 24 bits. A Reset needs 12
    octets after the generic header, not 4. A CCID 3 Loss Intervals option
    was added to the options tree instead of its own.

    The IPv4 address of an MP_ADDADDR option was shown in little endian
    byte order, and the follow stream filter used the field dcp.stream.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 381391912b172329a8d2f264130083248a09c2d1)

commit 63724ac738
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Nov 1 21:10:21 2026 -0500

    CP2179: Do not read more requested points than the request had

    A request keeps the points it asked for, but not how many there are,
    and the special calculation range and 16-bit analog scan responses
    indexed that list with a counter that only stopped at the end of the
    response. A response with more values than the request read past the
    heap buffer, and a request with no points has a NULL list, which
    crashed tshark. Both are reached with an Exported PDU naming "cp2179".
    Keep the number of points and stop there. Also do not wrap the count
    when the end point of a scan is below the start point.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 516257e00cbb49da91d479fe26cc999558ef4d8c)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 3daa4a5e59
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Nov 1 21:09:20 2026 -0500

    CSN.1: Do not write more elements than a variable array holds

    The element count of an M_VAR_TARRAY or M_VAR_TARRAY_OFFSET comes from
    the stream, and the elements were stored in the decoded structure with
    no check against the array size. In GSM RLC/MAC, an EXT Measurement
    Report can claim 31 measurements for an array of 9, and NC, CCN and 3G
    measurement reports 7 or 8 for arrays of 6, so the decoder wrote past
    the array into the rest of the uplink message structure. Store the array
    size in the descriptor, as M_REC_TARRAY does, and stop with an error
    when the count is larger. In a union, a count of 0 for the offset
    variant also no longer wraps to 65535.

    An M_REC_ARRAY element was always read with tvb_get_bits8(), which
    asserts for more than 8 bits. GSM RLC/MAC uses M_REC_ARRAY for 9 to 16
    bit PCIDs, PSCs, UTRAN FDD and TDD ARFCNs and EARFCNs, which are kept in
    uint16_t arrays, so a Packet Cell Change Order or Packet Measurement
    Order with any of them stopped with a dissector bug. Store those
    elements in the uint16_t array.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 1911eb17b02bc4af4273884ae57d108408af447c)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 4fad117218
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sat Oct 31 10:18:49 2026 +1000

    DCT2000: Do not crash or hang on malformed records

    A PDCP key comment that matches the sscanf conversions but omits the
    "key: " literal still returns 3. strstr() is then NULL and g_strdup()
    of that pointer crashes. Hex after a MAC-NR, NRUP, or RawTraffic '$'
    was decoded two digits at a time while testing only the first byte, so
    an odd length stepped over the NUL and read or wrote past the buffer.
    With no '$', the MAC-NR decoder treated the header text as payload and
    the NRUP decoder started one byte past the NUL.

    An Exported PDU record can name this dissector with any encapsulation
    byte and without a Catapult DCT2000 pseudo-header. An unknown byte hit
    DISSECTOR_ASSERT_NOT_REACHED(). ATM and ISDN read that pseudo-header,
    and PPP took its direction from the same union. The direction now comes
    from the record, and the inner header is used only for a real DCT2000
    file. ISDN without that header is handed to LAPD.

    Any fp_r* name other than fp_r4 through fp_r8 took the same assert.
    Channel, transport-block, transport-format, and paging counts were
    copied into fp_info and used as loop bounds by the FP dissector, with
    or without a tree. A count above INT_MAX became negative and skipped
    the channel cap. Counts are limited before they are stored: 64
    channels, 512 transport blocks, and sizes that cannot overflow the FP
    length arithmetic.

    An IPPrim remote-address TLV whose length was not 4 or 16 bytes was
    stored as an IPv6 address of that length. Only 4-byte and 16-byte
    addresses are installed.

    Assisted-by: Grok
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 1ac7497d30985d87b0eaa287809fc413625aca0c)

commit 071564b3e2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:35:10 2026 +1000

    CMIP: Do not use an object identifier from an earlier packet

    The object identifier of an action, event, error or extension type is
    kept in a file scope pointer to a string in packet scope, and the
    information that follows uses it to pick a dissector even when the
    packet did not set it, for example for an action type in local form.
    It then read the string of an earlier packet after it was freed.
    Keep it with the packet instead, clear it at the start of each CMIP
    PDU, and restore the one of an outer PDU at the end.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 57f4fdf56b9f07eba486daf44d0e5f18dc65b987)

commit 1cdded5fb0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:22:27 2026 +1000

    CIP Safety: Do not read the destination of a non-IPv4 packet as IPv4

    dissect_cip_safety_data() read the destination address as an IPv4
    address to tell multicast from unicast connections. Called without an
    IP layer, for example from an Exported PDU file, the address has no
    data and the dissector crashed. Treat only an IPv4 destination as
    possibly multicast.

    Also take the Configure Request OUNID from the OUNID, not the TUNID,
    and put the Application Data Path subtree at the path.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 27127c89fb44fd15b834a3b8adf83671abe85c38)

commit d17f061b5c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:39:54 2026 +1000

    CDT: Do not use the tree of an earlier packet, limit the nesting

    The compressed data dissector kept the tree it was called with in file
    scope pointers, which only dissect_cdt() set. Reached by object
    identifier, for example from a CAMEL extension, it used the trees of
    the last packet that P_Mul decoded as CDT, after they were freed, or
    showed nothing. Register a dissector for the object identifier that
    sets them, and restore those of an enclosing CDT at the end.

    Content compressed in P1 content that is itself compressed nested
    until the layer limit asserted. Stop before the limit.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 5b25e9165d70416e12dc605662e65cbfae7719ce)

commit a133b60f36
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 16:33:53 2026 +1000

    CIP: Keep the embedded and multiple service requests apart

    The request info of a CIP request kept the requests of a Multiple
    Service Packet and the request embedded in an Unconnected Send in the
    same pointer. ENIP passes the request's info to a frame in the other
    direction, so a frame that carried an Unconnected Send request after a
    Multiple Service Packet request used the 16 octet multiple service
    info as a 64 octet request info and wrote past its end. Keep the two
    in separate fields.

    The same frame order with a Forward Open request asserted that the
    request had no connection yet. Keep the connection of the first
    request instead.

    Do not allocate request info for more services than the packet has
    offsets for, and take the size of the additional status of a PCCC
    response from the additional status size instead of the general
    status.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit b25bf68af33839c1fca2a8dfbb7bf81741e46fba)

commit e6a32dced7
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Nov 1 10:16:28 2026 +0000

    [Automatic update for 2026-11-01]

    Update manuf, services, enterprise numbers, translations, and other items.

commit 6a9255fab1
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Nov 1 06:29:59 2026 -0500

    Ceph: Do not let lengths move the offset back, fix field types

    A PDU length or authorizer length that wrapped the end of the PDU back
    to its start made dissect_ceph() dissect the same PDU forever, saving
    new state each time. Stop at a PDU end that does not move forward.

    An encoded struct whose size ran past the end of the data, or a
    snapinfo that returned its encoded size instead of its end, moved the
    offset back, which asserted with a tree and looped over the same bytes
    up to four billion times without one. Throw when the size does not
    fit, and continue after the end of the snapinfo.

    When the first pass stopped at the start of a frame, no state was
    saved for it and later passes asserted. Start from a new state then.

    Adding the high flags of a pool from an empty list of fields was a
    dissector bug for every OSD map with a pool, and an object locator
    with both a key and a hash added an expert info without a tvb, which
    asserted. Add the high flags as a plain value, give the low flags
    their own subtree, and add the expert info at the locator.

    The redirect object name and the PG stats category are strings, and
    the Pool Op Reply data is a container; their field types asserted or
    were reported as malformed. Also label the snapinfo instead of
    replacing its text, and end each mon subscription item.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit cc642ef940bb861f57b3604b6c15bc111dd617c8)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit e3db0bcaf0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 10:17:19 2026 +1000

    BT HFP: Do not copy the rest of the packet for each command

    dissect_at_command() copied and upper-cased everything from the start
    of the command to the end of the (reassembled) data, for every command.
    In the HS role a single ';' is a command, so a 64 KB message allocated
    about 2 GB and took seconds per pass. Make the upper-cased copy once per
    packet and point into it.

    An AG +CIEV response whose first parameter is empty dereferenced the
    NULL indicator index when it reached the second parameter.

    Also find the end of a line with a non-ASCII octet in it (the ASCII
    string had three octets for each such octet, so the CR or LF was not
    found at its offset), show the commands after the first ';' in the HS
    role, and leave the delimiter out of an unknown command name.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 7dc910a11adb634ca1a4c96f2695cc46274e40fa)

commit fa6cd4c0eb
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 10:36:39 2026 +1000

    BZR: Do not let a prefixed length wrap the body offset

    dissect_body() added "4 + length" from a 32-bit length prefix to the
    offset. A length near 0xFFFFFFFF wrapped the offset back to the same
    kind octet, so the loop never ended, on the default TCP port and from
    an Exported PDU. Throw an exception when the addition wraps.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 703d1cacb116b45d61bb72e5c9502e9a9f2cf2d2)

commit 755ebd25bf
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 10:27:52 2026 +1000

    BT-uTP: Do not dereference missing uTP proto data

    utp_dissect_pdus() is called by the BitTorrent uTP dissector, whose
    "bittorrent.utp" handle is registered by name. Called from an Exported
    PDU there is no uTP layer and no uTP proto data, and adding the PDU
    size to the uTP tree dereferenced NULL. Add it to the given tree
    instead.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit efe57889f8ab2bb90b867a816fa9012598acdf44)

commit e36241912a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Sun Oct 25 23:17:43 2026 +1000

    AFS: Format the ACL text once and fix the get-xstats timestamp

    GETSTR formatted the whole rest of the packet into the packet pool for
    every ACL entry, while only the local offset moved through the entries.
    A Store ACL request with a large positive entry count and about 64 KB of
    data made one packet allocate several GB. Format the text once and step
    through it, and stop the positive entries at the end of the ACL as the
    negative ones already did.

    The get-xstats reply timestamp was added with OUT_TIMESECS() to an
    FT_UINT32 field, which hit DISSECTOR_ASSERT_NOT_REACHED() for every such
    reply when a tree was built. Make it an FT_ABSOLUTE_TIME like the other
    fields added that way.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 20c246c92ed6cd66665d8082380816f3ef5f430e)

commit e407b9b4e7
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 10:19:57 2026 +1000

    BER: Check the file data and the field of a reassembled OCTET STRING

    The "ber_file" dissector is registered by name and read the file name
    from its data argument without checking it, so a crafted Exported PDU
    file naming it crashed tshark on a NULL pointer. Use the default
    syntax when there is no data.

    A constructed OCTET STRING of two or more segments added the
    reassembled string with the caller's hf_id, which LDAP
    (AttributeValue), CredSSP, GSM MAP, X.509IF and Z39.50 pass as -1 to
    get only the tvb; an LDAP AttributeValue split in two segments on TCP
    389 failed the "Unregistered hf" assertion. Only add the item when
    there is a field.

    Also step over the contents of an unknown NULL, step over a UTCTime
    "+hhmm"/"-hhmm" offset (a valid time with an offset was always
    reported as malformed, and X.509 dropped it), fix the plural in that
    error, and use an unsigned shift for the 32nd SET element.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit c304808d9421b6d07725baac9d1c638be17d6bef)

commit 5d53a84cec
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 05:40:37 2026 +1000

    BT HCI EVT: Do not use a missing local name, fix the MWS and group offsets

    A successful Command Complete of Change Local Name used the name of
    the command even when it could not be read, an uninitialised pointer
    that then crashed the next BT HCI ISO packet. It also saved the name
    for the dissection only when the Bluetooth device tap had a listener,
    so the dissection depended on whether the Bluetooth Devices dialog was
    open. Skip it when there is no name and save it in any case.

    Get MWS Transport Layer Configuration restarted at octet 1 of the
    event, and the LE minimum connection interval groups were all labelled
    with the number of groups.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit c2869bd064799389a51635418b02bca211feda6b)

commit 04cbe0b4c6
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 05:40:37 2026 +1000

    BT HCI CMD: Check the EIR/AD data argument, fix several AD entry lengths

    "btcommon.eir_ad.ad" and "btcommon.eir_ad.eir" are registered handles,
    so an Exported PDU file can call them without bluetooth_eir_ad_data_t,
    which failed DISSECTOR_ASSERT(). Use the default interface and adapter
    instead. The Broadcom and Android vendor dissectors passed a
    bluetooth_data_t as bluetooth_eir_ad_data_t, so the device tap of the
    Bluetooth Devices dialog copied a BD_ADDR from a 4-octet counter; pass
    the right structure.

    Zero the saved command data, so a Command Complete for a Change Local
    Name command whose name could not be read does not use an
    uninitialised name pointer.

    AD entries: the TDS entry stepped over the entry length instead of the
    transport data length, an undecoded type 0x10 entry did not step over
    its data at all, and the LE Supported Features entry read up to the
    end of the AD data; each made the following entries malformed. Allocate
    the proto data that only lives for the packet from pinfo->pool.

    Also read the clock offset of Create Connection and Truncated Page at
    its offset (Truncated Page was malformed), step over the Write Page
    Scan Mode parameter, and append "(No data)" to the right Burst Number
    of LE Set CIG Parameters Test.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 72f2331aff3488b0caca2f9b1c841808576ee431)

commit 91b1ab55f0
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 05:37:29 2026 +1000

    BT BR/EDR RF: Check the FHS data argument, fix the continuation payload

    "btbredr_fhs" is a registered handle, so an Exported PDU file can call
    it without btbredr_fhs_data_t; it then dereferenced NULL. Return 0
    instead.

    A continuation fragment without a known connection was handed to
    L2CAP as a complete PDU, starting at the payload header, because
    next_tvb still held the whole payload. Recognise CPB EDR 3 payloads,
    which compared with the ACL value instead, and read the flags only if
    the header has room for them.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit fbf60bf4e9013f91c335604f76aea4ad73f01d56)

commit fbe1f19b5d
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 05:42:19 2026 +1000

    BT AVRCP: Reassemble only in the end frame, fix the response time

    Every end packet after a completed message on the same channel and
    operation reassembled that message again, a copy of the whole message
    per packet (quadratic time), and on a later pass an end packet before
    the message was continued decoded it too (tshark -2 differed).
    Reassemble only in the frame that completed the message.

    The seconds of the response time were divided by 1000 instead of
    multiplied, and the 2-octet CharacterSetID of the attribute and value
    text responses was added with length 1.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit e6bcb4bd5df064ec9b54deb6bd504c765b0fa7bf)

commit 146ba95d3c
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Fri Oct 30 05:45:38 2026 +1000

    BT ATT: Check the attribute data length, the data argument and handle 0

    A Read By Type or Read By Group Type Response with an attribute data
    length shorter than its handles added an item with length -1, or a
    subset with a reported length of 0xffffffff, for every octet, each
    covering the rest of the PDU: a 60 KB response took 2-4 GB with a
    tree. Report the length instead.

    "btgatt" and the "btgatt.uuid0x..." handles are registered, so an
    Exported PDU file can call them without btatt_data_t, which failed
    DISSECTOR_ASSERT(). Return 0 instead. dissect_btgatt() passes handle
    0, so Decode As of handle 0 to a BT GATT dissector recursed until the
    stack overflowed; handle 0 is reserved, do not look it up.

    Also take the Find By Type Value Request UUID from its own offset
    instead of the ending handle, keep the length of a Signed Write
    Command value over 255 octets, and check the response value, not the
    octet after it, for the supported sensor locations of the SC Control
    Point.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (cherry picked from commit 8d72f20923d8e955725f9036ebd311ce51996fcb)

commit eb7f00237e
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 30 21:40:11 2026 -0400

    dumpcap: Reset the options parser before parsing the log options

    The ws_getopt_long() loop that looks for -Z and the application flavor
    goes through all of the arguments, and ws_log_parse_args() was called
    after it without resetting the options parser, so it didn't see any
    arguments. --log-level, --log-file and the other log options had no
    effect, and invalid values for them were silently accepted, e.g.
    "dumpcap --log-level=bogus -v" printed the version and exited with 0.

    Set ws_optreset and ws_optind to 1 first, as tshark does.

    Assisted-by: Claude, for noticing the error when running some tests

    (cherry picked from commit b02b58b389ed67416cb966709439ab243817cb4f)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit b27afa02c1
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 30 08:45:46 2026 -0400

    ADB Service: Don't carry over a logcat entry longer than 5 KiB

    When a logcat message is not complete, the bytes received so far are
    kept in file scope and put in front of the next message's data. Each
    message that still doesn't complete the entry keeps its own copy of
    everything received so far, so memory grows with the square of the
    number of messages. A payload length of 0xFFFF with one byte in each
    following message made a 6 MB capture use 2.3 GB.

    Android limits a log entry, header included, to 5 KiB
    (LOGGER_ENTRY_MAX_LEN). Report a longer entry as malformed and don't
    keep it for the next message.

    Assisted-by: Claude

    (backported from commit b669f03cd5f814f83e944a661a5ec094f75a1134)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit a1a2b4db10
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 10:17:47 2026 +1000

    BBLog: Only dissect a Netflix custom block record

    The "bblog" dissector is registered by name and read the record's
    block as Netflix custom block data without checking it. A crafted
    Exported PDU file naming it hands it an ordinary packet record, whose
    block has no mandatory data, and tshark crashed on the NULL pointer.
    Return 0 unless the record is a Netflix custom block.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 010a8b62fd616ebe7da9138baa265c2a30b63604)

commit 551387fe31
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 30 08:57:19 2026 -0400

    BT HCI ISO: Check the data argument, fix stray and timestamped fragments

    "bthci_iso_data" is a registered handle, so an Exported PDU file can
    call it without iso_data_info_t; it then dereferenced NULL. Return 0
    instead, and drop the dead assertion in the main dissector.

    A reassembled SDU was decoded with the TS flag of the last fragment,
    which is always clear, so the time stamp was shown as the sequence
    number and SDU length; keep the flag of the first fragment. A
    continuation or last fragment after a completed SDU, or after a first
    fragment that was rejected, attached itself to that SDU and decoded it
    again; ignore an SDU that was completed in an earlier frame. Read the
    SDU length of a first fragment on every pass, so a short fragment is
    malformed with tshark -2 as well.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 6ba444337e77bfdbc623fdceda2bdb35bac4a8be)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 167cf87343
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 30 08:51:02 2026 -0400

    Bluetooth: Check the URB info of the USB dissector, the GAEN length

    The "bluetooth.usb" dissector is registered by name and passed its
    data argument on to HCI USB as the URB info without checking it, and
    HCI USB read through it. A crafted Exported PDU file naming
    "bluetooth.usb" crashed tshark on the NULL pointer. Return 0 when
    there is no URB info.

    The GAEN dissector took the last 20 octets of the captured length,
    which went negative for shorter data; use the reported length and
    decline shorter data.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit e5e6cf9c930a98aa906bf6bc32501dd871449306)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 2d8230ff69
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 29 23:56:51 2026 -0400

    ACN: Build the universe discovery list in a wmem_strbuf

    The sACN universe discovery list was built in a 97-byte buffer, one
    universe after another, but every snprintf() call was given the size
    of the whole buffer rather than the space left after the write
    pointer. No universe list wrote past the end, but a build with
    _FORTIFY_SOURCE=3, which CMakeLists.txt sets for all but Debug builds,
    checks the size against the space left and aborts ("buffer overflow
    detected") on the second universe. A discovery packet listing two or
    more universes made tshark abort.

    Append the universes to a wmem_strbuf instead, so no size has to be
    tracked.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 7a8aafe4170ec51ab188b53760dca8ec15cdf189)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit ce2da16286
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 29 21:17:37 2026 -0400

    ACN: Bound the DMP address ranges shown for a single data item

    A DMP range address with a single data item (D_RS) shows the item once
    for every address in the range, without advancing through the packet,
    so a 32-bit count taken from the packet decided how many strings were
    allocated. A 170-byte packet made tshark allocate memory until it was
    killed. A range with an array of equal size items (D_RE) whose count
    was larger than its data length gave an item size of zero and the same
    loop.

    Show at most 1024 addresses of a D_RS range, in both data and reason
    codes, and add an expert info when the range is longer. Decode a D_RE
    range item by item only when every item has at least one byte. Stop
    mapping virtual addresses when one doesn't advance the offset, and
    show the range increment and count as unsigned.

    Assisted-by: Claude

    (cherry picked from commit 20dbcf1c171e580272d043ec3da5675ce0def2ad)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 055aa4af3f
Author: Darius Davis <darius-wireshark@free-range.com.au>
Date:   Sat Aug 23 15:42:33 2025 +1000

    cllog: Fix a potential buffer read overrun.

    Require a non-NUL separator early during parsing.

    If parseColumnHeaderFields is called with no separator yet found, things go
    quickly off the rails when we try iterating through the field list using
    strchr(str, '\0')... and we end up scanning beyond the end of the string.

    (cherry picked from commit 78d43062f3a918517881b48aa413c831d1b00525)

commit 4f2f8bfdb7
Author: Darius Davis <darius-wireshark@free-range.com.au>
Date:   Tue Nov 26 10:58:32 2024 +1000

    cllog: Avoid a stack read overrun.

    If parseFieldLength received a length of greater than 8 and parseFieldData was
    not subsequently called, the memcpy at the end of cllog_read_common could read
    beyond the end of the eight-byte buffer in logEntry.data on the stack.  This
    condition could arise if the column header in a malicious input file stipulated
    that the Length field appeared _after_ the Data field, or perhaps if the Data
    field was absent.

    The result would most likely be just some garbage data from the stack appearing
    in the packet contents, but there's a chance it might also cause a crash if
    length is large enough to read beyond the top of stack.

    Clean up a related magic number "8" and an unnecessary loop counter.

    Ping #19937

    (cherry picked from commit 9a1005cd4ff3a7543ec226eb9c48472cc4002848)

commit 57c6cba7a2
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Oct 28 09:21:45 2026 +1000

    Ascend: Clear the user name before parsing each packet

    The scanner sets the user name of a WAN session packet only when it is
    empty, and nothing cleared it, so every packet showed the user of the
    first one. When the file is opened, the trial parse also tested the
    uninitialized user name of a record on the stack. Clear it with the
    called number before each parse.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit cbd62b83574457b75e0f2c9011a56abf4d97a887)

commit 4a8b0f7589
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Thu Oct 29 01:58:59 2026 +1000

    AT: Do not dereference a missing +CIEV indicator index

    A +CIEV parameter starting with CR is not dissected but still counted,
    so the next one was dissected as the indicator value and dereferenced
    the indicator index that the first one should have stored, which was
    NULL. A USB bulk transfer with "\r\n+CIEV:\rX,1\r\n", found by the
    default heuristic, crashed tshark. Add an expert info instead.

    Also decode the +CMUX port speed from its digits rather than from its
    first character, do not append the 8 skipped octets of a +CMGL or
    +CMGR SMS PDU to its end as zeros, and show the commands after the
    first ";" of a command line, which were added to no tree.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 01a8a8f78bce9fdd267437bfc6153675cb303236)

commit 724b0228d5
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 29 07:24:00 2026 -0400

    AppleTalk: Check the data argument of the short DDP dissector

    The short DDP dissector takes the LLAP node addresses from its data
    argument without checking it. It is registered by name, so an Exported
    PDU naming "ddp_short" calls it with no data, and a file with one such
    PDU crashed tshark. Reject the packet when there is no data.

    Also mask the short DDP datagram length with 0x3ff like the long
    header does, so set reserved bits do not give a bogus length warning
    and leave the padding in the payload, and do not set the ASP columns
    for a packet ASP rejects. And an RTMP Response with no routing tuples
    is valid, so do not peek at a missing first tuple.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 38e1cd4a62d62d59b8472e4ca913287a1ae4f8ba)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit bb0be18a4a
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 29 07:22:34 2026 -0400

    ATM: Do not dereference or assert on a missing pseudo-header

    The LANE dissector in the "atm.aal5.type" table takes the ATM
    pseudo-header from its data argument without checking it. An Exported
    PDU naming the table calls it with no data, so a file with one such
    PDU crashed tshark. The ATM and ATM pseudowire dissectors registered by
    name failed a DISSECTOR_ASSERT() in the same case. Reject the packet
    when there is no pseudo-header.

    An AAL5 PDU shorter than its 8-octet trailer made the unsigned trailer
    offsets wrap, so the trailer items threw with a tree and a bogus
    length was used without one. Only look for a trailer in 8 octets or
    more.

    Also pass the tree and item of dissect_reassembled_pdu() in the right
    order, drop the doubled parentheses around the CRC-10 results, and do
    not show the AAL3/4 Length Indicator with the segment type names.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit d1f1aa4c44908398196d163d6e18028ac08137f2)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 4e702cc7ed
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Oct 28 12:24:24 2026 +1000

    Exported PDU: Give TLS heuristic dissectors their TLS information

    A heuristic dissector named by an Exported PDU is called with the TCP
    information when it is in the "tcp" list and with no data otherwise.
    The dissectors in the "tls" list take a struct tlsinfo and several
    dereference it unconditionally (HTTP, HTTP2, WebSphere MQ, FIX and 5G
    LI, which store their handle through app_handle), so a file with one
    Exported PDU naming any of them crashed tshark. Such files are what
    "tshark -U" writes for TLS application data found by heuristics.

    Give the "tls" heuristic dissectors an empty struct tlsinfo, with an
    app_handle they can set.

    Fixes #21702

    Reported-by: Doha Nam
    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 576969db4d92c06b2f78871124dc38f6be398548)

commit 92572bea27
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Oct 28 18:08:00 2026 -0400

    AMQP: Store prop_flags in a uint32_t to prevent UB on Windows

    Since the C integer promotions turn uint16_t into int, retrieving
    the prop_flags and storing it in a uint16_t, and then repeatedly
    left-shifting to test against 0x8000 is UB, as left-shifting a
    negative signed number is UB.

    This is the shortest fix; it would probably be better to avoid the
    left shifts entirely, use defines for the masks, and also add
    individual fields for the flag bits in the header for clarity.

    (cherry picked from commit 3c812870a53f06e136e798b57aa3fad97cfa698f)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit c321c87234
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Oct 28 18:06:43 2026 -0400

    ARINC 615A: Do not dereference missing TFTP information

    The heuristic dissector takes the file name from the struct tftpinfo
    that TFTP passes as its data argument, without checking it. An Exported
    PDU naming the "a615a" heuristic dissector calls it with no data, so a
    file with one such 8-octet PDU crashed tshark. Reject the packet when
    there is no TFTP information.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 976c2c5d496b4e39aa82ecfa005a2fa7df84fc57)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit f740fe895a
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Mon Oct 26 09:02:52 2026 +1000

    Exported PDU: Only give the TCP information to TCP dissectors

    The TCP information tag of an exported PDU is turned into a struct
    tcpinfo that was passed to whichever dissector the PDU names. A
    dissector reached through the name, heuristic or table tags with a
    data argument of its own type then reads and writes the small tcpinfo
    allocation as that type. AMR, for example, takes it as a struct
    _rtp_info, reads a map pointer from past its end and stores to it, so
    a two tag capture that names "amr" crashes tshark.

    Pass the TCP information only to a dissector that TCP calls: one
    registered in the "tcp.port" table, one of a protocol with a "tcp"
    heuristic dissector (which set conversation dissectors such as
    PROXY's), a heuristic dissector of the "tcp" list, or the "tcp.port"
    table itself. When the tag is missing, give these dissectors empty TCP
    information instead of NULL: several dereference it unconditionally
    (the RTMPT heuristic dissector, for one), and TCP's own export code
    does when PROXY passes it on, so "tshark -U" of such a capture
    crashed. PDUs exported from TCP dissect as before.

    Two more paths ended in an assertion: a dissector table tag naming a
    string table aborted tshark in the uint lookup, and a port type the
    dissector did not map, including iWARP MPA and MCTP, which the export
    code writes itself, failed a DISSECTOR_ASSERT(). Only look up integer
    tables, map those two port types and treat other values as none.

    Check the length of the tags of a fixed size before reading them, so
    that a short tag gives the "Unexpected tag length" warning instead of
    reading the tags after it, and keep the type string of a 3GPP UE ID
    inside its tag.

    4.4 doesn't have the 3GPP UE ID or link direction tags, and its tag
    length is an int.

    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    Assisted-by: Claude
    (backported from commit 656881e3e05e730a992f2d4df35d3fe15836429b)

commit a688fda2fd
Author: Gerald Combs <gerald@wireshark.org>
Date:   Tue Oct 27 13:38:28 2026 -0700

    Release notes: Add a note about brotli

commit 0af2d309ff
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Oct 26 17:26:12 2026 -0400

    TLS: Check for transcript length overflow

    The handshake transcript (kept in [D]TLS 1.2 for the Extended
    Master Secret) is appended to with a 32-bit length and no overflow
    check. A connection with more than 4 GiB of handshake messages (e.g.,
    many handshake messages before the master secret is known, which is
    always the case when the keys aren't available) wraps the length, and
    the next copy writes past the end of the buffer. Check for that, and
    don't append to a transcript that would overflow.

    4.4 doesn't have the Encrypted Client Hello transcript from #21693.

    Related to #21693

    Assisted-by: Claude Opus 5.5

    (backported from commit eb6fdfd82527d9ef6755d204262fa297610e8c38)

commit bfaadd96ba
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Oct 25 16:17:06 2026 -0700

    NSIS: Fix a version recommendation

    (cherry picked from commit 1ff2f0656cb243aba5bc1d67b35d5bc0a19fd258)

    Conflicts:
            packaging/nsis/wireshark.nsi

commit 8f618016a2
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Oct 25 16:18:18 2026 -0700

    Release notes: This will probably be the final 4.4 release

commit 310fac68fd
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 25 15:19:39 2026 -0400

    BT HCI ACL, BT HCI ISO: Grow reassembly buffers as data arrives

    The first fragment of a PDU allocated a file scope buffer for the
    entire length claimed in its header (up to 64 KiB for ACL), even if no
    more fragments arrive, so a capture of tiny first fragments uses far
    more memory than its size. Grow the buffer as fragments are added
    instead, doubling in size up to the total length.

    Also, the ACL total length (L2CAP length + 4) was stored in 16 bits,
    so the largest L2CAP PDUs could never be reassembled.

    Thanks to AISLE Security for the report

    Fix #21656

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit b99adb99fb6e257a5a1bbc7fdb599cdc59c629d5)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit f579de9249
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 25 14:05:43 2026 -0400

    BT HCI ACL: Do not overflow in reassembly

    libwiretap packet size limits might prevent this from happening in
    real captures (or might not), but it's good to check here.

    Fix #21560.

    Thanks to Shivaraj Ramjali Nepali for the report.

    (cherry picked from commit bae955afe40101f4ac0c4091dd7e0373db845ef0)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 40d923914b
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Oct 25 10:16:47 2026 +0000

    [Automatic update for 2026-10-25]

    Update manuf, services, enterprise numbers, translations, and other items.

commit daae9bec43
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 25 00:11:48 2026 -0400

    RTMPT: Update documentation links

    Several of the links to the RTMP, AMF and FLV specifications and to
    other descriptions of the protocol no longer work. Link to the copies
    of Adobe's specifications that the Enhanced RTMP project keeps, and to
    Wayback Machine archives of the other pages.

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit cda8842f1faa0f8ae2a9e115ed99d9b36a5055bc)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 7dbc4832fe
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 25 00:11:08 2026 -0400

    RTMPT: Free the data of a message replaced before it is complete

    When a new message header arrives on a chunk stream whose previous
    message is still incomplete, the previous message is abandoned: it is
    never added to the list of complete messages, so its data is never
    dissected, but its buffer stayed allocated in file scope until the
    capture file was closed. Nothing limits how many such messages there
    can be.

    Free the buffer when the message is replaced, and mark the message as
    complete so that nothing is appended to it later.

    Ping #21677

    Assisted-by: Claude Opus 5.5

    (cherry picked from commit c0a0d515060a305e492cf3beea5845f8f2adf364)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 3a47906a5b
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Oct 24 23:42:55 2026 -0400

    RTMPT: Allocate only the header when starting to unchunk a message

    Each RTMP message whose body is longer than the chunk size started with
    a 32 KiB file scope allocation (or the declared length, if smaller).
    Every message on a chunk stream can declare a 24-bit body length and
    send as little as one byte of body before the next message header, so
    a 12 byte message header with a Set Chunk Size of 1 allocated 32 KiB
    that was never freed, about 2500 times the input.

    Allocate just the message header to start. The buffer is already grown
    as the body arrives, doubling each time, so the memory now grows with
    the bytes actually received.

    Fix #21677

    Assisted-by: Claude Opus 5.5

    (cherry picked from commit dc7d00150a4219c26c8b21ff27c7589c08e510e4)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit d77a77b905
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Oct 24 23:23:30 2026 -0400

    9P: Stop building a Twalk path after the first 250 names

    The Twalk loop stopped advancing the offset after 250 path elements
    but kept appending to the fid path for every one of the up to 65535
    elements in nwname. Each extra pass re-read the 251st element, so a
    Twalk of about 60 KB with a large last name grew the path by gigabytes
    on the first pass and aborted tshark when the allocation failed.

    Stop the whole loop at 250 elements, as Rwalk already does.

    Assisted-by: Claude

    (cherry picked from commit df2c9ebfadaa7a168ec4d4712fa36ee95100458e)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 4f6d9abeac
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Oct 24 21:40:13 2026 -0400

    tools: Update BROTLI_VERSION in macos-setup.sh

    Update the default Brotli version in light of recent CVE reports.
    This isn't the preferred way to install libraries on 4.6 anymore,
    but it is still present, and it's used in 4.4

    (cherry picked from commit 0881c3fa1018bd2131b469a95cc912a787a17912)

commit c7d46557cf
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Oct 24 21:52:39 2026 -0400

    Windows: Update brotli to 1.2.1

    Brotli 1.2.1 fixes CVE-2026-111497, an out-of-bounds write in the
    decoder which might affect us. Use the same packages as
    FetchArtifacts.cmake on 4.6 and later (10a6f4ab06).

commit 8f84ab33e2
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Oct 24 18:53:33 2026 -0400

    BGP: Finish dissecting path attributes after an invalid length

    dissect_bgp_path_attr() returning early for an invalid path attribute
    length skips decrementing the dissection depth. That's not ideal, though
    generally a low impact issue. Mildly more impactful if skipping the
    postponed dissection of a LINK_STATE attribute which came before the
    invalid one. (At worst, with the Protocol ID from the NLRI missing this
    doesn't dissect as well, but it still dissects something, just with
    less information.)

    We can break out of the for loop instead (we're at the first layer of
    nesting, so we don't need a goto and label.)

    Thanks to AISLE Research for reporting.

    (cherry picked from commit a3b4c9f448b1eb78c2b3734b5ebce945af702da6)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit b1a06824a6
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Oct 21 18:05:59 2026 +1000

    SMB2: Fix memory safety issues found in an audit

    - SRT tap: the opcode is a 16-bit value from the packet but indexed a
      256 entry table. add_srt_table_data() only checks the index with
      ws_assert(), so release builds wrote past the end of the heap
      buffer. Skip out-of-range opcodes.
    - Compression: refuse a compression transform nested inside
      decompressed data, which could multiply the decompression memory
      use for every nesting level. Add an expert info for it.
    - Fix the chained compression check that tested a constant instead of
      the algorithm.
    - Zero-initialize export object file info, and remember it when it is
      created on a file id lookup.
    - Use the offset returned by dissect_windows_sockaddr_storage() in the
      share redirect error context instead of adding it.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 0effc439d28551ea9073f93e99f2aac9c06b8c3d)

commit e3987ad6ca
Author: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Date:   Wed Oct 21 18:05:59 2026 +1000

    SMB: Fix memory safety issues found in an audit

    - SRT tap: the Trans2 and NT Transact subcommands are 16-bit values
      from the packet but indexed a 256 entry table. add_srt_table_data()
      only checks the index with ws_assert(), so release builds wrote past
      the end of the heap buffer. Skip out-of-range subcommands.
    - Export objects: avoid a division by zero when a file has length 0.
    - AndX chains recursed through dissect_smb_command() without a depth
      limit. Use increment_dissection_depth().
    - User quota lists: a quota entry offset could point backwards and
      make the loop run forever when no tree is built. Require each entry
      to move forward and stay within the byte count.
    - DFS referrals: report an expert info and stop when a referral
      string, or its offset, extends past the end of the transaction
      data, instead of letting the byte count wrap around.
    - FIND_FIRST2/FIND_NEXT2: stop when an entry does not advance.
    - Zero-initialize file-scope FID and conversation structures whose
      fields could otherwise be read uninitialized.

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
    (backported from commit 04f1fa8cb816126e91109d033516291d99b78116)

commit d95c695a28
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 23 09:33:42 2026 -0400

    wslua: Check that TvbRanges have not expired before using them

    The TreeItem methods add(), add_le(), add_packet_field() and
    add_tvb_expert_info() accepted a TvbRange without checking whether it
    had expired, as did TvbRange:le_float() and the TvbRange __eq
    metamethod. A Lua script that kept a TvbRange from a previous packet
    (e.g., in a global or upvalue) and passed it to one of these would
    read from the tvbuff of the earlier packet after it was freed.

    Raise an "expired tvb" error in those cases, as the other TvbRange
    methods already do.

    Assisted-by: Claude Opus 5.5

    (cherry picked from commit 31f3fcd117a0322545f1b9c39fd499b0dab6ac3f)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit b8ff038824
Author: John Thacker <johnthacker@gmail.com>
Date:   Tue Oct 20 20:48:57 2026 -0400

    rawshark: Call ws_buffer_increase_length

    Keep the Buffer's known length internally consistent.

    (backported from commit e6ba4faf2f0d91377f0678a66736f9252fa23470)

commit 088dff4362
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 23 08:58:49 2026 -0400

    systemd journal: Decode binary-safe string fields as UTF-8

    String fields serialized in the binary-safe format were added with
    ENC_NA, i.e. ASCII, so every non-ASCII byte was shown as a replacement
    character. Fields that contain a newline or other control character
    are serialized that way even when they are valid UTF-8, so e.g. a
    multi-line message with an accented character was mangled. Use
    ENC_UTF_8 as for text fields.

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit 3ed7ad87893e2e2b8f31c98ebe6e6c18b81d248a)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit e218b546fb
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Oct 14 17:20:36 2026 -0400

    sdjournal: Serialize entries with control codepoints as binary-safe

    The systemd journal export format, in addition to serializing journal
    fields which are not valid UTF-8 in a binary-safe manner, also uses the
    binary-safe method for fields with control codepoints. journalctl
    -o export treats as control codepoints C0 other than TAB (so including
    newline), plus DEL and C1 (U+007F to U+009F). The format documentation
    only mentions values below 32 other than TAB, but systemd 259 also
    exports fields with DEL or C1 codepoints as binary.

    https://systemd.io/JOURNAL_EXPORT_FORMATS/

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>
    (backported from commit 3a986b7686258218c633b04eb050fbe069d86cc2)

commit 85011bc28b
Author: Guy Harris <gharris@sonic.net>
Date:   Fri Oct 23 00:08:27 2026 +0000

    rpcap: dissect server byte-order magic in authenticaton reply.

    I added that a while ago, but forgot to add it to the protocol
    dissector.

    (cherry picked from commit 2f78c380e03ac382992dd2a64af41dfc59ab8fac)

    Co-authored-by: Guy Harris <gharris@sonic.net>

commit 97e82468b2
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 22 09:57:05 2026 -0400

    SMB Pipe: Bound LANMAN descriptor repeat counts

    A repeat count in a LANMAN descriptor string such as "B4294967295"
    overflowed get_count() and wrapped to -1. The data dissector then
    moved the offset backwards, so the auxiliary data loops could run
    billions of iterations without advancing.

    - Cap descriptor repeat counts at 0xFFFF.
    - Stop the auxiliary data loops when an entry does not advance the
      offset.
    - Dissect auxiliary response data with the auxiliary data descriptor
      instead of the main data descriptor.
    - Stop dissecting response entries when an entry consumes no data,
      for example when the data descriptor contains only null pointers,
      and add an expert info. Otherwise every one of up to 65535 entries
      added items at the same offset.

    Fixes #21482

    Assisted-by: Claude
    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit e0c23bb5979edea79a30ab5a686d1c9a6d4d90df)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 954af87215
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Oct 21 17:07:27 2026 -0400

    OpenWire: Add dissection depth checks for complex types

    The OpenWire complex types can have various nested recursions, and since
    OpenWire is dissected over TCP, large desegmented PDUs can possibly
    exhaust the default 8 MiB stack when dissecting with a NULL tree.

    Fix #21676

    (cherry picked from commit 3823253aaa88f9fb482501538d4b33ba465d9179)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit c10ced89a8
Author: John Thacker <johnthacker@gmail.com>
Date:   Tue Oct 20 20:43:50 2026 -0400

    IEEE 1905: Prevent possible buffer overflow calculating fragment hash

    The packet src and dst address when dissecting Bluetooth can be a string
    which is the concatenation of the string representation of the BDADDR
    (up to 64 bytes if name resolution is enabled) and the Bluetooth
    friendly device name (up to 248 bytes). Two of those addresses combined
    (plus 5 more bytes for other parts of the key) can overflow a uint8_t.
    Store the buffer length in a 32-bit unsigned integer so that overflow
    can never occur.

    Fix #21683

    Thanks to Anthropic's Claude and Ada Logics for the report and PoC.

    (backported from commit e099bf71c44046ef1ba39c6d0d827928a840664b)

commit dae2d4c755
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Oct 21 09:21:43 2026 -0400

    Wireguard: Always clear the last ephemeral key on reset

    wg_keylog_last_ekey points to a wmem_file_scope allocation, but
    wg_keylog_reset only cleared it when a key log file was open. Keys
    from embedded secrets (DSB) are processed without one, so the pointer
    outlived the capture file and a later PRESHARED_KEY line was linked to
    freed memory. Clear it unconditionally.

    Add a sharkd test that loads a capture with embedded secrets, then a
    second capture with a key log that starts with a PSK.

    Assisted-by: Claude Code

    (cherry picked from commit 549c410d088bbacf333cd2969717da41b94c841a)

    Co-authored-by: Mar <mardgez13@gmail.com>

commit 32cac07a8d
Author: John Thacker <johnthacker@gmail.com>
Date:   Tue Oct 20 12:01:52 2026 -0400

    ERF: Skip PAD blocks properly

    The ERF reader wants to completely skip PAD blocks. On the sequential
    pass, don't read the bytes from the pad blocks into the record Buffer,
    just discard them (it's still a short file if they end prematurely.)
    Unref the block created by erf_read_header for the pad block to avoid a leak.
    Update the record offset in erf_read at the start of each loop iteration,
    so that PAD blocks are ignored on the random access as well.

    It might be simpler logic in some ways to skip the pad blocks inside
    erf_read_header (it could avoid creating blocks to unref), but then
    we'd have to do more to retrieve the proper offset and avoid having to
    skip again on random access.

    Fix #21682

    Thanks to Anthropic's Claude and Ada Logics for the report and PoC.

    (backported from commit 055f2a8f5f06d93a42c27d101f414807cb466fc8)

commit 6702678ce6
Author: John Thacker <johnthacker@gmail.com>
Date:   Tue Oct 20 20:57:52 2026 -0400

    Kafka: Use TVB subsets to avoid overflow

    Instead of tracking a passed in end offset everywhere, take a TVB subset
    to avoid overflow when adding 32-bit unsigned quantities from packet
    data to existing offsets. Instead, we can pass the value in to
    tvb_reported_length_remaining.

    Fix #21684

    Thanks to Anthropic's Claude and Ada Logics for the report and PoC.

    (backported from commit e5d6aeb7193c87d6fb147bf1363635de2a151a94)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit eb9104aaee
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Oct 18 16:35:42 2026 -0700

    DNS: Add a couple of NULL checks

    This doesn't appear to affect this branch, but it won't hurt anything
    either.

    (cherry picked from commit 7e49d4f8fa1317257fb7f7e7069b96bf1fe09065)

commit 19f2c90064
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Oct 19 17:51:20 2026 -0400

    Matter: Increment dissection depth when recursively dissecting TLVs

    Matter TLVs can be carried within Bluetooth ATT, and it appears
    possible for unusually large attributes after reassembly at lower
    levels to exhaust the stack if the dissection depth isn't incremented
    before recursing.

    Fix #21675

    (cherry picked from commit 390f3482d19645065013052effe7bb5edfb4cef5)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit d7f0aca47b
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Oct 18 10:16:55 2026 +0000

    [Automatic update for 2026-10-18]

    Update manuf, services, enterprise numbers, translations, and other items.

commit b8ef578cab
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 16 18:58:58 2026 -0400

    TFTP: Don't process option negotiation after transaction begins

    TFTP is a lockstep protocol. The option extension requires option
    negotiation prior to beginning the file transfer.[1] In particular,
    the block size should not change during the transaction. The window
    size can change in a Microsoft extension, but that is not sent in a
    OACK opcode but in ACKs.

    If an OACK appears after the data has begun, flag it as an expert info
    and don't update the conversation info data such as the blocksize.
    Also add a DISSECTOR_ASSERT to make sure that this works and we don't
    overflow the buffer.

    Also, dupe the file-scoped data in the TShark single-pass case, because
    commit 722f9eb492e9bf72e9dae50a2dea806832eb8c5c means that eo data is
    g_free'd (like how the GUI does), and we can't free file-scoped data
    that way.

    Fix #21611

    Thanks to Lia04-WD for reporting.

    1 - https://www.rfc-editor.org/info/rfc2347/

    (cherry picked from commit 36edcd1d103ceb0be6a5f9fdaee1467fc4d7dda8)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 2284d8a57e
Author: Gerald Combs <gerald@wireshark.org>
Date:   Tue Oct 13 13:21:10 2026 -0700

    IrDA: Don't overflow ett arrays

    These might cause crashes, but that would require user interaction
    analyzing a historical protocol.

    Assisted-by: Claude Sonnet 5
    (cherry picked from commit da53e91f24085cde6b07f3f19f8d7f2029095a84)

commit 0900e5ed03
Author: Gerald Combs <gerald@wireshark.org>
Date:   Tue Oct 13 13:11:19 2026 -0700

    IrDA: Make sure we have valid conversation pointers

    Don't assume that conversation_get_proto_data will return a valid
    pointer. Zero-initialize our structs.

    Fixes #21658

    (cherry picked from commit 729dfb0048a17b4366259f4f557579cb77b39012)

commit 3c14e52beb
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 15 08:29:12 2026 -0400

    sdjournal: Don't read past the end of a binary field

    A binary (non-UTF-8) field was written in the export format by copying
    one byte more than the field's data, reading past the end of the data
    returned by sd_journal_enumerate_data(), and that byte was written
    where the newline that ends the field belongs. Copy just the data and
    write the newline.

    Fix #21664

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit 48fe356485535c0691829a2189465adcca9c1221)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 36ede88958
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 15 08:26:47 2026 -0400

    vwr: Require the original RF header to be present

    The IxVeriWave RF header has 100 bytes; we only supply 76 of them as
    part of the packet record, skipping over 20 bytes in the middle and
    4 at the end. When checking that the record is long enough to be valid,
    use the expected original length, not the length of what we supply after
    skipping some fields. This prevents accepting short records and filling
    the port fields with uninitialized bytes.

    Thanks to AISLE Security for reporting

    Fix #21663

    (cherry picked from commit bc671a906bc0b8fe1b3790482c0d98128e46be03)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit d95a611941
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 15 08:24:35 2026 -0400

    BT HSP, BT HFP: Bound AT command reassembly

    An AT command or response that never ends kept growing the same
    reassembly chain without limit. The unsigned offset of each fragment
    could in principle wrap (after 4 GiB), making the reassembly buffer
    smaller than the fragments copied into it, and every terminator
    character in a frame walked back through the entire chain to find its
    first fragment, which is quadratic in the length of the chain.

    Start a new message once the reassembled length would exceed 64 KiB,
    and store the first fragment of each chain instead of searching for it.
    V.250 and the HSP/HFP specifications don't limit the length of an AT
    command or response; this limit is arbitrary, well above any real
    command line, chosen to keep the offsets from overflowing.

    Thanks to AISLE Security for the report

    Fix #21653

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit d277f7c5c21215dd16b06a59d941e54193095679)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 4f07e6dcd6
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Oct 14 21:14:35 2026 -0400

    radiotap: Don't allow remaining length of EHT TLV to wrap

    Retrieve the EHT TLV length (an unsigned 16 bit integer) but store it
    in a signed int so that the remaining length becomes negative instead
    of wrapping (whether the original length is too short or if the length
    isn't a multiple of 4.) This stops the loop at the check instead of
    dissecting later TLVs and the 802.11 frame itself as EHT user info
    fields.

    We don't implement true proper handling of short TLVs according to the radiotap
    documentation, which says "Parsers shall be able to parse TLV fields with
    partial data present as indicated by the length of the data, i.e. producers
    can omit bytes they don’t fill at the end of the data. Such bytes are assumed
    to be 0." It's not clear any producers do that, though.

    Ping #21661

    (cherry picked from commit aae6da75c524dc8525373ae1a2bd634592e0e278)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 5d893f339e
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Oct 12 23:49:45 2026 -0400

    Aethra: Discard the data of skipped records on the first pass

    The Aethra wiretap module intends to ignore records that are not
    LAPD frames. However, when reading sequentially, the skipped records
    data are read into the record Buffer and kept, so the next frame
    returned has a Buffer which starts with all the skipped record data
    concatenated before its own (but with a caplen in the record taken
    from its own value.) On later passes, when wtap_seek_read is used,
    the record is retrieved correctly because the file_offset is still
    updated for each record, so this is most visible on one-pass TShark
    output, which differs from two-pass TShark and Wireshark view.

    Clean the buffer at the end of the loop if not breaking out of the loop.

    Thanks to AISLE Security for observing the issue (it is not really a
    security issue, because the amount of excess memory allocated to the
    first record is still no greater than what is contained in the file.)

    Fix #21654

    (backported from commit f1c00f80182d6ea20793736a83e9569282df985d)

commit 610917bbaa
Author: John Thacker <johnthacker@gmail.com>
Date:   Tue Oct 13 08:30:18 2026 -0400

    NR RRC: add recursion depth checks to recursive DCCH segments

    Similar to what was done in 7b809476 for LTE RRC dissector

    (cherry picked from commit c4863020c319986f26a73eb8e0592d972430976c)

    Co-authored-by: Pascal Quantin <pascal@wireshark.org>

commit b52dce88a5
Author: John Thacker <johnthacker@gmail.com>
Date:   Tue Oct 13 08:29:23 2026 -0400

    LTE RRC: Add recursion depth checks to recursive DCCH segments

    When LTE RRC DCCH segment reassembly is enabled (it is not by default),
    the resulting reassembly is parsed again as a full DCCH message. This
    recursion is not detected by clang-tidy or by the checks in asn2wrs.py,
    perhaps because of the complex interaction with the conformance file.
    It's not clear if the tree depth checks are sufficient, or if it's actually
    possible to construct a PoC that exhausts a 8MiB stack size. However,
    we can add increment/decrement dissection depth checks as a safeguard.

    Fix #21647

    (cherry picked from commit 7b809476730c2aedcd1b596893e5320dd1db3ce6)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit b87b352aaf
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Oct 12 21:25:20 2026 -0400

    BT L2CAP: Don't underflow the remaining LE SDU length

    In LE Credit Based Flow Control mode, if the first frame of an SDU
    declares an SDU length shorter than its payload, or a later frame is
    longer than the remaining length, the unsigned remaining length
    wrapped. Every later frame on the channel was then treated as a
    continuation of an SDU that never completes, and added to the reassembly.

    Treat a first frame whose payload is at least the SDU length as a
    complete SDU (with an expert info if the SDU length is too short),
    and a continuation frame that is at least the remaining length as the
    last one.

    A first frame whose L2CAP length is less than 2 is too short for the
    SDU length field. Previously, if more bytes were captured than the
    length, it started a segmented SDU and the 16-bit length minus 2
    wrapped, so the following frames were swallowed as continuations.
    If fewer were captured, reading the SDU length threw on the first pass
    before the frame data was stored, so the second pass dissected the frame
    differently. Mark such frames as too short, end any SDU in progress,
    and show them as payload with an expert info, before reading anything
    from the tvb.

    Thanks to AISLE Security for the report.

    Fix #21648

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit 828c0f2b42a2259af29c7444af9b57d026256d18)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit a7d654df7f
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 11:28:53 2026 -0400

    Kerberos: Check the PAC signature buffer lengths

    When verifying a PAC, the checksum type was read from the first four
    bytes of the server, KDC, ticket and full signature buffers without
    checking their length. krb5_pac_parse() accepts a zero length buffer,
    and krb5_pac_get_buffer() then returns a one byte allocation, so a
    malformed PAC caused a heap over-read.

    Only use a signature buffer that is long enough to hold the checksum
    type. Later code already checks the length of the checksum itself.

    This requires Kerberos decryption to be enabled and a build with MIT
    Kerberos PAC verification.

    Thanks to AISLE Research for the report

    Fix #21644

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>
    (backported from commit f57fdc80f23b0fc4f2bdffbc03b754632eb0c814)

commit 61ddb9deba
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Oct 12 20:59:10 2026 -0400

    BT SDP: Don't copy all the continuation data for each response

    Each response in a continuation sequence allocated a new file scope
    buffer for all the attribute list bytes so far plus its own, and kept
    all the previous ones, so memory grew quadratically with the number of
    responses (even responses with no attribute list bytes made a new copy
    of everything). The total length was also an unchecked 32-bit sum, and
    the length was stored before the copy from the tvb, which could throw
    and leave uninitialized bytes to be dissected on later passes.

    Store each response's bytes once, linked to the previous responses',
    and assemble the attribute list in packet scope when dissecting.

    Limit the total to 1 MiB. Each response's AttributeListByteCount is
    16 bits, but the specification does not limit the total across
    continuation responses; this limit is arbitrary, chosen to keep the
    total length from overflowing and to bound the buffer allocated when
    the final response is dissected.

    Thanks to Aisle Research for reporting.

    Fix #21649

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit d7fefcbe390db22d3bc93ed33e7ff36f2f1c8b5a)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit b1f7b7bba5
Author: Guy Harris <gharris@sonic.net>
Date:   Mon Oct 12 11:13:44 2026 -0700

    dumpcap: pass the netmask to pcap_compile() in host byte order.

    The documentation for pcap_lookupnet() says it returns the netmask in
    network byte order.

    The documentation for pcap_compile says the netmask should be provided
    in *host* byte order.

    tcpdump does this correctly, by calling ntohl() on the netmask to
    convert it to host byte order.

    Make dumpcap do the same thing using g_ntohl().

    Fix !26881.

    (cherry picked from commit 647b08553b9ff4bf15bc1878834dd8ac30d89b4b)

commit aa0111b52d
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Oct 12 10:52:20 2026 -0400

    E2AP: Fix gNB assignment used where lookup intended

    Fix #21646

    (cherry picked from commit 2e8f5dbebc35b0b6a91c930b8ec23ae283397617)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 5c29fb727f
Author: John Thacker <johnthacker@gmail.com>
Date:   Mon Oct 12 07:41:09 2026 -0400

    NTLMSSP: Compute the NTLMv2 proof over the whole response

    create_ntlmssp_v2_key() copied the NTLMv2 response after the NT proof
    into a 768 byte stack buffer, using MIN() so that the copy fit, but
    then passed the untrimmed length to ws_hmac_buffer(). A response
    longer than 776 bytes (they can be up to NTLMSSP_BLOB_MAX_SIZE)
    therefore made the HMAC read past the end of the buffer. It also meant
    that legitimate responses with long AV pair lists could never be
    verified, as the proof was computed over the wrong data.

    Build the server challenge and the rest of the response in a buffer of
    the right size, once, before trying each credential.

    This requires NT credentials or an RC4 keytab entry to be configured.

    Thanks to Yuyao Jiang for reporting.

    Fix #21643

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit 622de8cda8205b5a7175431b52cf6d0d9f773c1c)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit c7adf78d16
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 21:09:42 2026 -0400

    SPNEGO: Reject an empty RC4-HMAC wrap payload when decrypting

    decrypt_arcfour() didn't check that the wrap token had any data after
    the header. With nothing there, the output buffer is a zero length
    allocation (NULL from wmem_alloc), and gssapi_verify_pad() read the
    padding length from the byte before it.

    Return an error for an empty payload, and make gssapi_verify_pad()
    reject a zero length too.

    This requires an RC4-HMAC key in the keytab that decrypts the token's
    sequence number correctly.

    Thanks to AISLE Security for the report.

    Fix #21642

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit b3c47dc2a8c2a70bd386cceb99b6177b4a3fb3f8)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 07260ac4a2
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 18:27:43 2026 -0400

    k12 text: Don't match an unlimited number of hours digits

    The Tektronix K12 text format stores a time of day field at the
    beginning of each frame. The lexer should not match an unlimited number
    of digits for the hours token, as then a file which matches the magic
    bytes at the beginning can read for an extremely long time, taking up
    memory and CPU time and creating a Denial of Service.

    Limit the hours fields to 2 digits at maximum, treating earlier digits
    as junk. The parser could be made stricter against errors (or could
    allow three or four digit hours, if that's actually possible in the
    format, so long as unlimited is not allowed.)

    Thanks to AISLE Research for the report and PoC.

    (cherry picked from commit 78ead179aa68c0e0adfb4e9ff6b81f65657e5655)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 2e36691493
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 15:41:34 2026 -0400

    IEEE 1609.2: Handle gmtime() failure when formatting times

    The Time32 and Time64 formatters passed the result of gmtime()
    straight to snprintf. gmtime() returns NULL for times outside the
    range the C library can represent; on Windows that is anything after
    the year 3000, which a Time64 from packet data easily exceeds, so the
    dissector crashed when building the tree.

    Use ws_gmtime_r() and show "Not representable" when it fails.

    Thanks to Aisle Security for the report.

    Fix #21640

    (cherry picked from commit e13e9f9046ddf2f2c66e193ff9cc733583efac81)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 8c18d52e03
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 11:24:43 2026 -0400

    logcat text: Don't match an unread line buffer when opening

    logcat_text_open() reads lines into a g_malloc'd buffer until it
    gets one that isn't short or special, and then matches the buffer
    against the formats. If no line could be read at all, as for an
    empty file, the buffer had never been written, and matching it read
    uninitialized memory, possibly past the end of the buffer. Return
    "not mine" (or the read error) in that case.

    Fix #21639

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (cherry picked from commit 4c0455a50f6b3b8d4d46b969b00df07bced554bc)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit aec9fb83be
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 11:14:40 2026 -0400

    snoop: Error before allocation if a pseudo-header size is too large

    In the Shomiti/Finisar Surveyor flavor of the snoop format, the
    pseudo-header is variable length and given in the header. If the
    pseudo-header length is larger than the total packet length, go
    ahead and fail with an error instead of allocating a very large
    Buffer after integer overflow, and then failing to read bytes.

    That Buffer is freed shortly afterwards with the failure, so this
    near-4GiB allocation is only a problem for systems that might fail
    to make the allocation.

    Use checked integer arithmetic, and make the header length unsigned.

    Thanks to AISLE Security for the report and PoC.

    (cherry picked from commit 76459b81349ecc2c1d82dec1de2e032ebf2f731f)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 66a70bb36a
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 11:11:36 2026 -0400

    dfilter: Check for macro syntax errors in argument numbers

    The argument numbers are 1-indexed. Using argument zero ($0) will
    crash if the macros is actually used. Argument numbers that
    overflow an int also crash. Fail when checking and creating the
    macro instead.

    Thanks to AISLE Security for their report.

    (cherry picked from commit cb72ee52df856a8a736d7e50bd1c6572ae966497)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 86391d0d67
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 11:08:42 2026 -0400

    wiretap: busmaster: Bound scanner tokens

    Each time Flex refills its buffer it copies the partial token to the
    start of the buffer, and refills are capped at 16 KB, so scanning an
    unbounded token takes time quadratic in its length. 10 MB of spaces
    took 23 seconds to read, and a 10 MB line in the channel section of
    the header took over a minute.

    Put an upper bound on every pattern. Skip the header's channel and
    database sections, and the text after ERR, in pieces up to the end
    of the line instead of matching the whole line as one token.

    Assisted-by: Claude Opus 5.5

    (cherry picked from commit c21af83800d43e1d79da08f6603b056cc1915c0f)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit b114a6c573
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 08:57:11 2026 -0400

    DCM: check for a potential overflow on the last packet

    Commit 17689584d70f15419d4ae9de53d57ceb60cdb351 added an overflow
    check for any packet where pdv_curr->is_last_fragment is not true,
    but forgot to check for the last memmove that comes right after
    the while loop.

    Related-To #21497
    Fixup 17689584d70f15419d4ae9de53d57ceb60cdb351
    AI-Assisted: no

    (cherry picked from commit f553d3f796ff62e82a06f7168f071b234ba91bc0)

    Co-authored-by: Dirk Müller <dirk@dmllr.de>

commit b1668088ec
Author: John Thacker <johnthacker@gmail.com>
Date:   Sun Oct 11 08:12:46 2026 -0400

    nettrace: Don't preallocate an INT_MAX-sized buffer

    nettrace_3gpp_32_423_file_open() reserved RINGBUFFER_START_SIZE ==
    INT_MAX for the streaming buffer. g_array_sized_new() rounds that up to
    the next power of two, so it is really a 2 GiB g_realloc(), for a buffer
    that read_until() then fills 1024 bytes at a time. glib treats an
    allocation failure as fatal, so where the request cannot be satisfied
    the process aborts during file open, before any packet is parsed. Any
    XML file with a <traceCollecFile> root reaches this.

    Reserve 64 KiB instead and let the append path grow the buffer.

    Impact on 64-bit systems is low: the mapping is lazy, so the untouched
    2 GiB never becomes resident, and such machines generally have well over
    2 GiB of RAM anyway. It matters on 32-bit builds and under an address
    space limit (RLIMIT_AS, cgroups) or without overcommit, where the
    allocation genuinely fails and glib aborts. It was also a waste for
    legitimate traces, as the reservation is never released.

    Thanks to Aisle Research for the report.

    Fixes #21507

    (cherry picked from commit 008a0dacdb78b3e05bdb30c6fc0b2228028faed8)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit 6ad8dce24e
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Oct 11 10:17:55 2026 +0000

    [Automatic update for 2026-10-11]

    Update manuf, services, enterprise numbers, translations, and other items.

commit 7f16cd62f0
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Oct 10 10:01:53 2026 -0400

    logcat text: Check binary entries before writing them as text

    When converting binary logcat records to a text format, the
    dumper trusted the entry header: it didn't check that the record
    held the header and the payload, it found the end of the tag with
    strlen, and it computed the message length as payload length minus
    the tag length. The binary reader only validates the first three
    entries of a file, so a later entry with a short payload or a tag
    without a terminating NUL made the dumper read past the record,
    or pass a negative length to g_strndup and abort.

    Check the record against the header and payload lengths, and find
    the end of the tag within the payload.

    Fix #21636

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>
    (backported from commit 87e96a1d1c1c66cfbbacec06c04883bc434272ca)

commit 75021b2094
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Oct 10 17:45:21 2026 -0400

    ERF: Use the Module snaplen as the interface default

    When an interface section doesn't have its snaplen set, the code checked
    to see if the module section had a snaplen. However, instead of using
    the snaplen from the module if set in that case, it interpreted the last
    tag of the interface section as the snaplen, which at best is the wrong
    value or at worst could be past the end of the record by 4 or fewer
    bytes. Correctly use the saved module snaplen, similar to what is done
    for the FCS length.

    Most of the time the negative effects are reading possibly uninitialzed
    bytes from the Buffer (also possibly leftover from the previous record),
    though it may be possible to construct a case that reads past the
    allocated Buffer data itself.

    Fix #21635

    Thanks to AISLE Security for the report and proposed fix.

    (backported from commit bec0686a0e50482bbac24861e603187ee42f8725)

commit 817f931028
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Oct 10 14:22:32 2026 -0400

    observer: Return -1 for failure when packet data cannot be read

    Similar to 7be61b1f22b7f713814b3952ed2e37a1ef1984f5 but for Observer.
    The calling functions expect a negative value for failure, not zero,
    which is treated as success with zero bytes (which have not been
    read into the buffer despite caplen having a positive value.)

    Most of the time this results in copying uninitialized data. It might
    be possible to craft a unusual file that tries to copy outside the
    Buffer, but current master or release-4.6 does have a check for this.

    Thanks to AISLE Security for the report.

    (backported from commit 279f438a04d2a75e75256725260edba7edde3574)

commit cc59e57a90
Author: John Thacker <johnthacker@gmail.com>
Date:   Sat Oct 10 08:16:25 2026 -0400

    NBAP, UMTS FP: Don't write past the per-channel arrays

    The NBAP DCH chain loops checked num_dch_in_flow < FP_maxNrOfDCHs
    before incrementing it and using it as an index, and then wrote one
    more entry after the loop, so a flow with the maximum number of DCHs
    wrote past dch_ids_in_flow_list and fp_dch_channel_info. Stop one
    entry earlier and remove the extra write, which nothing reads.

    The number of UL and DL transport formats per channel was incremented
    for each item, but not every TransportFormatSet resets it, so it could
    exceed MAX_FP_CHANS. Set it from the (bounded) item count instead.

    UMTS FP's per-channel arrays only have MAX_FP_CHANS entries, fewer
    than the number of DCHs NBAP allows in a flow, so check the count in
    the consumer too, and clamp the channel count read from K12 files.

    Thanks to AISLE Research for the report.

    Fix #21630

    Assisted-by: Claude Opus 5.5 <noreply@anthropic.com>

    (backported from commit 6d2bd70dc4a6f46bf094a5888b33596b003c9b63)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 4e94aa6774
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 9 20:54:04 2026 -0400

    peektaged:  return -1 for error in peektagged_read_packet

    Some error paths returned false/0 instead of <0
    and trigger callers to believe the buffer was filled with
    packet data when it was not.

    Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

    (cherry picked from commit 7be61b1f22b7f713814b3952ed2e37a1ef1984f5)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit fc30b73244
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 9 20:49:06 2026 -0400

    RTPS: Fix OOB write in DATA_BATCH sample info list

    dissect_RTPS_DATA_BATCH() allocates sample_info_flags[] and
    sample_info_length[] with sample_info_max entries, then fills them in the
    sample-info loop indexed by sample_info_count. The loop's bounds check
    only fired when the max_batch_samples_dissected preference was > 0:

      if (rtps_max_batch_samples_dissected > 0 && ... >= rtps_max_batch_samples_dissected)

    When the preference is set to 0 (a valid value meaning "show up to 1024"),
    sample_info_max is set to 1024 for the allocation, but the guard is
    disabled entirely. A crafted DATA_BATCH submessage with many minimal
    sample-info records (~8 bytes each) then drives sample_info_count past
    1024, writing sample_info_flags[] and sample_info_length[] out of bounds
    in packet-scope (wmem) memory, and the stale entries are read again when
    the batched samples are dissected.

    Guard on sample_info_max instead, which equals the preference when it is
    > 0 and 1024 when it is 0, so the loop can never exceed the allocation in
    either case.

    This requires the non-default preference value of 0; with the default
    (16) the OOB write is not reachable.

    Reported by Aisle Research.

    Fixes #21508

    (cherry picked from commit 716a200295d2f91629211fd8660e8cce37c13884)

    Co-authored-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>

commit a57511ba14
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 9 20:47:11 2026 -0400

    ERF: Bound metadata dump scan by captured length, not wlen

    erf_dump_priv_compare_capture_comment() walked the Provenance (META)
    record's tags using the record's on-wire length (pseudo_header->erf.phdr.wlen)
    as the remaining-bytes bound, while pd points to a buffer holding only
    caplen (= rlen - header) bytes. A record whose wlen exceeds its captured
    payload therefore made erf_meta_read_tag() read past the end of the
    buffer. wlen is attacker-controlled (16-bit, up to 65535) and is not
    constrained to caplen on read.

    Use the captured length instead, matching the read side
    (populate_summary_info() already parses metadata using the rlen-derived
    packet_size).

    The read path is safe, and stock tshark/editcap do not reach this writer
    code when re-exporting ERF (they copy IDBs, which the ERF writer rejects,
    aborting before any packet record is written). It is reachable from any
    libwiretap consumer that writes ERF without copying IDBs, e.g. via
    wtap_dump_params_init_no_idbs(): feeding a crafted ERF META record
    (caplen 8, wlen 65535, one Capture Section tag) through wtap_dump() to an
    ERF dumper produces tens of thousands of invalid reads under Valgrind in
    erf_meta_read_tag(); none after this change.

    Thanks to AISLE Security for reporting

    Fixes: #21629

    Assisted-by: Claude Opus 4.8

    (cherry picked from commit 7cb4b5ca791f8ed2457392ca90e6cfa9f275be13)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 036df7a350
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Oct 9 16:02:46 2026 -0400

    Kerberos: Trim the saved key's keylength when truncating it

    While the largest key length we expect to get in valid data is 32 bytes
    (e.g., AES-256), this is obtained from cleartext packet data and a BER
    ASN.1 OCTET STRING with no constraint  If we truncate the key bytes when
    storing them to KRB_MAX_KEY_LENGTH, record the truncated length in the
    stored key, similar to what is done in keytab_file_read in read_keytab_file.c
    The keylength can be used for hashing and comparisons using the bytes,
    and thus risks memory-safety errors otherwise.

    Fixup 20bfb9ffb90cac37892630e1c24743c8b6800471
    Fix #21625

    Thanks to Yuyao Jiang for reporting

    (cherry picked from commit 78fd235858bc6195ab34114c0f2a2a880cc68a48)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit ebc86cf10f
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 8 09:34:30 2026 -0400

    wiretap: ascend: Bound scanner tokens and line length

    The scanner's YY_INPUT supplies one byte at a time, and each time Flex
    refills its buffer it copies the partial token to the start of the
    buffer, so scanning an unbounded token takes time quadratic in its
    length. Several patterns were unbounded (hex numbers, the character
    display after the hex bytes, quoted task names, the WDD call number
    and type, and others); a 400 KB run of hex digits took over nine
    minutes to read. Put an upper bound on every pattern, and consume the
    character display in pieces up to the end of the line.

    Also limit lines to 1024 characters (real lines are around 80), and
    return an error for a longer one.

    When the parser was converted from Bison to Lemon, run_ascend_parser()
    went from returning 0 on success to returning true, but its callers
    were not updated. ascend_open() could never fail on an I/O error, and
    parse_ascend() reported "no data returned by parse" for parse errors
    and vice versa. Fix both.
    Fixup f21cd2e23f372bd42e4792ef0b4294d2db881a92

    Thanks to AISLE security for a report.

    Assisted-by: Claude Opus 5.5
    (backported from commit 08ed4cf790de9f9c7bb1ef17bbf15b2d1308f25f)

commit d77a7e8039
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Oct 8 21:19:40 2026 -0400

    wiretap: RTPDump: Limit the length of the header text

    The text after the magic number was appended to a GString until a
    newline, so a file starting with the magic number followed by a long
    run of printable characters would use memory proportional to the file
    size. Return an error once it exceeds RTP_HEADER_MAX_LEN.

    Thanks to AISLE Security for a report.

    Assisted-by: Claude Opus 5.5

    (cherry picked from commit 1543d27ca60c7ac0d6ba3f62b572848b86b8f43c)

    Co-authored-by: John Thacker <johnthacker@gmail.com>

commit 10649f5784
Author: Pascal Quantin <pascal@wireshark.org>
Date:   Tue Oct 6 09:51:15 2026 +0200

    NRPPa: add dissection of TRP-MeasurementReportList IE

    Also fix the dissection of Cell-ID IE.

    Fixes #21615

    (cherry picked from commit 2c348f1e5b57f6f9f20c783f0248358d65f5c196)

    Conflicts:
            epan/dissectors/asn1/nrppa/nrppa.cnf
            epan/dissectors/packet-nrppa.c

commit 02a4ae0bae
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Oct 4 10:22:04 2026 +0000

    [Automatic update for 2026-10-04]

    Update manuf, services, enterprise numbers, translations, and other items.

commit 052b5d3c1a
Author: John Thacker <johnthacker@gmail.com>
Date:   Wed Sep 30 10:42:35 2026 -0400

    K12 Text: Use a GString for writing

    We already set a limit on the maximum frame we are willing to write,
    based on WTAP_MAX_PACKET_SIZE_STANDARD, but the fixed output buffer
    allocated is not large enough to handle such a size. Compounding the
    issue, the return values for snprintf are not correctly handled in
    the case of insufficient room.

    Instead of a fixed buffer and snprintf, allocate a GString with
    sufficient capacity based upon the caplen and the header and footer.
    libwiretap is able to read this maximum length; if this is too large
    for some third party purposes, the maximum should be reduced elsewhere.

    Fix #21606

    (backported from commit 3729b649680e3ff878e128a6e39d236020d870fb)

commit 6cdb6dcd0f
Author: Anders Broman <a.broman58@gmail.com>
Date:   Mon Sep 28 11:26:34 2026 +0200

    OPCUA: Use built-in ExtensionObject parsers only for ns 0

    getExtensionObjectType() returned only the numeric part of the TypeId
    and dropped the namespace, so an ExtensionObject of a custom type whose
    numeric ID matches a standard encoding ID was decoded with the wrong
    built-in parser, e.g. ns=1;i=886 silently as Range. Return 0 (no
    built-in parser) for TypeIds outside namespace 0; such bodies are then
    shown as ByteString, like string, Guid and ByteString TypeIds already
    are.

    Also fix the NodeId encoding comments (0x04 is Guid, 0x05 ByteString,
    there is no 0x06) and add the first OPC UA suite test.

    Assisted-by: Claude Code

    (cherry picked from commit a3bdb60df6cf6b7fff8f0c3fa38ed17a7f4315cd)

    Co-authored-by: Michael Kirsche <michael.kirsche@codewerk.de>

commit 7c1920ee11
Author: Jaap Keuter <jaap.keuter@xs4all.nl>
Date:   Mon Sep 28 11:16:56 2026 +0000

    Qt: Switch endpoint map to OpenStreetMap

    Switch tile server to OSM, while adding the necessary HTTP header to
    the requests made to it.

    Fixes #21578
    Assisted-by: Mistral.ai

    (cherry picked from commit 7198c957032d531e081d11673bbbbded1fea7a56)

    Co-authored-by: Jaap Keuter <jaap.keuter@xs4all.nl>

commit 1a125e5c1b
Author: Jaap Keuter <jaap.keuter@xs4all.nl>
Date:   Sun Sep 27 20:50:10 2026 +0000

    capture: initialize start_time to prevent uninit var warning

    Also, move start_time setting closer to actual call to measure, i.e.
    sync_pipe_run_command_actual()

    (cherry picked from commit 20011cf089f08fe4e6c4c2930c6a3e11535e59e1)

    Co-authored-by: Jaap Keuter <jaap.keuter@xs4all.nl>

commit 3203b12d07
Author: Gerald Combs <gerald@wireshark.org>
Date:   Sun Sep 27 10:35:52 2026 +0000

    [Automatic update for 2026-09-27]

    Update manuf, services, enterprise numbers, translations, and other items.

    manuf failed.

commit b3609ef11c
Author: John Thacker <johnthacker@gmail.com>
Date:   Fri Sep 25 21:53:03 2026 -0400

    Fix inverted return value check

    Issue #21520 introduced a inversion of the caller return value check
    without actually changing the return value of the function, so this
    appears to be an oversight.

    If decrypt_lorawan_frame_payload actually fails and returns false,
    leads to a NULL pointer or uninitialized data dereference.

    (cherry picked from commit 6ca46d12eda1035be15d6d0d6c58e829a2814f6a)

    Co-authored-by: Dirk Müller <dirk@dmllr.de>

commit 6873368472
Author: John Thacker <johnthacker@gmail.com>
Date:   Thu Sep 24 08:50:01 2026 -0400

    RTSE: Global proto_tree considered harmful

    The proto_tree objects are automatically freed at the end of a frame;
    having them dissector level globals leads to a possibility of use after
    free when a dissector has multiple entry points, some of which are PDUs
    called by other dissectors where top_tree is intended to be NULL. The
    old top tree might not get set to NULL when an exception occurs.

    Set it in proto_data for the protocol, which cleans up automatically at
    the correct time.  Make the indentation somewhat more consistent.

    Fix #21585. Thanks to Antti Levomäki & Christian Jalio, Forcepoint for
    the report and PoC.

    (backported from commit 8574da1dc1d7250c69eb76213c07c649b6d5a704)

commit eda2cc9e61
Author: Gerald Combs <gerald@wireshark.org>
Date:   Wed Sep 23 13:59:16 2026 -0700

    Version: 4.4.19 → 4.4.20 [skip ci]
