Vulnerabilidades de Linux

In the Linux kernel, the following vulnerability has been resolved: nvmet: fix NULL pointer dereference in nvmet_execute_identify_nslist() When a host issues an Identify command with CNS 07h (Active Namespace ID List for a specific I/O Command Set), nvmet_execute_identify_nslist() is called with match_css set. The command-set filter dereferences req->ns, but this handler never calls nvmet_req_find_ns(), so req->ns is always NULL (nvmet_req_init() resets it to NULL). As soon as an enabled namespace with an NSID greater than the requested value exists, req->ns->csi dereferences a NULL pointer and oopses. Besides the crash, the comparison is logically wrong: to filter the list by command set it must test the command set of the namespace being iterated, not a single fixed value. Use the loop variable ns->csi.

In the Linux kernel, the following vulnerability has been resolved: ALSA: dummy: Check card index validity at probe snd_dummy_probe() blindly trusts that the given devptr->id value is within the proper card index range. It's OK for the devices the driver itself creates at the module probe time, but if the device is bound manually via sysfs interface, this could be -1 as "none", and this leads to OOB access for index[] and other parameters. Add a sanity check for the card index and warn/correct it if it's a value out of the range.

In the Linux kernel, the following vulnerability has been resolved: io_uring/cmd: fix iovec leak when the async cmd is not recycled An io_async_cmd carries an iovec array in ->vec.iovec, allocated when the vec has to grow and kept across recycling through ctx->cmd_cache. On two paths nothing frees it and io_clean_op()'s kfree(req->async_data) drops the io_async_cmd without it. io_req_uring_cleanup() clears the async data flags only when io_alloc_cache_put() succeeds, and the cache holds IO_ALLOC_CACHE_MAX == 128 entries, so once it is full the put fails and the vec is left behind. An NVMe passthrough workload gets there without doing anything unusual: nvme_uring_cmd_io() returns -EIOCBQUEUED, so the io_async_cmd stays attached for the lifetime of the command and the live object count tracks the queue depth. Above 128 the puts start failing. ->cleanup is the last chance to free an inherited vec, since io_req_uring_cleanup() returns early for an io-wq issued command and is not called at all for one completed without ever being issued. But io_clean_op() calls ->cleanup only if REQ_F_NEED_CLEANUP is set, and for uring_cmd that happens only where the vec has to grow, so a command reusing a large enough cached vec never sets it. io_rw_alloc_async() and io_msg_alloc_async() flag an inherited vec for exactly this reason; io_uring_cmd_prep() does not. Flag an inherited vec in io_uring_cmd_prep(), and free the vec when the cache put fails, as io_req_rw_cleanup() does. The leak is invisible under KASAN, where io_alloc_cache_vec_kasan() frees the vec unconditionally.

In the Linux kernel, the following vulnerability has been resolved: io_uring/rsrc: fix folio size overflow in io_vec_fill_bvec() io_vec_fill_bvec() computes the folio size with a plain int 1: unsigned long folio_size = 1 << imu->folio_shift; imu->folio_shift is unsigned int and comes from folio_shift() of the folio backing the registered buffer, so it can be 32 or more on a 64 bit kernel. Shifting int 1 that far is undefined, and on x86 and arm64 the count is taken modulo 32, so a shift of 34 yields 4 rather than 16G. Every other folio_shift shift in this file already uses 1UL. The result is that the segment estimate and the fill loop disagree. io_estimate_bvec_size() sizes the bvec array with the real shift: max_segs += (iov[i].iov_len >> shift) + 2; so a 1M iovec on a 16G folio is charged 2 segments, while io_vec_fill_bvec() then walks the same iovec in folio_size chunks of 4 bytes and writes res_bvec[bvec_idx] a quarter of a million times, past the end of the array it was given. src_bvec is advanced once per iteration as well, so imu->bvec is read past its end at the same time. validate_fixed_range() only checks that the range is inside the registered buffer and does not bound the segment count. Reaching it needs a folio with a shift of at least 32, which means a gigantic hugetlb page: 16G on arm64 with 64K pages, where CONT_PMD_SHIFT is 34 and hugetlb_add_hstate(CONT_PMD_SHIFT - PAGE_SHIFT) registers that size, and likewise on powerpc. x86_64 tops out at 1G, so a shift of 30, which still fits in int and is unaffected. Use 1UL, as the rest of the file does.

In the Linux kernel, the following vulnerability has been resolved: ocfs2: fix missing metadata reservation for large xattrs [BUG] lsetxattr() panics the kernel when setting a large xattr value on a fragmented filesystem where the file already has an external xattr block. [CAUSE] ocfs2_calc_xattr_set_need() never reserves metadata blocks for a new xattr value's extent tree when the file already has an external xattr block. The not_found path leaves meta_add at zero, so meta_ac is NULL when ocfs2_xattr_extend_allocation() runs. A new value root has room for a single extent record. On a fragmented filesystem, the allocator cannot satisfy the xattr value in one contiguous run, so each non-contiguous run requires its own extent record. When the value root's extent list is full and meta_ac is NULL, ocfs2_add_clusters_in_btree() returns RESTART_META, and ocfs2_xattr_extend_allocation() hits BUG_ON(why == RESTART_META). [FIX] The case where no xattr block exists yet already calls ocfs2_extend_meta_needed(&def_xv.xv.xr_list) to reserve value tree metadata. Add the same reservation to the case where an xattr block already exists, making the two cases consistent. Replace the BUG_ON with a -ENOSPC return so that if RESTART_META is returned despite the reservation, the error propagates to userspace instead of panicking the kernel.

In the Linux kernel, the following vulnerability has been resolved: ext4: stop retrying saturated xattr cache entries ext4_xattr_block_set() retries when a cache entry selected for reuse has a saturated reference count after taking the buffer lock. The retry returns to the mbcache lookup without making that entry ineligible, so it can select the same unusable entry indefinitely. A task spinning there can hold the parent directory's i_rwsem and leave concurrent rmdir callers blocked. Normally a reusable entry has a reference count below EXT4_XATTR_REFCOUNT_MAX because the count and MBE_REUSABLE_B are updated under the same buffer lock. A corrupted filesystem can violate that invariant. The syzbot reproducer reports allocator and xattr corruption before triggering this retry loop. Check the untrusted on-disk count before incrementing it, avoiding overflow, and clear MBE_REUSABLE_B when it is already saturated. The next lookup then skips the entry that was just proven unusable. This mirrors the normal transition at EXT4_XATTR_REFCOUNT_MAX; the release path marks the entry reusable again on the exact 1024-to-1023 transition. Using the same QEMU harness and guest parameters, current unpatched Linux hung in 6 of 8 420-second trials with the do_rmdir signature; representative NMI backtraces caught the owner spinning in ext4_xattr_block_set(). The patched kernel completed 28 of 28 trials without a hung-task report; the final twelve trials exercised the reviewed overflow-safe form of the change. syzbot's patch testing also completed without reproducing the hang.

In the Linux kernel, the following vulnerability has been resolved: nilfs2: reject invalid block index in GC ioctl Syzbot reported list corruption caused by a double list_add_tail() call on bh->b_assoc_buffers within nilfs_lookup_dirty_data_buffers(). Analysis revealed that the root cause was the insertion of a page/folio with a page index of ULONG_MAX into the page cache via the GC ioctl. filemap_get_folios_tag(), called by nilfs_lookup_dirty_data_buffers(), repeatedly detects a dirty folio with a page index of ULONG_MAX due to index wrap-around, leading to duplicate processing of dirty buffers. As a preparatory step, the GC ioctl loads the page/folio of the block to be moved during GC and inserts it into the page cache based on information in the nilfs_vdesc structure passed as an argument. Normally, this does not cause issues because the user-space GC library configures the nilfs_vdesc structure properly. However, since there is no range check on the parameters determining the page index, a request with artificially crafted parameters -- such as those generated by Syzbot -- can result in a page/folio being inserted with a page index of ULONG_MAX, triggering the above problem. This resolves the issue by checking the ranges of 'vd_offset' and 'vd_vblocknr' in the nilfs_vdesc structure that determine the page index, thereby preventing the invalid page/folio insertions.

In the Linux kernel, the following vulnerability has been resolved: ext4: don't enable DAX on new encrypted files Currently, when a new encrypted regular file is created, the call to ext4_set_inode_flags(inode, init=true) in __ext4_new_inode() is made before EXT4_INODE_ENCRYPT is set. As a result, it can set S_DAX if the filesystem is mounted with "-o dax=always". EXT4_INODE_ENCRYPT then actually gets set a bit later in __ext4_new_inode(), when it calls fscrypt_set_context() which calls ext4_set_context(). ext4_set_context() sets EXT4_INODE_ENCRYPT and calls ext4_set_inode_flags(inode, init=false) to set S_ENCRYPTED too. This was intended to clear S_DAX as well. However, this was broken by commit 043546e46dc7 ("fs/ext4: Only change S_DAX on inode load"). This causes data written to the file to bypass encryption, also causing xfstests failures such as generic/548 (when "-o dax=always" is used). Fix this by simplifying the flow by making __ext4_new_inode() set EXT4_INODE_ENCRYPT earlier. This makes it take effect in ext4_set_inode_flags(inode, init=true), making S_DAX never be set. Similarly, make EXT4_STATE_MAY_INLINE_DATA never be set in the first place on new encrypted inodes. Then it doesn't need to be cleared. As a result of these simplifications, ext4_set_context() no longer needs to change inode flags or state when 'handle != NULL'. Remove that too.

In the Linux kernel, the following vulnerability has been resolved: xfs: validate attr entry pointer before field access xfs_attr3_leaf_verify_entry() accesses lentry/rentry fields (namelen, valuelen) before checking if the entry pointer itself is within bounds. If nameidx is crafted to point near the end of the buffer, these field accesses can read out-of-bounds before the bounds check at name_end > buf_end is performed. Add explicit bounds checks for entry pointers before accessing their fields. Use offsetof() to check that the start of the flexible array member (nameval/name) is within bounds, which ensures all preceding fields are safe to access.

In the Linux kernel, the following vulnerability has been resolved: xfs: restore nofs context unconditionally in xfs_trans_roll When __xfs_trans_commit() fails in xfs_trans_roll(), the NOFS context is cleared but only restored in the success path. This leaves the error path without nofs protection, causing a circular lock dependency between xfs_nondir_ilock_class and fs_reclaim: CPU0 CPU1 ---- ---- lock(&xfs_nondir_ilock_class); lock(fs_reclaim); lock(&xfs_nondir_ilock_class); lock(fs_reclaim); Fix this by moving xfs_trans_set_context() before the error check so that nofs context is always restored on the new transaction.

In the Linux kernel, the following vulnerability has been resolved: nfc: digital: clamp SENSF_RES length to the destination buffer digital_in_recv_sensf_res() memcpy()s resp->len bytes from a remote NFC-F device response into the NFC_SENSF_RES_MAXSIZE-byte target.sensf_res field without an upper-bound check. A nearby malicious NFC-F device can send an oversized SENSF_RES response to overflow the stack-local struct nfc_target. Clamp resp->len to NFC_SENSF_RES_MAXSIZE before the copy. Found by 0sec automated security-research tooling (https://0sec.ai).

In the Linux kernel, the following vulnerability has been resolved: nfc: fdp: bound the device-reported read length and fix an skb leak fdp_nci_i2c_read() takes the next packet length from two device-supplied bytes and never validates it. The value is a u16 used as the i2c_master_recv() count into a 261-byte on-stack buffer: a malicious, counterfeit or malfunctioning controller (or an i2c bus interposer) can drive it far past the buffer for a stack out-of-bounds write that clobbers the canary and return address, or below the minimum frame size (directly, or by truncating the computed sum) so the header/LRC strip and the next length read run past a short receive. Reject a length outside [FDP_NCI_I2C_MIN_PAYLOAD, FDP_NCI_I2C_MAX_PAYLOAD], as a corrupted packet already is, and force resynchronization. The same loop allocates one data skb per iteration and assumes a length packet followed by a data packet; a device that sends two data packets in one call leaks the first skb when the second allocation overwrites it. Free a previously allocated skb before allocating the next.

In the Linux kernel, the following vulnerability has been resolved: nfc: microread: validate target discovery payload lengths microread_target_discovered() parses target discovery payloads from skb->data according to the HCI gate. The fixed field offsets and UID copies were checked only against the destination nfc_target buffers, not against the actual skb length. Validate that each gate-specific payload contains the fixed fields and UID bytes before reading or copying them.

In the Linux kernel, the following vulnerability has been resolved: nfc: llcp: bound the connect_sn TLV walk to the skb Commit 27256cdb290e ("nfc: llcp: bound SNL TLV parsing to the skb and add length checks") fixed the unbounded TLV walk in nfc_llcp_recv_snl(), and commit d8bd2dedbde5 ("nfc: llcp: fix OOB read and u8 offset wrap in TLV parsers") subsequently bounded nfc_llcp_parse_gb_tlv() and nfc_llcp_parse_connection_tlv(). One sibling parser sharing the same pattern remains unbounded: nfc_llcp_connect_sn(). nfc_llcp_connect_sn() walks a TLV list, reading a two-byte header (type, length) followed by length bytes of value, without checking that the two header bytes or the declared length stay within the buffer. It returns a pointer to a service name of up to 255 bytes that may point past the end of the skb; it is subsequently consumed by memcmp() in nfc_llcp_sock_from_sn(). In addition tlv_array_len was computed as "skb->len - LLCP_HEADER_SIZE" in size_t, so a CONNECT/CC frame shorter than the LLCP header underflows to a huge length and the walk runs far past the buffer. nfc_llcp_connect_sn() is reachable from nfc_llcp_recv_connect() and nfc_llcp_recv_cc(), i.e. from received CONNECT and CC PDUs. A nearby NFC device can reach this without authentication; LLCP link activation happens automatically after NFC-DEP, and the nfc_llcp_rx_skb() dispatcher applies no minimum-length guard. Walk the TLV list by pointer, bounded by skb_tail_pointer(skb), and validate each declared length before use, matching the approach already used for nfc_llcp_recv_snl(). Starting the walk at &skb->data[LLCP_HEADER_SIZE] against the tail pointer also removes the size_t underflow for short frames. Found by 0sec automated security-research tooling (https://0sec.ai).

In the Linux kernel, the following vulnerability has been resolved: nfc: llcp: fix OOB read and u8 offset wrap in TLV parsers nfc_llcp_parse_gb_tlv() and nfc_llcp_parse_connection_tlv() contain three related bugs in their TLV parsing loops: 1. 'offset' is declared u8 but tlv_array_len is u16. When TLV data advances offset past 255 it silently wraps to zero, causing infinite loops or double-processing of buffer data. 2. Before reading tlv[0] (type) and tlv[1] (length) there is no check that offset+2 <= tlv_array_len. A truncated TLV causes an OOB read of one byte past the buffer end. 3. After reading the length field, the value bytes are accessed without checking offset+2+length <= tlv_array_len. A crafted length=0xFF on a short buffer causes up to 255 bytes of OOB read past the buffer end. Both functions are reachable without authentication via nfc_llcp_set_remote_gb() which feeds remote LLCP general bytes directly into nfc_llcp_parse_gb_tlv() with no additional validation. Fix all three issues by widening offset from u8 to u16 and adding bounds checks for both the TLV header and value field before each access.

In the Linux kernel, the following vulnerability has been resolved: nfc: llcp: reject PDUs shorter than the LLCP header Every LLCP PDU begins with a two-byte header (DSAP/SSAP + PTYPE), but the receive path never checked that a frame is at least LLCP_HEADER_SIZE bytes before parsing it. nfc_llcp_rx_skb() reads the header via nfc_llcp_ptype()/nfc_llcp_dsap()/ nfc_llcp_ssap(), which dereference pdu->data[0] and pdu->data[1], and a CONNECT or CC PDU then computes tlv_array_len = skb->len - LLCP_HEADER_SIZE; as a size_t and hands it to the TLV walk. When the frame is shorter than the header the subtraction wraps to a huge value and the walk runs far past the buffer, an out-of-bounds read. A nearby NFC device can reach this without authentication; LLCP link activation happens automatically after NFC-DEP. Guard the common receive choke point __nfc_llcp_recv(), shared by both the target (nfc_llcp_data_received()) and initiator (nfc_llcp_recv()) paths, so a short skb is dropped before the rx_work worker parses it. Use pskb_may_pull() rather than a skb->len test so the two header bytes are guaranteed to sit in the skb linear area even for a non-linear skb, matching how the sibling NCI and HCI receive paths validate their headers. Reproduced with a KFENCE out-of-bounds read via /dev/virtual_nci on linux-next. Found by 0sec automated security-research tooling (https://0sec.ai).

In the Linux kernel, the following vulnerability has been resolved: nfc: pn533: purge fragmented skbs during cleanup pn53x_common_clean() purges resp_q before freeing the common PN533 state, but it leaves fragment_skb untouched. The fragmentation helpers queue transmit fragments there while sending large initiator or target-mode frames, and those skbs remain owned by the driver until they are sent or discarded. If the device is removed while fragments are still queued, the common cleanup path frees the PN533 state without releasing the queued fragment skbs, leaking them. Purge fragment_skb during cleanup alongside resp_q.

In the Linux kernel, the following vulnerability has been resolved: nfc: nci: add data_len bound checks to activation parameter extractors nci_extract_activation_params_iso_dep() and nci_extract_activation_params_nfc_dep() read an inner length byte from the NCI RF_INTF_ACTIVATED_NTF payload and use it to memcpy() into fixed kernel buffers, but neither function receives the caller-validated activation_params_len. A crafted NCI notification with activation_params_len=1 and an inner length byte of up to 20 (NFC-A) or 50 (NFC-B) causes memcpy() to read that many bytes past the one valid byte in the activation params region -- a slab out-of-bounds read of kernel memory adjacent to the NCI skb. The sibling nci_extract_rf_params_*() family was given equivalent protection by commit 571dcbeb8e63 ("net: nfc: nci: Fix parameter validation for packet data"), but the two activation parameter extractors were not updated at that time. Add a data_len parameter to both functions, guard against an empty region before consuming the inner length byte, decrement the remaining count after consuming it, and clamp the copy length to what is actually available. Update both call sites to pass ntf.activation_params_len, which is already validated against the skb at ntf.c:801.

In the Linux kernel, the following vulnerability has been resolved: nfc: nci: fix out-of-bounds write in nci_target_auto_activated() nci_target_auto_activated() appends a target to the fixed-size array ndev->targets[NCI_MAX_DISCOVERED_TARGETS] and increments ndev->n_targets without first checking the array is full; unlike its sibling nci_add_new_target(), which bails out when n_targets already equals NCI_MAX_DISCOVERED_TARGETS. ndev->n_targets is only cleared by nci_clear_target_list(), so an NFCC that repeatedly re-runs discovery (RF_DISCOVER_RSP, which re-enters NCI_DISCOVERY without clearing the target list) and reports an auto-activated target (RF_INTF_ACTIVATED_NTF) drives n_targets past the limit. The append then writes a struct nfc_target past the end of the array (a slab out-of-bounds write), and nfc_targets_found() goes on to walk the array with the inflated count: BUG: KASAN: slab-out-of-bounds in nci_add_new_protocol+0x94/0x2ac [nci] Write of size 2 at addr ffff0000c7299a18 by task kworker/u8:0/12 Workqueue: nfc0_nci_rx_wq nci_rx_work [nci] Call trace: nci_add_new_protocol+0x94/0x2ac [nci] nci_ntf_packet+0xddc/0x11a0 [nci] nci_rx_work+0x15c/0x1e0 [nci] process_one_work+0x2dc/0x500 worker_thread+0x240/0x460 kthread+0x1c0/0x1d0 ret_from_fork+0x10/0x20 The buggy address belongs to the cache kmalloc-2k of size 2048 The buggy address is located 1024 bytes to the right of allocated 1560-byte region [ffff0000c7299000, ffff0000c7299618) Guard nci_target_auto_activated() with the same check used by nci_add_new_target().

In the Linux kernel, the following vulnerability has been resolved: nfc: nci: fix uninit-value in the RF discover/activated NTF handlers nci_rf_discover_ntf_packet() and nci_rf_intf_activated_ntf_packet() each parse a notification into an on-stack struct (nci_rf_discover_ntf / nci_rf_intf_activated_ntf) that is not initialised. The RF technology-specific parameters are only extracted when rf_tech_specific_params_len is non-zero, so a notification that reports a zero length leaves the rf_tech_specific_params union uninitialised - and both handlers then pass it to nci_add_new_protocol(), which reads it: - discover: nci_add_new_target() -> nci_add_new_protocol(); - activated: nci_target_auto_activated() -> nci_add_new_protocol(). nci_add_new_protocol() uses nfca_poll->nfcid1_len as both a branch condition and a memcpy() length and copies nfcid1/sens_res/sel_res into ndev->targets, which is later exposed to user space via NFC_CMD_GET_TARGET. BUG: KMSAN: uninit-value in nci_add_new_protocol+0x624/0x6c0 nci_add_new_protocol+0x624/0x6c0 nci_ntf_packet+0x25b2/0x3c30 nci_rx_work+0x318/0x5d0 process_scheduled_works+0x84b/0x17a0 worker_thread+0xc10/0x11b0 kthread+0x376/0x500 Local variable ntf.i created at: nci_ntf_packet+0xbc2/0x3c30 Zero-initialise both on-stack notifications so the union reads back as zero when no technology-specific parameters are present.

In the Linux kernel, the following vulnerability has been resolved: ipv4: reject undersized MTUs in ip_do_fragment() ip_do_fragment() subtracts the IPv4 header length from the effective MTU and passes the resulting payload MTU to ip_frag_next(). If the effective MTU is smaller than hlen + 8, ip_frag_next() rounds the fragment payload length down to zero. The fragmentation state then never makes forward progress: state->left, state->ptr and state->offset stay unchanged while ip_do_fragment() keeps allocating and transmitting header-only fragments until the softlockup detector fires. This is reproducible with a route installed using "mtu lock 20", but it is also reproducible without route MTU lock, for example by forwarding a packet to a device whose MTU is 20. Fix it in ip_do_fragment() by rejecting mtu < hlen + 8 with -EMSGSIZE, matching the existing IPv6 fragmentation check.

In the Linux kernel, the following vulnerability has been resolved: ipv6: fix use-after-free in ip6_finish_output2() ip6_finish_output2() caches a pointer to the IPv6 destination address (daddr) before invoking lwtunnel_xmit(). The LWT-BPF transmit path or other encapsulation operations within lwtunnel_xmit() can reallocate the skb head, freeing the memory that daddr points to. When lwtunnel_xmit() returns LWTUNNEL_XMIT_CONTINUE, the function continues to use the stale daddr pointer to compute the nexthop and to look up or create the neighbour entry. This results in a use-after-free read, which can leak sensitive kernel data, pollute the neighbour table with arbitrary values, misdirect traffic, or crash the system. Fix this by re-fetching the IPv6 header and the destination address pointer after lwtunnel_xmit() returns LWTUNNEL_XMIT_CONTINUE, ensuring that the subsequent nexthop computation and neighbour lookup operate on valid memory.

In the Linux kernel, the following vulnerability has been resolved: nvmet-auth: zero the AUTH_RECEIVE response buffer nvmet_execute_auth_receive() allocates the response buffer with kmalloc() sized by the host-supplied AUTH_RECEIVE allocation length, but the DH-HMAC-CHAP builders write only a fixed-size message into it. The full allocation length is then copied to the wire by nvmet_copy_to_sgl(), so a remote initiator receives the bytes past the built message -- up to nearly a page of uninitialized slab -- during the pre-authentication handshake. Allocate the buffer with kzalloc() so the unwritten tail is zeroed before it is sent; conforming responses are unaffected.

In the Linux kernel, the following vulnerability has been resolved: nvmet-fc: fix invalid free in LS IOD error path nvmet_fc_alloc_ls_iodlist() advances iod while initializing the LS IOD array. If an rqstbuf allocation or response buffer DMA mapping fails, the unwind loop decrements iod past the start of the array. The final kfree(iod) therefore frees an address before the allocated object. This can be reproduced with nvme-fcloop and failslab by setting fail-nth to 6 before creating a target port. KASAN reports: BUG: KASAN: invalid-free in nvmet_fc_register_targetport Free of addr ffff88816cf8ff48 by task nvmet_fail_nth/9552 Free the original allocation base stored in tgtport->iod instead. With this fix applied, the same sysfs write with fail-nth=6 returns -ENOMEM without any KASAN report.

In the Linux kernel, the following vulnerability has been resolved: nvmet-tcp: bound SGL data length before allocating command buffers nvmet_tcp_map_data() reads the host-controlled 32-bit sgl->length and, for the in-capsule offset descriptor (type 0x01), checks it against port->inline_data_size before use. Any other SGL descriptor type -- including the non-inline transport SGL data-block descriptor (type (NVME_TRANSPORT_SGL_DATA_DESC << 4) | NVME_SGL_FMT_TRANSPORT_A, the type a real host uses for out-of-capsule writes) skips that check entirely and falls straight through to: cmd->req.sg = sgl_alloc(len, GFP_KERNEL, &cmd->req.sg_cnt); with len taken directly from the wire, unbounded up to 4 GiB. nvmet_req_init() only parses the command and never inspects sgl->length, and nvmet_check_transfer_len() -- the only other place transfer_len is validated -- runs later, from req->execute(), after the allocation has already happened. For a write command the target responds with an R2T and parks the command waiting for the host to send the data; if the host (or an unauthenticated peer that simply never follows up) never does, the sgl_alloc() buffer stays resident for the life of the command. NVMe/TCP has no mandatory authentication in the default configuration, so any peer able to reach the target portal and complete a Fabrics connect can drive this with a single crafted command, repeatable across queues and connections for amplification. This is unbounded kernel memory allocation triggered by a remote, effectively unauthenticated peer. Validate len against the same NVMET_TCP_MAXH2CDATA ceiling this file already uses to bound per-PDU H2C data, for every SGL descriptor type, before doing any allocation. This closes the gap for the non-inline descriptor while leaving the existing, tighter inline_data_size check in place for the in-capsule case. Runtime-verified on a v6.19 KASAN stand: with this bound in place, a crafted write command carrying an oversized non-inline SGL length is rejected before sgl_alloc() runs, where the same request previously drove an unbounded ~256 MiB kernel allocation (up to 4 GiB) that stayed resident pending an R2T the host never satisfies.