-
CVE-2026-97570
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
bnxt_en: Bound SW TPA IDs to prevent crashes
FW supports up to 1024 concurrent TPAs, so the FW TPA ID is in the range
0..1023 (see commit ec4d8e7cf024 ("bnxt_en: Add TPA ID mapping logic for
57500 chips.")). bnxt_alloc_agg_idx is intended to wrap the FW ID down to a
software ID which is used to index rxr-rx_tpa, and to generate a mapping
between FW IDs and the wrapped software ID.
On a 57608 with firmware version 233, the firmware advertises 32
concurrent TPAs. As of the commit under fixes, bp->max_tpa on this NIC
is set to 32.
If the software ID from bnxt_alloc_agg_idx is above 31, this results in
an invalid address being loaded on this line:
tpa_info = &rxr->rx_tpa[agg_id];
because rx_tpa is allocated with only bp->max_tpa (32) entries. Writes
to tpa_info later in the code are out of bounds.
This bug results in a crash at boot:
Oops: general protection fault, kernel NULL pointer dereference 0x8: 0000 [#1] SMP NOPTI
RIP: 0010:bnxt_rx_pkt+0xc0/0x1560
RSP: 0018:ffffc900009b8c78 EFLAGS: 00010246
RAX: 0000000000000000 RBX: 0000000000000048 RCX: 0000000206682516
RDX: ffffc900009b8db4 RSI: 0000000000000000 RDI: 01ffffff038fe1c0
RBP: ffffc9006e687480 R08: ffffc9006e687000 R09: 0000000000003048
R10: 0000000000000480 R11: ffff8881c6083900 R12: 0000000006682516
R13: ffff8881c6095400 R14: 0000000000000016 R15: ffff8881c6b66680
FS: 0000000000000000(0000) GS:ffff88fef3c77000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007fc8bda40584 CR3: 000000807c812001 CR4: 0000000008772ef0
PKRU: 55555554
Call Trace:
IRQ>
? __netif_receive_skb_list_core+0x1ca/0x250
__bnxt_poll_work+0x152/0x280
bnxt_poll_p5+0x1cd/0x480
__napi_poll+0x30/0x180
net_rx_action+0x20b/0x3b0
? note_gp_changes+0x53/0xe0
? tick_setup_sched_timer+0x180/0x180
? __napi_schedule+0x9a/0xb0
? bnxt_msix+0x24/0x30
handle_softirqs+0xdd/0x2c0
__irq_exit_rcu.llvm.3171231171502365008+0x47/0xf0
common_interrupt+0x85/0x90
asm_common_interrupt+0x22/0x40
This stack trace is from a crash triggered when an out of bounds rx_tpa
is dereferenced. The invalid write mentioned above is silent in this
particular crash.
Fix this by allocating rx_tpa with bp->max_tpa rounded up to the next
power of 2 (bp->max_tpa_roundup_size) entries and masking the FW TPA ID
with that size, so the wrapped ID can never index past the end of the
array.
-
CVE-2026-97567
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
mptcp: prevent race between disconnect() and rtx
Sashiko noted that the two event can race, leading to inconsistent
status. Prevent the race using the synchronous timer stop operation.
-
CVE-2026-97566
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
mptcp: pm: kernel: drop pending ADD_ADDR when removing ID0
The in-kernel MPTCP path manager can leave a stale ADD_ADDR announcement
entry alive when removing the id 0 endpoint. This happens because the id 0
removal path does not tear down pending announcements, unlike the non-zero
id path.
When the PM later reselects id 0 after adding another signal endpoint, it
finds the stale anno_list entry and hits WARN_ON_ONCE(mptcp_pm_is_kernel())
in mptcp_pm_announced_alloc().
Root cause: asymmetry between removal paths.
- Non-zero id path: mptcp_nl_remove_subflow_and_signal_addr() calls
mptcp_pm_remove_announced() to clean up.
- Id 0 path: mptcp_nl_remove_id_zero_address() skips cleanup entirely.
Fix by making the id 0 path symmetric: call mptcp_pm_announced_remove()
and decrement add_addr_signaled before queuing the RM_ADDR.
Subtle detail: signal endpoints are stored in anno_list with port 0, but
msk_local carries the connection's local port. In other words, entries
linked to ID0 paths should have port == 0. A follow-up patch will ensure
that. mptcp_pm_announced_remove() uses use_port=true for comparison. So
clear the port before the lookup.
-
CVE-2026-97565
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
smb: client: reject short READ responses in CIFSSMBRead()
CIFSSMBRead() reads DataLengthHigh, DataLength and DataOffset out of
the READ_RSP returned by the server without first checking that a
whole READ_RSP was actually received. The length of the response is
recorded in rsp_iov.iov_len, but nothing constrains it to be at least
read_rsp_size before those fields are dereferenced.
A malicious or compromised SMB1 server can return a response shorter
than the READ_RSP header, so that parsing the header itself reads past
the end of the receive buffer. SMB1 is not negotiated by default;
reaching this code requires an explicit vers=1.0 mount.
Reject the response unless it is at least read_rsp_size bytes long.
-
CVE-2026-97562
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
smb: client: pin DFS superblock in iterator callback
tcon_super_cb() stores a raw superblock pointer, but __cifs_get_super()
takes its active reference only after iterate_supers_type() has dropped
s_umount and its passive reference. Concurrent DFS automount expiry can
therefore free the superblock before cifs_sb_active() uses it.
A deterministic KASAN test reproduces the race as:
BUG: KASAN: slab-use-after-free in cifs_sb_active+0x77/0x80
The same test passes with this change applied.
Take the active reference in the callback while iterate_supers_type()
still holds s_umount shared. cifs_put_tcp_super() remains the matching
release.
-
CVE-2026-97559
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
smb: client: fail DACL rewrite when the new DACL exceeds 64K
replace_sids_and_copy_aces() and set_chmod_dacl() accumulate the size of
the DACL they build in a u16. That accumulator can wrap.
validate_dacl() caps num_aces at (dacl_size - sizeof(struct smb_acl)) /
20, i.e. 3276 for a maximally sized DACL, while each rewritten ACE can
grow to sizeof(struct smb_ace) (76 bytes) once its SID is replaced with
one carrying SID_MAX_SUB_AUTHORITIES sub-authorities. The worst case is
therefore sizeof(struct smb_acl) + 3276 * 76 = 248984 bytes, far beyond
what a u16 can hold. A wraparound is reached with 863 ACEs.
After the wraparound, ndacl_ptr-size becomes meaningless and the offset
will point anywhere in the ACE array. As a result, we will see
corruption of the DACL, which then gets sent to the server. This is not
an out-of-bounds write as the allocation now covers the worst-case
expansion, so writes will always go into the buffer.
Adjust the code to use a u32 internally and return -EOVERFLOW in the
overflow case. The operation must be refused, because a DACL can only
hold 2^16-1 bytes on the wire and larger DACLs cannot be represented.
set_chmod_dacl() carries the same pattern and is fixed the same way. It
only wraps once the source DACL comes within roughly 380 bytes of the
64K ceiling, but the failure mode is identical.
-
CVE-2026-97556
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
smb: client: avoid leaking refcount when cifs_sb_tlink() fails
cifs_oplock_break() takes over the reference that
cifs_queue_oplock_break() acquired when it queued the work, and drops it
with _cifsFileInfo_put() once the break has been processed.
Only in setups with "-o multiuser", cifs_sb_tlink() may fail, at which
point cifs_oplock_break() returns without putting the file reference,
mirroring the reference leak we already fixed in the companion patch to
cifs_queue_oplock_break().
This would trigger a crash due to busy inodes on the next unmount:
BUG: Dentry ... still in use (1) [unmount of cifs cifs]
VFS: Busy inodes after unmount of cifs (cifs)
Drop the reference on that path as well. Doing so before the out label
mirrors the normal path, which also puts the reference before
cifs_done_oplock_break().
Found by Sashiko code review. The failure path was not exercised at
runtime.
-
CVE-2026-97552
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
xfs: initialise error in xfs_defer_finish_one()
xfs_defer_finish_one() declares error without an initialiser and only
assigns it inside the loop over dfp-dfp_work. When that list is empty
the loop body never runs, control falls through to the "Done with the
dfp, free it" path, and the function returns an indeterminate value.
An item-less pending item reaches this through xfs_defer_add_barrier(),
which xfs_reap_ag_blocks() uses on any CONFIG_XFS_ONLINE_REPAIR kernel.
xfs_defer_finish_noroll() treats any non-EAGAIN return as fatal, so a
non-zero stack value turns a successful barrier into a
SHUTDOWN_CORRUPT_INCORE in the middle of a repair. Zero is the correct
result: reaching the free path means the item loop drained without a
non-zero error.
-
CVE-2026-97544
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
xfs: don't leak dqacct if rhashtable insertion fails
LOLLM observes that xqcheck_mod_live_ino_dqtrx doesn't free the newly
allocated dqa object if rhashtable insertion fails. Fix this leak.
-
CVE-2026-97536
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
scsi: qla2xxx: Fix use-after-free of qpair work on queue teardown
The response queue MSI-X handler qla2xxx_msix_rsp_q() schedules
qla_do_work() via queue_work(ha-wq, &qpair->q_work). qla_do_work()
dereferences the qpair (vha, rsp) and takes qpair->qp_lock.
During teardown, qla2xxx_delete_qpair() deletes the response queue, which
calls free_irq() in qla25xx_free_rsp_que(), and then frees the queue and
the qpair. free_irq() waits for running hardirq handlers but does not
cancel work already placed on ha->wq. A still-pending q_work then runs
qla_do_work() against the freed qpair and response queue, causing a
use-after-free. This is especially likely during full adapter teardown,
where destroy_workqueue(ha->wq) forces pending work to run after the queue
pairs have been freed.
Flush the work item with cancel_work_sync() in qla25xx_free_rsp_que()
after free_irq() has released the interrupt (so no new work can be
queued) and before the response queue and qpair memory are freed (so the
flushed handler still sees valid memory). Guard on rsp->qpair and ha->wq
to match the INIT_WORK() condition and avoid operating on an
uninitialized work_struct.
-
CVE-2026-97526
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
s390/pai: Support CPU hotplug for PMU PAI
The command 'perf stat -e pai_crypto/CRYPTO_ALL/ -- command'
crashes the kernel when CPUs are hotplug added during that run.
Root cause is the missing allocation of per-CPU data structures
for that new CPU. The allocation is dynamic and the first
event that has task context creates such a structure for
each online CPU. This is not sufficient. CPUs may be offline
during event creation and can be set online during the
perf run time. For example commands
# echo 0 > /sys/devices/system/cpu/cpu1/online
# perf stat -e cycles -i -- stress-ng -t10s --matrix X
# sleep 1
# echo 1 > /sys/devices/system/cpu/cpu1/online
Currently without a CPU hotplug handler, that new CPU has no
per-CPU data infrastructure. The scheduler runs PMU call back
function pai_add() to install the PMU support for that CPU before
the task is being scheduled on that new CPU.
In pai_add() instructions
mp = this_cpu_ptr(pai_root[idx].mapptr);
cpump = mp->mapptr;
return a NULL pointer and the result is a kernel panic as variable
cpump is used inside that function.
Add CPU hotplug support for CPU add and delete and create
the necessary per-CPU data infrastructure during CPU hotplug
add processing. Same for CPU hotplug remove.
This is done when the CPU is offline to ensure the data structures
are available when CPU is made online and tasks are scheduled on it.
[hca@linux.ibm.com: fixup error path in pai_init()]
-
CVE-2026-97525
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
x86/mm/pat: Allocate split page tables as kernel page tables
A PTE is allocated directly without going through the standard page table
allocation routines (such as pte_alloc_one_kernel()) when the CPA code
splits a large page (__split_large_page()).
This means the page table constructor is never called nor is the page table
marked as a kernel page table.
The former results in the folio associated with the page table not being
marked as a page table (__pagetable_ctor() is never called thus neither is
__folio_set_pgtable()) nor are statistics updated to reflect
it (lruvec_stat_add_folio() is never called).
The latter issue of failing to mark the page table as a kernel page
table (ptdesc_set_kernel() is never called) is far more problematic.
Since commit:
5ba2f0a15564 ("mm: introduce deferred freeing for kernel page tables")
kernel page table freeing has been batched and since the
subsequent commit:
e37d5a2d60a3 ("iommu/sva: invalidate stale IOTLB entries for kernel address space")
IOTLB cache entries for kernel page tables have been invalidated upon
being freed.
Since split page tables are freed without this invalidation, the IOTLB
can contain stale entries for them.
Resolve the issue by using the ordinary PTE allocation API at split time.
This results in these kernel page tables invoking a page table constructor,
and thus requires a page table destructor.
Destructors are not always present, like for early allocated direct map
page tables). Conditionally call pagetable_dtor_free() if the PG_table
folio flag for the ptdesc is set, otherwise we free the page table via
pagetable_free().
Regardless of which path is taken page tables marked as kernel page tables,
which now includes split page tables, take the correct route through
pagetable_free_kernel().
There is a user-visible side effect in that split page tables will appear
in nr_page_table_pages in /proc/vmstat (as do other kernel page tables
allocated after early boot), however this is a positive change.
This issue started being markedly problematic after commit:
5ba2f0a15564 ("mm: introduce deferred freeing for kernel page tables")
so choose this as the Fixes target.
[ dhansen: rephrase in imperative mood ]
-
CVE-2026-97524
•
published on September 25, 2026
In the Linux kernel, the following vulnerability has been resolved:
mptcp: avoid unneeded actions on subflow reset
Once in a blue moon, the mptcp receive path can recursively call
mptcp_data_ready() via state change under unlucky error conditions, and
then try to hold the data lock again.
Break the recursion loop explicitly checking for the exceptional
condition.
Add a new flag instead of using an existing one like 'closing', to exit
early in subflow_state_change(), and explicitly flush the RX queue at
reset time.
This avoids unneeded processing to check for available data -- calling
get_mapping_status() and more on a dying subflow -- but also in error
reporting and worker scheduling.
Note that we must consume the currently peeked skb before invoking
mptcp_dss_corruption to avoid consuming it again after the eventual
reset has freed it.
-
CVE-2026-97508
•
published on September 24, 2026
In the Linux kernel, the following vulnerability has been resolved:
thunderbolt: Set tb-root_switch to NULL when domain is stopped
Similarly what we do with the firmware connection manager. This makes
tb_xdp_handle_request() return error to the remote host. However, we
need to make sure we keep the uuid alive so that we can reply until the
whole domain is released.
-
CVE-2026-97506
•
published on September 24, 2026
In the Linux kernel, the following vulnerability has been resolved:
crypto: ixp4xx - fix buffer chain unwind on allocation failure
chainup_buffers() builds a linked list of buffer descriptors for a
scatterlist. If dma_pool_alloc() fails while constructing the list, the
current code sets buf to NULL and later dereferences it unconditionally
at the end of the function:
buf-next = NULL;
buf->phys_next = 0;
This can lead to a null-pointer dereference on allocation failure.
If the failure happens after part of the descriptor chain has already
been allocated and DMA-mapped, the partially constructed chain also
needs to be released.
Fix this by terminating the partially constructed chain on allocation
failure and letting the callers unwind it via their existing cleanup
paths. Also fix ablk_perform() to preserve the hook pointers before
checking for failure, so partially built chains can be freed correctly.
-
CVE-2026-97502
•
published on September 24, 2026
In the Linux kernel, the following vulnerability has been resolved:
mmc: davinci: avoid NULL deref of host-data in IRQ handler
mmc_davinci_irq() returns early only when both host->cmd and
host->data are NULL:
if (host->cmd == NULL && host->data == NULL) {
...
return IRQ_NONE;
}
So we may legitimately reach the rest of the handler with
host->data == NULL (and therefore data == NULL). The DATDNE branch
already guards against this with an explicit "if (data != NULL)"
check, but the subsequent TOUTRD ("read data timeout") and
CRCWR/CRCRD ("data CRC error") branches dereference data
unconditionally:
if (qstatus & MMCST0_TOUTRD) {
data->error = -ETIMEDOUT; -- NULL deref
...
davinci_abort_data(host, data);
}
if (qstatus & (MMCST0_CRCWR | MMCST0_CRCRD)) {
data->error = -EILSEQ; <-- NULL deref
...
}
If either bit is set in qstatus while host->data is NULL, the kernel
will crash inside the IRQ handler. smatch flags this:
drivers/mmc/host/davinci_mmc.c:933 mmc_davinci_irq() error: we
previously assumed 'data' could be null (see line 914)
Gate both branches on a non-NULL data, matching the existing pattern
used by the DATDNE branch.
No functional change for callers where data is non-NULL, which is
the only case in which these branches did meaningful work before
this change.
-
CVE-2026-97501
•
published on September 24, 2026
In the Linux kernel, the following vulnerability has been resolved:
pinctrl: mediatek: paris: bypass pinctrl GPIO layer in set GPIO direction
pinctrl_gpio_direction_input() / pinctrl_gpio_direction_output() take
the pinctrl mutex. This causes a gpiochip operations to need to sleep.
Worse yet, the .can_sleep field in the gpiochip is not set. This causes
the shared GPIO proxy to trip over, as it uses gpiod_cansleep() to check
whether it can use a spinlock or needs a mutex. In this case, it ends
up taking a spinlock, then calls pinctrl_gpio_direction_output(), which
takes a mutex. This causes a huge warning.
While this class of Mediatek hardware does not have separate clear/set
registers, the pinctrl context has a spinlock that is taken whenever
a register read-modify-write is done. Also, once the GPIO function is
selected / muxed in, further GPIO operations do not involve pinctrl
operations or state. The GPIO direction and level values do not require
toggling the pinmux or any other pin config options.
Switch to directly calling mtk_pinmux_gpio_set_direction() in the GPIO
set direction callbacks to avoid taking the pinctrl mutex. Drop the
.gpio_set_direction field in mtk_pmxops to signal we are no longer using
the pinctrl GPIO layer for setting the direction.
-
CVE-2026-97500
•
published on September 24, 2026
In the Linux kernel, the following vulnerability has been resolved:
wifi: rtw89: phy: check length before parsing PHY status IE
Hardware might report PHY status IE with unexpected length, and parser
might access out of range. Check the length ahead.
-
CVE-2026-97499
•
published on September 24, 2026
In the Linux kernel, the following vulnerability has been resolved:
coresight: perf: Retrieve path and source from event data
ETM perf callbacks currently use the per-CPU csdev_src pointer, which
can race with updates during device registration and unregistration.
The AUX setup already builds and stores the path in the event data.
Use this path to retrieve the source instead of csdev_src to avoid
the race.
Export coresight_get_source() and add etm_event_get_ctxt_path() to
retrieve the context's path and its source with READ_ONCE() /
WRITE_ONCE() accessors. Give the comments to explain why this
approach is safe when pause or resume callbacks preempt the disable
callback (e.g. via NMI).
-
CVE-2026-97496
•
published on September 24, 2026
In the Linux kernel, the following vulnerability has been resolved:
drm/amdkfd: Fix OOB memory exposure in get_wave_state()
The get_wave_state() function for v9 trusts cp_hqd_cntl_stack_size and
cp_hqd_cntl_stack_offset values read directly from the MQD, which are
written by GPU microcode and fully attacker-controlled on the
CRIU-restore path (via AMDKFD_IOC_RESTORE_PROCESS with H3).
this leads to an unbounded copy_to_user() that can leak adjacent
GTT/kernel memory. If offset size, integer underflow produces a ~4 GiB
read length, if size is set to 1 MiB against a 4 KiB allocation, we leak
1 MiB of adjacent kernel memory (other queues' MQDs, ring buffers, KASLR
pointers).
Fix by clamping both cp_hqd_cntl_stack_size to the actual allocated
buffer size (q->ctl_stack_size) and cp_hqd_cntl_stack_offset to the
clamped size before performing arithmetic and copy_to_user().
This ensures we never read beyond the allocated kernel BO regardless of
attacker-supplied MQD field values.