| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| In the Linux kernel, the following vulnerability has been resolved:
usb: gadget: f_fs: serialize DMABUF cancel against request completion
ffs_epfile_dmabuf_io_complete() calls usb_ep_free_request() on the
completed request but leaves priv->req, the back-pointer that
ffs_dmabuf_transfer() set on submission, pointing at the freed
memory. A later FUNCTIONFS_DMABUF_DETACH ioctl or
ffs_epfile_release() on the close path still sees priv->req
non-NULL under ffs->eps_lock:
if (priv->ep && priv->req)
usb_ep_dequeue(priv->ep, priv->req);
so usb_ep_dequeue() is called on a freed usb_request.
On dummy_hcd the dequeue path only walks a live queue and
pointer-compares, so the freed pointer reads without faulting and
KASAN requires an explicit check at the FunctionFS call site to
surface the use-after-free. On SG-capable in-tree UDCs the
dequeue path dereferences the supplied request immediately:
* chipidea's ep_dequeue() does
container_of(req, struct ci_hw_req, req) and reads
hwreq->req.status before acquiring its own lock.
* cdnsp's cdnsp_gadget_ep_dequeue() reads request->status first.
The narrower option of clearing priv->req via cmpxchg() in the
completion does not close the race: the completion runs without
eps_lock, so a cancel path holding eps_lock can still observe
priv->req non-NULL, race a concurrent completion that clears and
frees, and pass the freed pointer to usb_ep_dequeue(). A slightly
longer fix that moves the free into the cleanup work is needed.
Same class of lifetime race as the recent usbip-vudc timer fix [1].
Take eps_lock in the sole place that mutates priv->req from the
callback direction by moving usb_ep_free_request() out of the
completion into ffs_dmabuf_cleanup(), the existing work handler
scheduled by ffs_dmabuf_signal_done() on
ffs->io_completion_wq. Clear priv->req there under eps_lock
before freeing, and only clear if priv->req still names our
request (a subsequent ffs_dmabuf_transfer() on the same
attachment may have queued a new one).
This keeps the existing dummy_hcd sync-dequeue invariant: the
completion callback is still invoked by the UDC without
eps_lock held (dummy_hcd drops its own lock before calling the
callback), and the callback now takes no f_fs lock at all.
Serialization against the cancel path happens in cleanup, which
runs from the workqueue with no f_fs lock held on entry.
The priv ref count protects the containing ffs_dmabuf_priv:
ffs_dmabuf_transfer() takes a ref via ffs_dmabuf_get(), cleanup
drops it via ffs_dmabuf_put(), so priv stays live for the
cleanup even after the cancel path's list_del + ffs_dmabuf_put.
The ffs_dmabuf_transfer() error path no longer frees usb_req
inline: fence->req and fence->ep are set before usb_ep_queue(),
so ffs_dmabuf_cleanup() (scheduled by the error-path
ffs_dmabuf_signal_done()) owns the free regardless of whether
the queue succeeded.
Reproduced under KASAN on both detach and close paths against
dummy_hcd with an observability hook
(kasan_check_byte(priv->req) immediately before usb_ep_dequeue)
at the two FunctionFS cancel sites to surface the stale-pointer
access; the hook is not part of this patch. The KASAN
allocator / free stacks in the captured splats identify the
same request: alloc in dummy_alloc_request, free in
dummy_timer, fault reached from ffs_epfile_release (close) and
from the FUNCTIONFS_DMABUF_DETACH ioctl (detach). With the
patch applied, both paths are silent under the same hook.
The bug is reached from the FunctionFS device node, which in
real deployments is owned by the privileged gadget daemon
(adbd, UMS, composite gadget services, etc.); it is not
reachable from unprivileged userspace or from a USB host on the
cable. FunctionFS mounts default to GLOBAL_ROOT_UID, but the
filesystem supports uid=, gid=, and fmode= delegation to a
non-root gadget daemon, so on real deployments the attacker may
be a less-privileged service rather than root. |
| In the Linux kernel, the following vulnerability has been resolved:
iommu/vt-d: Fix race condition during PASID entry replacement
The Intel VT-d PASID table entry is 512 bits (64 bytes). When replacing
an active PASID entry (e.g., during domain replacement), the current
implementation calculates a new entry on the stack and copies it to the
table using a single structure assignment.
struct pasid_entry *pte, new_pte;
pte = intel_pasid_get_entry(dev, pasid);
pasid_pte_config_first_level(iommu, &new_pte, ...);
*pte = new_pte;
Because the hardware may fetch the 512-bit PASID entry in multiple
128-bit chunks, updating the entire entry while it is active (Present
bit set) risks a "torn" read. In this scenario, the IOMMU hardware
could observe an inconsistent state — partially new data and partially
old data — leading to unpredictable behavior or spurious faults.
Fix this by removing the unsafe "replace" helpers and following the
"clear-then-update" flow, which ensures the Present bit is cleared and
the required invalidation handshake is completed before the new
configuration is applied. |
| In the Linux kernel, the following vulnerability has been resolved:
Bluetooth: fix UAF in l2cap_sock_cleanup_listen() vs l2cap_conn_del()
bt_accept_dequeue() unlinks a not-yet-accepted child from the parent
accept queue and release_sock()s it before returning, so the returned
sk has no caller reference and is unlocked.
l2cap_sock_cleanup_listen() walks these children on listening-socket
close. A concurrent HCI disconnect drives hci_rx_work ->
l2cap_conn_del() which runs l2cap_chan_del() + l2cap_sock_kill() and
frees the child sk and its l2cap_chan; cleanup_listen() then uses both:
BUG: KASAN: slab-use-after-free in l2cap_sock_kill
l2cap_sock_kill / l2cap_sock_cleanup_listen / __x64_sys_close
Freed by: l2cap_conn_del -> l2cap_sock_close_cb -> l2cap_sock_kill
This is distinct from the two fixes already in this area: commit
e83f5e24da741 ("Bluetooth: serialize accept_q access") serialises the
accept_q list/poll and takes temporary refs inside bt_accept_dequeue(),
and CVE-2025-39860 serialises the userspace close()/accept() race by
calling cleanup_listen() under lock_sock() in l2cap_sock_release().
Neither covers l2cap_conn_del() running from hci_rx_work, so this UAF
still reproduces on current bluetooth/master.
Take the reference at the source: bt_accept_dequeue() does sock_hold()
while sk is still locked, before release_sock(); callers sock_put().
cleanup_listen() pins the chan with l2cap_chan_hold_unless_zero() under
a brief child sk lock (serialising vs l2cap_sock_teardown_cb()), drops
it before l2cap_chan_lock(), and skips a duplicate l2cap_sock_kill() on
SOCK_DEAD. conn->lock is not taken here: cleanup_listen() runs under
the parent sk lock and that would invert
conn->lock -> chan->lock -> sk_lock (lockdep).
KASAN/SMP: an unprivileged listen/close vs HCI-disconnect race produced
12 use-after-free reports per run before this change; 0, and no lockdep
report, over 1600+ raced iterations after it on bluetooth/master. |
| In the Linux kernel, the following vulnerability has been resolved:
net: annotate data-races around sk->sk_{data_ready,write_space}
skmsg (and probably other layers) are changing these pointers
while other cpus might read them concurrently.
Add corresponding READ_ONCE()/WRITE_ONCE() annotations
for UDP, TCP and AF_UNIX. |
| In the Linux kernel, the following vulnerability has been resolved:
f2fs: protect extension_list reading with sb_lock in f2fs_sbi_show()
In f2fs_sbi_show(), the extension_list, extension_count and
hot_ext_count are read without holding sbi->sb_lock. If a concurrent
sysfs store modifies the extension list via f2fs_update_extension_list(),
the show path may read inconsistent count and array contents, potentially
leading to out-of-bounds access or displaying stale data.
Fix this by holding sb_lock around the entire extension list read
and format operation. |
| In the Linux kernel, the following vulnerability has been resolved:
misc: fastrpc: fix use-after-free of fastrpc_user in workqueue context
There is a race between fastrpc_device_release() and the workqueue
that processes DSP responses. When the user closes the file descriptor,
fastrpc_device_release() frees the fastrpc_user structure. Concurrently,
an in-flight DSP invocation can complete and fastrpc_rpmsg_callback()
schedules context cleanup via schedule_work(&ctx->put_work). If the
workqueue runs fastrpc_context_free() in parallel with or after
fastrpc_device_release() has freed the user structure, it dereferences
the freed fastrpc_user. Depending on the state of the context at the
time of the race, any one of the following accesses can be hit:
1. fastrpc_buf_free() calls fastrpc_ipa_to_dma_addr(buf->fl->cctx, ...)
to strip the SID bits from the stored IOVA before passing the
physical address to dma_free_coherent().
2. fastrpc_free_map() reads map->fl->cctx->vmperms[0].vmid to
reconstruct the source permission bitmask needed for the
qcom_scm_assign_mem() call that returns memory from the DSP VM
back to HLOS.
3. fastrpc_free_map() acquires map->fl->lock to safely remove the
map node from the fl->maps list.
The resulting use-after-free manifests as:
pc : fastrpc_buf_free+0x38/0x80 [fastrpc]
lr : fastrpc_context_free+0xa8/0x1b0 [fastrpc]
fastrpc_context_free+0xa8/0x1b0 [fastrpc]
fastrpc_context_put_wq+0x78/0xa0 [fastrpc]
process_one_work+0x180/0x450
worker_thread+0x26c/0x388
Add kref-based reference counting to fastrpc_user. Have each invoke
context take a reference on the user at allocation time and release it
when the context is freed. Release the initial reference in
fastrpc_device_release() at file close. Move the teardown of the user
structure — freeing pending contexts, maps, mmaps, and the channel
context reference — into the kref release callback fastrpc_user_free(),
so that it runs only when the last reference is dropped, regardless of
whether that happens at device close or after the final in-flight
context completes. |
| In the Linux kernel, the following vulnerability has been resolved:
ipv6: anycast: insert aca into global hash under idev->lock
syzbot reported a splat [1]: a slab-use-after-free in
ipv6_chk_acast_addr(), which walks the global inet6_acaddr_lst[] hash
under RCU and dereferences a struct ifacaddr6 that has already been
freed while still linked in the hash, so a later reader walks into a
dangling node.
In __ipv6_dev_ac_inc() the aca is allocated with refcount 1, then
aca_get() bumps it to 2 to keep it alive across the unlocked region.
It is published to idev->ac_list under idev->lock, but
ipv6_add_acaddr_hash() runs after write_unlock_bh(). A concurrent
teardown (ipv6_ac_destroy_dev() from addrconf_ifdown(), under RTNL)
can slip into that window:
CPU0 __ipv6_dev_ac_inc CPU1 ipv6_ac_destroy_dev (RTNL)
------------------------------ ------------------------------------
aca_alloc() refcnt 1
aca_get() refcnt 2
write_lock_bh(idev->lock)
add aca to ac_list
write_unlock_bh(idev->lock)
write_lock_bh(idev->lock)
pull aca off ac_list
write_unlock_bh(idev->lock)
ipv6_del_acaddr_hash(aca)
hlist_del_init_rcu() is a no-op,
aca is not in the hash yet
aca_put() refcnt 2->1
ipv6_add_acaddr_hash(aca)
aca now inserted into the hash
aca_put() refcnt 1->0
call_rcu(aca_free_rcu) -> kfree(aca)
The hash removal becomes a no-op because the insertion has not
happened yet, so once CPU0 inserts and drops the last reference, the
aca is freed while still linked in inet6_acaddr_lst[], and readers
dereference freed memory after the slab slot is reused.
This window opened once RTNL stopped serializing the join path against
device teardown. Move ipv6_add_acaddr_hash() inside the idev->lock
section so the ac_list and hash insertions are atomic with respect to
teardown: a racing remover now either misses the aca entirely or finds
it in both lists.
acaddr_hash_lock is now nested under idev->lock, which is acquired in
softirq context, so switch all acaddr_hash_lock sites to spin_lock_bh()
to avoid the irq lock inversion reported in [2].
[1] https://syzkaller.appspot.com/bug?extid=a01df04303c131efbf3a
[2] https://lore.kernel.org/netdev/6a194ef7.ba3b1513.1890b4.0000.GAE@google.com/ |
| In the Linux kernel, the following vulnerability has been resolved:
memcg: use round-robin victim selection in refill_stock
Harry Yoo reported that get_random_u32_below() is not safe to call in the
nmi context and memcg charge draining can happen in nmi context.
More specifically get_random_u32_below() is neither reentrant- nor
NMI-safe: it acquires a per-cpu local_lock via local_lock_irqsave() on the
batched_entropy_u32 state. An NMI that lands on a CPU mid-update of the
ChaCha batch state and recurses into the random subsystem would corrupt
that state. The memcg_stock local_trylock prevents re-entry on the percpu
stock itself, but cannot protect an unrelated subsystem's per-cpu lock.
Replace the random pick with a per-cpu round-robin counter stored in
memcg_stock_pcp and serialized by the same local_trylock that already
guards cached[] and nr_pages[]. No atomics, no random calls, no extra
locks needed. |
| In the Linux kernel, the following vulnerability has been resolved:
misc: fastrpc: fix use-after-free race in fastrpc_map_create
fastrpc_map_lookup returns a raw pointer after releasing fl->lock. The
caller fastrpc_map_create then calls fastrpc_map_get (kref_get_unless_zero)
on this unprotected pointer. A concurrent MEM_UNMAP can free the map
between the lock release and the kref operation, resulting in a
use-after-free on the freed slab object.
Restore the take_ref parameter to fastrpc_map_lookup so the reference
is acquired atomically under fl->lock before the pointer is exposed to
the caller. |
| In the Linux kernel, the following vulnerability has been resolved:
net: bcmgenet: fix racing timeout handler
The bcmgenet_timeout handler tries to take down all tx queues when
a single queue times out. This is over zealous and causes many race
conditions with queues that are still chugging along. Instead lets
only restart the timed out queue. |
| In the Linux kernel, the following vulnerability has been resolved:
powerpc/pgtable-frag: Fix bad page state in pte_frag_destroy
powerpc uses pt_frag_refcount as a reference counter for tracking it's
pte and pmd page table fragments. For PTE table, in case of Hash with
64K pagesize, we have 16 fragments of 4K size in one 64K page.
Patch series [1] "mm: free retracted page table by RCU"
added pte_free_defer() to defer the freeing of PTE tables when
retract_page_tables() is called for madvise MADV_COLLAPSE on shmem
range.
[1]: https://lore.kernel.org/all/7cd843a9-aa80-14f-5eb2-33427363c20@google.com/
pte_free_defer() sets the active flag on the corresponding fragment's
folio & calls pte_fragment_free(), which reduces the pt_frag_refcount.
When pt_frag_refcount reaches 0 (no active fragment using the folio), it
checks if the folio active flag is set, if set, it calls call_rcu to
free the folio, it the active flag is unset then it calls pte_free_now().
Now, this can lead to following problem in a corner case...
[ 265.351553][ T183] BUG: Bad page state in process a.out pfn:20d62
[ 265.353555][ T183] page: refcount:0 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x20d62
[ 265.355457][ T183] flags: 0x3ffff800000100(active|node=0|zone=0|lastcpupid=0x7ffff)
[ 265.358719][ T183] raw: 003ffff800000100 0000000000000000 5deadbeef0000122 0000000000000000
[ 265.360177][ T183] raw: 0000000000000000 c0000000119caf58 00000000ffffffff 0000000000000000
[ 265.361438][ T183] page dumped because: PAGE_FLAGS_CHECK_AT_FREE flag(s) set
[ 265.362572][ T183] Modules linked in:
[ 265.364622][ T183] CPU: 0 UID: 0 PID: 183 Comm: a.out Not tainted 6.18.0-rc3-00141-g1ddeaaace7ff-dirty #53 VOLUNTARY
[ 265.364785][ T183] Hardware name: IBM pSeries (emulated by qemu) POWER10 (architected) 0x801200 0xf000006 of:SLOF,git-ee03ae pSeries
[ 265.364908][ T183] Call Trace:
[ 265.364955][ T183] [c000000011e6f7c0] [c000000001cfaa18] dump_stack_lvl+0x130/0x148 (unreliable)
[ 265.365202][ T183] [c000000011e6f7f0] [c000000000794758] bad_page+0xb4/0x1c8
[ 265.365384][ T183] [c000000011e6f890] [c00000000079c020] __free_frozen_pages+0x838/0xd08
[ 265.365554][ T183] [c000000011e6f980] [c0000000000a70ac] pte_frag_destroy+0x298/0x310
[ 265.365729][ T183] [c000000011e6fa30] [c0000000000aa764] arch_exit_mmap+0x34/0x218
[ 265.365912][ T183] [c000000011e6fa80] [c000000000751698] exit_mmap+0xb8/0x820
[ 265.366080][ T183] [c000000011e6fc30] [c0000000001b1258] __mmput+0x98/0x300
[ 265.366244][ T183] [c000000011e6fc80] [c0000000001c81f8] do_exit+0x470/0x1508
[ 265.366421][ T183] [c000000011e6fd70] [c0000000001c95e4] do_group_exit+0x88/0x148
[ 265.366602][ T183] [c000000011e6fdc0] [c0000000001c96ec] pid_child_should_wake+0x0/0x178
[ 265.366780][ T183] [c000000011e6fdf0] [c00000000003a270] system_call_exception+0x1b0/0x4e0
[ 265.366958][ T183] [c000000011e6fe50] [c00000000000d05c] system_call_vectored_common+0x15c/0x2ec
The bad page state error occurs when such a folio gets freed (with
active flag set), from do_exit() path in parallel.
... this can happen when the pte fragment was allocated from this folio,
but when all the fragments get freed, the pte_frag_refcount still had some
unused fragments. Now, if this process exits, with such folio as it's cached
pte_frag in mm->context, then during pte_frag_destroy(), we simply call
pagetable_dtor() and pagetable_free(), meaning it doesn't clear the
active flag. This, can lead to the above bug. Since we are anyway in
do_exit() path, then if the refcount is 0, then I guess it should be
ok to simply clear the folio active flag before calling pagetable_dtor()
& pagetable_free(). |
| In the Linux kernel, the following vulnerability has been resolved:
s390/cio: use generic driver_override infrastructure
When a driver is probed through __driver_attach(), the bus' match()
callback is called without the device lock held, thus accessing the
driver_override field without a lock, which can cause a UAF.
Fix this by using the driver-core driver_override infrastructure taking
care of proper locking internally.
Note that calling match() from __driver_attach() without the device lock
held is intentional. [1] |
| In the Linux kernel, the following vulnerability has been resolved:
batman-adv: stop tp_meter sessions during mesh teardown
TP meter sessions remain linked on bat_priv->tp_list after the netlink
request has already finished. When the mesh interface is removed,
batadv_mesh_free() currently tears down the mesh without first draining
these sessions.
A running sender thread or a late incoming tp_meter packet can then keep
processing against a mesh instance which is already shutting down.
Synchronize tp_meter with the mesh lifetime by stopping all active
sessions from batadv_mesh_free() and waiting for sender threads to exit
before teardown continues. |
| In the Linux kernel, the following vulnerability has been resolved:
mm/damon/core: fix damon_call() vs kdamond_fn() exit race
Patch series "mm/damon/core: fix damon_call()/damos_walk() vs kdmond exit
race".
damon_call() and damos_walk() can leak memory and/or deadlock when they
race with kdamond terminations. Fix those.
This patch (of 2);
When kdamond_fn() main loop is finished, the function cancels all
remaining damon_call() requests and unset the damon_ctx->kdamond so that
API callers and API functions themselves can know the context is
terminated. damon_call() adds the caller's request to the queue first.
After that, it shows if the kdamond of the damon_ctx is still running
(damon_ctx->kdamond is set). Only if the kdamond is running, damon_call()
starts waiting for the kdamond's handling of the newly added request.
The damon_call() requests registration and damon_ctx->kdamond unset are
protected by different mutexes, though. Hence, damon_call() could race
with damon_ctx->kdamond unset, and result in deadlocks.
For example, let's suppose kdamond successfully finished the damon_call()
requests cancelling. Right after that, damon_call() is called for the
context. It registers the new request, and shows the context is still
running, because damon_ctx->kdamond unset is not yet done. Hence the
damon_call() caller starts waiting for the handling of the request.
However, the kdamond is already on the termination steps, so it never
handles the new request. As a result, the damon_call() caller threads
infinitely waits.
Fix this by introducing another damon_ctx field, namely
call_controls_obsolete. It is protected by the
damon_ctx->call_controls_lock, which protects damon_call() requests
registration. Initialize (unset) it in kdamond_fn() before letting
damon_start() returns and set it just before the cancelling of remaining
damon_call() requests is executed. damon_call() reads the obsolete field
under the lock and avoids adding a new request.
After this change, only requests that are guaranteed to be handled or
cancelled are registered. Hence the after-registration DAMON context
termination check is no longer needed. Remove it together.
Note that the deadlock will not happen when damon_call() is called for
repeat mode request. In tis case, damon_call() returns instead of waiting
for the handling when the request registration succeeds and it shows the
kdamond is running. However, if the request also has dealloc_on_cancel,
the request memory would be leaked.
The issue is found by sashiko [1]. |
| In the Linux kernel, the following vulnerability has been resolved:
net/smc: avoid early lgr access in smc_clc_wait_msg
A CLC decline can be received while the handshake is still in an early
stage, before the connection has been associated with a link group.
The decline handling in smc_clc_wait_msg() updates link-group level sync
state for first-contact declines, but that state only exists after link
group setup has completed. Guard the link-group update accordingly and
keep the per-socket peer diagnosis handling unchanged.
This preserves the existing sync_err handling for established link-group
contexts and avoids touching link-group state before it is available. |
| In the Linux kernel, the following vulnerability has been resolved:
wifi: rsi: fix kthread lifetime race between self-exit and external-stop
RSI driver use both self-exit(kthread_complete_and_exit) and external-stop
(kthread_stop) when killing a kthread. Generally, kthread_stop() is called
first, and in this case, no particular issues occur.
However, in rare instances where kthread_complete_and_exit() is called
first and then kthread_stop() is called, a UAF occurs because the kthread
object, which has already exited and been freed, is accessed again.
Therefore, to prevent this with minimal modification, you must remove
kthread_stop() and change the code to wait until the self-exit operation
is completed. |
| In the Linux kernel, the following vulnerability has been resolved:
RDMA/mlx4: Fix mis-use of RCU in mlx4_srq_event()
Sashiko points out the radix_tree itself is RCU safe, but nothing ever
frees the mlx4_srq struct with RCU, and it isn't even accessed within the
RCU critical section. It also will crash if an event is delivered before
the srq object is finished initializing.
Use the spinlock since it isn't easy to make RCU work, use
refcount_inc_not_zero() to protect against partially initialized objects,
and order the refcount_set() to be after the srq is fully initialized. |
| In the Linux kernel, the following vulnerability has been resolved:
xfs: stop reclaim before pushing AIL during unmount
The unmount sequence in xfs_unmount_flush_inodes() pushed the AIL while
background reclaim and inodegc are still running. This is broken
independently of any use-after-free issues - background reclaim and
inodegc should not be running while the AIL is being pushed during
unmount, as inodegc can dirty and insert inodes into the AIL during the
flush, and background reclaim can race to abort and free dirty inodes.
Reorder xfs_unmount_flush_inodes() to stop inodegc and cancel background
reclaim before pushing the AIL. Stop inodegc before cancelling
m_reclaim_work because the inodegc worker can re-queue m_reclaim_work
via xfs_inodegc_set_reclaimable. |
| In the Linux kernel, the following vulnerability has been resolved:
PCI: endpoint: pci-epf-vntb: Stop cmd_handler work in epf_ntb_epc_cleanup
Disable the delayed work before clearing BAR mappings and doorbells to
avoid running the handler after resources have been torn down.
Unable to handle kernel paging request at virtual address ffff800083f46004
[...]
Internal error: Oops: 0000000096000007 [#1] SMP
[...]
Call trace:
epf_ntb_cmd_handler+0x54/0x200 [pci_epf_vntb] (P)
process_one_work+0x154/0x3b0
worker_thread+0x2c8/0x400
kthread+0x148/0x210
ret_from_fork+0x10/0x20 |
| In the Linux kernel, the following vulnerability has been resolved:
mm: blk-cgroup: fix use-after-free in cgwb_release_workfn()
cgwb_release_workfn() calls css_put(wb->blkcg_css) and then later accesses
wb->blkcg_css again via blkcg_unpin_online(). If css_put() drops the last
reference, the blkcg can be freed asynchronously (css_free_rwork_fn ->
blkcg_css_free -> kfree) before blkcg_unpin_online() dereferences the
pointer to access blkcg->online_pin, resulting in a use-after-free:
BUG: KASAN: slab-use-after-free in blkcg_unpin_online (./include/linux/instrumented.h:112 ./include/linux/atomic/atomic-instrumented.h:400 ./include/linux/refcount.h:389 ./include/linux/refcount.h:432 ./include/linux/refcount.h:450 block/blk-cgroup.c:1367)
Write of size 4 at addr ff11000117aa6160 by task kworker/71:1/531
Workqueue: cgwb_release cgwb_release_workfn
Call Trace:
<TASK>
blkcg_unpin_online (./include/linux/instrumented.h:112 ./include/linux/atomic/atomic-instrumented.h:400 ./include/linux/refcount.h:389 ./include/linux/refcount.h:432 ./include/linux/refcount.h:450 block/blk-cgroup.c:1367)
cgwb_release_workfn (mm/backing-dev.c:629)
process_scheduled_works (kernel/workqueue.c:3278 kernel/workqueue.c:3385)
Freed by task 1016:
kfree (./include/linux/kasan.h:235 mm/slub.c:2689 mm/slub.c:6246 mm/slub.c:6561)
css_free_rwork_fn (kernel/cgroup/cgroup.c:5542)
process_scheduled_works (kernel/workqueue.c:3302 kernel/workqueue.c:3385)
** Stack based on commit 66672af7a095 ("Add linux-next specific files
for 20260410")
I am seeing this crash sporadically in Meta fleet across multiple kernel
versions. A full reproducer is available at:
https://github.com/leitao/debug/blob/main/reproducers/repro_blkcg_uaf.sh
(The race window is narrow. To make it easily reproducible, inject a
msleep(100) between css_put() and blkcg_unpin_online() in
cgwb_release_workfn(). With that delay and a KASAN-enabled kernel, the
reproducer triggers the splat reliably in less than a second.)
Fix this by moving blkcg_unpin_online() before css_put(), so the
cgwb's CSS reference keeps the blkcg alive while blkcg_unpin_online()
accesses it. |
