| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| This CVE ID has been rejected or withdrawn by its CVE Numbering Authority. |
| In the Linux kernel, the following vulnerability has been resolved:
btrfs: zoned: fix deadlock between metadata writeback and transaction commit
When writing out metadata extent buffers in a zoned filesystem,
btree_writepages() holds fs_info->zoned_meta_io_lock across the whole
writeback loop, including the call to btrfs_check_meta_write_pointer() ->
check_bg_is_active().
For the tree-log block group, check_bg_is_active() may fail to activate
the zone and fall back to btrfs_zone_finish_one_bg() to free an active
zone. That path waits for the running transaction to commit while still
holding zoned_meta_io_lock, but the committer needs that same lock to
write out the tree extents, so the two tasks deadlock:
Task A (kworker, metadata writeback) Task B (fsstress, transaction commit)
------------------------------------ -------------------------------------
wb_workfn() btrfs_commit_transaction(T)
btree_writepages() btrfs_write_and_wait_transaction()
btrfs_zoned_meta_io_lock() btrfs_write_marked_extents()
btrfs_check_meta_write_pointer() btree_writepages()
check_bg_is_active() [treelog_bg] btrfs_zoned_meta_io_lock()
btrfs_zone_finish_one_bg() <blocks on zoned_meta_io_lock,
btrfs_zone_finish() held by Task A>
do_zone_finish()
btrfs_inc_block_group_ro()
btrfs_wait_for_commit()
<blocks waiting for commit
of transaction T, done by
Task B>
The sibling branch in check_bg_is_active() already drops zoned_meta_io_lock
around do_zone_finish() for this exact reason. Do the same in the tree-log
branch: release the lock around btrfs_zone_finish_one_bg() and re-acquire
it afterwards. The lock only protects fs_info->active_{meta,system}_bg,
which this branch does not touch, and ctx->zoned_bg keeps a reference to
the block group across the unlock, so nothing is lost while the lock
is dropped.
This hang occasionally reproduces with fstests generic/475 on a zoned
btrfs filesystem. |
| In the Linux kernel, the following vulnerability has been resolved:
scsi: scsi_debug: Fix REPORT ZONES alloc_len underflow OOB write
resp_report_zones() sizes the reply buffer from the CDB allocation
length. The v3 fix rounds alloc_len up with ALIGN() before deriving the
descriptor count:
rep_max_zones = (ALIGN((u64)alloc_len, RZONES_DESC_HD) -
RZONES_DESC_HD) >> ilog2(RZONES_DESC_HD);
arr_len = (u64)RZONES_DESC_HD * (rep_max_zones + 1);
For alloc_len in 0xFFFFFFC1..0xFFFFFFFF, ALIGN() rounds up to
0x100000000, so arr_len is 4 GB. On 32-bit, kzalloc()'s size_t is 32-bit
and truncates 0x100000000 to 0; kzalloc(0) returns ZERO_SIZE_PTR, which
passes the !arr check, and desc = arr + 64 is then dereferenced in the
loop -> out-of-bounds write / panic.
Clamp rep_max_zones to devip->nr_zones. The loop already stops at
sdebug_capacity (after nr_zones zones), so a report can never hold more
than nr_zones descriptors; the clamp does not change the report, it only
bounds arr_len to (nr_zones + 1) * RZONES_DESC_HD, a real device
property that can never reach 0x100000000. |
| In the Linux kernel, the following vulnerability has been resolved:
rxrpc: Fix the reception of a reply packet before data transmission
Fix rxrpc_receiving_reply() to handle the reception of an apparent reply
DATA packet before rxrpc has had a chance to send any request DATA packets
on a client call by checking to see if the call has been exposed yet by
sending the first packet.
Without this, rxrpc_rotate_tx_window() might oops.
Also fix rxrpc_rotate_tx_window() to handle the Tx queue being empty by
changing the do...while loop into a while loop, just in case a call is
abnormally terminated by an early reply before the last request packet is
transmitted. |
| In the Linux kernel, the following vulnerability has been resolved:
iommu/vt-d: Fix RB-tree corruption in probe error path
The info->node RB-tree member is zero-initialized via kzalloc. If
a device does not support ATS, the device_rbtree_insert() call is
skipped. If a subsequent probe step fails, the error path jumps to
device_rbtree_remove(), which misinterprets the zeroed node as
a tree root and corrupts the device RB-tree.
Fix this by explicitly initializing the RB-node as empty using
RB_CLEAR_NODE() during initialization and guarding the removal with
RB_EMPTY_NODE(). |
| In the Linux kernel, the following vulnerability has been resolved:
lockd: Avoid hashing uninitialized bytes in nlm4svc_lookup_file()
file_hash() digests the first LOCKD_FH_HASH_SIZE bytes of
nfs_fh.data when bucketing nlm_files[], independent of fh.size.
Commit 3de744ee4e45 ("lockd: Use xdrgen XDR functions for the
NLMv4 TEST procedure") set .pc_argzero to zero for the converted
procedures and moved file-handle population into
nlm4svc_lookup_file(), which copies only xdr_lock->fh.len bytes
into lock->fh.data.
When an NLMv4 client presents a file handle shorter than
LOCKD_FH_HASH_SIZE, bytes fh.len..31 retain whatever the argument
buffer held from an earlier request. The same wire handle then
hashes to different buckets across calls; nlm_lookup_file() misses
the existing nlm_file entry, and lock-state lookups fail.
Zero only the tail bytes that file_hash() would otherwise consume.
Handles of LOCKD_FH_HASH_SIZE or larger already populate every byte
that file_hash() reads. |
| In the Linux kernel, the following vulnerability has been resolved:
bpf: Tighten cgroup storage cookie checks for prog arrays
The fix in commit abad3d0bad72 ("bpf: Fix oob access in cgroup local
storage") is still incomplete. The prog-array compatibility check
treats a program with no cgroup storage as compatible with any stored
storage cookie. This allows a storage-less program to bridge a tail
call chain between an entry program and a storage-using callee even
though cgroup local storage at runtime still follows the caller's
context, that is, A -> B(no storage) -> C(storage) path.
Requiring exact cookie equality would break the legitimate case of a
storage-less leaf program being tail called from a storage-using one.
Instead, only accept a zero storage cookie if the program cannot
perform tail calls itself. This keeps A -> B(no storage) working
while rejecting the A -> B(no storage) -> C(storage) bridge. |
| In the Linux kernel, the following vulnerability has been resolved:
net/sched: sch_codel: Do not call qdisc_tree_reduce_backlog during peek before restoring qlen
Whenever codel drops packets during peek, it calls
qdisc_tree_reduce_backlog. An issue arises because it calls
qdisc_tree_reduce_backlog before it reincrements the qlen. If qlen drops
to zero, but peek returns an skb, the parent's qlen_notify callback will
be executed even though codel still has 1 packet on the queue and, thus,
will mistakenly deactivate the parent's class causing issues like a wild
memory access when qfq has codel as a child:
[ 36.339843][ T370] Oops: general protection fault, probably for non-canonical address 0xfbd59c0000000024: 0000 [#1] SMP KASAN NOPTI
[ 36.340408][ T370] KASAN: maybe wild-memory-access in range [0xdead000000000120-0xdead000000000127]
[ 36.340737][ T370] CPU: 2 UID: 0 PID: 370 Comm: tc Not tainted 7.1.0-rc5-00287-g66e13b626592 #87 PREEMPT(full)
[ 36.341113][ T370] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[ 36.341357][ T370] RIP: 0010:qfq_deactivate_agg (include/linux/list.h:1029 (discriminator 2) include/linux/list.h:1043 (discriminator 2) net/sched/sch_qfq.c:1369 (discriminator 2) net/sched/sch_qfq.c:1395 (discriminator 2)) sch_qfq
[ 36.342221][ T370] RSP: 0018:ffff8881100ef370 EFLAGS: 00010216
[ 36.342422][ T370] RAX: 0000000000000000 RBX: ffff8881058a9568 RCX: dffffc0000000000
[ 36.342664][ T370] RDX: 1ffff11021064dc3 RSI: ffff888108326e00 RDI: dffffc0000000000
[ 36.342905][ T370] RBP: ffff8881058a8280 R08: dead000000000122 R09: 1bd5a00000000024
[ 36.343140][ T370] R10: fffffbfff2940329 R11: fffffbfff2940329 R12: 0000000000000000
[ 36.343383][ T370] R13: dead000000000100 R14: ffff8881058a9580 R15: ffff8881058a9578
[ 36.343631][ T370] FS: 00007fc04b0ca780(0000) GS:ffff888184fef000(0000) knlGS:0000000000000000
[ 36.343911][ T370] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 36.344116][ T370] CR2: 0000557c02c02000 CR3: 000000010e0ba000 CR4: 0000000000750ef0
[ 36.344359][ T370] PKRU: 55555554
[ 36.344481][ T370] Call Trace:
...
[ 36.345054][ T370] qfq_reset_qdisc (net/sched/sch_qfq.c:357 net/sched/sch_qfq.c:1487) sch_qfq
[ 36.345222][ T370] qdisc_reset (net/sched/sch_generic.c:1057)
[ 36.345503][ T370] __qdisc_destroy (net/sched/sch_generic.c:1096)
[ 36.345677][ T370] qdisc_graft (net/sched/sch_api.c:1062 net/sched/sch_api.c:1053 net/sched/sch_api.c:1159)
[ 36.346335][ T370] tc_get_qdisc (net/sched/sch_api.c:1528 net/sched/sch_api.c:1556)
Fix this by only calling qdisc_tree_reduce_backlog in peek after the
qlen is restored. |
| In the Linux kernel, the following vulnerability has been resolved:
net: watchdog: fix refcount tracking races
Blamed commit converted the untracked dev_hold()/dev_put() calls
in the watchdog code to use the tracked dev_hold_track()/dev_put_track()
(which were later renamed/interfaced to netdev_hold() and netdev_put()).
By introducing dev->watchdog_dev_tracker to store the
reference tracking information without adding synchronization
between netdev_watchdog_up() and dev_watchdog(), it enabled the
race condition where this pointer could be overwritten or freed
concurrently, leading to the list corruption crash syzbot reported:
list_del corruption, ffff888114a18c00->next is NULL
kernel BUG at lib/list_debug.c:52 !
Oops: invalid opcode: 0000 [#1] SMP KASAN PTI
CPU: 1 UID: 0 PID: 91 Comm: kworker/u8:5 Not tainted syzkaller #0 PREEMPT(lazy)
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 05/09/2026
Workqueue: events_unbound linkwatch_event
RIP: 0010:__list_del_entry_valid_or_report.cold+0x22/0x2a lib/list_debug.c:52
Call Trace:
<TASK>
__list_del_entry_valid include/linux/list.h:132 [inline]
__list_del_entry include/linux/list.h:246 [inline]
list_move_tail include/linux/list.h:341 [inline]
ref_tracker_free+0x1a7/0x6c0 lib/ref_tracker.c:329
netdev_tracker_free include/linux/netdevice.h:4491 [inline]
netdev_put include/linux/netdevice.h:4508 [inline]
netdev_put include/linux/netdevice.h:4504 [inline]
netdev_watchdog_down net/sched/sch_generic.c:600 [inline]
dev_deactivate_many+0x28c/0xfe0 net/sched/sch_generic.c:1363
dev_deactivate+0x109/0x1d0 net/sched/sch_generic.c:1397
linkwatch_do_dev net/core/link_watch.c:184 [inline]
linkwatch_do_dev+0xd3/0x120 net/core/link_watch.c:166
__linkwatch_run_queue+0x3a5/0x810 net/core/link_watch.c:240
linkwatch_event+0x8f/0xc0 net/core/link_watch.c:314
process_one_work+0xa0e/0x1980 kernel/workqueue.c:3314
process_scheduled_works kernel/workqueue.c:3397 [inline]
worker_thread+0x5ef/0xe50 kernel/workqueue.c:3478
kthread+0x370/0x450 kernel/kthread.c:436
ret_from_fork+0x69a/0xc80 arch/x86/kernel/process.c:158
ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:245
This patch has three coordinated parts:
1) Add dev->watchdog_lock and dev->watchdog_ref_held to serialize watchdog operations.
2) Remove netdev_watchdog_up() call from netif_carrier_on():
This ensures netdev_watchdog_up() is only called from process/BH context
(via linkwatch workqueue dev_activate()), allowing us to use
spin_lock_bh() for synchronization.
3) Synchronize watchdog up and watchdog timer:
Protect netdev_watchdog_up() with tx_global_lock and watchdog_lock.
Only allocate a new tracker in netdev_watchdog_up() if one is
not already present.
In dev_watchdog(), ensure we don't release the tracker if the
timer was rescheduled either by dev_watchdog() itself or concurrently
by netdev_watchdog_up(). |
| In the Linux kernel, the following vulnerability has been resolved:
netfilter: nf_conncount: prevent connlimit drops for early confirmed ct
Commit 69894e5b4c5e ("netfilter: nft_connlimit: update the count if add
was skipped") introduced a regression where packets for valid
connections are dropped when using connlimit for soft-limiting
scenarios.
The issue occurs when a new connection reuses a socket currently in
the TIME_WAIT state. In this scenario, the connection tracking entry
is evaluated as already confirmed. Previously, __nf_conncount_add()
assumed that if a connection was confirmed and did not originate from
the loopback interface, it should skip the addition and return -EEXIST.
Skipping the addition triggers a garbage collection run that cleans up
the TIME_WAIT connection. Consequently, the active connection count
drops to 0, which xt_connlimit mishandles, leading to the false rejection
of the perfectly valid new connection.
Fix this by replacing the interface check with protocol-agnostic state
checks. We now skip the tree insertion and preserve the lockless garbage
collection optimization only if the connection is IPS_ASSURED. This
allows early-confirmed setup packets (such as reused TIME_WAIT sockets
or locally generated SYN-ACKs) to be properly evaluated and counted
without falsely dropping. The goto check_connections path is maintained
to ensure these setup packets are deduplicated correctly.
This has been tested with slowhttptest and HTTP server configured
locally to ensure we are not breaking soft-limiting scenarios for local
or external connections. In addition, it was tested with a OVS zone
limit too. |
| In the Linux kernel, the following vulnerability has been resolved:
afs: Fix callback service message parsers to pass through -EAGAIN
The AFS filesystem client uses an rxrpc server to listen for callback
notifications. Each callback call type handler has a delivery function
that parses the incoming request stream, and this should return -EAGAIN the
last packet hasn't yet been seen, but all currently queued received data is
consumed. afs_extract_data() does this, but the -EAGAIN return is switched
to 0 inadvertantly
Fix callback service message parsers to pass through -EAGAIN |
| In the Linux kernel, the following vulnerability has been resolved:
afs: Fix the volume AFS_VOLUME_RM_TREE is set on
Fix afs_insert_volume_into_cell() to set AFS_VOLUME_RM_TREE on the volume
replaced, not the new volume, as it's now removed from the cell's volume
tree. This will cause the old volume to be removed from the tree twice and
the new volume never to be removed. |
| In the Linux kernel, the following vulnerability has been resolved:
net/mlx5: LAG, Fix off-by-one in single-FDB error rollback
On failure at index i, the reverse cleanup loop in
mlx5_lag_create_single_fdb() starts from i, so the failed index
itself is rolled back. That can operate on uninitialized state or
double-tear-down a rule the add_one path already self-rolled-back.
Start the rollback from i - 1 so only successfully-installed entries
are undone. |
| In the Linux kernel, the following vulnerability has been resolved:
net/mlx5e: Fix HV VHCA stats agent registration race
mlx5e_hv_vhca_stats_create() registers the stats agent through
mlx5_hv_vhca_agent_create(). The helper publishes the agent in
hv_vhca->agents[type] under agents_lock and immediately schedules an
asynchronous control invalidation on the HV VHCA workqueue before
returning to mlx5e.
The asynchronous invalidation invokes the control agent's invalidate
callback, which reads the hypervisor control block and forwards the
command to mlx5e_hv_vhca_stats_control(). That callback may either:
- call cancel_delayed_work_sync(&priv->stats_agent.work), or
- call queue_delayed_work(priv->wq, &sagent->work, sagent->delay).
However, the delayed_work and priv->stats_agent.agent are only
initialized after mlx5_hv_vhca_agent_create() returns to mlx5e:
agent = mlx5_hv_vhca_agent_create(...); /* publish + invalidate */
...
priv->stats_agent.agent = agent; /* too late */
INIT_DELAYED_WORK(&priv->stats_agent.work, ...); /* too late */
If the asynchronous control path runs before the two assignments
above, it can:
- Operate on an uninitialized delayed_work whose timer.function is
NULL. queue_delayed_work() calls add_timer() unconditionally, so
when the timer expires the timer softirq invokes a NULL function
pointer.
- Re-initialize the timer later through INIT_DELAYED_WORK() while
the timer is already enqueued in the timer wheel, corrupting the
hlist (entry.pprev cleared while the previous bucket node still
points at this entry).
- When the worker eventually runs, mlx5e_hv_vhca_stats_work() reads
sagent->agent (NULL) and dereferences it inside
mlx5_hv_vhca_agent_write().
Fix this by:
- Initializing priv->stats_agent.work before invoking
mlx5_hv_vhca_agent_create(), so the work is always in a valid
state when the control callback observes it.
- Adding a struct mlx5_hv_vhca_agent **ctx_update out-parameter
to mlx5_hv_vhca_agent_create(). The helper writes the agent
pointer to *ctx_update before publishing into hv_vhca->agents[]
and triggering the agents_update flow, so any callback
subsequently invoked from that flow already sees a valid
priv->stats_agent.agent. This avoids having the control
callback participate in agent initialization.
While at it, access priv->stats_agent.agent with
READ_ONCE()/WRITE_ONCE() for the cross-CPU access with the worker, and
clear priv->stats_agent.buf on the agent_create() failure path. |
| In the Linux kernel, the following vulnerability has been resolved:
smb: client: fix busy dentry warning on unmount after DIO
Commit c68337442f03 ("cifs: Fix busy dentry used after unmounting") fixed
the issue in cifs where deferred close of a file led to a dentry reference
count not being released in umount, by flushing deferredclose_wq in
cifs_kill_sb() to solve it.
However, the cifs DIO path suffers from the same busy-dentry problem caused
by a delayed dentry reference-count release:
[dio] [cifsd] [close + umount]
netfs_unbuffered_write_iter_locked
...
cifs_demultiplex_thread
netfs_unbuffered_write
cifs_issue_write
netfs_wait_for_in_progress_stream [1]
...
netfs_write_subrequest_terminated
netfs_subreq_clear_in_progress
netfs_wake_collector // wake [1]
netfs_put_subrequest
netfs_put_request
queue_work(system_dfl_wq, xxx) [2]
// dio write return cifs_close
_cifsFileInfo_put
// cfile->count 2->1
--cfile->count [3]
// umount
cifs_kill_sb
kill_anon_super
// warning triggered!
shrink_dcache_for_umount [4]
[system_dfl_wq] [5]
netfs_free_request
...
_cifsFileInfo_put
// cfile->count 1->0
--cfile->count
queue_work(fileinfo_put_wq, xxx)
[fileinfo_put_wq] [6]
cifsFileInfo_put_work
cifsFileInfo_put_final
dput
If the umount path is triggered before [5], it results warning:
BUG: Dentry 00000000eab1f070{i=9a917b66ae404fec,n=test} still in use (1)
[unmount of cifs cifs]
The existing per-inode ictx->io_count wait in cifs_evict_inode() does not
help: it lives in the inode eviction path, which runs after
shrink_dcache_for_umount() has already warned about the busy dentries.
Fix it by adding a per-superblock outstanding-rreq counter that is
incremented in cifs_init_request() and decremented in cifs_free_request().
In cifs_kill_sb(), before kill_anon_super(), wait for this counter to reach
0 - which guarantees that all cleanup_work for this sb have run and thus
all relevant cfile puts are queued on fileinfo_put_wq or serverclose_wq.
Then drain the workqueue so the dentry refs are dropped.
This is a targeted wait, not a flush of the system-wide system_dfl_wq. |
| In the Linux kernel, the following vulnerability has been resolved:
KVM: x86: Ignore pending PV EOI if the vCPU has since disabled PV EOIs
Ignore KVM's internal "service pending PV EOI" request if the vCPU has
disabled PV EOIs since the request was made. Asserting that PV EOIs are
enabled can fail if reading guest memory in pv_eoi_get_user() fails, i.e.
if pv_eoi_test_and_clr_pending() bails early, *and* the vCPU also disables
PV EOIs.
kernel BUG at arch/x86/kvm/lapic.c:3338!
Oops: invalid opcode: 0000 [#1] SMP
CPU: 4 UID: 1000 PID: 890 Comm: pv_eoi_test Not tainted 7.0.0-d585aa5894d8-vm #337 PREEMPT
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
RIP: 0010:kvm_lapic_sync_from_vapic+0x12b/0x140 [kvm]
Call Trace:
<TASK>
kvm_arch_vcpu_ioctl_run+0x1075/0x1c30 [kvm]
kvm_vcpu_ioctl+0x2d5/0x980 [kvm]
__x64_sys_ioctl+0x8a/0xd0
do_syscall_64+0xb5/0xb40
entry_SYSCALL_64_after_hwframe+0x4b/0x53
</TASK>
Modules linked in: kvm_intel kvm irqbypass
---[ end trace 0000000000000000 ]--- |
| In the Linux kernel, the following vulnerability has been resolved:
ntfs: skip extent mft records in writeback to prevent deadlock
This patch fixes the ABBA deadlock between extent_lock and extent
mrec_lock triggered by xfstests generic/113, that occurs since the commit
6994acf33bae ("ntfs: use base mft_no when looking up base inode for
extent record").
Path A (inode writeback):
VFS writeback
-> ntfs_write_inode()
-> __ntfs_write_inode()
-> mutex_lock(&ni->extent_lock)
-> mutex_lock(&tni->mrec_lock)
Path B (MFT folio writeback):
VFS writeback of $MFT dirty folios
-> ntfs_mft_writepages()
-> ntfs_write_mft_block()
-> ntfs_may_write_mft_record()
-> holds one extent mrec_lock from a previous iteration
-> tries to acquire another base inode extent_lock
By removing all extent_lock and extent mrec_lock acquisition from the MFT
folio writeback path, the ABBA lock ordering is eliminated:
Path A: __ntfs_write_inode(): extent_lock -> mrec_lock
Path B (removed): ntfs_write_mft_block(): mrec_lock -> extent_lock
Path B is always redundant for extent records because:
1. mark_mft_record_dirty(ext_ni) does NOT dirty the MFT folio.
It only sets NInoDirty(ext_ni) and marks the base VFS inode dirty
via __mark_inode_dirty(I_DIRTY_DATASYNC), which triggers Path A.
Therefore, normal extent modifications never create a situation where
the MFT folio is dirty and Path B is not scheduled.
2. The MFT folio only gets dirtied via ntfs_mft_mark_dirty() inside
ntfs_mft_record_alloc(). But all identified callers in attrib.c
(ntfs_attr_add, ntfs_attr_record_move_away,
ntfs_attr_make_non_resident, ntfs_attr_record_resize) follow through
with mark_mft_record_dirty(), which triggers Path A to write the
complete record.
3. ntfs_evict_big_inode() calls ntfs_commit_inode() before freeing extent
inodes, ensuring all dirty extents are flushed via Path A before the
base inode leaves the icache. |
| In the Linux kernel, the following vulnerability has been resolved:
dmaengine: tegra: Fix burst size calculation
Currently, the Tegra GPC DMA hardware requires the transfer length to
be a multiple of the max burst size configured for the channel. When a
client requests a transfer where the length is not evenly divisible by
the configured max burst size, the DMA hangs with partial burst at
the end.
Fix this by reducing the burst size to the largest power-of-2 value
that evenly divides the transfer length. For example, a 40-byte
transfer with a 16-byte max burst will now use an 8-byte burst
(40 / 8 = 5 complete bursts) instead of causing a hang.
This issue was observed with the PL011 UART driver where TX DMA
transfers of arbitrary lengths were stuck. |
| In the Linux kernel, the following vulnerability has been resolved:
platform/x86: ISST: Restore SST-PP control to all domains
The SST-PP control offset is only restored to power domain 0 after
resume. During suspend, control values are read and stored for all
power domains.
Use pd_info->sst_base instead of power_domain_info->sst_base, which
only points to power domain 0 base address. |
| In the Linux kernel, the following vulnerability has been resolved:
bpf: Reset register bounds before narrowing retval range in check_mem_access()
When the BPF verifier processes a context load of an LSM hook return
value, it calls __mark_reg_s32_range() to narrow the register to the
hook's valid range. However, __mark_reg_s32_range() intersects the new
range with the register's existing bounds using max_t()/min_t() rather
than replacing them.
If the destination register carries stale bounds from a prior instruction
(e.g. BPF_MOV64_IMM), the intersection can produce a range narrower than
reality. The verifier then believes it knows the register's exact value,
while at runtime the actual hook return value is loaded, creating a
verifier/runtime mismatch that can be used to bypass BPF memory safety
checks.
The else branch already calls mark_reg_unknown() to reset register state
before any narrowing. Apply the same reset in the is_retval path so
stale bounds are cleared before __mark_reg_s32_range() intersects. |
