| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| In the Linux kernel, the following vulnerability has been resolved:
iommu/amd: Wait for completion instead of returning early in iommu_completion_wait()
need_sync is a per-IOMMU flag shared by all domains and devices behind
that IOMMU. It is set whenever a command is queued with sync == true and
cleared when a completion-wait (CWAIT) command is queued. However, a
cleared need_sync only means that a covering CWAIT has been queued, not
that all previously queued commands have actually completed in hardware.
iommu_completion_wait() read need_sync locklessly and returned early
when it was false. This breaks the "block until all previously queued
commands have completed" contract in a multi-CPU scenario:
CPU2: queue inv-B => need_sync = true
CPU1: queue CWAIT(N); need_sync = false; then wait_on_sem(N)
CPU2: read need_sync == false => return 0 (no wait!)
CPU2 returns without waiting for any sequence number even though its
inv-B may not have completed yet (CWAIT(N), queued after inv-B, has not
been signaled). CPU2 then proceeds to, for example, free page-table
pages while the IOMMU can still walk stale translations, opening a
use-after-free window. This is a logical race in the meaning of the
flag, not a memory-visibility issue, so barriers alone do not help.
Fix it without losing the optimization of avoiding redundant CWAIT
commands: take iommu->lock before testing need_sync, and when it is
false do not return early but wait for the last allocated sequence
number (cmd_sem_val). Since need_sync == false implies no sync command
was queued after the last CWAIT, that CWAIT is FIFO-ordered after every
not-yet-completed command, so waiting for its sequence number guarantees
all prior commands (possibly queued by another CPU) have completed. The
common path with pending work is unchanged and no extra hardware command
is issued. |
| In the Linux kernel, the following vulnerability has been resolved:
media: chips-media: wave5: Move src_buf Removal to finish_encode
During encoder processing, there is a case where the IRQ response could
return the buffer back to userspace via v4l2_m2m_buf_done call. In this
time, userspace could queue up this same buffer before start_encode removes
the index from the ready queue. This would then lead to a case where the
buffer in the ready queue could be a self loop due to the
WRITE_ONCE(prev->next, new) call in __list_add.
When __list_del is finally called, the loop is already made so nothing
points back to ready queue list head and pointers are poisoned.
A buffer should not be marked as DONE before the buffer is removed from
m2m ready queue. Move removal entirely to finish_encode. |
| In the Linux kernel, the following vulnerability has been resolved:
ublk: wait on ublk_dev_ready() instead of ub->completion
ub->completion is only re-armed by a successful START_USER_RECOVERY. If
the ublk server sends END_USER_RECOVERY without one - e.g. its START
failed with -EBUSY and the error was ignored - the wait is satisfied by
the stale completion of the previous recovery cycle, and the device is
marked LIVE and the requeue list kicked while the FETCH stream is still
running and ubq->canceling is still set. The kick redispatches a
previously requeued request, __ublk_queue_rq_common() sees ->canceling
and parks it again via __ublk_abort_rq(), and after the last FETCH
clears ->canceling nothing ever kicks the requeue list again: the
request is stranded there while holding its tag. If it is the flush
machinery's flush_rq, every subsequent fsync piles up in uninterruptible
sleep and teardown hangs on tag draining. This matches a report of a
lost PREFLUSH with ext4 on top of ublk after daemon crash recovery.
ub->completion is an edge-triggered latch used as a proxy for the level
condition "every queue has fetched all I/O commands", which can regress
(F_BATCH's UNPREP, daemon death) and whose re-arm can be skipped. Drop
it and wait on the real condition instead: the new helper
ublk_wait_dev_ready_and_lock() waits on ublk_dev_ready() via
wait_var_event_interruptible(), woken from ublk_mark_io_ready(), then
re-checks it under ub->mutex, waiting again on regression, and returns
with the mutex held and readiness guaranteed.
Readiness becomes true in the same ub->mutex critical section that
clears the last queue's ->canceling, so END_USER_RECOVERY marks the
device LIVE and kicks the requeue list strictly after ->canceling
clears. The wait stays interruptible, so a server whose daemon died can
still be signalled out. For ublk_ctrl_start_dev() this replaces the
fail-fast -EINVAL on an F_BATCH ready->UNPREP regression with waiting
until the device is ready again. |
| In the Linux kernel, the following vulnerability has been resolved:
drm/amdgpu/vcn4: avoid rereading IB param length
Reuse the parameter length returned by
vcn_v4_0_enc_find_ib_param() instead of rereading it from
the IB.
This avoids a potential TOCTOU issue if the IB contents
change between reads.
(cherry picked from commit dbb02b4755f8c1f3773263f2d779872c1c0c073a) |
| In the Linux kernel, the following vulnerability has been resolved:
drm/amdgpu: reject mapping a reserved doorbell to a new queue
When creating an user-queue, the user space
provides a doorbell BO handle and an offset within
the bo to obtain a doorbell.
However current implementation using xa_store_irq()
to store a doorbell, which allows a later queue created
with the same BO and offset parameters to overwrite an
existing queue and doorbell mapping.
This can cause problems like misrouting fence IRQ
processing to a wrong queue, and mislead the cleanup
process of one queue erasing the mapping of another queue.
This commit fixes this issue by replacing xa_store_irq with
xa_insert_irq, which rejects mapping a reserved
doorbell to a newly created queue
(cherry picked from commit 6244eae22966350db52faf9c1369d3b2ffc5de4e) |
| vLLM is an inference and serving engine for large language models. From 0.20.2rc0 until 0.26.0, safe_load_prompt_embeds in vllm/renderers/embed_utils.py uses torch.sparse.check_sparse_tensor_invariants, whose process-global save, enable, and restore state can be raced by concurrent prompt_embeds parts submitted to POST /v1/chat/completions through AsyncMultiModalItemTracker.resolve_items, asyncio.gather, and the default executor, allowing an invalid sparse tensor to reach tensor.to_dense despite the CVE-2025-62164 guard when enable_prompt_embeds is enabled. This issue is fixed in version 0.26.0. |
| In the Linux kernel, the following vulnerability has been resolved:
i2c: core: fix adapter deregistration race
Adapters can be looked up by their id using i2c_get_adapter() which
takes a reference to the embedded struct device.
Remove the adapter from the IDR before tearing it down during
deregistration (and on registration failure) to make sure its resources
are not accessed after having been freed (e.g. the device name). |
| In the Linux kernel, the following vulnerability has been resolved:
KVM: SVM: Disable AVIC IPI virtualization on Hygon Family 18h (erratum #1235)
Hygon Family 18h CPUs are derived from AMD Family 17h (Zen1) silicon and
share the same erratum #1235: hardware may read a stale IsRunning=1 bit
during ICR write emulation and silently fail to generate an
AVIC_IPI_FAILURE_TARGET_NOT_RUNNING VM-Exit on the sending vCPU.
The absence of the VM-Exit causes KVM to miss the required wakeup of
blocking target vCPUs, leading to hung vCPUs and unbounded delays in
guest execution.
Extend the existing AMD Family 17h erratum #1235 workaround to also cover
Hygon Family 18h. With IPI virtualization disabled, KVM never sets
IsRunning=1 in the Physical ID table, so every non-self IPI generates a
VM-Exit and is correctly emulated. |
| An authorization bypass vulnerability in LXD due to a timing flaw during configuration merging allows an authenticated attacker to bypass target project restrictions during cross-project instance copies. When copying an instance to a target project, LXD performs restriction checks before configuration merging is complete, creating a time-of-check to time-of-use (TOCTOU) condition. An attacker can exploit this flaw to copy instances with disallowed high-privilege configurations into restricted projects, bypassing security controls. |
| In the Linux kernel, the following vulnerability has been resolved:
netfs, afs: Fix write skipping in dir/link writepages
Fix netfs_write_single() and afs_single_writepages() to better handle a
write that would be skipped due to lock contention and WB_SYNC_NONE by
returning 1 from netfs_write_single() if it skipped and making
afs_single_writepages() skip also. If a skip occurs, the inode must be
re-marked as the VFS may have cleared the mark.
This is really only theoretical for directories in netfs_write_single() as
the only path to that is through afs_single_writepages() that takes the
->validate_lock around it, thereby serialising it. |
| In the Linux kernel, the following vulnerability has been resolved:
s390/ap: use generic driver_override infrastructure
When the AP masks are updated via apmask_store() or aqmask_store(),
ap_bus_revise_bindings() is called after ap_attr_mutex has been
released.
This calls __ap_revise_reserved(), which accesses the driver_override
field without holding any lock, racing against a concurrent
driver_override_store() that may free the old string, resulting in a
potential UAF.
Fix this by using the driver-core driver_override infrastructure, which
protects all accesses with an internal spinlock.
Note that unlike most other buses, the AP bus does not check
driver_override in its match() callback; the override is checked in
ap_device_probe() and __ap_revise_reserved() instead.
Also note that we do not enable the driver_override feature of struct
bus_type, as AP - in contrast to most other buses - passes "" to
sysfs_emit() when the driver_override pointer is NULL. Thus, printing
"\n" instead of "(null)\n".
Additionally, AP has a custom counter that is modified in the
corresponding custom driver_override_store(). |
| In the Linux kernel, the following vulnerability has been resolved:
drm/dp_mst: Handle torn-down topology gracefully in drm_dp_mst_topology_queue_probe()
A hotplug or link-loss event can tear down the MST topology
(setting mgr->mst_state = false and mgr->mst_primary = NULL) concurrently
with a caller invoking drm_dp_mst_topology_queue_probe(). Since the check
is already performed under mgr->lock, the condition is not a programming
error but a valid race -- the topology was valid when the caller decided
to call this function, but was torn down before the lock was acquired.
Replace the drm_WARN_ON() with a graceful early return. This eliminates
spurious kernel warnings and the resulting compositor crashes observed
when connecting/disconnecting DP MST monitors, while keeping the correct
behavior of doing nothing when MST is not active. A drm_dbg_mst() trace
is added so the skipped probe remains observable under MST debug logging.
The existing WARN_ON(mgr->mst_primary) in drm_dp_mst_topology_mgr_set_mst()
already catches the case where the topology is initialized twice, so no
diagnostic coverage is lost. |
| In the Linux kernel, the following vulnerability has been resolved:
xsk: cache csum_start/csum_offset to fix TOCTOU in xsk_skb_metadata()
The TX metadata area resides in the UMEM buffer which is memory-mapped
and concurrently writable by userspace. In xsk_skb_metadata(),
csum_start and csum_offset are read from shared memory for bounds
validation, then read again for skb assignment. A malicious userspace
application can race to overwrite these values between the two reads,
bypassing the bounds check and causing out-of-bounds memory access
during checksum computation in the transmit path.
Fix this by reading csum_start and csum_offset into local variables
once, then using the local copies for both validation and assignment.
Note that other metadata fields (flags, launch_time) and the cached
csum fields may be mutually inconsistent due to concurrent userspace
writes, but this is benign: the only security-critical invariant is
that each field's validated value is the same one used, which local
caching guarantees. |
| Time-of-check time-of-use race condition for the Intel(R) NPU Driver for Windows for all versions within Ring 1: Device Drivers may allow a denial of service. Unprivileged software adversary with an authenticated user combined with a high complexity attack may enable denial of service. This result may potentially occur via local access when attack requirements are present without special internal knowledge and requires no user interaction. The potential vulnerability may impact the confidentiality (none), integrity (low) and availability (high) of the vulnerable system, resulting in subsequent system confidentiality (none), integrity (none) and availability (none) impacts. |
| In the Aimeos Pagible content management system prior to version 0.10.4, the administrative proxy route (`cmsproxy`) is vulnerable to a Server-Side Request Forgery (SSRF) attack via DNS Rebinding. A Time-of-Check to Time-of-Use (TOCTOU) race condition exists between the URL validation phase and the actual HTTP request phase, allowing attackers to access internal network resources and cloud metadata endpoints. Version 0.10.4 fixes the issue. |
| In the Linux kernel, the following vulnerability has been resolved:
mm/khugepaged: write all dirty file folios when collapsing
[There is no upstream commit, as this code was removed by upstream
commit 044925f9b565 ("mm: fs: remove filemap_nr_thps*() functions and their users")]
As-is, khugepaged and writable-file opening exclude each other. A file
cannot be open writeable and have THPs (because the filesystem is not aware
of them). khugepaged will never collapse file pages for files that are
opened writeable. On an open(O_RDWR/O_WRONLY), the page cache for that
particular file is dropped. This is fine because nothing could've been
dirtied.
However, there is an edge-case: collapse_file() might not be able to
coexist with concurrent writers, but it can coexist with dirty folios
(from previous writers). Therefore, the following can happen:
open(file, O_RDWR)
write(file)
close(file)
madvise(file_mapping, MADV_COLLAPSE, some non-dirty range)
open(file, O_RDWR)
nr_thps > 0
truncate_inode_pages()
/* THPs are cleared out, but so are the dirty folios */
When this edge-case happens, there is data loss, as the dirty folios are
fully discarded.
Fix it by fully writing back the page cache (and waiting) when collapsing
file THPs. Doing so provides the guarantee that no dirty folio will be
observed while there are active THPs. To fully ensure this is safe, the
invalidate_lock needs to be held while doing the writeout, so that
do_dentry_open()'s page cache truncation excludes this write-and-wait.
As a side effect, move the nr_thps counter bumping outside the i_pages
lock. This is correct since the counter itself is an atomic_t and the
producer <-> consumer correctness is provided by a full memory barrier:
smp_mb() in collapse_file()/memory barrier implied by full ordering in
get_write_access() -> atomic_inc_unless_negative(). |
| In the Linux kernel, the following vulnerability has been resolved:
debugobjects: Plug race against a concurrent OOM disable
syzbot reported a puzzling splat:
WARNING: kernel/time/hrtimer.c:443 at stub_timer+0xa/0x20
stub_timer() is installed as timer callback function in
hrtimer_fixup_assert_init(), which is invoked when
debug_object_assert_init() can't find a shadow object. In that case debug
objects emits a warning about it before invoking the fixup.
Though the provided console log lacks this warning and instead has the
following a few seconds before the splat:
ODEBUG: Out of memory. ODEBUG disabled
So the object was looked up in debug_object_assert_init() and the lookup
failed due a concurrent out of memory situation which disabled debug
objects and freed the shadow objects:
debug_object_assert_init()
if (!debug_objects_enabled)
return; obj = alloc();
if (!obj) {
// Out of memory
debug_objects_enabled = false;
free_objects();
obj = lookup_or_alloc();
// The lookup failed because the other side
// removed the objects, so this returns
// an error code as the object in question
// is not statically initialized
if (!IS_ERR_OR_NULL(obj))
return;
if (!obj) {
debug_oom();
return;
}
print(...)
if (!debug_objects_enabled)
return;
fixup(...)
The debug object splat is skipped because debug_objects_enabled is false,
but the fixup callback is invoked unconditionally, which makes the timer
disfunctional.
This is only a problem in debug_object_assert_init() and
debug_object_activate() as both have to handle statically initialized
objects and therefore must handle the error pointer return case
gracefully. All other places only handle the found/not found case and the
NULL pointer return is a signal for OOM. Otherwise they get a valid shadow
object.
Plug the hole by checking whether debug objects are still enabled before
invoking the print and fixup function in those two places. |
| In the Linux kernel, the following vulnerability has been resolved:
afs: Fix afs_edit_dir_remove() to get, not find, block 0
Fix afs_edit_dir_remove() to use afs_dir_get_block() to get block 0 rather
than afs_dir_find_block() as the latter caches the found block in the
afs_dir_iter and may[*] switch out the page it's on if another
afs_dir_find_block() is done. This parallels what afs_edit_dir_add() does.
[*] There's more than one block per page. |
| Kernel software installed and running inside a Host VM may post improper commands to the GPU Firmware to trigger a memory write outside the permitted range of memory for the host kernel.
A TOCTOU bug existed where a malicious driver could modify values in memory after firmware validation but before use. |
| Diffusers is the a library for pretrained diffusion models. Prior to 0.38.0, Diffusers' DiffusionPipeline.from_pretrained flow can bypass the trust_remote_code guard because download() validates model_index.json and custom pipeline code before later loading from a cached folder that can change, allowing a Hub repository with custom .py pipeline code to execute through the custom pipeline flow without passing custom_pipeline or trust_remote_code=True. This issue is fixed in version 0.38.0. |
