| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| In the Linux kernel, the following vulnerability has been resolved:
cxl/region: Block region delete during region creation
Expand the range lock, rename it "regions_lock", to disable region deletion
in the critical period between construct_region() and attach_target(), as
well as the period between device_add() and registering the remove actions.
Otherwise, userspace can confuse the kernel. It can violate the assumption
the region stays registered through the completion of cxl_add_to_region().
It can violate the assumption that devm_add_action_or_reset() is working
with a live 'struct cxl_region'.
It is ok for the region to disappear outside of those windows as that
mirrors device hotplug flows where the proper locks are held. |
| In the Linux kernel, the following vulnerability has been resolved:
cxl/region: Resolve region deletion races
Sungwoo noticed that the sysfs trigger to delete a region may try to delete
a region multiple times. It also has no exclusion relative to the kernel
releasing the region via CXL root device teardown.
Instead of installing new cxl root devres actions per region, use the
existing root decoder unregistration event to remove all remaining regions.
An xarray of regions replaces a devres list of regions.
This handles 3 separate issues with the old approach:
1/ sysfs users racing to delete the same region: no longer possible now
that the regions_lock is held over the lookup and deletion.
2/ multiple actions triggering deletion of the same region: solved by
erasing regions while holding @regions_lock, and only proceeding on
successful erasure.
3/ userspace racing devres_release_all() to trigger the devres not found
warning: solved by sysfs unregistration not requiring a release action |
| In the Linux kernel, the following vulnerability has been resolved:
coresight: tmc-etr: Fix race condition between sysfs and perf mode
When trying to run perf and sysfs mode simultaneously, the WARN_ON()
in tmc_etr_enable_hw() is triggered sometimes:
WARNING: CPU: 42 PID: 3911571 at drivers/hwtracing/coresight/coresight-tmc-etr.c:1060 tmc_etr_enable_hw+0xc0/0xd8 [coresight_tmc]
[..snip..]
Call trace:
tmc_etr_enable_hw+0xc0/0xd8 [coresight_tmc] (P)
tmc_enable_etr_sink+0x11c/0x250 [coresight_tmc] (L)
tmc_enable_etr_sink+0x11c/0x250 [coresight_tmc]
coresight_enable_path+0x1c8/0x218 [coresight]
coresight_enable_sysfs+0xa4/0x228 [coresight]
enable_source_store+0x58/0xa8 [coresight]
dev_attr_store+0x20/0x40
sysfs_kf_write+0x4c/0x68
kernfs_fop_write_iter+0x120/0x1b8
vfs_write+0x2c8/0x388
ksys_write+0x74/0x108
__arm64_sys_write+0x24/0x38
el0_svc_common.constprop.0+0x64/0x148
do_el0_svc+0x24/0x38
el0_svc+0x3c/0x130
el0t_64_sync_handler+0xc8/0xd0
el0t_64_sync+0x1ac/0x1b0
---[ end trace 0000000000000000 ]---
Since the enablement of sysfs mode is separeted into two critical regions,
one for sysfs buffer allocation and another for hardware enablement, it's
possible to race with the perf mode. Fix this by double check whether
the perf mode's been used before enabling the hardware in sysfs mode.
mode:
[sysfs mode] [perf mode]
tmc_etr_get_sysfs_buffer()
spin_lock(&drvdata->spinlock)
[sysfs buffer allocation]
spin_unlock(&drvdata->spinlock)
spin_lock(&drvdata->spinlock)
tmc_etr_enable_hw()
drvdata->etr_buf = etr_perf->etr_buf
spin_unlock(&drvdata->spinlock)
spin_lock(&drvdata->spinlock)
tmc_etr_enable_hw()
WARN_ON(drvdata->etr_buf) // WARN sicne etr_buf initialized at
the perf side
spin_unlock(&drvdata->spinlock)
With this fix, we retain the check for CS_MODE_PERF in get_etr_sysfs_buf.
This ensures we verify whether the perf mode's already running before we
actually allocate the buffer. Then we can save the time of
allocating/freeing the sysfs buffer if race with the perf mode. |
| In the Linux kernel, the following vulnerability has been resolved:
xfs: close crash window in attr dabtree inactivation
When inactivating an inode with node-format extended attributes,
xfs_attr3_node_inactive() invalidates all child leaf/node blocks via
xfs_trans_binval(), but intentionally does not remove the corresponding
entries from their parent node blocks. The implicit assumption is that
xfs_attr_inactive() will truncate the entire attr fork to zero extents
afterwards, so log recovery will never reach the root node and follow
those stale pointers.
However, if a log shutdown occurs after the leaf/node block cancellations
commit but before the attr bmap truncation commits, this assumption
breaks. Recovery replays the attr bmap intact (the inode still has
attr fork extents), but suppresses replay of all cancelled leaf/node
blocks, maybe leaving them as stale data on disk. On the next mount,
xlog_recover_process_iunlinks() retries inactivation and attempts to
read the root node via the attr bmap. If the root node was not replayed,
reading the unreplayed root block triggers a metadata verification
failure immediately; if it was replayed, following its child pointers
to unreplayed child blocks triggers the same failure:
XFS (pmem0): Metadata corruption detected at
xfs_da3_node_read_verify+0x53/0x220, xfs_da3_node block 0x78
XFS (pmem0): Unmount and run xfs_repair
XFS (pmem0): First 128 bytes of corrupted metadata buffer:
00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000040: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000050: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000060: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000070: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
XFS (pmem0): metadata I/O error in "xfs_da_read_buf+0x104/0x190" at daddr 0x78 len 8 error 117
Fix this in two places:
In xfs_attr3_node_inactive(), after calling xfs_trans_binval() on a
child block, immediately remove the entry that references it from the
parent node in the same transaction. This eliminates the window where
the parent holds a pointer to a cancelled block. Once all children are
removed, the now-empty root node is converted to a leaf block within the
same transaction. This node-to-leaf conversion is necessary for crash
safety. If the system shutdown after the empty node is written to the
log but before the second-phase bmap truncation commits, log recovery
will attempt to verify the root block on disk. xfs_da3_node_verify()
does not permit a node block with count == 0; such a block will fail
verification and trigger a metadata corruption shutdown. on the other
hand, leaf blocks are allowed to have this transient state.
In xfs_attr_inactive(), split the attr fork truncation into two explicit
phases. First, truncate all extents beyond the root block (the child
extents whose parent references have already been removed above).
Second, invalidate the root block and truncate the attr bmap to zero in
a single transaction. The two operations in the second phase must be
atomic: as long as the attr bmap has any non-zero length, recovery can
follow it to the root block, so the root block invalidation must commit
together with the bmap-to-zero truncation. |
| In the Linux kernel, the following vulnerability has been resolved:
md: suspend array while updating raid_disks via sysfs
In raid1_reshape(), freeze_array() is called before modifying the r1bio
memory pool (conf->r1bio_pool) and conf->raid_disks, and
unfreeze_array() is called after the update is completed.
However, freeze_array() only waits until nr_sync_pending and
(nr_pending - nr_queued) of all buckets reaches zero. When an I/O error
occurs, nr_queued is increased and the corresponding r1bio is queued to
either retry_list or bio_end_io_list. As a result, freeze_array() may
unblock before these r1bios are released.
This can lead to a situation where conf->raid_disks and the mempool have
already been updated while queued r1bios, allocated with the old
raid_disks value, are later released. Consequently, free_r1bio() may
access memory out of bounds in put_all_bios() and release r1bios of the
wrong size to the new mempool, potentially causing issues with the
mempool as well.
Since only normal I/O might increase nr_queued while an I/O error occurs,
suspending the array avoids this issue.
Note: Updating raid_disks via ioctl SET_ARRAY_INFO already suspends
the array. Therefore, we suspend the array when updating raid_disks
via sysfs to avoid this issue too. |
| A flaw was found in the admin backend of gvfs. The privileged gvfsd-admin daemon changes the ownership of newly created private D-Bus sockets by calling the link-following chown() function on a pathname inside a user-controlled directory. A local attacker can exploit this via a Time-of-Check Time-of-Use (TOCTOU) race condition and exchange the socket pathname with a symbolic link pointing to an arbitrary root-owned file (such as /etc/pam.d/su). The daemon subsequently follows the symlink and changes the ownership of the targeted root-owned file to the attacker's user ID. This allows an authenticated local attacker to modify critical system files, leading to a full local privilege escalation to root. |
| Improper privilege management in the log rotation mechanism of the Skylight Workspace Config Service in Amazon WorkSpaces for Windows before 2.6.2034.0 allows a local non-admin authenticated user to place arbitrary files into arbitrary locations bypassing file system permission protections, leading to local privilege escalation to SYSTEM. |
| Improper enforcement of single-use authorization code semantics in the JPA OAuth2 authorization code grant provider in Apache CXFallows a remote attacker to obtain multiple valid access tokens from a single authorization code via concurrent token exchange requests that race the non-atomic find-then-delete operation against a shared relational database under READ_COMMITTED isolation. Users are recommended to upgrade to versions 4.2.4 or 4.1.9 or 3.6.13, which fixes this issue. |
| Use after free in Windows Background Intelligent Transfer Service (BITS) allows an unauthorized attacker to execute code over a network. |
| In the Linux kernel, the following vulnerability has been resolved:
dm thin metadata: fix metadata snapshot consistency on commit failure
__reserve_metadata_snap() and __release_metadata_snap() modify the
superblock's held_root directly in the block_manager's buffer. If the
subsequent metadata commit fails, the held_root gets flushed to disk
through the abort_transaction path, resulting in inconsistent metadata.
Reproducer 1: __reserve_metadata_snap()
1. Create a 2 MiB metadata device and make the region after the 14th
block inaccessible, to trigger metadata commit failure in the
subsequent reserve_metadata_snap operation. The 14th block will be
the shadow destination for the index block.
dmsetup create tmeta --table "0 112 linear /dev/sdc 0
112 3984 error"
2. Create a 16 MiB thin-pool
dmsetup create tdata --table "0 32768 zero"
dd if=/dev/zero of=/dev/mapper/tmeta bs=4k count=1
dmsetup create tpool --table "0 32768 thin-pool /dev/mapper/tmeta \
/dev/mapper/tdata 128 0 1 skip_block_zeroing"
3. Take a metadata snapshot to trigger metadata commit failure and
transaction abort. However, the held_root is written to disk,
breaking metadata consistency.
dmsetup message tpool 0 "reserve_metadata_snap"
thin_check v1.2.2 result:
Bad reference count for metadata block 6. Expected 2, but space map contains 1.
Bad reference count for metadata block 7. Expected 2, but space map contains 1.
Bad reference count for metadata block 13. Expected 1, but space map contains 0.
Reproducer 2: __release_metadata_snap()
1. Create a 2 MiB metadata device and make the region after the 16th
block inaccessible, to trigger metadata commit failure in the
subsequent release_metadata_snap operation. The 16th block will be
the shadow destination for the index block.
dmsetup create tmeta --table "0 128 linear /dev/sdc 0
128 3968 error"
2. Create a 16 MiB thin-pool
dmsetup create tdata --table "0 32768 zero"
dd if=/dev/zero of=/dev/mapper/tmeta bs=4k count=1
dmsetup create tpool --table "0 32768 thin-pool /dev/mapper/tmeta \
/dev/mapper/tdata 128 0 1 skip_block_zeroing"
3. Reserve then release the metadata snapshot, to trigger metadata
commit failure and transaction abort. The held_root gets removed
from the on-disk superblock, causing inconsistent metadata.
dmsetup message tpool 0 "reserve_metadata_snap"
dmsetup message tpool 0 "release_metadata_snap"
thin_check v1.2.2 result:
Bad reference count for metadata block 6. Expected 1, but space map contains 2.
Bad reference count for metadata block 7. Expected 1, but space map contains 2.
1 metadata blocks have leaked.
Fix by deferring the held_root update to commit time.
Additionally, move the existing-snapshot check in __reserve_metadata_snap
before the shadow operation to avoid unnecessary work. In
__release_metadata_snap, clear pmd->held_root before btree deletion so
partial failure leaks blocks rather than leaving a stale reference, and
unlock the snapshot block before decrementing its refcount. |
| In the Linux kernel, the following vulnerability has been resolved:
drm/xe/vf: Fix VF CCS attach/detach race with in-flight BO moves
xe_bo_move() attaches VF CCS read/write batch buffers (BBs) to a BO
after it transitions NULL/SYSTEM -> TT, and detaches them after it
transitions TT -> SYSTEM. Both operations were done synchronously on
the CPU immediately after building the move's copy/clear fence,
without waiting for that fence to signal. This creates two races with
VF migration:
- Attach happens too late relative to the copy job it is meant to
protect. If the copy job is submitted before the CCS BBs are
attached, a VF migration event that pauses execution mid-copy can
observe partially copied CCS metadata without the attach state
needed to correctly save/restore it.
- Detach happens too early relative to the copy job that moves data
out of TT. The CCS BBs are torn down right after the copy fence is
obtained, while the actual blit may still be in flight. A VF
migration event that pauses execution mid-copy can then race the
save/restore path against the still-running blit, and the CCS BBs
it would need to make sense of the paused state have already been
removed.
Fix both races:
- Move the attach call to before the copy/clear job is submitted, so
the CCS BBs are already registered by the time the copy runs. On
attach failure, unwind and bail out of the move. xe_migrate_ccs_rw_copy()
now takes the destination resource explicitly, since bo->ttm.resource
is not updated to the new resource until after the move commits.
- Detach only after explicitly waiting for the copy fence to signal,
instead of tearing down the CCS BBs immediately after obtaining it.
While here, also fix xe_sriov_vf_ccs_attach_bo() to properly unwind and
propagate errors: the per-context loop previously never broke out on
error, silently discarding earlier failures. Unwind by clearing each
attached context directly via xe_migrate_ccs_rw_copy_clear() instead of
reusing xe_sriov_vf_ccs_detach_bo(), which requires both contexts to be
attached before it will clean up either one.
(cherry picked from commit d45ad0aa7a1eb5d7288b5ed948b05695611dc39e) |
| In the Linux kernel, the following vulnerability has been resolved:
net/packet: avoid fanout hook re-registration after unregister
packet_set_ring() temporarily detaches a socket from packet delivery while
reconfiguring its ring. It records the previous running state, clears
po->num, unregisters the protocol hook when needed, drops po->bind_lock,
and later restores po->num and re-registers the hook from the saved
was_running value.
That unlocked window can race with NETDEV_UNREGISTER. The notifier can
observe the socket as not running, skip __unregister_prot_hook(), and
invalidate the per-socket binding by setting po->ifindex to -1 and clearing
po->prot_hook.dev. A one-member fanout group can still retain its shared
fanout hook device pointer. When packet_set_ring() resumes, re-registering
solely from the stale was_running state can re-add the fanout hook after
the device has been unregistered.
Treat po->ifindex == -1 as an invalidated binding after reacquiring
po->bind_lock. This is distinct from ifindex 0, the normal
unbound/wildcard state: ifindex -1 marks an existing device binding that
was invalidated when the device was unregistered. Restore po->num as
before, but do not re-register the hook if device unregister already
detached the socket. |
| 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) |
| In the Linux kernel, the following vulnerability has been resolved:
proc: protect ptrace_may_access() with exec_update_lock (FD links)
proc_pid_get_link() and proc_pid_readlink() currently look up the task from
the pid once, then do the ptrace access check on that task, then look up
the task from the pid a second time to do the actual access.
That's racy in several ways.
To fix it, pass the task to the ->proc_get_link() handler, and instead of
proc_fd_access_allowed(), introduce a new helper call_proc_get_link() that
looks up and locks the task, does the access check, and calls
->proc_get_link(). |
| In the Linux kernel, the following vulnerability has been resolved:
KVM: x86: Ensure vendor's exit handler runs before fastpath userspace exits
Move the handling of fastpath userspace exits into vendor code to ensure
KVM runs vendor specific operations that need to run before userspace gains
control of the vCPU. E.g. for VMX (and soon to be for SVM as well), KVM
needs to flush the PML buffer prior to exiting to userspace, otherwise any
memory written by the final KVM_RUN might never be flagged as dirty.
Note, waiting to snapshot CR0 and CR3 until svm_handle_exit() is flawed in
general, as that risks consuming stale state in a fastpath handler. That
will be addressed in a future change. |
| In the Linux kernel, the following vulnerability has been resolved:
rbd: eliminate a race in lock_dwork draining on unmap
Given how rbd_lock_add_request() and rbd_img_exclusive_lock() are
written, lock_dwork may be (re)queued more than it's actually needed:
for example in case a new I/O request comes in while we are in the
middle of rbd_acquire_lock() on behalf of another I/O request. This is
expected and with rbd_release_lock() preemptively canceling lock_dwork
is benign under normal operation.
A more problematic example is maybe_kick_acquire():
if (have_requests || delayed_work_pending(&rbd_dev->lock_dwork)) {
dout("%s rbd_dev %p kicking lock_dwork\n", __func__, rbd_dev);
mod_delayed_work(rbd_dev->task_wq, &rbd_dev->lock_dwork, 0);
}
It's not unrealistic for lock_dwork to get canceled right after
delayed_work_pending() returns true and for mod_delayed_work() to
requeue it right there anyway. This is a classic TOCTOU race.
When it comes to unmapping the image, there is an implicit assumption
of no self-initiated exclusive lock activity past the point of return
from rbd_dev_image_unlock() which unlocks the lock if it happens to be
held. This unlock is assumed to be final and lock_dwork (as well as
all other exclusive lock tasks, really) isn't expected to get queued
again. However, lock_dwork is canceled only in cancel_tasks_sync()
(i.e. later in the unmap sequence) and on top of that the cancellation
can get in effect nullified by maybe_kick_acquire(). This may result
in rbd_acquire_lock() executing after rbd_dev_device_release() and
rbd_dev_image_release() run and free and/or reset a bunch of things.
One of the possible failure modes then is a violated
rbd_assert(rbd_image_format_valid(rbd_dev->image_format));
in rbd_dev_header_info() which is called via rbd_dev_refresh() from
rbd_post_acquire_action().
Redo exclusive lock task draining to provide saner semantics and try
to meet the assumptions around rbd_dev_image_unlock(). |
| When an audio channel's buffer needs to be resized, the sound(4) driver checks that the buffer is not memory-mapped and then drops the channel lock to allocate a new buffer. An mmap(2) call on another thread can complete during this window, after which the resize path frees the now-mapped buffer without rechecking.
An unprivileged local user with access to a sound device can exploit this race to trigger a use-after-free in the kernel, potentially escalating privileges. |
| In the Linux kernel, the following vulnerability has been resolved:
drm/gem: Try to fix change_handle ioctl, attempt 4
[airlied: just added some comments on how to reenable]
On-list because the cat is out of the bag and we're clearly not good
enough to figure this out in private. The story thus far:
5e28b7b94408 ("drm: Set old handle to NULL before prime swap in
change_handle") tried to fix a race condition between the gem_close and
gem_change_handle ioctls, but got a few things wrong:
- There's a confusion with the local variable handle, which is actually
the new handle, and so the two-stage trick was actually applied to the
wrong idr slot. 7164d78559b0 ("drm/gem: fix race between
change_handle and handle_delete") tried to fix that by adding yet
another code block, but forgot to add the error handling. Which meant
we now have two paths, both kinda wrong.
- dc366607c41c ("drm: Replace old pointer to new idr") tried to apply
another fix, but inconsistently, again because of the handle confusion
- this would be the right fix (kinda, somewhat, it's a mess) if we'd
do the two-stage approach for the new handle. Except that wasn't the
intent of the original fix.
We also didn't have an igt merged for the original ioctl, which is a big
no-go. This was attempted to address off-list in the original bugfix,
and amd QA people claimed the bug was fixed now. Very clearly that's not
the case. Here's my attempt to sort this out:
- Rename the local variable to new_handle, the old aliasing with
args->handle is just too dangerously confusing.
- Merge the gem obj lookup with the two-stage idr_replace so that we
avoid getting ourselves confused there.
- This means we don't have a surplus temporary reference anymore, only
an inherited from the idr. A concurrent gem_close on the new_handle
could steal that. Fix that with the same two-stage approach
create_tail uses. This is a bit overkill as documented in the comment,
but I also don't trust my ability to understand this all correctly, so
go with the established pattern we have from other ioctls instead for
maximum paranoia.
- Adjust error paths. I've tried to make the error and success paths
common, because they are identical except for which handle is removed
and on which we call idr_replace to (re)install the object again. But
that made things messier to read, so I've left it at the more verbose
version, which unfortunately hides the symmetry in the entire code
flow a bit.
- While at it, also replace the 7 space indent with 1 tab.
And finally, because I flat out don't trust my abilities here at all
anymore:
- Disable the ioctl until we have the igt situation and everything else
sorted out on-list and with full consensus.
v2:
Sashiko noticed that I didn't handle the error path for idr_replace
correctly, it must be checked with IS_ERR_OR_NULL like in
gem_handle_delete. So yeah, definitely should just the existing paths
1:1 because this is endless amounts of tricky.
Also add the Fixes: line for the original ioctl, I forgot that too. |
| In the Linux kernel, the following vulnerability has been resolved:
Revert "wireguard: device: enable threaded NAPI"
This reverts commit 933466fc50a8e4eb167acbd0d8ec96a078462e9c which is
commit db9ae3b6b43c79b1ba87eea849fd65efa05b4b2e upstream.
We have had three independent production user reports in combination
with Cilium utilizing WireGuard as encryption underneath that k8s Pod
E/W traffic to certain peer nodes fully stalled. The situation appears
as follows:
- Occurs very rarely but at random times under heavy networking load.
- Once the issue triggers the decryption side stops working completely
for that WireGuard peer, other peers keep working fine. The stall
happens also for newly initiated connections towards that particular
WireGuard peer.
- Only the decryption side is affected, never the encryption side.
- Once it triggers, it never recovers and remains in this state,
the CPU/mem on that node looks normal, no leak, busy loop or crash.
- bpftrace on the affected system shows that wg_prev_queue_enqueue
fails, thus the MAX_QUEUED_PACKETS (1024 skbs!) for the peer's
rx_queue is reached.
- Also, bpftrace shows that wg_packet_rx_poll for that peer is never
called again after reaching this state for that peer. For other
peers wg_packet_rx_poll does get called normally.
- Commit db9ae3b ("wireguard: device: enable threaded NAPI")
switched WireGuard to threaded NAPI by default. The default has
not been changed for triggering the issue, neither did CPU
hotplugging occur (i.e. 5bd8de2 ("wireguard: queueing: always
return valid online CPU in wg_cpumask_choose_online()")).
- The issue has been observed with stable kernels of v5.15 as well as
v6.1. It was reported to us that v5.10 stable is working fine, and
no report on v6.6 stable either (somewhat related discussion in [0]
though).
- In the WireGuard driver the only material difference between v5.10
stable and v5.15 stable is the switch to threaded NAPI by default.
[0] https://lore.kernel.org/netdev/CA+wXwBTT74RErDGAnj98PqS=wvdh8eM1pi4q6tTdExtjnokKqA@mail.gmail.com/
Breakdown of the problem:
1) skbs arriving for decryption are enqueued to the peer->rx_queue in
wg_packet_consume_data via wg_queue_enqueue_per_device_and_peer.
2) The latter only moves the skb into the MPSC peer queue if it does
not surpass MAX_QUEUED_PACKETS (1024) which is kept track in an
atomic counter via wg_prev_queue_enqueue.
3) In case enqueueing was successful, the skb is also queued up
in the device queue, round-robin picks a next online CPU, and
schedules the decryption worker.
4) The wg_packet_decrypt_worker, once scheduled, picks these up
from the queue, decrypts the packets and once done calls into
wg_queue_enqueue_per_peer_rx.
5) The latter updates the state to PACKET_STATE_CRYPTED on success
and calls napi_schedule on the per peer->napi instance.
6) NAPI then polls via wg_packet_rx_poll. wg_prev_queue_peek checks
on the peer->rx_queue. It will wg_prev_queue_dequeue if the
queue->peeked skb was not cached yet, or just return the latter
otherwise. (wg_prev_queue_drop_peeked later clears the cache.)
7) From an ordering perspective, the peer->rx_queue has skbs in order
while the device queue with the per-CPU worker threads from a
global ordering PoV can finish the decryption and signal the skb
PACKET_STATE_CRYPTED out of order.
8) A situation can be observed that the first packet coming in will
be stuck waiting for the decryption worker to be scheduled for
a longer time when the system is under pressure.
9) While this is the case, the other CPUs in the meantime finish
decryption and call into napi_schedule.
10) Now in wg_packet_rx_poll it picks up the first in-order skb
from the peer->rx_queue and sees that its state is still
PACKET_STATE_UNCRYPTED. The NAPI poll routine then exits e
---truncated--- |
