| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| In the Linux kernel, the following vulnerability has been resolved:
null_blk: serialize configfs attribute stores with the lock
The NULLB_DEVICE_ATTR _store takes no lock: apply_fn attributes
(submit_queues, poll_queues) get dev->NAME written again after apply_fn
returns, outside its lock; APPLY=NULL attributes are entirely lockless.
configfs only serializes stores per-open-file, so concurrent stores on
separate fds race.
For apply_fn attributes, once one store's apply_fn has reconfigured the
hardware, a second (losing) store can still overwrite dev->NAME
afterwards. This leaves dev->submit_queues out of sync with the live
queue count, which is later caught by the WARN_ON_ONCE() in
null_map_queues().
For !apply_fn attributes, power_store()'s null_add_dev() validates and
builds the device under "lock" but only sets CONFIGURED afterwards. A store
slipping in during this window can change a field mid-setup -- for example,
zone_nr_conv can be pushed above nr_zones after it has already been
clamped, leading to an out-of-bounds dev->zones[] access.
Take "lock" in the macro around the apply_fn call, the CONFIGURED test and
the field write, and move it out of nullb_apply_submit_queues()/
nullb_apply_poll_queues() so both paths are covered once. This serializes
stores with power_store's setup and with each other. |
| In the Linux kernel, the following vulnerability has been resolved:
RDMA/erdma: Hold CQ references when processing EQ events
EQ handlers look up CQs from dev->cq_xa and invoke CQ completion or
error callbacks outside the xarray lock. erdma_destroy_cq() can erase the
CQ from the xarray and free its queue buffer and doorbell record while a
previously scheduled EQ handler is still using the CQ.
Add a CQ refcount and take a reference under the xarray lock with
refcount_inc_not_zero(). Remove the CQ from the xarray before dropping
the destroy-path reference, then wait for in-flight EQ users before
releasing CQ resources. |
| In the Linux kernel, the following vulnerability has been resolved:
bpf: Disallow bpf_{g,s}etsockopt() in cgroup UNIX getname hooks
_bpf_setsockopt() and _bpf_getsockopt() call sock_owned_by_me() for
full sockets, so these helpers expect the socket lock to be held.
BPF_CGROUP_UNIX_GETPEERNAME and BPF_CGROUP_UNIX_GETSOCKNAME run BPF
programs without acquiring the socket lock. A program attached to
either hook can therefore trigger the sock_owned_by_me() warning by
calling bpf_setsockopt() or bpf_getsockopt().
Disallow bpf_setsockopt() and bpf_getsockopt() for CGROUP_UNIX_GETPEERNAME
and CGROUP_UNIX_GETSOCKNAME. |
| In the Linux kernel, the following vulnerability has been resolved:
ntfs: serialize resident iomap reads with mrec_lock
ntfs_read_iomap_begin_resident() walks the MFT record through
ntfs_attr_lookup() -> ntfs_attr_find() without taking ni->mrec_lock,
while ntfs_attr_record_resize(), ntfs_make_room_for_attr() and
ntfs_resident_attr_record_add() memmove() the same base_ni->mrec buffer
under that lock. map_mft_record() only takes a reference and does not
serialize, so the reader can observe torn attribute length and offset
fields while a writer is relocating the records.
KCSAN reports the race between the mmap read fault path and both link()
and unlink():
BUG: KCSAN: data-race in ntfs_attr_find / ntfs_attr_record_resize
write to 0xffff888100af1018 of 4 bytes by task 96 on cpu 1:
ntfs_attr_record_resize+0xd2/0x130
ntfs_attr_record_rm+0xad/0x530
ntfs_delete+0x224/0x640
ntfs_unlink+0x14d/0x280
vfs_unlink+0x157/0x520
read to 0xffff888100af1018 of 4 bytes by task 95 on cpu 0:
ntfs_attr_find+0x104/0x5b0
ntfs_attr_lookup+0x39c/0x10c0
ntfs_read_iomap_begin_resident+0xc6/0x230
ntfs_read_iomap_begin+0x5d/0xa0
iomap_iter+0x2e2/0x6e0
iomap_read_folio+0x147/0x2a0
ntfs_read_folio+0x108/0x170
filemap_read_folio+0x35/0x100
filemap_fault+0x993/0x1000
value changed: 0x00000250 -> 0x000001f0
The address is mrec + 0x18, i.e. mft_record.bytes_in_use, and the change
is the 96 bytes of one $FILE_NAME attribute being removed.
Keep base_ni->mrec_lock from the resident read iomap lookup through
iomap_end(). This protects both the attribute walk and the subsequent copy
from iomap->inline_data, which points into the MFT record. The non-resident
path is left alone: ntfs_lookup() already holds the directory inode's
mrec_lock when it reads an index folio through read_mapping_folio(), and
taking the lock in the shared wrapper deadlocks there with recursive locking
on mrec_lock. The comment above the read_mapping_folio() call in
fs/ntfs/dir.c notes the same hazard.
The seek path uses the same lookup helper but does not dereference
iomap->inline_data. Release the lock before returning from that path,
whereas the regular read path records base_ni in iomap->private and releases
the lock from its iomap_end() callback.
Tested with a reproducer that faults in a 16-byte resident file while
another thread runs link()/unlink() on it. Before: 40 KCSAN reports in
about one second. After: no reports in 180 seconds over 206,090 read
iterations and 423,540 link/unlink cycles. A PROVE_LOCKING build shows no
lockdep splat with the same reproducer running for 60 seconds. |
| In the Linux kernel, the following vulnerability has been resolved:
wifi: mt76: fix RXDMAD_C buffer recycling race
The RXDMAD_C buffers come from the RRO data queues' page pools, which are
bound to a different NAPI, so the direct page-pool recycle used here could
race the owning NAPI; take the non-direct path as is already done for WED
RX queues. |
| In the Linux kernel, the following vulnerability has been resolved:
null_blk: serialize configfs attribute updates with device setup
The attribute store methods generated with NULLB_DEVICE_ATTR() refuse to
change the configuration of a live device by testing
NULLB_DEV_FL_CONFIGURED, but that flag is only set by
nullb_device_power_store() after null_add_dev() has returned, and the
store methods take no lock at all. configfs only serializes writes to
the same open file (buffer->mutex), so a write to any attribute can run
concurrently with null_add_dev() and change the device configuration
while it is being used.
null_add_dev() reads the configuration several times, e.g. dev->zoned is
read once to set up the queue limits and once to initialize the zone
resources:
CPU0: echo 1 > nullb0/power CPU1: echo 1 > nullb0/zoned
nullb_device_power_store()
mutex_lock(&lock)
null_add_dev()
if (dev->zoned) -> false
/* no BLK_FEAT_ZONED */ nullb_device_zoned_store()
test_bit(FL_CONFIGURED) -> 0
dev->zoned = true
blk_mq_alloc_disk()
/* queue is not zoned */
if (nullb->dev->zoned) -> true
null_register_zoned_dev()
blk_revalidate_disk_zones()
blk_revalidate_disk_zones() is then called for a queue that does not
have BLK_FEAT_ZONED set, which triggers its WARN_ON_ONCE() and fails the
device setup with -EIO:
WARNING: CPU: 2 PID: 322 at block/blk-zoned.c:2357 blk_revalidate_disk_zones+0x4c/0x560
Clearing dev->zoned in the same window is worse: the queue is created
with BLK_FEAT_ZONED but the zone resources are never initialized, so
add_disk() succeeds for a zoned disk that has no zones. And a store that
lands after the last dev->zoned test leaves dev->zoned set while
dev->zones is still NULL, which null_process_zoned_cmd() dereferences on
the first write.
Fix this by taking the global lock, which nullb_device_power_store()
already holds across null_add_dev() and null_del_dev(), around both the
NULLB_DEV_FL_CONFIGURED test and the update of the device configuration.
The submit_queues and poll_queues apply callbacks are now called with
that lock held, so remove the locking they did themselves.
Since the store methods can run as soon as configfs_register_subsystem()
returns, that is, before null_init() gets to mutex_init(&lock), also
initialize the lock statically with DEFINE_MUTEX(). |
| In the Linux kernel, the following vulnerability has been resolved:
ALSA: seq: midi: Serialize input teardown with event_input
snd_midi_input_event() must not be running while a rawmidi substream is
closing, since this can lead to the trigger state becoming out-of-step
through this sequence in snd_rawmidi_input_trigger():
snd_rawmidi_input_trigger(up=0)
snd_midi_input_event()
-> snd_rawmidi_kernel_read()
-> snd_rawmidi_input_trigger(up=1)
-> cancel_work_sync()
which ends with the underlying device being active unexpectedly.
When this is called from close_substream(), further input can re-trigger
the input event leaving it running after rawmidi_release_priv() has set
rfile->rmidi to NULL which leads to:
Unable to handle kernel NULL pointer dereference at virtual address 00000000000000b0
Call trace:
snd_midi_input_event+0x3c/0x134 [snd_seq_midi] (P)
snd_rawmidi_input_event_work+0x1c/0x2c
process_one_work+0x150/0x3a4
worker_thread+0x190/0x318
Apply a similar approach to commit ef7607ab1c8ad ("ALSA: seq: midi:
Serialize output teardown with event_input") which fixed the same issue
in the output direction, but updated to use RCU following Takashi Iwai's
proposed follow-on patch [1].
With this change in place, midisynth_unsubscribe() clears the input file
so snd_midi_input_event() will not re-trigger the stream and will be
quiesced by the cancel_work_sync() in snd_rawmidi_input_trigger().
[1] https://lore.kernel.org/linux-sound/20260813144224.753399-1-tiwai@suse.de/ |
| In the Linux kernel, the following vulnerability has been resolved:
sunrpc: xprtsock: annotate shared socket callbacks with READ_ONCE/WRITE_ONCE
xprtsock replaces and restores sk->sk_data_ready and
sk->sk_write_space on live sockets with plain stores, and
xs_udp_do_set_buffer_size() invokes sk->sk_write_space via a plain
load. These callback pointers are shared with generic socket and
protocol paths that may read or invoke them concurrently, so xprtsock
needs the same READ_ONCE()/WRITE_ONCE() callback visibility contract
that the validated 4022 family applied elsewhere.
When SUNRPC takes over an AF_LOCAL, UDP, or TCP socket and later
restores the lower-socket callbacks during teardown, another CPU may
still hold an earlier callback snapshot. The plain replace/restore
pattern leaves the same visibility hole as the validated 4022 family,
so a stale snapshot can still invoke xs_data_ready() or
xs_udp_write_space() after the live callback fields have already been
restored to the lower-socket handlers.
Use WRITE_ONCE() for the shared sk_data_ready and sk_write_space
stores in xs_local_finish_connecting(), xs_udp_finish_connecting(),
xs_tcp_finish_connecting(), and xs_restore_old_callbacks(). Use
READ_ONCE() for the direct sk_write_space invocation in
xs_udp_do_set_buffer_size(). This matches the required callback
visibility contract while leaving adjacent sk_state_change and
sk_error_report handling unchanged. |
| In the Linux kernel, the following vulnerability has been resolved:
iommu/vt-d: Clear Present bit before tearing down copied context entry
copied_context_tear_down() zeroes the 128-bit context entry with
context_clear_entry() while the Present bit is still set, and only then
issues the context-cache and IOTLB invalidations. This leaves a window
in which hardware can fetch a torn entry, with some fields already zeroed
while Present is still set, leading to unpredictable behaviour or
spurious faults. While x86 provides strong write ordering, the compiler
may reorder the writes to the two 64-bit halves of the entry, and the
hardware fetch is not guaranteed to be atomic with respect to multiple
CPU writes.
There is no cacheline flush before the invalidation either, so on an
IOMMU without coherent access to the context table the zeroed entry may
not be visible to hardware at the point the invalidation is submitted.
Apply the same ownership handshake described in the VT-d spec, Section
6.5.3.3 ("Guidance to Software for Invalidations"): clear only the Present
bit, flush it out to the IOMMU, perform the invalidations, and only then
zero the remainder of the entry. |
| In the Linux kernel, the following vulnerability has been resolved:
iommu/tegra241-cmdqv: Synchronize the error ISR against VINTF (de)init
A user VINTF is torn down by tegra241_cmdqv_deinit_vintf(), which runs from
the destroy callback and from the init-failure unwind in the alloc handler.
It clears the cmdqv->vintfs[] slot and lets the iommufd core free it, but
nothing serializes that against the error interrupt: tegra241_cmdqv_isr()
reads cmdqv->vintfs[idx] and dereferences the vintf. A concurrent error can
make the ISR read a slot mid-clear (a NULL deref) or use a vintf which is
about to be freed (a use-after-free).
deinit_vintf() also returns idx to the IDA before clearing the slot, so a
concurrent create that reuses idx can publish its new vintf into the slot,
only for this teardown to erase it again with the stale NULL store.
On the other end, tegra241_cmdqv_init_vintf() publishes a new vintf with a
plain store to the cmdqv->vintfs[] slot, and the ISR dereferences fields of
a published vintf such as vintf->base. A plain store gives no ordering on a
weakly-ordered CPU, and a stale VINTF_ERR_MAP bit on a reused idx can make
the ISR pick a vintf the moment it is published, before its fields are set
or tegra241_vintf_hw_init() runs.
The cmdqv->vintfs[0] slot stays NULL until tegra241_cmdqv_init_structures()
first creates VINTF0, so the slot 0 read needs the same NULL check.
Publish every slot with an smp_store_release(), and read each slot in the
ISR with an smp_load_acquire() under a NULL check, so the ISR always sees
a fully built vintf or NULL. Also make deinit_vintf() clear the slot, and
synchronize_irq() prior to returning idx to the IDA, so no vintf is freed
under a running handler and no reused idx is clobbered. |
| In the Linux kernel, the following vulnerability has been resolved:
batman-adv: bla: avoid CRC corruption due to parallel claim add
batadv_bla_add_claim() is used to add claims and modify the backbone of
claims for CLAIM frames from remote backbones and local packets. When it
handles a claim, it needs to either
* add the new claim's CRC to the backbone CRC
* remove the already existing claim's CRC from the old backbone and add it
to the new backbone
But when the "new" claim code was running in parallel to the "change
backbone" code, it can happen that the CRC was invalid because the
backbone_gw of the claim was changed twice in the "new" claim code path:
* CPU0 creates the claim for gateway A and publishes it in the claim
hash. The crc16 of the address has not yet been added to A's crc at
this point.
* CPU1 processes a claim frame of gateway B for the same client, finds
the just published claim, and performs the ownership change: it
switches the pointer to B, removes the crc16 from A's crc - which
never contained it - and adds it to B's crc.
* CPU0 continues behind the creation branch, unconditionally switches
the pointer back to A without compensating B's crc (its remove_crc
is false for the creation path), and finally adds the crc16 to A's
crc
The CRC is then wrong for both:
* claim belongs to A: but CRC is not part of backbone A's CRC
* claim doesn't belong to B: CRC is still part of backbone B's CRC
This wrong CRC is never recomputated from the stored claims. For local
backbone claims, this can also not recovered using syncs.
To avoid this, split the functionality in clear separate parts:
* new claim which always adds claim CRC to the backbone CRC (but never
changes the already set backbone_gw of the claim back)
* update of existing claim which automatically changes the backbone_gw
entry and only updates both backbone CRCs when there was an actual change |
| In the Linux kernel, the following vulnerability has been resolved:
nfc: nci: fix double completion race in nci_data_exchange_complete
nci_close_device() and nci_rx_work can both call
nci_data_exchange_complete() concurrently. After commit 4527025d440ce8
("nfc: nci: fix circular locking dependency in nci_close_device") moved
flush_workqueue(ndev->rx_wq) after mutex_unlock(&ndev->req_lock),
rx_work is no longer serialized with the explicit completion call in the
close path. Both callers read the non-NULL callback pointer and invoke
rawsock_data_exchange_complete(), which calls sock_put() -- but only one
sock_hold() was taken, so the second sock_put() underflows the refcount
and frees the socket while it is still in use.
Replace the bare clear_bit(NCI_DATA_EXCHANGE) with
test_and_clear_bit() so that only the first caller proceeds to invoke
the callback. |
| In the Linux kernel, the following vulnerability has been resolved:
iommu/dma: Restore locking around msi_page_list
Unlike a group's default domain, which is always freshly allocated
and privately owned (iommu_group_alloc_default_domain()), VFIO type1's
legacy container merges any newly attached group into an existing
domain whenever their iommu_ops and cache-coherency enforcement match.
iommu_dma_get_msi_page() only asserts the caller's own group mutex is
held (iommu_group_mutex_assert()). On an IOMMU that publishes
IOMMU_RESV_SW_MSI, e.g. ARM SMMU, a VM with two such devices assigned
through the legacy container can have their guest drivers probe and
allocate MSIs in parallel; each host-side VFIO_DEVICE_SET_IRQS lands
on a different device fd and group mutex, but both devices' domains
are the same merged domain, so both can enter
iommu_dma_get_msi_page() concurrently and corrupt msi_page_list.
commit 288683c92b1a ("iommu: Make iommu_dma_prepare_msi() into a
generic operation") dropped the prior msi_prepare_lock on the
reasoning that "each iommu_domain is unique to a group," which holds
for default domains but not this VFIO type1 case. Restore the static
lock, since it's only guarding a corner case and will likely never
be contended.
iommufd avoids the equivalent problem by having its own callers
(iommufd_sw_map_msi()) take a ctx-wide sw_msi_lock before ever
reaching the shared list. VFIO type1 can't mirror that since it
dispatches to iommu_dma_sw_msi() which is outside VFIO's jurisdiction. |
| In the Linux kernel, the following vulnerability has been resolved:
net/sched: cls_api: fix teardown of an adopted proto on insert-race loss
In tc_new_tfilter() the create branch sets tp_created = 1 before calling
tcf_chain_tp_insert_unique(). When the caller loses the race (another
request inserted a proto at the same chain/prio first), insert_unique()
destroys the caller's own tp_new and returns the winner's proto with an
extra reference. tp_created was never cleared, so the loser's errout
path treated the winner's live proto as its own and called
tcf_chain_tp_delete_empty() on it, silently unlinking an active
classifier that the winning request already advertised via
RTM_NEWTFILTER.
Track the outcome of the insert step in a single tri-state variable so
each errout path reacts correctly:
- TP_NOT_CREATED: no proto created; pursue the old path.
- TP_CREATED: proto inserted successfully; same code path as before.
- TP_NOT_OWNED: New - lost the insert race; tp is another request's proto
(chain ref already released by tp_new's destroy)
Both errout reactions are single expressions derived from the state.
This fix is motivated by the Sashiko's automated review of Patch
(net/sched: cls_api: Always acquire rtnl_lock when destroying locked
classifiers) [1][2]. The review identified the silent-unlink behaviour of
an adopted proto's teardown when a request loses the
tcf_chain_tp_insert_unique() race.
[1] https://sashiko.dev/#/patchset/20260801125632.360365-1-jhs%40mojatatu.com
[2] https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260801125632.360365-1-jhs%40mojatatu.com |
| In the Linux kernel, the following vulnerability has been resolved:
regulator: core: use system_freezable_wq for init complete work
schedule_delayed_work() uses system_wq, which is non-freezable, allowing
regulator_init_complete_work to run concurrently with system suspend. This
work fires ~30s after boot to disable unused regulators via I2C. When it
races with PM suspend, the I2C adapter may already be suspended, triggering
a -ESHUTDOWN warning in __i2c_transfer():
WARNING: ... at __i2c_transfer+0x36c/0x3c8
Call trace:
__i2c_transfer
i2c_transfer
regmap_i2c_write
_regmap_update_bits
regulator_disable_regmap
_regulator_do_disable
regulator_late_cleanup
regulator_init_complete_work_function
process_one_work
Switch to system_freezable_wq so the work is frozen before any device
is suspended, eliminating the race. |
| In the Linux kernel, the following vulnerability has been resolved:
wifi: mt76: mt7915: clear wcid mask under mutex after RCU pointer clear
mt7915_remove_interface() cleared the wcid mask bit with no lock held and
before clearing the RCU wcid pointer. The mask is a non-atomic RMW shared
with the allocators, which all run under dev->mt76.mutex; on DBDC the two
wiphys share one mt76_dev, so this raced add_interface/sta_add on the
other band and could leak or double-hand-out a wcid. Clearing the bit
before the RCU pointer also let a concurrent allocation reuse the index
and publish its wcid, which the subsequent NULL assignment then wiped.
Move the clear into the existing mutex section, after the RCU pointer is
cleared. |
| In the Linux kernel, the following vulnerability has been resolved:
spi: davinci: switch to managed controller allocation
The controller is allocated with the non-managed spi_alloc_host() while
the interrupt is registered with devm_request_threaded_irq(). During
removal, spi_bitbang_stop() only unregisters the controller; the
subsequent spi_controller_put() then frees the controller together with
its embedded davinci_spi devdata, which is the IRQ handler's dev_id.
The devm_request_threaded_irq() release action (free_irq()), which
drains the handler, does not run until after .remove() returns. A late
or latched interrupt can therefore reach davinci_spi_irq() and
dereference already-freed memory.
Switch to devm_spi_alloc_host() so that the devres LIFO order releases
the controller only after free_irq() has drained the handler, and drop
the now-redundant spi_controller_put() from .remove(). The probe error
path is simplified to direct returns.
The clock is acquired with devm_clk_get_enabled(), which is registered
after the IRQ and thus released before it by the devres LIFO order.
Drain the interrupt explicitly with devm_free_irq() before disabling the
controller so that a late interrupt cannot access the registers of a
clock-gated controller.
This issue was found by an in-house static analysis tool. |
| In the Linux kernel, the following vulnerability has been resolved:
remoteproc: Prevent crash handling to race with rproc_del()
There's no synchronization between rproc_crash_handler_work() and
rproc_del(), as such it's possible for a driver to be removed while
crash-handler work is scheduled, or even executing - resulting in
use-after-free issues.
To avoid this the scheduled work need to be cancelled and synchronized
against before the removal proceeds.
In order to ensure that this doesn't race with the reporting, and
thereby scheduling new work, a "deleting" flag is introduced. This is
similar to the RPROC_DELETE state that was introduced to ensure that
"start" didn't race with rproc_del(), but the existing mechanism can not
be used as it's valid to call rproc_report_crash() in atomic context -
and the "state" is protected by a mutex.
In the event that work is cancelled the pm_stay_awake() is left
unbalanced and need to be unrolled.
The blocking and cancelling of crash-handler work prior to the actual
rproc_shutdown() call does have the explicit side-effect that crashes
resulting from the shutdown process will not enter the crash-handling
path, and as such will not generate devcoredumps etc. Due to the
existing mutual exclusion between these code paths there's no concrete
reduction in functionality, but further work would be needed to handle
this case. |
| In the Linux kernel, the following vulnerability has been resolved:
ext4: use fsdata to track inline data write state and fix race
Instead of checking the live inode state (ext4_has_inline_data(inode)
and ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) in the
write_end handlers, use the fsdata parameter of the address space
operations to explicitly pass down the state in which write_begin
prepared the write.
A concurrent thread (such as ext4_page_mkwrite()) can convert the
inline data to an extent between write_begin and write_end. If this
happens, the write_end handlers would previously miss the inline
write_end path and fall through to extent-based write_end logic.
However, since block buffers were never allocated in write_begin,
this resulted in NULL pointer dereferences or data loss because
folio_buffers(folio) was NULL.
Define EXT4_WRITE_DATA_INLINE (4) as a bit flag (Bit 2), treating
fsdata as bitwise flags rather than mutually exclusive enums to keep
states of the write path independent. Communicate this state via
fsdata:
1) ext4_write_begin() and ext4_da_write_begin() set the
EXT4_WRITE_DATA_INLINE bit in *fsdata via bitwise OR when an inline
write is successfully prepared.
2) On entry, ext4_write_begin() clears the EXT4_WRITE_DATA_INLINE bit
to safely handle VFS retries (where generic_perform_write() bypasses
the fsdata initialization on its retry jump).
3) The write_end handlers perform a bitwise AND to check if the
EXT4_WRITE_DATA_INLINE bit is set and invoke the inline write_end
helper accordingly.
Furthermore, during a buffered write, ext4_write_inline_data_end()
acquires the xattr lock after preparing the write. If a concurrent
page fault (ext4_page_mkwrite()) converts the inline data to an extent
after the write_end handlers check the state but before
ext4_write_inline_data_end() acquires the xattr write lock, the
subsequent check will trigger a kernel panic via
BUG_ON(!ext4_has_inline_data(inode)).
To keep git history working and bisectability clean, replace the
BUG_ON check in ext4_write_inline_data_end() with a graceful error-
handling retry path in this same commit. If the inline data is cleared
after locking the xattr, we safely release all resources (releasing
iloc.bh, unlocking/putting the folio, stopping the active journal
transaction handle) and return 0 (VFS retry) to let the generic write
path retry the operation safely. |
| In the Linux kernel, the following vulnerability has been resolved:
ext4: drain in-flight DIO before buffered write fallback
generic/746 started failing intermittently on ext3 (no-extent inodes).
The test triggers 'Page cache invalidation failure on direct I/O'
warnings and subsequent fsync returns -EIO. Adding a 50ms delay
between ext4_buffered_write_iter() and filemap_write_and_wait_range()
in ext4_dio_write_iter() makes the race almost always reproducible.
On no-extent inodes, DIO writes to holes cannot use unwritten extents,
so ext4_iomap_alloc() leaves m_flags=0 and ext4_map_blocks() returns 0.
The iomap layer then returns -ENOTBLK, causing fallback to buffered I/O.
The fallback path in ext4_dio_write_iter() calls
ext4_buffered_write_iter() which dirties pages, then does flush and
invalidate. However, there's an unprotected window between
ext4_buffered_write_iter() returning (with inode lock released) and
the subsequent flush+invalidate.
Concurrent async DIO completions from other threads can run
kiocb_invalidate_post_direct_write() during this window. If pages have
been re-dirtied, post-invalidation finds dirty pages and triggers the
warning, setting -EIO in the error sequence.
Consider a file with two 4k extents: [hole][written]. Thread A does
DIO to the written extent, while thread B does DIO spanning both:
kworker A (4k DIO, allocated block) kworker B (8k DIO, fallback)
----------------------------------- ----------------------------
inode_lock_shared() inode_lock_shared()
iomap_dio_rw(): iomap_dio_rw():
kiocb_invalidate_pages -> clean iomap_begin -> -ENOTBLK
submit_bio (async) dio->size = 0
inode_unlock_shared() inode_unlock_shared()
[bio pending in block layer] /* fallback: lock released */
ext4_buffered_write_iter()
inode_lock(exclusive)
generic_perform_write()
-> dirty pages [0, 8k]
inode_unlock(exclusive)
/* pages dirty, no lock */
[bio completes] filemap_write_and_wait_range()
iomap_dio_complete() -> flush dirty pages
kiocb_invalidate_post_direct_write() invalidate_mapping_pages()
invalidate_inode_pages2_range()
-> finds dirty page!
-> dio_warn_stale_pagecache()
-> errseq_set(-EIO)
This issue can be triggered through normal I/O paths, not just
intentionally overlapping DIO writes from userspace. For example,
generic/746 uses a loop device where multiple kworkers issue concurrent
I/O to the backing file. Additionally, when block_size < folio_size,
non-overlapping DIO writes that share a large folio can also trigger
the race.
Add inode_dio_wait() in ext4_buffered_write_iter() before
ext4_write_checks() to drain all in-flight DIO. This ensures that
all DIO clears existing pages before submitting IO (via
kiocb_invalidate_pages()), all BIO waits for all DIO to complete
(via inode_dio_wait()), and ext4_write_checks() observes the inode
size after all completed DIO so that ext4_block_zero_eof() does not
race with in-flight DIO, thus eliminating the race. |
