| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| In the Linux kernel, the following vulnerability has been resolved:
nfsd: fix layout fence worker double-reference race
The workqueue core clears WORK_STRUCT_PENDING before the callback
is invoked, so delayed_work_pending() in lm_breaker_timedout() can
return false while the fence worker is already running. This lets
the breaker take a duplicate sc_count reference and schedule a new
worker that coalesces with the in-progress one. The extra reference
is never put, leaking the layout stateid.
Replace the racy delayed_work_pending() check with an
ls_fence_inflight boolean set atomically with
refcount_inc_not_zero() under ls_lock, and cleared under ls_lock
before the final nfs4_put_stid() on the dispose path; the retry
path intentionally retains it. Remove the self-rearm
mod_delayed_work() at the top of the worker. |
| In the Linux kernel, the following vulnerability has been resolved:
nfsd: don't free session slots that are still in use
nfsd4_sequence() can free the very slot it is currently processing.
When the session shrinker has reduced se_target_maxslots below
se_fchannel.maxreqs, the shrink path checks three conditions before
calling free_session_slots():
1. se_target_maxslots < maxreqs (shrink was advertised)
2. slot->sl_generation == se_slot_gen (slot is up-to-date)
3. seq->maxslots <= se_target_maxslots (client acknowledges)
However, seq->slotid is never checked against se_target_maxslots.
A client using a slot in the range [se_target_maxslots, maxreqs) can
satisfy all three conditions: its slot has the current generation
(set by a prior SEQUENCE), and it sends sa_highest_slotid <=
se_target_maxslots to acknowledge the reduction.
free_session_slots() then kfrees every slot at index >=
se_target_maxslots, including the caller's own slot. The function
continues to write sl_seqid, sl_flags, sl_generation, and stores the
dangling pointer in cstate->slot. Later, nfsd4_store_cache_entry()
copies up to maxresp_cached bytes of the compound reply into the freed
sl_data[] array, corrupting whatever slab object now occupies that
address.
Additionally, a concurrent thread processing SEQUENCE on a different
high-numbered slot can have its slot freed out from under it.
NFSD4_SLOT_INUSE is set under nn->client_lock before the lock is
released, so any concurrent thread past SEQUENCE will have its slot
marked. However, free_session_slots() does not check NFSD4_SLOT_INUSE
before freeing.
Fix both problems by:
1. Checking that the current request's slotid is below the shrink
boundary.
2. Scanning slots in the to-be-freed range for NFSD4_SLOT_INUSE and
deferring the shrink if any are active. |
| In the Linux kernel, the following vulnerability has been resolved:
scsi: fnic: Use GFP_ATOMIC for VLAN alloc under spinlock
fnic_fcoe_process_vlan_resp() allocates a VLAN descriptor with
kzalloc_obj() (default GFP_KERNEL) while holding vlans_lock via
spin_lock_irqsave(). GFP_KERNEL may sleep, which is not allowed in this
atomic context and can trigger a sleeping-from-invalid-context warning
or deadlock.
Pass GFP_ATOMIC so the allocation is safe under the IRQ-safe spinlock. |
| In the Linux kernel, the following vulnerability has been resolved:
SUNRPC: reject duplicate CREDS_VALUE options
gssx_dec_option_array() walks the wire-supplied option array and, for
every entry whose name matches CREDS_VALUE, calls
gssx_dec_linux_creds() on the same struct svc_cred. That helper
unconditionally installs a fresh groups_alloc() result into
creds->cr_group_info without releasing whatever pointer was already
there:
for (i = 0; i < count; i++) {
... decode name ...
if (length == sizeof(CREDS_VALUE) &&
memcmp(p, CREDS_VALUE, sizeof(CREDS_VALUE)) == 0) {
err = gssx_dec_linux_creds(xdr, creds);
...
}
}
A reply that carries two CREDS_VALUE entries therefore overwrites
cr_group_info on the second iteration and orphans the group_info
allocated by the first call. The earlier free_creds path only
releases the last cr_group_info via free_svc_cred(), so the first
allocation's refcount stays at one and its kvmalloc-backed storage
is leaked. No in-tree caller of gssp_accept_sec_context_upcall()
expects more than one CREDS_VALUE per reply.
Fix by tracking whether a CREDS_VALUE option has already been
decoded and returning -EINVAL on any subsequent match, so the
free_creds path releases the single group_info that was installed. |
| In the Linux kernel, the following vulnerability has been resolved:
SUNRPC: always drain cache_cleaner before destroying a cache_detail
sunrpc_destroy_cache_detail() only cancels the global cache_cleaner
delayed_work when cache_list is empty. During per-netns teardown
cache_list is never empty because init_net's caches remain registered,
so the cancel never fires. After unlink, the caller proceeds to
cache_destroy_net() which kfrees the cache_detail while cache_clean()
may still hold a dangling pointer to it. The result is a
use-after-free: cache_dequeue() takes cd->queue_lock on freed memory,
and cache_put() dereferences cd->cache_put as a function pointer from
freed slab.
Drop the list_empty guard so that cancel_delayed_work_sync() always
runs, ensuring any in-flight cache_clean() completes before the
cache_detail is freed. Re-arm the cleaner afterwards if other caches
are still registered. |
| In the Linux kernel, the following vulnerability has been resolved:
sunrpc: route to a populated pool in svc_pool_for_cpu()
svc_set_num_threads() spreads the requested threads evenly across the
service's pools (base = nrservs / sv_nrpools). When a service runs
fewer threads than it has pools -- e.g. an nfsd configured with fewer
threads than the host has NUMA nodes while running in "pernode" or
"percpu" mode -- the trailing pools are left with no threads at all.
svc_xprt_enqueue() selects a pool from the CPU servicing the transport,
queues the transport on that pool's sp_xprts, and only wakes a thread
from the same pool. Each thread services exclusively its own pool, so a
transport that lands on a threadless pool is enqueued on sp_xprts and
never picked up: the connection hangs indefinitely.
Have svc_pool_for_cpu() skip pools that currently have no threads,
falling back to the next populated pool. This trades NUMA locality for
a guarantee that the work is actually serviced. sp_nrthreads is only
updated under the service mutex; the lockless read here is a best-effort
routing hint, so annotate it with data_race(). |
| In the Linux kernel, the following vulnerability has been resolved:
nvmet-tcp: Do not WARN on remotely-controlled oversized SGL allocations
When fuzzing the nvme target code, I tripped a kernel warning in
nvmet_tcp_map_data() because the length passed into the allocator is
controlled by the remote initiator.
A remote initiator that sends a command with an SGL claiming a huge
number, can create a scatterlist and iovec allocation of over 1 million
entries, which causes the backing kmalloc call to exceed MAX_PAGE_ORDER
and then the page allocator will trip on a WARN_ON_ONCE_GFP() message:
WARNING: mm/page_alloc.c:5280 __alloc_frozen_pages_noprof
Workqueue: nvmet_tcp_wq nvmet_tcp_io_work
...
sgl_alloc_order
nvmet_tcp_map_data
nvmet_tcp_try_recv_pdu
As it's never good to trip a kernel warning remotely due to many systems
having panic-on-warn enabled, let's silence it by just add GFP_NOWARN to
the allocation flags. |
| In the Linux kernel, the following vulnerability has been resolved:
drm/amd/display: Skip Update HDCP Config In Transition State
Transition state does not have a valid dm_stream_ctx that should skip
configuring HDCP routine. The routine is valid to go through only when
a valid stream is created. |
| In the Linux kernel, the following vulnerability has been resolved:
btrfs: write-protect folios during data writeback
commit 095be159f3eb ("btrfs: unify folio dirty flag clearing") replaced
the folio_clear_dirty_for_io() call in extent_write_cache_pages() with a
plain folio_test_dirty() check. Besides clearing the dirty flag,
folio_clear_dirty_for_io() also calls folio_mkclean(), which write-protects
the shared mmap PTEs mapping the folio. Note that we still do call
folio_clear_dirty_for_io() later in submit_one_sector() when we clear
dirty on the last sector of the folio (the only sector for non-subpage
cases). But we lost this early call in extent_write_cache_pages().
Without the extra write-protection, a process with the file mmap-ed can
modify a sector while it is being used by writeback in a way that
expects a stable folio (checksumming, compressing, copying, etc...)
without faulting, which manifests as a handful of concrete bugs.
1. For large folios or subpage sectorsize, it is possible to submit a bio
which does not cover the whole folio. When this happens, we will have a
bio in flight for a folio that we have *not* called
folio_clear_dirty_for_io() on. If a task with an existing mmap-ed PTE
writes (without faulting..) in this window, it can result in
corruptions. If the write arrives while the checksumming or writing itself
is underway, this can result in an invalid checksum and later corruption
reports on read. If the write arrives after checksumming/writing is done
but before the last sector dirty is cleared, then the write is present
in page cache but doesn't affect the dirty tracking and will be lost
when the folio is fully finished being submitted and the dirty bit
is cleared. This results in losing the write even if fsync() is called.
2. For zoned submissions which are done in batch separate from the main
extent_writepage() loop, we also risk csum violations for those
submissions. Zoned writes are clamped to max_zone_append_size and are
not aligned with folios, so a submission can span two folios. The first
folio being processed in extent_write_cache_pages() will call
extent_write_locked_range() which will submit the partial range of the
next folio, while the rest of that folio could still be dirty. So
clearing dirty on the submitted sectors doesn't call
folio_clear_dirty_for_io() and we have the same issue. Since
extent_write_cache_pages() skips these batch submitted folios (they are
already marked for writeback from submission by the preceding folio), we
must add the extra write protection in lock_delalloc_folios().
3. For inline extents this will subtly risk losing writes that happen
after/while we copy the inline extent but before we clear dirty on
the folio.
4. For folios spanning EOF, mmap could tamper with the zeroed bytes past
EOF and cause them to be persisted where future faults would improperly
see them instead of zeros.
5. Finally, for compressed extents, we risk modifying the folios while we
work on compressing them which will result in corrupted compressed data.
Specifically, in run_delalloc_compressed() we queue up work to do
compress_file_range() in BTRFS_COMPRESSION_CHUNK_SIZE (512K) chunks which
will call btrfs_folio_clamp_clear_dirty() on the range. For non-subpage,
this will always clear the whole folio, safely. For subpage, we risk a
partial clear here as well. In particular, imagine a 2M folio broken up
into 512K chunks of work which might start compression work on one chunk
before all the chunks compress_file_range() workers have gotten far
enough to finish clearing all the dirty bitmaps of the folio and getting
to folio_clear_dirty_for_io(). Large folios on the edges of submission
ranges are similarly at risk to be only partly cleared.
This particular gap was introduced by a second patch in the same series:
commit a4ef54dbb576 ("btrfs: make extent_range_clear_dirty_for_io() to handle sector size < page size cases")
We cannot simply restore the call to folio_clear
---truncated--- |
| In the Linux kernel, the following vulnerability has been resolved:
ring-buffer: Fix subbuf resize race with ring buffer readers
trace_buffer subbuf_size is read lockless in ring_buffer_read_page() and
ring_buffer_read_start(), while it can simultaneously be resized with
ring_buffer_subbuf_order_set().
Instead of trace_buffer::subbuf_size, use bpage::order in
ring_buffer_read_start() and ring_buffer_read_page().
In ring_buffer_read_start(), even with resize_disabled, there is still a
possibility of a race with a buffer modification. Hold the trace_buffer
mutex to synchronise with any pending ring buffer order modification.
trace_buffer::subbuf_size is now actually useless, remove it. Also,
create accessors rb_subbuf_capacity() and rb_page_capacity() which
return the actual size available for storing events, while
rb_subbuf_size() returns the actual subbuf page-size. |
| In the Linux kernel, the following vulnerability has been resolved:
clocksource/drivers/nxp-pit: Fix IRQ leak on cpuhp_setup_state error path
When cpuhp_setup_state fails after pit_clockevent_per_cpu_init has
successfully called request_irq, the error handling jumps directly to
out_pit_clocksource_unregister without freeing the registered IRQ.
This leaks the IRQ line and, since kfree(pit) follows, leaves a
dangling pointer registered as the interrupt handler's dev_id,
potentially leading to a use-after-free if the IRQ fires afterwards.
Fix it by calling pit_clockevent_per_cpu_exit to properly release the
IRQ before falling through to the existing cleanup chain. |
| In the Linux kernel, the following vulnerability has been resolved:
fs: fix user path of nested backing files
backing_file_open() derives the path to be stored in the new backing
file from user_file->f_path. This is incorrect when user_file itself
is a backing file, which is the case for nested stacking filesystems,
e.g. overlayfs mounts where the lowerdir of one overlayfs is the merged
directory of another. Since commit def3ae83da02 ("fs: store real path
instead of fake path in backing file f_path") the f_path of a backing
file holds the real path of the intermediate layer, not the path that
the user opened.
Commit 924577e4f6ca ("ovl: Fix nested backing file paths") fixed this
for such configurations by passing file_user_path() from
ovl_open_realfile(). However, commit 6af36aeb147a ("lsm: add
backing_file LSM hooks") changed the first argument of
backing_file_open() from the user path back to the user file and
derived the path from user_file->f_path again, silently re-introducing
the problem.
As a result, files mapped through a nested overlayfs show the wrong
path in /proc/<pid>/maps and in perf/ftrace mmap records. For example,
with two nested overlayfs mounts:
mkdir -p /ovl/{lower,upper,work,merged} /ovl/nested
echo hello > /ovl/lower/foo
mount -t overlay overlay \
-o lowerdir=/ovl/lower,upperdir=/ovl/upper,workdir=/ovl/work \
/ovl/merged
# at least two lowerdirs are needed when upperdir is nonexistent
mount -t overlay overlay \
-o lowerdir=/ovl/merged:/ovl/lower /ovl/nested
mapping /ovl/nested/foo shows a disconnected path instead of the user
path:
# readlink /proc/self/fd/3
/ovl/nested/foo
# grep foo /proc/self/maps
7f6e2c100000-7f6e2c101000 r--s 00000000 00:24 15813027 /foo
The bogus path is derived from the f_path of the intermediate backing
file, whose mount is a private clone that d_path() cannot resolve.
Fix this by using file_user_path(), which returns the outermost
user-visible path for backing files and falls back to
&user_file->f_path for regular files. This restores the behavior of
commit 924577e4f6ca ("ovl: Fix nested backing file paths") for
overlayfs and also fixes the same problem for the other
backing_file_open() callers, fuse passthrough and erofs ishare, when
their user file is itself a backing file.
backing_tmpfile_open() has the same pattern but is not affected: it is
only called by ovl_create_tmpfile() for the upper layer, and another
overlayfs is rejected as upperdir by the DCACHE_OP_REAL check in
ovl_mount_dir_check(), so its user_file can never be a backing file. |
| In the Linux kernel, the following vulnerability has been resolved:
pidfd: hold exec_update_lock around namespace ioctl
The PIDFD_GET_*_NAMESPACE ioctls in pidfd_ioctl() perform a filesystem
credentials ptrace access check before handing out a namespace file
descriptor. The accompanying comment states that the code "mirrors nsfs
behavior", but, unlike the corresponding procfs paths, it does so without
holding the target task's exec_update_lock.
proc_ns_get_link() and proc_ns_readlink() both take exec_update_lock for
reading around the ptrace check and the namespace lookup, so that the
credentials used for the access decision match those of the task when its
namespace is read. Without it, a caller can pass the check against the
target's old credentials and then read the namespace after the target has
execve()'d a setuid binary and committed new credentials -- accessing
namespace information it should have been denied.
Hold exec_update_lock for reading around the ptrace check and the
namespace lookup so that pidfd truly mirrors nsfs behavior, as the comment
already claims. open_namespace() itself runs outside the lock: once a
namespace reference is obtained it carries its own refcount and is opened
with the caller's own credentials, so a concurrent execve() on the target
can no longer affect the outcome. |
| In the Linux kernel, the following vulnerability has been resolved:
timers/itimer: Zero-init old itimerval before copy to userspace
On native sparc64, struct __kernel_old_timeval contains a four-byte hole
after tv_usec because tv_sec is 64-bit while __kernel_suseconds_t is 32-bit.
put_itimerval() fills only the named fields in a stack-allocated
__kernel_old_itimerval and copies the entire object to userspace, so
getitimer() can expose the two padding holes.
Zero-initialize the aggregate before assigning the fields so implicit
padding is deterministic before it crosses the user/kernel boundary. |
| In the Linux kernel, the following vulnerability has been resolved:
KEYS: trusted: Fix TPM teardown ordering
trusted_tpm_exit() drops the TPM chip reference and frees the digest
array before unregistering the trusted key type. key_type_lookup()
holds key_types_sem for reading until the key operation finishes, while
unregister_key_type() takes it for writing. It therefore provides the
synchronization point that must precede backend teardown.
The current order permits this interleaving:
CPU 0 CPU 1
trusted_tpm_exit() key_type_lookup("trusted")
put_device(&chip->dev) trusted_tpm_seal()
kfree(digests) pcrlock()
unregister_key_type() tpm_pcr_extend(..., digests)
CPU 1 can consequently dereference the freed digest array. The chip can
also be released before callbacks stop using it.
KASAN reported:
BUG: KASAN: slab-use-after-free in tpm_pcr_extend+0x1f0/0x200
Read of size 2 at addr ffff88810872d000 by task poc/89
Call Trace:
tpm_pcr_extend+0x1f0/0x200
pcrlock+0x42/0x70 [trusted]
trusted_tpm_seal+0x1b6/0x570 [trusted]
trusted_instantiate+0x293/0x340 [trusted]
__key_instantiate_and_link+0xb2/0x2b0
__key_create_or_update+0x61e/0xb50
__do_sys_add_key+0x1b8/0x310
Allocated by task 88:
__kmalloc_noprof+0x1a7/0x490
do_one_initcall+0xa1/0x390
do_init_module+0x2df/0x840
Freed by task 90:
kfree+0x131/0x3c0
trusted_tpm_exit+0x59/0xa0 [trusted]
__do_sys_delete_module+0x346/0x510
Move unregister_key_type() before releasing either resource. This stops
new lookups and waits for in-flight key operations to finish before the
backend state is destroyed. |
| In the Linux kernel, the following vulnerability has been resolved:
apparmor: fix out-of-bounds write when null terminating a label vec
aa_vec_unique() null terminates at vec[n - dups] when VEC_FLAG_TERMINATE
is passed. If the components are all distinct no duplicates are dropped,
dups is 0 and the terminator goes to vec[n], so the caller has to provide
room for n + 1 entries.
aa_label_strn_parse() sets up its vector with vec_setup(profile, vec, len,
gfp) and then calls aa_vec_unique(vec, len, VEC_FLAG_TERMINATE), but
vec_setup() does not reserve the terminator entry. Up to LOCAL_VEC_ENTRIES
it uses the local array of LOCAL_VEC_ENTRIES pointers, above that it
allocates exactly len pointers. The terminator therefore lands one entry
past the end of the local array when len is LOCAL_VEC_ENTRIES, and one
entry past the end of the allocation when len is larger.
len comes from the number of "//&" separated components in the label name
and label_count_strn_entries() does not bound it. An unprivileged task
reaches the parse by writing to /proc/self/attr/apparmor/current or through
lsm_set_self_attr(2), both of which go through do_setattr(), and the name
is parsed before the change_profile permission is checked.
The query_label() path behind the securityfs .access file, which is
mode 0666, performs no permission check at all. Every component has to
resolve to a loaded profile, so a system with policy loaded is required.
The other two VEC_FLAG_TERMINATE users work on a label vec that
aa_label_alloc() has already sized with "+ 1 for null terminator entry on
vec". Reserve the same entry in vec_setup() and DEFINE_VEC(). Passing
len + 1 from the caller instead would move len == LOCAL_VEC_ENTRIES out of
the local array and into kzalloc(). |
| In the Linux kernel, the following vulnerability has been resolved:
mm, swap: don't free a hibernation slot that is in the swap cache
A slot with a folio in the swap cache is freed when the folio leaves the
cache, not when its count drops. swap_put_entries_cluster() follows that
rule. swap_free_hibernation_slot() does not, it calls
__swap_cluster_free_entries() whether or not a folio sits on the slot.
Cluster readahead can put one there. It walks a raw page_cluster sized
window of offsets around the faulting entry, and a hibernation slot passes
__swap_cache_add_check() because it is not a folio and its count is not
zero. Freeing the slot then clears the entry under that folio.
The folio is now unreachable from the swap table, and the offset goes back
to the allocator. The folio is still on the LRU though, so reclaim can
pick it up later. It then takes the old offset out of folio->swap and
overwrites the table entry there, which by then may belong to someone
else.
This bug can trigger silent memory corruption, process crashes, or data
instability across completely unrelated userspace applications - typically
occurring when uswsusp is preparing the hibernation image.
I found this while working on giving hibernation slots their own marker in
the swap table, which I had discussed with Kairui.
(https://lore.kernel.org/linux-mm/abp7aDgYLrxF3Me8@KASONG-MC4/) As far as
I know there are no reports, so there is no Reported-by/Closes to add.
Check for a cached folio before freeing. The slot is then left in the
ordinary state where only the swap cache holds it, and it is freed when
the folio leaves the cache, either through the reclaim below or through
normal reclaim later. |
| In the Linux kernel, the following vulnerability has been resolved:
mm/mempolicy: skip non-present PMDs when queueing folios
Patch series "mm: handle device-private PMDs in walk callbacks", v3.
Since commit 368076f52ebe ("mm/huge_memory: add device-private THP support
to PMD operations") a PMD may hold a device-private swap entry whenever an
HMM-based GPU driver migrates an anonymous THP folio to device memory via
migrate_vma_pages().
pmd_trans_huge_lock() succeeds for such PMDs (pmd_is_huge() returns true
for any non-present, non-none huge PMD), so several MM walk callbacks that
used to assume present THP or migration entry are now reachable with a
device-private PMD. The results range from a VM_BUG_ON() firing on debug
kernels, to an oops on a bogus vmemmap dereference, to silently isolating
an unrelated live folio from LRU in the aliasing case.
This patch (of 3):
queue_folios_pmd() is called under pmd_trans_huge_lock(), whose
pmd_is_huge() check returns true for any non-present, non-none PMD
softleaf. Passing such a PMD to pmd_folio() treats the softleaf encoding
as a hardware PFN and can return a bogus folio pointer.
Mirror queue_folios_pte_range(): handle non-present entries before looking
up a folio. Keep migration entries counted as failures, but skip other
non-present PMDs such as device-private entries.
Potential trigger: an HMM-based GPU driver migrates an anonymous THP folio
to device memory via migrate_vma_pages(), leaving a device-private PMD.
Userspace then calls mbind(), migrate_pages() or set_mempolicy_home_node()
on that range. |
| In the Linux kernel, the following vulnerability has been resolved:
mm/mglru: fix and remove redundant unevictable folio handling
sort_folio() has a shortcut for moving folios that are no longer evictable
but are still sitting on a generation list. However, this shortcut is
buggy. It does not follow the PG_lru usage convention, and it has a more
serious issue.
Unevictable folios are not threaded on lists[LRU_UNEVICTABLE], so that
folio->lru can be reused to hold folio->mlock_count (see the comment in
lruvec_init()). Hence lruvec_add_folio() skips the list_add() for them,
and every other place that turns a folio unevictable initialises
mlock_count explicitly: lru_add() sets it to 0, __mlock_folio() and
__mlock_new_folio() set it to !!folio_test_mlocked(folio). sort_folio()
sets nothing, and the lru_gen_del_folio() right above it may have already
poisoned folio->lru via list_del(), so mlock_count ends up aliasing
LIST_POISON2, which reads as 0x122, i.e. 290. The result is user
visible. On munlock, __munlock_folio() decrements that bogus count, finds
it still non-zero and bails out before clearing PG_mlocked, so the folio
remains unevictable and the Mlocked accounting stays inflated until the
folio is freed.
The shortcut also touches the LRU flags in the wrong order. It calls
lru_gen_del_folio() while PG_lru is still set, so a concurrent
folio_test_clear_lru() (e.g. compaction, folio_isolate_lru()) can succeed
on a folio that has already been taken off the generation list, which may
lead to unexpected behavior.
So fix it by isolating them as common folios and letting the generic
shrink path cull them. This matches the classical LRU behavior, and there
should be no visible effect on the generic eviction or isolation behavior.
There is no performance concern either, such a folio goes through this
once, and then it is off the generation lists for good. |
| In the Linux kernel, the following vulnerability has been resolved:
mm/migrate: report RCU-tasks quiescent states in migrate_pages_batch()
migrate_pages_batch() unmaps each folio before moving it, and every
unmap runs the mmu_notifier invalidate callbacks. On KVM hosts
try_to_migrate() ends up in kvm_mmu_notifier_invalidate_range_start() ->
tdp_mmu_zap_leafs(), which is expensive, so unmapping a large batch keeps
the CPU busy for a long time.
The loop already calls cond_resched(), but on PREEMPTION kernels that is
a no-op, and involuntary preemption is not a Tasks-RCU quiescent state.
A long batch therefore never reports a quiescent state, and the
migrating task (e.g. kcompactd) becomes a Tasks-RCU holdout, stalling the
Tasks-RCU grace period for minutes, which is common at Meta fleet:
INFO: rcu_tasks detected stalls on tasks:
0000000055349ecc: .. nvcsw: 1157401/1157401 holdout: 1 idle_cpu: -1/56 task:kcompactd0 state:R running task
Call Trace:
tdp_mmu_zap_leafs
tdp_mmu_next_root
gfn_to_pfn_cache_invalidate_start
kvm_mmu_notifier_invalidate_range_start
__mmu_notifier_invalidate_range_start
try_to_migrate_one
try_to_migrate
migrate_pages_batch
migrate_pages
compact_zone
compact_node
kcompactd
kthread
Use cond_resched_tasks_rcu_qs() so a quiescent state is reported even
when cond_resched() does nothing.
This has also been discussed at [1] |
