| CVE |
Vendors |
Products |
Updated |
CVSS v3.1 |
| In the Linux kernel, the following vulnerability has been resolved:
netfilter: nft_counter: serialize reset with spinlock
Add a global static spinlock to serialize counter fetch+reset
operations, preventing concurrent dump-and-reset from underrunning
values.
The lock is taken before fetching the total so that two parallel
resets cannot both read the same counter values and then both
subtract them.
A global lock is used for simplicity since resets are infrequent.
If this becomes a bottleneck, it can be replaced with a per-net
lock later. |
| In the Linux kernel, the following vulnerability has been resolved:
perf/x86/intel/uncore: Fix die ID init and look up bugs
In snbep_pci2phy_map_init(), in the nr_node_ids > 8 path,
uncore_device_to_die() may return -1 when all CPUs associated
with the UBOX device are offline.
Remove the WARN_ON_ONCE(die_id == -1) check for two reasons:
- The current code breaks out of the loop. This is incorrect because
pci_get_device() does not guarantee iteration in domain or bus order,
so additional UBOX devices may be skipped during the scan.
- Returning -EINVAL is incorrect, since marking offline buses with
die_id == -1 is expected and should not be treated as an error.
Separately, when NUMA is disabled on a NUMA-capable platform,
pcibus_to_node() returns NUMA_NO_NODE, causing uncore_device_to_die()
to return -1 for all PCI devices. As a result,
spr_update_device_location(), used on Intel SPR and EMR, ignores the
corresponding PMON units and does not add them to the RB tree.
Fix this by using uncore_pcibus_to_dieid(), which retrieves topology
from the UBOX GIDNIDMAP register and works regardless of whether NUMA
is enabled in Linux. This requires snbep_pci2phy_map_init() to be
added in spr_uncore_pci_init().
Keep uncore_device_to_die() only for the nr_node_ids > 8 case, where
NUMA is expected to be enabled. |
| In the Linux kernel, the following vulnerability has been resolved:
ext4: move ext4_percpu_param_init() before ext4_mb_init()
When running `kvm-xfstests -c ext4/1k -C 1 generic/383` with the
`DOUBLE_CHECK` macro defined, the following panic is triggered:
==================================================================
EXT4-fs error (device vdc): ext4_validate_block_bitmap:423:
comm mount: bg 0: bad block bitmap checksum
BUG: unable to handle page fault for address: ff110000fa2cc000
PGD 3e01067 P4D 3e02067 PUD 0
Oops: Oops: 0000 [#1] SMP NOPTI
CPU: 0 UID: 0 PID: 2386 Comm: mount Tainted: G W
6.18.0-gba65a4e7120a-dirty #1152 PREEMPT(none)
RIP: 0010:percpu_counter_add_batch+0x13/0xa0
Call Trace:
<TASK>
ext4_mark_group_bitmap_corrupted+0xcb/0xe0
ext4_validate_block_bitmap+0x2a1/0x2f0
ext4_read_block_bitmap+0x33/0x50
mb_group_bb_bitmap_alloc+0x33/0x80
ext4_mb_add_groupinfo+0x190/0x250
ext4_mb_init_backend+0x87/0x290
ext4_mb_init+0x456/0x640
__ext4_fill_super+0x1072/0x1680
ext4_fill_super+0xd3/0x280
get_tree_bdev_flags+0x132/0x1d0
vfs_get_tree+0x29/0xd0
vfs_cmd_create+0x59/0xe0
__do_sys_fsconfig+0x4f6/0x6b0
do_syscall_64+0x50/0x1f0
entry_SYSCALL_64_after_hwframe+0x76/0x7e
==================================================================
This issue can be reproduced using the following commands:
mkfs.ext4 -F -q -b 1024 /dev/sda 5G
tune2fs -O quota,project /dev/sda
mount /dev/sda /tmp/test
With DOUBLE_CHECK defined, mb_group_bb_bitmap_alloc() reads
and validates the block bitmap. When the validation fails,
ext4_mark_group_bitmap_corrupted() attempts to update
sbi->s_freeclusters_counter. However, this percpu_counter has not been
initialized yet at this point, which leads to the panic described above.
Fix this by moving the execution of ext4_percpu_param_init() to occur
before ext4_mb_init(), ensuring the per-CPU counters are initialized
before they are used. |
| In the Linux kernel, the following vulnerability has been resolved:
tcp: fix potential race in tcp_v6_syn_recv_sock()
Code in tcp_v6_syn_recv_sock() after the call to tcp_v4_syn_recv_sock()
is done too late.
After tcp_v4_syn_recv_sock(), the child socket is already visible
from TCP ehash table and other cpus might use it.
Since newinet->pinet6 is still pointing to the listener ipv6_pinfo
bad things can happen as syzbot found.
Move the problematic code in tcp_v6_mapped_child_init()
and call this new helper from tcp_v4_syn_recv_sock() before
the ehash insertion.
This allows the removal of one tcp_sync_mss(), since
tcp_v4_syn_recv_sock() will call it with the correct
context. |
| In the Linux kernel, the following vulnerability has been resolved:
net: af_key: zero aligned sockaddr tail in PF_KEY exports
PF_KEY export paths use `pfkey_sockaddr_size()` when reserving sockaddr
payload space, so IPv6 addresses occupy 32 bytes on the wire. However,
`pfkey_sockaddr_fill()` initializes only the first 28 bytes of
`struct sockaddr_in6`, leaving the final 4 aligned bytes uninitialized.
Not every PF_KEY message is affected. The state and policy dump builders
already zero the whole message buffer before filling the sockaddr
payloads. Keep the fix to the export paths that still append aligned
sockaddr payloads with plain `skb_put()`:
- `SADB_ACQUIRE`
- `SADB_X_NAT_T_NEW_MAPPING`
- `SADB_X_MIGRATE`
Fix those paths by clearing only the aligned sockaddr tail after
`pfkey_sockaddr_fill()`. |
| In the Linux kernel, the following vulnerability has been resolved:
bridge: mrp: reject zero test interval to avoid OOM panic
br_mrp_start_test() and br_mrp_start_in_test() accept the user-supplied
interval value from netlink without validation. When interval is 0,
usecs_to_jiffies(0) yields 0, causing the delayed work
(br_mrp_test_work_expired / br_mrp_in_test_work_expired) to reschedule
itself with zero delay. This creates a tight loop on system_percpu_wq
that allocates and transmits MRP test frames at maximum rate, exhausting
all system memory and causing a kernel panic via OOM deadlock.
The same zero-interval issue applies to br_mrp_start_in_test_parse()
for interconnect test frames.
Use NLA_POLICY_MIN(NLA_U32, 1) in the nla_policy tables for both
IFLA_BRIDGE_MRP_START_TEST_INTERVAL and
IFLA_BRIDGE_MRP_START_IN_TEST_INTERVAL, so zero is rejected at the
netlink attribute parsing layer before the value ever reaches the
workqueue scheduling code. This is consistent with how other bridge
subsystems (br_fdb, br_mst) enforce range constraints on netlink
attributes. |
| In the Linux kernel, the following vulnerability has been resolved:
net: usb: cdc_ncm: add ndpoffset to NDP16 nframes bounds check
cdc_ncm_rx_verify_ndp16() validates that the NDP header and its DPE
entries fit within the skb. The first check correctly accounts for
ndpoffset:
if ((ndpoffset + sizeof(struct usb_cdc_ncm_ndp16)) > skb_in->len)
but the second check omits it:
if ((sizeof(struct usb_cdc_ncm_ndp16) +
ret * (sizeof(struct usb_cdc_ncm_dpe16))) > skb_in->len)
This validates the DPE array size against the total skb length as if
the NDP were at offset 0, rather than at ndpoffset. When the NDP is
placed near the end of the NTB (large wNdpIndex), the DPE entries can
extend past the skb data buffer even though the check passes.
cdc_ncm_rx_fixup() then reads out-of-bounds memory when iterating
the DPE array.
Add ndpoffset to the nframes bounds check and use struct_size_t() to
express the NDP-plus-DPE-array size more clearly. |
| In the Linux kernel, the following vulnerability has been resolved:
net: usb: cdc_ncm: add ndpoffset to NDP32 nframes bounds check
The same bounds-check bug fixed for NDP16 in the previous patch also
exists in cdc_ncm_rx_verify_ndp32(). The DPE array size is validated
against the total skb length without accounting for ndpoffset, allowing
out-of-bounds reads when the NDP32 is placed near the end of the NTB.
Add ndpoffset to the nframes bounds check and use struct_size_t() to
express the NDP-plus-DPE-array size more clearly.
Compile-tested only. |
| In the Linux kernel, the following vulnerability has been resolved:
ceph: fix oops due to invalid pointer for kfree() in parse_longname()
This fixes a kernel oops when reading ceph snapshot directories (.snap),
for example by simply running `ls /mnt/my_ceph/.snap`.
The variable str is guarded by __free(kfree), but advanced by one for
skipping the initial '_' in snapshot names. Thus, kfree() is called
with an invalid pointer. This patch removes the need for advancing the
pointer so kfree() is called with correct memory pointer.
Steps to reproduce:
1. Create snapshots on a cephfs volume (I've 63 snaps in my testcase)
2. Add cephfs mount to fstab
$ echo "samba-fileserver@.files=/volumes/datapool/stuff/3461082b-ecc9-4e82-8549-3fd2590d3fb6 /mnt/test/stuff ceph acl,noatime,_netdev 0 0" >> /etc/fstab
3. Reboot the system
$ systemctl reboot
4. Check if it's really mounted
$ mount | grep stuff
5. List snapshots (expected 63 snapshots on my system)
$ ls /mnt/test/stuff/.snap
Now ls hangs forever and the kernel log shows the oops. |
| In the Linux kernel, the following vulnerability has been resolved:
ALSA: aloop: Fix racy access at PCM trigger
The PCM trigger callback of aloop driver tries to check the PCM state
and stop the stream of the tied substream in the corresponding cable.
Since both check and stop operations are performed outside the cable
lock, this may result in UAF when a program attempts to trigger
frequently while opening/closing the tied stream, as spotted by
fuzzers.
For addressing the UAF, this patch changes two things:
- It covers the most of code in loopback_check_format() with
cable->lock spinlock, and add the proper NULL checks. This avoids
already some racy accesses.
- In addition, now we try to check the state of the capture PCM stream
that may be stopped in this function, which was the major pain point
leading to UAF. |
| In the Linux kernel, the following vulnerability has been resolved:
of: unittest: Fix memory leak in unittest_data_add()
In unittest_data_add(), if of_resolve_phandles() fails, the allocated
unittest_data is not freed, leading to a memory leak.
Fix this by using scope-based cleanup helper __free(kfree) for automatic
resource cleanup. This ensures unittest_data is automatically freed when
it goes out of scope in error paths.
For the success path, use retain_and_null_ptr() to transfer ownership
of the memory to the device tree and prevent double freeing. |
| In the Linux kernel, the following vulnerability has been resolved:
cpuset: fix warning when disabling remote partition
A warning was triggered as follows:
WARNING: kernel/cgroup/cpuset.c:1651 at remote_partition_disable+0xf7/0x110
RIP: 0010:remote_partition_disable+0xf7/0x110
RSP: 0018:ffffc90001947d88 EFLAGS: 00000206
RAX: 0000000000007fff RBX: ffff888103b6e000 RCX: 0000000000006f40
RDX: 0000000000006f00 RSI: ffffc90001947da8 RDI: ffff888103b6e000
RBP: ffff888103b6e000 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000000001 R11: ffff88810b2e2728 R12: ffffc90001947da8
R13: 0000000000000000 R14: ffffc90001947da8 R15: ffff8881081f1c00
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007f55c8bbe0b2 CR3: 000000010b14c000 CR4: 00000000000006f0
Call Trace:
<TASK>
update_prstate+0x2d3/0x580
cpuset_partition_write+0x94/0xf0
kernfs_fop_write_iter+0x147/0x200
vfs_write+0x35d/0x500
ksys_write+0x66/0xe0
do_syscall_64+0x6b/0x390
entry_SYSCALL_64_after_hwframe+0x4b/0x53
RIP: 0033:0x7f55c8cd4887
Reproduction steps (on a 16-CPU machine):
# cd /sys/fs/cgroup/
# mkdir A1
# echo +cpuset > A1/cgroup.subtree_control
# echo "0-14" > A1/cpuset.cpus.exclusive
# mkdir A1/A2
# echo "0-14" > A1/A2/cpuset.cpus.exclusive
# echo "root" > A1/A2/cpuset.cpus.partition
# echo 0 > /sys/devices/system/cpu/cpu15/online
# echo member > A1/A2/cpuset.cpus.partition
When CPU 15 is offlined, subpartitions_cpus gets cleared because no CPUs
remain available for the top_cpuset, forcing partitions to share CPUs with
the top_cpuset. In this scenario, disabling the remote partition triggers
a warning stating that effective_xcpus is not a subset of
subpartitions_cpus. Partitions should be invalidated in this case to
inform users that the partition is now invalid(cpus are shared with
top_cpuset).
To fix this issue:
1. Only emit the warning only if subpartitions_cpus is not empty and the
effective_xcpus is not a subset of subpartitions_cpus.
2. During the CPU hotplug process, invalidate partitions if
subpartitions_cpus is empty. |
| 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:
ovl: fix double end_creating() on the casefold-mismatch path
ovl_create_real() releases the new dentry twice when the casefold
consistency check fails. The S_IFDIR branch calls end_creating() and
sets err, then falls through to the common out: label which calls
end_creating() on the same dentry again:
case S_IFDIR:
newdentry = ovl_do_mkdir(ofs, dir, newdentry, attr->mode);
err = PTR_ERR_OR_ZERO(newdentry);
if (!err && ofs->casefold != ovl_dentry_casefolded(newdentry)) {
pr_warn_ratelimited(...);
end_creating(newdentry); /* first */
err = -EINVAL;
}
break;
...
if (err)
goto out;
...
out:
if (err) {
end_creating(newdentry); /* second, same dentry */
return ERR_PTR(err);
}
end_creating() is end_dirop(), which does inode_unlock() on the parent
and dput() on the dentry, so the parent directory's i_rwsem is unlocked
twice and the dentry is put twice. The second unlock releases a lock
that is not held, which is what wedges every later creation under that
parent, and the second dput() drops a reference that was never taken.
The branch was added by commit dfc7da402ccc ("ovl: Check for casefold
consistency when creating new dentries") as a bare dput(), which already
released the reference twice; commit fe497f0759e0 ("VFS: change
vfs_mkdir() to unlock on failure.") converted both sites to
end_creating(), adding the double unlock.
This is reachable by an unprivileged user. The casefold consistency of
the layers is validated at mount time in ovl_parse_layer(), and again on
every lookup in ovl_lookup_single(), but ofs->workdir is the internal
"work" subdirectory created inside the user-supplied workdir, and that
subdirectory is not re-checked. Marking it casefolded after the mount
therefore makes every ovl_create_temp() inherit the wrong state - and
that path reaches ovl_create_real() through ovl_start_creating_temp(),
which uses start_creating() with a generated name and so never runs the
lookup-time check.
unshare -Urm
mount -t tmpfs -o casefold=utf8-12.1.0 tmpfs mnt
mkdir -p mnt/lower/d mnt/upper mnt/work mnt/merged
mount -t overlay ovl -o lowerdir=mnt/lower,\
upperdir=mnt/upper,workdir=mnt/work mnt/merged
chattr +F mnt/work/work
mkdir mnt/merged/d/sub # directory copy-up
overlayfs: wrong inherited casefold (work/#5)
and the next copy-up blocks forever on the parent's i_rwsem:
mkdir D start_creating+0x65/0xb0
ovl_start_creating_temp+0xb0/0xe0 [overlay]
ovl_create_temp+0xa3/0x1d0 [overlay]
ovl_copy_up_one+0x1f1c/0x21c0 [overlay]
ovl_copy_up_flags+0xf5/0x140 [overlay]
ovl_create_object+0xb7/0x220 [overlay]
ovl_mkdir+0x23/0x40 [overlay]
Drop the end_creating() from the branch and let out: own the cleanup,
which is what every other error path in this function already does. |
| In the Linux kernel, the following vulnerability has been resolved:
rust: devres: fix race between concurrent revokers
There is a potential race condition when two paths try to revoke a
Devres concurrently.
The driver core's devres_release_all() calls Revocable::revoke() via the
release callback, while Devres::drop() calls revoke_nosync() on another
CPU.
The revoker that does not claim the is_available swap returns
immediately, but the revoker that did may still be executing
drop_in_place() on the inner data. This can cause a use-after-free when
the other revoker's caller proceeds to drop adjacent resources that
drop_in_place() still references (e.g., Devres<DmaMappedSgt> racing with
SGTable freeing the backing sg_table and pages).
Fix this by adding a Completion. The release callback signals the
Completion after revoke() finishes, and Devres::drop() waits for it when
it loses the is_available swap. This ensures the wrapped object is fully
torn down before Devres::drop() returns. |
| In the Linux kernel, the following vulnerability has been resolved:
apparmor: fix cred UAF caused by begin_current_label_crit_section()
AppArmor's begin_current_label_crit_section() is a scary function called
from lots of LSM hooks (in particular VFS/socket-related ones) that checks
if the label referenced by the current creds is marked FLAG_STALE, and if
so, attempts to use aa_replace_current_label() to replace the creds with an
updated version that uses a new label.
The first problem with this is that it would directly lead to UAF of
`struct cred` if anything in the kernel takes a pointer to the current
creds and accesses these past a security hook invocation that replaces
creds, like so:
```
const struct cred *cred = current_cred();
alloc_file_pseudo(...);
uid_t uid = cred->euid;
```
I don't know if anything in the kernel actually does this, but I think it
is very surprising that this pattern could lead to UAF.
The second problem is that things go wrong when aa_replace_current_label()
runs with overridden credentials. aa_replace_current_label() bails out if
`current_cred() != current_real_cred()` (mirroring the check in
proc_pid_attr_write()), but this check can't actually reliably detect
overridden credentials because the overridden creds can be the same as the
objective creds.
So in approximately the following scenario, things go wrong:
1. task begins with <creds A> (as both objective and subjective creds),
with refcount=2
2. task grabs an extra reference on <creds A> for overriding
3. task calls override_creds(<creds A>), which returns a pointer to the old
subjective creds (<creds A>)
4. task enters AppArmor LSM hook
5. AppArmor checks that objective/subjective creds are equal
6. AppArmor replaces both cred pointers with <creds B> and drops 2 refs on
<creds A>
7. task leaves AppArmor LSM hook
8. task calls revert_creds(<creds A>)
9. now task->cred is <creds A> while task->real_cred is <creds B>, but the
task_struct logically holds two references to <creds B>
10. another task drops the extra reference on <creds A> that was used for
overriding, refcount drops to 0
11. now task->real_cred points to freed creds
At this point, any access to current_cred() will be UAF.
I have a test case where I run aa-disable on a profile while a process
using that profile is blocked on splice() from a FUSE passthrough file into
a full pipe; after the profile update, the pipe becomes empty, splice()
resumes, the credentials go out of sync, and a subsequent getuid() syscall
results in a KASAN UAF splat.
To fix this, instead of directly replacing creds, do it via task_work that
will run at the end of the current syscall. (The point in time at which the
cred replacement happens should have no correctness impact; it is just a
performance optimization to avoid unnecessarily touching the refcount of
the new label.)
Note that AppArmor still performs direct cred replacements in the
sb_pivotroot LSM hook after this change, and that direct cred replacements
can still happen in VFS ->write() callbacks via proc_pid_attr_write().
There are two options for what to do with aa_dup_task_ctx(): Either
explicitly reset new->label_replacement_pending after the entire
aa_task_ctx has been copied, or switch to manually copying members over.
I am switching to manually copying members over because that should make
bugs more obvious. |
| 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. |
