In the Linux kernel, the following vulnerability has been resolved:
Bluetooth: RFCOMM: connect the session socket without rfcomm_mutex
An RFCOMM connect() issued while a BR/EDR link is being authenticated
makes lockdep report a circular dependency, and the reported cycle is a
real AB/BA between rfcomm_mutex and hdev->lock.
rfcomm_security_cfm() is called from the HCI event path, which already
holds hdev->lock:
hci_rx_work()
hci_event_packet()
hci_cc_read_enc_key_size() [hdev->lock]
hci_encrypt_cfm() [hci_cb_list_lock]
rfcomm_security_cfm() [rfcomm_mutex]
while an RFCOMM connect() from userspace takes the same two locks the
other way round:
rfcomm_sock_connect()
rfcomm_dlc_open() [rfcomm_mutex]
__rfcomm_dlc_open()
rfcomm_session_create()
kernel_connect()
l2cap_sock_connect()
l2cap_chan_connect() [hdev->lock]
WARNING: possible circular locking dependency detected
kworker/u131:1/1128 is trying to acquire lock:
rfcomm_mutex, at: rfcomm_security_cfm+0x31/0x3e0 [rfcomm]
but task is already holding lock:
hci_cb_list_lock, at: hci_cc_read_enc_key_size+0x1d2/0xcc0
Chain exists of:
rfcomm_mutex --> &hdev->lock --> hci_cb_list_lock
hci_auth_complete_evt() and hci_encrypt_change_evt() reach the callback
the same way.
Both orders have to be seen in the same boot, which is why a BR/EDR
connection alone is not enough to show it: a session set up by the
remote side is created by rfcomm_accept_connection() in krfcommd, which
calls kernel_accept() and never takes hdev->lock under rfcomm_mutex.
Connecting a device that authenticates and encrypts the link and then
calling connect() on an RFCOMM socket towards any address - the connect
does not have to succeed, the order is recorded before the page timeout
- reports it every time.
Only the session socket has to be connected with the lock held, and it
does not: nothing else can see the socket before it is put on the
session list. So connect it first and take rfcomm_mutex afterwards,
which removes the rfcomm_mutex -> hdev->lock order for good, rather
than keeping the HCI event path out of rfcomm_mutex.
rfcomm_session_create() becomes rfcomm_session_connect(), which returns
the connected socket without touching the session list, and
rfcomm_dlc_open() adds the session once it holds the lock again. If
another opener added a session for the same pair while this socket was
connecting, that session is used and this socket is dropped.
__rfcomm_dlc_open() now takes the session it should use, and its state
check runs after the lock is re-acquired, so a DLC that was opened or
closed in the meantime is still handled.
Over an existing ACL link the connection can complete before the
session reaches the list, and the wakeup from the socket callback is
then lost, so krfcommd is woken once the session is visible.
Connecting without the lock opens a window in which the socket can be
closed. The DLC is not on a session yet, so rfcomm_dlc_close() finds
nothing to do and returns without touching it, and rfcomm_dlc_open()
would then attach a DLC whose owner is gone. rfcomm_dlc_close() now
marks such a DLC with RFCOMM_CLOSED, and rfcomm_dlc_open() checks the
mark once it holds rfcomm_mutex again and drops the socket instead of
attaching the DLC. The mark is cleared when an open starts, so a DLC
that is reused is not affected. This covers both callers of
rfcomm_dlc_open(), the socket and the tty layer.
In the Linux kernel, the following vulnerability has been resolved:
Bluetooth: RFCOMM: Fix NULL tty_dev dereference in rfcomm_dev_shutdown
syzbot hit a NULL pointer dereference in rfcomm_dev_shutdown() when an
RFCOMM tty is closed.
rfcomm_dev_add() puts the new device on rfcomm_dev_list first and only
then registers the tty device and saves it in dev->tty_dev. If that
registration fails, dev->tty_dev is never set at all. The device can be
looked up in the meantime, so an open() of /dev/rfcommN gets through
rfcomm_tty_install() and attaches to it. When that tty is closed again,
rfcomm_dev_shutdown() uses dev->tty_dev without checking it:
Oops: general protection fault, probably for non-canonical address
KASAN: null-ptr-deref in range [0x0000000000000040-0x0000000000000047]
RIP: 0010:rfcomm_dev_shutdown+0x48/0xb0 net/bluetooth/rfcomm/tty.c:131
Call Trace:
tty_port_shutdown+0x1bf/0x220 drivers/tty/tty_port.c:372
tty_port_close+0x4e/0x150 drivers/tty/tty_port.c:705
tty_release+0x359/0x1670 drivers/tty/tty_io.c:1718
__fput+0x418/0xa50 fs/file_table.c:512
rfcomm_dev_destruct() already copes with a NULL dev->tty_dev. Add the
same check to rfcomm_dev_shutdown(), so the tty device is only moved
when it was actually registered.
In the Linux kernel, the following vulnerability has been resolved:
drm/fb-helper: defer setup when fbdev_probe() fails, not just on -EAGAIN
__drm_fb_helper_initial_config_and_unlock() calls
drm_client_modeset_probe() first, which already fills in each
mode_set->mode/num_connectors for every connected output, and only
afterwards calls drm_fb_helper_single_fb_probe() to create
fb_helper->fb and wire it into those mode_sets via
drm_setup_crtcs_fb().
If drm_fb_helper_single_fb_probe() fails with anything other than
-EAGAIN (e.g. dev->driver->fbdev_probe() itself fails), the function
bails out before drm_setup_crtcs_fb() runs and before setting
fb_helper->deferred_setup. The client is left with mode_sets that
have a mode and connectors but mode_set->fb == NULL, and nothing
marks this fb_helper as needing a retry.
__drm_fb_helper_restore_fbdev_mode_unlocked() only skips committing
when fb_helper->deferred_setup is set, so a later .restore() call
(drm_lastclose() on process exit, in this report) commits the stale
mode_set through drm_client_modeset_commit() ->
drm_client_modeset_commit_atomic() -> __drm_atomic_helper_set_config(),
which hits WARN_ON(!set->fb).
Mark fb_helper->deferred_setup on any drm_fb_helper_single_fb_probe()
failure, not just -EAGAIN, so restore keeps skipping the commit until
a later hotplug event successfully re-probes and re-populates
mode_set->fb. The real error code is still returned to the caller
unchanged.
In the Linux kernel, the following vulnerability has been resolved:
drm/i915/gem: Prevent overstepping exec array boundary
It is possible for eb_relocate_parse_slow() to call kvfree() on
entries outside the original exec array during following
scenario inside eb_relocate_parse_slow():
1) eb_copy_relocations() allocates as many as eb->buffer_count
copies for exec entries;
2) eb_parse() appends up to two VMAs, right after incrementing
eb->buffer_count;
3) cleanup under "out" label uses this incremented buffer_count
to release relocation pointers from exec array via kvfree(),
not from VMA array.
To amend this problem, ensure that relocation cleanup uses the
original exec object count, excluding the VMAs appended by the
command parser.
(cherry picked from commit 2a0158925fb943c7ba8f3216e2734952128defc9)
In the Linux kernel, the following vulnerability has been resolved:
drm/i915/gt: Unmask interrupts only when ACTIVE is true
There is a small race window between intel_rps_boost() reading
a stale ACTIVE state and queuing rps->work, where ACTIVE may be
changed via asynchronous calls triggered by wakeref drop.
This means that rps_work() could be triggered after the GT has
been parked, so unmasking RPS PM interrupts would happen without
holding PM wakelock. This sequence of events may be observed
during live_rc6_ctx_wa() selftest:
1) live_rc6_ctx_wa()
-> intel_engine_pm_put()
-> wakeref_put()
ASYNC1) wakeref_put() callback
-> __engine_park()
-> intel_gt_pm_put_async()
-> intel_gt_pm_put_async_untracked()
-> intel_wakeref_put_async()
ASYNC2) wakeref_put_async() callback
-> __gt_park()
-> intel_rps_park()
-> intel_rps_clear_active() [clear ACTIVE]
WITH INTERRUPTS:
-> rps_disable_interrupts()
-> cancel_work_sync(rps->work)
-> rps_reset_interrupts() [rps->pm_iir = 0]
WITHOUT INTERRUPTS (TIMER mode):
-> rps_stop_timer()
-> cancel_work_sync(rps->work)
2) live_rc6_ctx_wa()
-> intel_gt_wait_for_idle()
-> intel_gt_retire_requests_timeout()
-> dma_fence_wait_timeout()
-> ops->wait() [i915_fence_wait()]
-> i915_request_wait_timeout()
-> intel_rps_boost()
-> intel_rps_is_active() [check ACTIVE]
[OPEN RACE WINDOW]
-> queue_work(rps->work)
-> retire_requests()
-> i915_request_retire()
-> intel_rps_dec_waiters() [from 1 to 0]
If ASYNC1 and ASYNC2 would execute during [OPEN RACE WINDOW],
then:
3) rps_work() executes from the workqueue and may satisfy
"(!pm_iir && !client_boost)" test (in TIMER mode rps->pm_iir
is zeroed by an earlier rps_work() call or never gets a new
value after previous intel_rps_unpark() event), so it jumps
to "out" label, bypassing intel_rps_is_active() guard and
executes gen6_gt_pm_unmask_irq();
4) GT is parked and runtime PM wakelock is dropped, so only
a raw reference in __gt_park() ->
intel_display_power_put_async() remains for 100ms, then the
MMIO write in gen6_gt_pm_unmask_irq() detects missing
wakelock via assert_rpm_wakelock_held() and emits
"RPM wakelock ref not held during HW access" tainting kernel.
Commit 3e7abf814193 ("drm/i915: Extract GT render power state management")
dropped the "if (rps->interrupts_enabled)" guard that
gen6_pm_rps_work() had around this unmask in i915_irq.c.
INTEL_RPS_ACTIVE brackets the same window, since it is set in
intel_rps_unpark() before rps_enable_interrupts() and
rps_start_timer() and cleared in intel_rps_park() before
rps_stop_timer() and rps_disable_interrupts().
Skip unmasking when RPS is no longer active (reaching unmask
operation is independent of INTERRUPT mode). This does not
prevent RPS operation from resuming on the next unpark.
In interrupt mode, intel_rps_unpark() calls
rps_enable_interrupts(), which enables and unmasks the PM
interrupts. In timer mode, it calls rps_start_timer(),
restarting periodic engine-busyness evaluation. The timer
generates software RPS events and queues rps_work() directly,
independently of PM interrupt delivery. Neither mode requires
the late worker to unmask interrupts while the GT is parked
(cherry picked from commit ca4a6deadad15b91ee083afe445fa6bdbfb68206)
In the Linux kernel, the following vulnerability has been resolved:
drm/imagination: Size page table preallocation by device address
Commit 0a8224058a58 ("drm/imagination: Fix page count for page table for
map() interface") passed the device address to
pvr_mmu_op_context_create(), but the preallocation is still sized from
device_addr + sgt_offset with an exclusive end. The offset into the
object's pages has no place in a device-virtual range, and the exclusive
end allocates one table too many when the range ends on a table
boundary.
Count the tables from device_addr and the inclusive end of the range.
The new calculation does not take page table indices into account. This
fixes another issue with the previous implementation, where the
difference between start and end indices could underflow when a mapping
crossed a page table boundary. This would lead to the driver
preallocating a huge number of MMU pages, exhausting system memory and
triggering an out of memory panic.
[Alessio: add paragraph to point out OOM bug being fixed, cc stable]
In the Linux kernel, the following vulnerability has been resolved:
drm/panthor: validate userspace queue count against initialized firmware slot count
In panthor_drv.c:panthor_ioctl_group_create(),
the queue count is validated against the hardcoded maximum:
if (!args->queues.count || args->queues.count > MAX_CS_PER_CSG)
return -EINVAL;
However, if the firmware reports fewer slots than MAX_CS_PER_CSG,
panthor_fw_init_ifaces() will only initialize that smaller number of
stream interfaces.
If unprivileged userspace requests more queues than the firmware reported
(but no more than MAX_CS_PER_CSG), cs_slot_prog_locked() fetches an
uninitialized stream interface and may crash when dereferencing
cs_iface->input->ringbuf_base.
In the Linux kernel, the following vulnerability has been resolved:
drm/vmwgfx: validate pitch coming from userspace
Validate the pitch, alongside the box dimensions before trying
to copy data from the underlying surface. Fixes possible
out of bounds reads with cursor snooping.
In the Linux kernel, the following vulnerability has been resolved:
drm/xe/i2c: cancel the client work on remove
xe_i2c_notifier() stores the DesignWare adapter in i2c->adapter and
schedules i2c->work when the adapter is registered under the xe I2C
platform device, and xe_i2c_client_work() then instantiates the AMC
client device on that adapter.
xe_i2c_remove() tears down the AMC, unregisters the client devices,
the bus notifier and the adapter platform device, but it never drains
i2c->work. A work item that is still queued or running when the
adapter is unregistered dereferences i2c->adapter in
i2c_new_client_device() after platform_device_unregister() has
released the adapter. The work item is also embedded in the
devm-allocated struct xe_i2c, so a work item still queued after the
drm device devm unwind frees that allocation runs its callback on
freed memory.
The bus notifier is the only thing that schedules this work, and it is
unregistered after the client devices. An instance that is still queued
when the teardown runs can therefore write
i2c->client[XE_I2C_CLIENT_AMC] while the loop is unregistering and
clearing the same array, and an AMC client it instantiates late is only
cleaned up by the adapter's own child sweep in i2c_del_adapter().
Move bus_unregister_notifier() in front of the client teardown loop
and cancel the work right after it, so no new instance can be
scheduled and a queued instance is drained before the client array is
touched. A running instance still finds a live adapter, since the
adapter is unregistered later.
(cherry picked from commit a23f27d5ec11c3f418c5e5531eea34ccda2b5b20)
In the Linux kernel, the following vulnerability has been resolved:
drm/xe: Fix shrinker accounting double-subtraction on nested external pins
xe_bo_pin_external() and xe_bo_unpin_external() support nested pins,
since the underlying ttm_bo_pin()/ttm_bo_unpin() maintain a pin_count
refcount rather than a boolean. However, the shrinker accounting calls
xe_ttm_tt_account_subtract() and xe_ttm_tt_account_add() are invoked
unconditionally on every pin_external()/unpin_external() call, instead
of only on the transition into and out of the pinned state.
An external BO (dma-buf) can be pinned more than once while already
pinned, for example when it has multiple attachments/importers, each
independently calling xe_bo_pin_external(). Each such nested pin
subtracts the BO's pages from the shrinker's accounting again, even
though the BO's pages were already removed from consideration by the
first pin. Symmetrically, each nested unpin adds them back again
before the BO is actually unpinned. This desynchronizes the
shrinker's page and object counts from reality, and since they are
declared as signed long but interpreted as unsigned long in
xe_shrinker_count(), the underflow can turn into an enormous
shrinkable/purgeable page count, causing the shrinker to be invoked
excessively under memory pressure.
Guard the accounting calls with the same pin-count transition checks
already used to guard the pinned_link list maintenance, so accounting
is only updated on the outermost pin and the final unpin.
While centralizing this logic into xe_bo_account_pin()/
xe_bo_account_unpin(), also close a second, related accounting gap
affecting imported dma-bufs. For an imported bo's ttm_tt, xe's
ttm_tt_populate hook (xe_ttm_tt_populate()) early-returns without
calling xe_ttm_tt_account_add(), since TTM_TT_FLAG_EXTERNAL is set
without TTM_TT_FLAG_EXTERNAL_MAPPABLE. However, TTM core's
ttm_tt_populate() wrapper unconditionally marks the tt as populated
on driver-hook success, so ttm_tt_is_populated() still reports true
for it. xe_bo_account_pin()/xe_bo_account_unpin() trusted
ttm_tt_is_populated() as a proxy for "these pages are accounted for
by the shrinker", and would subtract/add pages for such a tt that
were never added in the first place, again underflowing the
shrinker's counts. This is reachable today via xe_bo_pin_external(),
and would have become newly reachable via the fb-pin path once this
patch routes __xe_pin_fb_vma() through xe_bo_pin_account() below, so
fix it here rather than in a later patch, to avoid a commit in this
series that introduces the regression before fixing it. Skip the
accounting for imported bos, using the same EXTERNAL &&
!EXTERNAL_MAPPABLE test xe_ttm_tt_populate()/xe_ttm_tt_unpopulate()
use to decide whether to act, via a new xe_ttm_bo_is_imported()
helper (hoisted up from its only prior use site).
v2:
- Fix the critical issue caught by Sashiko AI review: the pin-count
transition check introduced here was still bypassed by the fb-pin
path (__xe_pin_fb_vma()/__xe_unpin_fb_vma()), which could leave a
stale pinned_link entry and a permanent shrinker accounting leak
when the fb-unpin performed the true last unpin. Addressed in a
separate follow-up patch in this series rather than folded in here,
since it is a distinct bug with its own Fixes: tag.
- Drop the unneeded local "last_unpin" variable in
xe_bo_unpin_external(), restoring the original inline
bo->ttm.pin_count == 1 check; it was a no-op rename not used by any
of the new accounting helpers, and was going to be deleted again by
the follow-up patch anyway.
v3:
- Exclude imported dma-bufs from the pin/unpin shrinker accounting,
fixing another underflow caught by Sashiko AI review.
- Add a short comment on the backup-object pin site noting it's
already populated/accounted at pin time, addressing (as a
false-positive clarification, no code fix needed) a separate
Medium-severity accounting-leak concern also caught by Sashiko AI
---truncated---
In the Linux kernel, the following vulnerability has been resolved:
drm/xe: Fix stale pinned_link entry when fb-pin performs the final unpin
xe_bo_pin_external() and xe_bo_unpin_external() maintain the bo's
pinned_link list membership in xe->pinned.late.external, based on
whether the current call is the outermost pin or the final unpin.
However, the same external bo's pin_count can also be raised and
lowered directly by __xe_pin_fb_vma()/__xe_unpin_fb_vma(), which pin
the bo as a display scanout buffer without going through
xe_bo_pin_external()/xe_bo_unpin_external() at all, and have no
notion of, or ownership over, pinned_link.
If a bo is pinned both externally (e.g. dma-buf export) and as an fb,
and the external unpin happens first, xe_bo_unpin_external() correctly
observes pin_count > 1 and leaves the bo on pinned_link. When the fb
unpin later performs the true last unpin (pin_count 1 -> 0), it never
touches pinned_link, leaving the bo linked on xe->pinned.late.external
indefinitely. Once the bo is subsequently freed, this stale list entry
points into freed memory, corrupting the list and risking a
use-after-free the next time the list is walked or spliced.
The backup object pin/unpin sites in xe_bo_notifier_prepare_pinned()/
xe_bo_notifier_unprepare_pinned() have the same bypass characteristic,
though they never add their bo to a pinned list, so are not affected
by this particular list-corruption issue.
Move the pinned_link removal into xe_bo_account_unpin(), which already
runs on every unpin path (kernel, external, framebuffer, backup
object) right before the true 1 -> 0 pin_count transition. Since
list_del_init() only operates on the node itself, this removal is
list-agnostic and safe to perform regardless of which list (external
or kernel_bo_present) the bo happens to be linked on, or which code
path is performing the final unpin. Drop the now-redundant explicit
list_del_init() calls in xe_bo_unpin_external() and xe_bo_unpin().
v2:
- New patch
v3:
- Rebased on changes to previous patch.
(cherry picked from commit ca58d36a1f93af586ee93b340c9f9fd81312cc01)
In the Linux kernel, the following vulnerability has been resolved:
drm/bridge: th1520-dw-hdmi: Fix remove() callback
This driver stores struct th1520_hdmi * in platform drvdata, but
th1520_dw_hdmi_remove() was reading it back as struct dw_hdmi *
and passing it to dw_hdmi_remove(), so teardown runs on the wrong
pointer.
Retrieve struct th1520_hdmi * from platform drvdata and pass
hdmi->dw_hdmi to dw_hdmi_remove().
In the Linux kernel, the following vulnerability has been resolved:
drm/amdgpu/userq: only accept doorbell BOs as queue doorbell
amdgpu_userq_get_doorbell_index() pins any BO passed as doorbell_handle
into the doorbell domain. For a regular BO, TTM moves it with
ttm_bo_move_null() and its contents are lost. Since an importer on the
same device gets the exporter's GEM object, a client can do this to a
buffer another client shared with it.
Reject BOs that were not created in the doorbell domain.
v2: drop the error message, use a local abo variable (Christian)
(cherry picked from commit f7497bbc19db4821c62b9d97b31c9086feb2d917)
In the Linux kernel, the following vulnerability has been resolved:
fs: don't silently unmount busy mounts
The propagate_mount_busy() helper exists to decide whether a synchronous
can suceed. It does that by looking at the copies of the victims in the
mounts its parent propagates to. One of those functions from propagation
hell.
It skips a copy that has child mounts except for the case of a single
child mount covering the copy's root. That algorithm used to be exact.
It isn't anymore.
Originally only childless copies were unmounted during propagation.
And 1064f874abc0 ("mnt: Tuck mounts under others instead of creating
shadow/side mounts.") ended up adding covered mounts to both sides in
one go.
Along came 99b19d16471e ("mnt: In propgate_umount handle visiting mounts
in any order") and that started to unmount more than before:
[1]: A copy of a mount is unmounted when each of its children is either
its overmount or a copy of the victim itself. In the former case
the overmount gets reparented. In the latter case the copy of the
victim will end up being unmounted.
So propagate_umount() ended up being rewritten without adjusting
propagate_mount_busy(). And currently propagate_mount_busy() doesn't
look at copies of the victim that have two or more children nor a single
child that is not the overmount.
So any such copy in [1] is not checked for
references. The result is that a synchronous unmount succeeds, pulling
out a mount that is still in use.
A container can run into this with its own mounts. The host shares a
tree with the container and the host has a mount in that tree:
mount -t tmpfs none /mnt
mount --make-shared /mnt
mount -t tmpfs none /mnt/a
The container's /mnt is a slave copy of the host mount. The container
now takes a detached copy of the tree and mounts that copy beneath its
propagated copy of /mnt/a and keeps the file descriptor to the tree
open:
unshare -m --propagation unchanged
mount --make-rslave /
fd = open_tree(AT_FDCWD, "/mnt", OPEN_TREE_CLONE | AT_RECURSIVE)
move_mount(fd, "", AT_FDCWD, "/mnt/a", MOVE_MOUNT_BENEATH)
grep /mnt /proc/self/mountinfo
775 355 ... /mnt ... master:646
777 775 ... /mnt/a ... master:646
776 777 ... /mnt/a ... master:647
778 777 ... /mnt/a/a ... master:647
The now attached tree (777) is now mounted beneath the container's copy
of /mnt/a. That's also where the host mount sits (as seen from the host
mount namespace). So the propagated copy (776) of the host's mount is on
top of its root and the copy of that (778) is inside it. Now the host
runs:
umount /mnt/a
and it succeeds. The moved tree tree and the copy inside of it are gone
from the container. Now only the propagated copy is left and it got
reparented to where it was before:
grep /mnt /proc/self/mountinfo
775 355 ... /mnt ... master:646
776 775 ... /mnt/a ...
But the file descriptor still refers to the moved tree. So a synchronous
umount detached a mount that is still in use.
Teach propagate_mount_busy() the same checks that propagate_umount()
does. It now checks for a single victim mount without children
for which it can assume that MNT_LOCKED is cleared on every copy like
propagate_umount() does. The copies form chains. IOW, when a copy itself
receives propagation from the victim's parent then the mount at the same
mountpoint inside that copy is a copy as well. Whether a mount receives
propagation from the victim's can simply be checked by walking its
masters. Walk each chain once from its top so we are sure that nested
copies stay linear. Check every reference once of every copy that gets
unmounted.
In the Linux kernel, the following vulnerability has been resolved:
namespace: reset the old parent's ->overmount in mnt_change_mountpoint()
__umount_mnt() resets ->overmount of the parent when it unmounts the
mount that overmounts it. mnt_change_mountpoint() doesn't. It moves the
mount to its new parent and leaves the old one pointing at it.
For a mount that is moved beneath in attach_recursive_mnt() that doesn't
matter. The new mount has already reset ->overmount of the old parent by
the time the old top mount is moved onto it.
It matters for reparent() though. propagate_umount() moves a surviving
overmount onto the first ancestor that isn't going to be unmounted. The
mount it was taken from is unmounted immediately afterwards. If a file
descriptor keeps that unmounted mount alive ->overmount points at the
moved mount. That's a dangling pointer and UAF as the moved mount is
unmounted and freed.
do_lock_mount() follows ->overmount from the target path under
mount_lock. With a file descriptor on such a mount as the target of a
mount beneath topmost_overmount() walks into the freed mount and
do_lock_mount() takes references on whatever it finds there.
So reset ->overmount of the old parent in mnt_change_mountpoint() the
same way __umount_mnt() does. A mount's ->overmount is then always a
mount that is attached to it.
In the Linux kernel, the following vulnerability has been resolved:
statmount: read the parent of an unmounted mount under mount_lock
statmount_mnt_basic() reads the ids of the parent through m->mnt_parent
with namespace_sem held for reading. That's fine for a mount that is
in a namespace. Everything above it is pinned by that namespace.
STATMOUNT_BY_FD also accepts a file descriptor on a mount that has
been unmounted on purpose and that only the file descriptor keeps alive.
If that mount stayed attached to its parent when it was unmounted the
parent is unmounted as well and nothing pins it for the child.
The final mntput() unmounts the children under mount_lock alone and
frees the parent afterwards. Concurrently, statmount_mnt_basic() reads
m->mnt_parent twice while that can happen and can easily dereference a
mount that's already gone.
Read the parent of an unmounted mount under mount_lock. The parent
can't be put while it is held and both ids come from the same mount. A
mounted mount's parent is pinned by the namespace and a mount can't
leave its namespace while namespace_sem is held, so the common case
takes no global lock.
In the Linux kernel, the following vulnerability has been resolved:
namespace: don't inherit MNT_UMOUNT in clone_mnt()
MNT_UMOUNT gets raised on a mount that umount_tree() or umount_one() has
taken out of its namespace. It's never cleared. __detach_mounts()
unhooks a mount that carries it instead of unmounting it.
disconnect_mount() lets an unmounted child stay attached only if its
parent carries it. propagate_umount() takes it as the sign that a mount
is going to be unmounted when it decides what else can be unmounted,
where a chain of locked mounts ends and where a for reparenting an
overmount.
clone_mnt() copies the flags of the source and only masks
MNT_INTERNAL_FLAGS which doesn't include MNT_UMOUNT. A mount that was
lazily unmounted but is kept alive by a file descriptor can still be
copied when its root is an nsfs or pidfs file because may_copy_tree()
accepts those before checking whether the source is mounted at all:
mount --bind /proc/self/ns/net /tmp/x
exec 3< /tmp/x
umount -l /tmp/x
mount --bind /proc/self/fd/3 /tmp/y
The mount on /tmp/y is attached to the caller's namespace with
MNT_UMOUNT already set and every copy of it inherits the flag again.
That causes a number of bugs:
- Unlink its mountpoint from a mount namespace where it isn't one.
__detach_mounts() takes it for a mount that was unmounted earlier,
unhooks it from its parent and drops the reference the namespace
holds on it. It stays in the namespace with one reference too few
and as its own parent, so it can't be unmounted through the file
descriptor anymore and the namespace never finds it again. When the
namespace goes away it's leaked and points to a freed struct
mnt_namespace that may_change_propagation() and anon_ns_root() read
on the next propagation change or mount_setattr() through the file
descriptor.
- Make it shared and unmount a peer. trace_transfers() follows the
flag from the peer into it and strips it from its peer group.
- Let a propagated umount reach it. gather_candidates() marks it with
T_UMOUNT_CANDIDATE but never puts it on the candidate list, so the
mark is never cleared and every later propagated umount skips it. If
it's the parent of a locked mount handle_locked() takes it as the
cutoff and commits the locked child that should have stayed and
disconnect_mount() leaves that child attached to it, unmounted under
a mounted parent, which is what disconnect_mount() exists to
prevent. When the namespace goes away umount_tree() visits the child
a second time. And once __detach_mounts() has made it its own parent
reparent() never finds the end of the chain of unmounted parents.
A clone is a new mount that has never been unmounted. Add MNT_UMOUNT
to MNT_INTERNAL_FLAGS so clone_mnt() never copies it. And warn when
umount_tree() reaches a mount that already has it set.
In the Linux kernel, the following vulnerability has been resolved:
namespace: don't copy an unmounted mount tree
may_copy_tree() accepts a mount whose root is an nsfs or pidfs file
before it checks whether that mount is mounted. This exception exists
for nsfs and pidfs mounts. They are part of no mount namespace so
neither check_mnt() nor is_mounted() would let them through and
bind-mounting a nsfsfd or a pidfd wouldn't work.
Problem is that exception also admits a bind mount of such a file when
it was lazily unmounted and is only kept alive by a file descriptor:
mount --bind /proc/self/ns/net /tmp/x
exec 3< /tmp/x
umount -l /tmp/x
mount --rbind /proc/self/fd/3 /tmp/y
With MS_REC or AT_RECURSIVE copy_tree() doesn't stop at the unmounted
mount. It copies every mount that stayed attached to it as well. So if
you have a bunch of overmounts over the nsfsfd or the pidfd the whole
unmounted stack is copied. That's a bug.
So refuse to walk an unmounted tree. Copying its root alone stays
possible. So a bind mount of a namespace that is gone can still be
recreated from a file descriptor on the old one. The file descriptor
pins that mount and its root and __has_locked_children() only reads the
direct children which don't leave the parent's list without
namespace_sem. The internal mounts carry MNT_NS_INTERNAL instead of NULL
and are still walked. Copying an unmounted tree recursively fails with
EINVAL.
In the Linux kernel, the following vulnerability has been resolved:
namespace: queue a mount only once for mount notifications
mnt_notify_add() puts a mount on notify_list. It doesn't check whether
the mount is on the list already. That's fine as long as every mount is
queued at most once per namespace_sem hold. It isn't.
propagate_umount() moves a surviving overmount off a stack of mounts
that are going away and queues it as moved. shrink_submounts() and
mark_mounts_for_expiry() call umount_tree() for several mounts under a
single namespace_sem hold. So the second umount_tree() can take down
exactly the mount the first one reparented, reparent it once more and so
end up queueing it a second time.
The second list_add_tail() cuts the mounts queued in between out of the
list while the head still points at the last of them. notify_mnt_list()
then only visits that one mount and the others are freed after the grace
period with notify_list still pointing at them. Every later mount
operation on the host walks freed memory:
BUG: KASAN: slab-use-after-free in __list_add_valid_or_report
Read of size 8 at addr ffff8881003d57e0 by task notify_dq/162
__list_add_valid_or_report+0x15c/0x1a0
mnt_add_to_ns+0x322/0x890
attach_recursive_mnt.isra.0+0xf2a/0x1a00
path_mount+0x139c/0x1d40
This needs a mount with MNT_SHRINKABLE to bind from. That's what
finish_automount() creates (submounts of NFS, AFS and CIFS and the
tracefs mount below debugfs). Bind mounts inherit the flag. A mount
namespace with a FAN_MARK_MNTNS mark does the rest and root in a user
namespace can have both.
Initialize to_notify in alloc_vfsmnt() and leave a mount alone that is
queued already. mnt_notify() looks at the state the mount has when the
list is drained so a single entry per mount is enough. The move event
for a mount that is unmounted under the same hold is lost. That's fine.
The detach is what matters.
In the Linux kernel, the following vulnerability has been resolved:
namespace: check a submount for references right before unmounting it
shrink_submounts() and mark_mounts_for_expiry() first collect all the
mounts they are allowed to unmount and then unmount them.
Whether a mount is busy is decided by propagate_mount_busy() on the
tree. But unmounting the first mount changes the tree that the second
one propagates into. propagate_umount()
moves a surviving overmount off a copy it unmounts and mounts it where
the copy was mounted. If that's where the propagated copy of the second
victim is looked up the overmount becomes a candidate of the second
umount_tree(). It's childless and so trim_one() commits it without
looking at its reference count.
So a synchronous umount pulls out a busy mount that is in use somewhere
else and marks it MNT_SYNC_UMOUNT while it has users:
umount2(/tmp/plshrink/p1, 0) = 0 errno 0
cwd is (unreachable)
The same two umounts requested one after the other fail with EBUSY.
This needs a mount with MNT_SHRINKABLE, i.e., automounted submounts of
NFS, AFS and CIFS or the tracefs mount below debugfs. Bind mounts
inherit the flag.
Check each mount right before it is unmounted under the same hold of
namespace_sem and mount_lock as the umount itself. Make sure that the
algorithm stays linear.
In the Linux kernel, the following vulnerability has been resolved:
namespace: check a recursive bind mount for mount namespace loops
do_loopback() refuses to bind mount the file of a mount namespace that
is as old as the caller's or older because a mount namespace that holds
a mount of its own file, or of an ancestor's, would cause a cycle. But
it only checks the dentry the bind mount starts from. With MS_REC
everything below it is copied and do_loopback() passes
CL_COPY_MNT_NS_FILE so mount namespace files below the source are
copied.
That's fine for a source in the caller's own mount namespace. Every
mount namespace file in there passed the same check when it was
bind-mounted. It isn't fine for a source in another mount namespace.
may_copy_tree() accepts a bind mount of any nsfs or pidfs file no matter
what mount namespace it lives in so that /proc/<pid>/ns/<ns> and pidfds
can be bind mounted. If we create a bind-mount stack of mount namespaces
file descriptors the kernel will copy them irrespective of their
ancestoral relationship to the mount namespace in question:
151 146 0:7 net:[4026531833] /tmp/nrc/y rw - nsfs nsfs rw
152 151 0:7 mnt:[4026532293] /tmp/nrc/y rw - nsfs nsfs rw
Mount 152 is a mount of the mount namespace file of mount namespace
4026532293 inside mount namespace 4026532293. That namespace and every
mount in it are leaked... All it takes is a process in an older mount
namespace that stacks the file and repeating it leaks without limit:
after control: Shmem: 380 kB
child: mount(/proc/self/fd/6, MS_BIND|MS_REC): ok
after cycle: Shmem: 65916 kB
do_move_mount() runs check_for_nsfs_mounts() over a detached tree for
exactly that reason. Do the same for the copy before it is grafted.
In the Linux kernel, the following vulnerability has been resolved:
namespace: keep covered mounts covered in OPEN_TREE_NAMESPACE
open_tree(OPEN_TREE_NAMESPACE) creates a new mount namespace from a
copy of the tree at the given path. The caller needs to be privileged
over its current user namespace because the new mount namespace will be
owned by it. It doesn't need to be privileged over the mount namespace
the tree is copied from. An unprivileged user just needs to create a
user namespace first.
The copy follows the rules of a detached bind mount. Without
AT_RECURSIVE only the mount itself is cloned and its children are left
out. With AT_RECURSIVE unbindable mounts are skipped. Both is fine when
the caller has privileges over the source mount namespace because it
could unmount those mounts anyway. Not so for an unprivileged user in
that mount namespace. For them a mount namespace copy is what unshare()
does. copy_mnt_ns() copies everything including unbindable mounts and
lock_mnt_tree() makes sure nothing can be unmounted in the copy. Nothing
that was covered gets revealed.
create_new_namespace() only does the locking and so the covered content
is right there in the new mount namespace:
# mount -t tmpfs none /tmp/otn; mkdir /tmp/otn/covered
# echo hidden > /tmp/otn/covered/under.txt
# mount -t tmpfs none /tmp/otn/covered
uid 1000, after unshare(CLONE_NEWUSER):
fd = open_tree(AT_FDCWD, "/tmp/otn", OPEN_TREE_NAMESPACE);
setns(fd, CLONE_NEWNS);
open("/covered/under.txt", O_RDONLY) -> "hidden"
Covering paths with mounts is how container runtimes mask parts of
/proc and /sys and how admins hide things.
When the caller's user namespace doesn't own the source mount namespace
copy the way unshare() does and refuse a non-recursive copy of a mount
that has anything mounted below the requested directory and copy
unbindable mounts in a recursive copy. A caller that is privileged over
the source mount namespace sees no change.
In the Linux kernel, the following vulnerability has been resolved:
namespace: look at the topmost mount for a mount namespace file
attach_recursive_mnt() preallocates a mountpoint for the root of the
topmost mount of the source so that a mount that already sits at the
destination can be put on top of the source or on top of one of its
propagated copies.
The copies are made without CL_COPY_MNT_NS_FILE so their chain of
overmounts is shorter than the source's. It ends below the first bind
mount of a mount namespace file. So while the loop walks up the chain of
the source it remembers the mountpoint of that mount in "shorter" and
the existing mount is put there for the copies.
But the loop ends at the topmost mount without ever looking at it. If
the topmost mount is the only mount namespace file in the chain
"shorter" stays NULL and mnt_change_mountpoint() attaches the existing
mount of the copy below the root of the mount namespace file. That's a
dentry of nsfs. No path walk in the copy's mount namespace ever gets
there. The mount is gone until its parent goes and the copy that took
its place can't be unmounted synchronously because it has a child:
S bind mount of a file
N bind mount of a mount namespace file on top of S
A shared mount, B a slave of A, Q mounted on B/file
move_mount(S, A/file)
cat B/file -> the content of S instead of Q
umount B/file -> EBUSY
Look at every mount of the chain including the topmost one.
In the Linux kernel, the following vulnerability has been resolved:
namespace: remove the fsnotify marks of a mount namespace in process context
mnt_ns_release() drops the last passive reference of a mount namespace
and removes its fanotify marks via fsnotify_mntns_delete(). That takes
the mutex of every group with a mark on the namespace and the spinlock
of the connector. Fine from process context. But mnt_ns_tree_remove()
hands the reference the namespace was allocated with to call_rcu() and
so the marks are removed from the RCU softirq:
BUG: sleeping function called from invalid context at kernel/locking/mutex.c:623
in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 0, name: swapper/3
__mutex_lock+0x113/0x24b0
fsnotify_destroy_marks+0x11b/0x3d0
mnt_ns_release_rcu+0x57/0xa0
rcu_core+0x6b6/0x1e40
and lockdep complains about the connector lock being taken from softirq
context. A fanotify group with a FAN_MARK_MNTNS mark on the mount
namespace of another task is all that's needed. Root in a user
namespace can do that for a mount namespace it owns. The task exits and
the namespace is freed with the mark still on it.
Remove the marks in free_mnt_ns() before the namespace is handed to RCU.
That runs in process context once the last active reference is gone. A
mark is added through a file descriptor to the namespace which holds an
active reference so no mark can show up after that.
In the Linux kernel, the following vulnerability has been resolved:
namespace: check the mounts before reading their parents in pivot_root()
path_pivot_root() reads the parents of new_root and of the caller's root
and checks whether they are shared before it checks that either mount is
in the caller's mount namespace. Only namespace_sem is held. That's fine
for a mount that is in the caller's mount namespace.
But new_root can be a file descriptor to a mount that has been unmounted
and that only the file descriptor keeps alive. If that mount stayed
attached to its parent when it was unmounted nothing pins the parent for
it. The final mntput() of the parent unhooks the children under
mount_lock alone and frees the parent afterwards:
pivot_root() close(fd), last ref on the parent
------------ ---------------------------------
ex_parent = new_mnt->mnt_parent
mntput_no_expire_slowpath()
__umount_mnt(new_mnt)
cleanup_mnt()
call_rcu()
IS_MNT_SHARED(ex_parent)
BUG: KASAN: slab-use-after-free in path_pivot_root+0xf1a/0x1840
Read of size 4 at addr ffff8881047382f8 by task pivot_widen/157
path_pivot_root+0xf1a/0x1840
__x64_sys_pivot_root+0x165/0x190
The same goes for the caller's root via chroot(). Both outcomes of the
check end in EINVAL so nothing but the read itself goes wrong. It's the
same thing commit bb4609405752 ("statmount: read the parent of an
unmounted mount under mount_lock") fixed for statmount().
Check that both mounts are in the caller's namespace before their
parents are read. Every path returns EINVAL either way.
In the Linux kernel, the following vulnerability has been resolved:
dcache: don't put a mountpoint on a dentry that's being removed
rmdir(), unlink() and rename() call dont_mount() on the victim and then
detach_mounts() with the victim's inode locked. do_lock_mount() takes
the inode lock of the mountpoint and checks cant_mount() so a mount
can't show up after detach_mounts().
But attach_recursive_mnt() makes a second mountpoint for the root of the
source mount so that the mounts already located at the destination can
be put on top of it. No inode is locked for that one and d_set_mounted()
only refuses a dentry that is unlinked. Between detach_mounts() and
d_delete() the victim is still hashed:
rmrace: b passed dont_mount() and detach_mounts(), sleeping
T2: move_mount(S, /tmp/plcant/x, BENEATH) = 0 errno 0 ()
T1: rmdir(/tmp/plcant/d/b) = 0 errno 0
109 107 0:61 /d/b//deleted /tmp/plcant/x rw,relatime - tmpfs tmpfs rw
108 109 0:63 / /tmp/plcant/x rw,relatime - tmpfs T rw
A bind mount of the directory that's being removed is moved beneath an
existing mount while the rmdir() is located between the two calls. The
mount ends up on the removed directory and nothing will ever detach it.
Check cant_mount() in d_set_mounted() as well. dont_mount() raises the
flag under d_lock before detach_mounts() runs so either the mountpoint
is set first and detach_mounts() finds the mount or the flag is seen and
the mount is refused.
In the Linux kernel, the following vulnerability has been resolved:
unshare: don't drop active namespace references that were never taken
Active references on the namespaces of an nsproxy are taken when the
nsproxy is installed into a task in switch_task_namespaces() and
copy_namespaces() and dropped again by put_nsproxy() through
deactivate_nsproxy().
But ksys_unshare() calls put_nsproxy() on an nsproxy that was never
installed when set_cred_ucounts() fails. The new namespaces of that
nsproxy go from zero to minus one and the namespaces shared with the
caller lose a reference that belongs to the nsproxy the caller keeps
using:
WARNING: kernel/nscommon.c:171 at __ns_ref_active_put+0x1cd/0x230
nsproxy_ns_active_put
deactivate_nsproxy
ksys_unshare
Afterwards the namespaces the caller lives in aren't listed by listns()
anymore. set_cred_ucounts() only fails when alloc_ucounts() can't
allocate and it only allocates when the real uid of the caller differs
from its effective uid.
Commit cefd55bd2159 ("nsproxy: fix free_nsproxy() and simplify
create_new_namespaces()") separated the two cases on purpose.
nsproxy_free() frees an nsproxy that was prepared but never installed
and that's what a failed setns() uses in put_nsset(). Export it and use
it for a failed unshare() as well.
In the Linux kernel, the following vulnerability has been resolved:
namespace: don't let a pseudo dentry become the root of a mount
d_alloc_pseudo() hands out dentries for pipes, sockets and other files
that are never anyone's child or parent. They carry DCACHE_NORCU and
dentry_free() frees them right away because no lockless path walk can
ever reach them. That holds as long as such a dentry isn't the root of
a mount. But bind mounting /proc/self/fd/<fd> of such a file does
exactly that.
Most callers of alloc_file_pseudo() put their files on kernel-internal
mounts and may_copy_tree() refuses those. bpf_token_create() doesn't. It
places the token file on the bpffs mount the caller handed it and that
mount is in the caller's mount namespace so the clone goes through:
mount --bind /proc/self/fd/<token> <file>
open_tree(tokfd, "", AT_EMPTY_PATH | OPEN_TREE_CLONE) + move_mount()
__follow_mount_rcu() then loads ->mnt_root of that mount, reads d_seq
and d_flags of the dentry and only then checks mount_lock. The dentry
can be gone by then. The final mntput() of an unmounted parent unhooks
a child that stayed attached to it under mount_lock alone. When the
child's holder does the final put right after that the root is dput()
and freed immediately while a walker that found the mount hashed still
looks at it.
Refuse to clone a mount with a DCACHE_NORCU dentry as its root. Nothing
sensible can be done with a bind mount of a bpf token anyway.
In the Linux kernel, the following vulnerability has been resolved:
namespace: unhash a dentry before detaching the mounts on it
unlink(), rmdir() and rename() remove the entry from the filesystem,
call detach_mounts() on the entry with the inode locked and then call
d_delete() once the inode is unlocked.
The thing is that between detach_mounts() and d_delete() the dentry is
still hashed and positive. But after detach_mounts() nothing covers the
dentry anymore. A lookup that finds this dentry in the dcache can
uncover what the mounts hid.
That's a problem when unlinking files or directories that are
mountpoints in other mount namespaces. Everybody who had the underlying
entry covered can race the detach_mount() call until d_delete() has run.
The race window isn't all that small. It encompasses namespace_unlock()
with a full synchronize_rcu_expedited() grace period and the
inode_unlock() of the entry.
In my experiments three walkers that kept trying read an overmounted
file in 11938 of the 16140 unlinks that removed its mountpoint from a
bind mount of the filesystem within a minute.
Fun fact, d_invalidate() has the same ordering problem but gets it
right. It unhashes the dentry first and detaches the mounts afterwards.
Let's do the same in __detach_mounts():
- A lookup that hasn't found the dentry yet misses it in the dcache and
waits for the directory lock that the caller holds until the name is
gone for good.
- A lockless lookup that found it already fails the mount_lock check
while the dentry still counts as a mountpoint and the d_seq check once
it doesn't, and retries.
This fixes the lockless path. We still need to fix the reference count
lookup in a follow-up patch.
So d_drop() the dentry. The dentry stays positive and held but can't be
found anymore. Then proceed with the detach and unlink.
Reproducer:
The reads were counted with the program below, three walkers for 60 s
in a VM with 4 CPUs. The race window is stretched with the debug patch
pasted here. fs.detach_race_walk_us sleeps in lookup_fast() once it
found a mountpoint dentry and before its mounts are crossed.
fs.detach_race_unlink_us sleeps in vfs_unlink() between detach_mounts()
and d_delete().
// SPDX-License-Identifier: GPL-2.0
/*
* unlink_covered: unlink a file that is a mountpoint in a detached copy of
* its mount, against lookups that open it through that copy.
*
* A is a tmpfs with the file f1 (content SECRET_F) and the plain file
* MARK_A. A' is an open_tree(OPEN_TREE_CLONE) copy of A with MARK_A bound
* on A'/f1, held through an O_PATH fd on its root once the tree fd is
* closed. Walkers open f1 through that fd while the driver unlinks A/f1,
* where nothing is mounted on it. A walker may read MARK_A or get ENOENT.
* A read of SECRET_F is a hit: the name was found after its mount was
* gone. The lockless walks use openat2(RESOLVE_CACHED).
*
* usage: unlink_covered [-t seconds] [-w walkers]
*/
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#include <errno.h>
#include <fcntl.h>
#include <pthread.h>
#include <stdatomic.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <sys/mount.h>
#include <sys/stat.h>
#include <sys/syscall.h>
#include <linux/openat2.h>
#ifndef __NR_open_tree
#define __NR_open_tree 428
#endif
#ifndef __NR_move_mount
#define __NR_move_mount 429
#endif
#ifndef __NR_openat2
#define __NR_openat2 437
#endif
#ifndef OPEN_TREE_CLONE
#define OPEN_TREE_CLONE 1
#endif
#ifndef OPEN_TREE_CLOEXEC
#define OPEN_TREE_CLOEXEC O_CLOEXEC
#endif
#ifndef MOVE_MOUNT_F_EMPTY_PATH
#define MOVE_MOUNT_F_EMPTY_PATH 0x00000004
#endif
#define WORK "/tmp/uc"
#define ADIR WORK "/A"
static int duration = 60, nwalkers = 3;
static atomic_int stop, writer_waiting;
static pthread_rwlock_t cur_lock = PTHREAD_RWLOCK_INITIALIZER;
static int cur_fd = -1; /* the root of A', -1 wh
---truncated---
In the Linux kernel, the following vulnerability has been resolved:
namei: don't reveal overmounted entries in refwalk
In rcuwalk the dentry is validated before it is accepted. step_into()
step_into() rechecks d_seq and __follow_mount_rcu() rechecks mount_lock.
An entry that gets unlinked in between causes the lookup to retry and
miss.
A refwalk doesn't do this. lookup_fast() takes a reference on the hashed
dentry and simply accepts it. So an unlink that happens after the
reference was taken isn't seen by refwalk. That's fine for a simple
file. We just happened to open it before it was unlinked, no problem.
For a mountpoint and specifically a locked mountpoint it very much
isn't. The unlink detaches all mounts and then removes the name. Any
refwalk that hasn't traversed the mounts yet simply reveals the
underlying entry. It's a very narrow window but it can be hit:
reads of a covered file in 60 s, 3 walkers, ~15000 unlinks
no widening 27
that step + 200 us 3741
See the appended patch for a more reliable reproducer.
So check the entry after step_into(). unlink(), rmdir() and rename()
mark the dentry with dont_mount() before they detach the mounts and
remove the dentry.
So a dentry that is marked with DCACHE_CANT_MOUNT and is unhashed by the
time its mounts were looked at is a name that was unlinked under the
refwalk. DCACHE_CANT_MOUNT is read after the mounts were looked up and
it is set before they are detached by detach_mounts(). A refwalk that
missed the mounts will see DCACHE_CANT_MOUNT.
Plain d_unlinked() is fine. A rename takes the dentry off its hash chain
but ___d_drop() leaves d_hash.pprev set. So only __d_drop() and a rename
over the dentry unhash it. The only move that flips IS_ROOT splices in a
disconnected alias. That doesn't have DCACHE_CANT_MOUNT set.
Add the to step_into_slowpath(). ".." and LOOKUP_DOWN may legitimately
land on an unhashed directory. A refwalk that crossed onto a mount has
path.mnt different from nd->path.mnt. A dentry that a filesystem dropped
on its own (d_invalidate(), d_drop()) doesn't have the flag set and is
treated as before.
// SPDX-License-Identifier: GPL-2.0
/*
* unlink_covered: unlink a file that is a mountpoint in a detached copy of
* its mount, against walkers that open it through that copy.
*
* A is a tmpfs with the file f1 (content SECRET_F) and the plain file
* MARK_A. A' is an open_tree(OPEN_TREE_CLONE) copy of A with MARK_A bound
* on A'/f1, held through an O_PATH fd on its root once the tree fd is
* closed. Walkers open f1 through that fd while the driver unlinks A/f1,
* where nothing is mounted on it. A walker may read MARK_A or get ENOENT.
* A read of SECRET_F is a hit: the name was found after its mount was
* gone. The lockless walks use openat2(RESOLVE_CACHED).
*
* usage: unlink_covered [-t seconds] [-w walkers]
*/
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#include <errno.h>
#include <fcntl.h>
#include <pthread.h>
#include <stdatomic.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <sys/mount.h>
#include <sys/stat.h>
#include <sys/syscall.h>
#include <linux/openat2.h>
#ifndef __NR_open_tree
#define __NR_open_tree 428
#endif
#ifndef __NR_move_mount
#define __NR_move_mount 429
#endif
#ifndef __NR_openat2
#define __NR_openat2 437
#endif
#ifndef OPEN_TREE_CLONE
#define OPEN_TREE_CLONE 1
#endif
#ifndef OPEN_TREE_CLOEXEC
#define OPEN_TREE_CLOEXEC O_CLOEXEC
#endif
#ifndef MOVE_MOUNT_F_EMPTY_PATH
#define MOVE_MOUNT_F_EMPTY_PATH 0x00000004
#endif
#define WORK "/tmp/uc"
#define ADIR WORK "/A"
static int duration = 60, nwalkers = 3;
static atomic_int stop, writer_waiting;
static pthread_rwlock_t cur_lock = PTHREAD_RWLOCK_INITIALIZER;
static int cur_fd = -1; /* the root of A', -1 while there is none */
static atomic_long n_unlink, n_walk, n_mark, n_enoent, n_othe
---truncated---
In the Linux kernel, the following vulnerability has been resolved:
namespace: refuse an automount below a mount that is in no namespace
It's possible to add autmounts even when the parent mount isn't in the
mount namespace of the caller. The only requirement we have is that the
parent mount namespace must not be NULL, i.e., unmounted.
Problem is that clone_private_mount() has MNT_NS_INTERNAL which makes
that trivially true. So that passes the test and attach_recursive_mnt()
accepts that as a mount point and funny enough, count_mounts()
dereferences MNT_NS_INTERNAL. The problem is it is an error pointer...
So we can reach this in userspace via fanotify. A filesystem mark on the
lower filesystem of an overlay reports paths on the layer clone and
reading the event hands out a descriptor on it. For example with debugfs
as the lower layer it goes kaboom:
openat(evfd, "tracing", O_DIRECTORY)
Oops: general protection fault
KASAN: null-ptr-deref in range [0x1d0-0x1d7]
RIP: 0010:count_mounts+0x35/0x200
attach_recursive_mnt
finish_automount
And since that sleeping beauty happens under namespace_sem held for
writing every mount operation on the system blocks from then on.
Congrats.
Use is_mounted() instead which rejects unmounted and internal mounts
alike. The open fails with EINVAL just as it did before
clone_private_mount() used MNT_NS_INTERNAL.
In the Linux kernel, the following vulnerability has been resolved:
namespace: never expire a locked mount
Locked mounts are special. They protect the underlying files and
directories from being revealed. do_umount() refuses to unmount locked
mounts but shrink_submounts() doesn't.
A shrinkable mount can become locked once the owner of a user namespace
puts it beneath a locked mount with MOVE_MOUNT_BENEATH. The locked
property now moves to the mount at the bottom making it possible to
unmount the top mount.
So umount() of an unlocked ancestor now expires the bottom mount first
and the covered directory is revealed:
move_mount(c -> x/hidden, MOVE_MOUNT_BENEATH) = 0
the cover after the lock moved: umount2(x/hidden) = 0
the holder of the lock: umount2(x/hidden) = EINVAL
the root of the copy, busy: umount2(x) = EBUSY
reads x/hidden/secret: "covered-by-root"
So leave a locked mount alone as it dies together with its parent. A
plain umount() of an unlocked mount with a locked one below it is EBUSY
from now on. It's the same for any other locked child. A lazy umount
still takes the whole tree.
mark_mounts_for_expiry() never sees a locked mount. lock_mnt_tree()
leaves a mount on an expiry list alone and nothing else puts a locked
one on such a list. Add an assert for this.
In the Linux kernel, the following vulnerability has been resolved:
namespace: handle mount locking for automounts correctly
If mounts are propagated across user namespaces, attach_recursive_mnt()
locks every copy of the source mount to protect overmounts from
vanishing and revealing the underlying files or directories.
The user namespace is taken from the caller's mount namespaces since
this is where the mounts end up. Except, that's not always true.
Automounts may legitimately get popped in by tasks located in a
different mount and user namespace during path lookup.
Then check doesn't make sense at that point. The copy in the namespace
of the parent - which may be the host's - is now locked and the host
cannot change the flags of its own mounts anymore.
Here's the reproducer:
- host has debugfs mounted nosuid,nodev,noexec and shared
- tracefs automount below it is not yet active
- hand a child process a directory descriptor on that mount
- child process enters auser and mount namespace
- child process' namespace now holds a locked copy of the host's debugfs
mount that receives propagation from it
- child process stats "tracing/." through the descriptor
- lookup runs on the host's mount so the automount lands below the host's mount
- propagation puts a copy below the child's copy
- both try to clear the flags on the mount they got, with a bind remount:
host, on its own automount: MS_REMOUNT|MS_BIND = EPERM
child, on the copy in its namespace: MS_REMOUNT|MS_BIND = 0
So the lock landed on the host's mount instead of the child's copy.
Congrats. So we need to compare with the owner of the namespace the
mount actually gets mounted on. For all regular cases that is the
caller's mount namespace and so nothing changes.
Detached trees in anonymous mount namespaces by be handed over via
SCM_RIGHTS or inherited in other ways on purpose so the attaching task's
mount namespace is authoritative, not the creator of the detached tree.
In the Linux kernel, the following vulnerability has been resolved:
namespace: keep the lock on a mount that a propagated copy is moved beneath
attach_recursive_mnt() transfers MNT_LOCKED from the top mount to the
mount that is moved beneath it with MOVE_MOUNT_BENEATH. This allows the
owner of a user namespace to replace its locked /proc or its root. The
mount beneath takes on the job of covering the underlying mountpoint
allowing the top mount to be unmounted.
Consider two mount namespaces:
(H) The host H has a shared mount P with a secret in P/d
(Z) Z is a user namespace made from M. Its copy of P receives
propagation from the host's P and its copy of M's cover on P/d is
locked Z's owner must not get to see P/d.
Now (H) mounts X on P/d. This propagates into (Z). The copy of X lands
beneath (Z)'s locked cover. The locked property is transfered from (Z)'s
cover to the copy of X propagated beneath it. The cover is now unlocked.
Now (H) unmounts X again. The copy of X in (Z) gets unmounted and the
covering mount is left unlocked on top of P/d. (Z) can now unmount it:
Z: umount2(P/d) = EINVAL /* the cover is locked */
H: mount X on P/d, umount X /* both propagate into Z /*
Z: umount2(P/d) = 0 /* the cover is now unlocked */
Z: read P/d/secret = "covered-by-root" /* secret revealed */
So only transfer the locked property to the mount beneath for mounts the
caller has placed. A propagated copy that lands beneath a locked mount
is locked as well so that the mount at the bottom of the stack carries a
lock the way every check expects. The mount on top of it remains locked
to ensure that it keeps covering even if the propagated mount is
unmounted again.
In the Linux kernel, the following vulnerability has been resolved:
xfs: stop exchanging reflink flags when exchanging mappings
When an exchange covers both files from offset 0 to i_disk_size,
xmi_can_exchange_reflink_flags() assumes every mapping moved and hands
the reflink flag over to the other file.
That assumption is wrong. A file can still own shared blocks past
i_disk_size, e.g. after EXCHANGE_RANGE_TO_EOF sets a smaller size
without unmapping anything. The flag is cleared anyway, and the next
write to such a block happens in place instead of CoW, corrupting the
other file that shares it.
Commit a23eca88448e ("xfs: fix exchange-range reflink flag clearing
issue with INO1_WRITTEN") fixed one instance of this. Rather than add
another special case, stop moving the flags.
xfs_exchmaps_ensure_reflink() already sets the flag on both inodes
before the exchange, so both just keep it.
Nothing sets CLEAR_INO{1,2}_REFLINK anymore. Intents logged by older
kernels may still carry them; recovery already masks them off, so they
stay in the log format mask and are ignored.
XFS_IOC_SWAPEXT has the same problem: it moves the reflink flag from
one inode to the other along with the extents. Just removing that is
not enough here: the exchange path flags both inodes up front in
xfs_exchmaps_ensure_reflink(), but swapext has no such step, so the
inode receiving the shared extents would be left without the flag.
Set it on both inodes instead, and make sure both have a CoW fork,
since an inode with the reflink flag is expected to have one.
The flag now stays until FALLOC_FL_UNSHARE_RANGE, xfs_scrub, or a
truncate to zero removes it.
In the Linux kernel, the following vulnerability has been resolved:
xfs: fix exchange-range-to-eof file size exchange
Norbert Szetei posted a patch containing an incomplete description of a
bug in XFS_IOC_EXCHANGE_RANGE's TO_EOF flag. The bug tricks
exchange-range into modifying a file so that it has shared blocks
starting beyond EOF, which is never allowed because files never have
written data beyond EOF, and you can only share written data blocks.
This is key to all the incorrect behavior that follows.
Eventually I got from Norbert a description of what's going wrong:
> Fair. Here it is with physical block numbers, 4k blocks. peer is the file
> that gets rewritten and file1 is the one that ends up with the flag
> cleared.
>
> Start, nothing shared:
>
> peer size 4096 -> [100]
> file1 size 12288 -> [200] [201] [202]
> donor1 size 4096 -> [300]
> donor2 size 4096 -> [400]
>
> 1. FICLONERANGE peer's block into file1's middle slot:
>
> peer size 4096 -> [100]
> file1 size 12288 -> [200] [100] [202] reflink flag set
> ^^^^^ both files own block 100
>
> 2. XFS_IOC_EXCHANGE_RANGE file1 against donor1, file1_offset 0,
> file2_offset 8192, length 0, TO_EOF. One block is exchanged, and the sizes
> are exchanged with it:
>
> file1 size 4096 -> [200] [100] [300] reflink flag set
> ^^^^^^^^^^^ still owned, now past the size
> donor1 size 12288 -> [202]
>
> file1 says it is one block long while it still owns three. Nothing was
> unmapped. file2_offset is block 2, so this exchange cannot clear a flag.
>
> 3. XFS_IOC_EXCHANGE_RANGE file1 against donor2, both offsets 0, length 4096.
> By i_disk_size this is two one-block files exchanging everything, so
> xmi_can_exchange_reflink_flags() agrees and the flag is cleared:
>
> file1 size 4096 -> [400] [100] [300] reflink flag CLEARED
> donor2 size 4096 -> [200]
>
> Block 100 is still shared with peer.
>
> 4. pwrite(file1, 4096, 4096), which is file1's second slot, block 100.
> xfs_is_cow_inode(file1) is false, so no CoW:
>
> peer size 4096 -> [100] same block, new contents
>
> Block 100 is the only one that matters here. PoC below, it reads a file
> called peer in the current directory. The filesystem needs reflink and
> exchange-range, which mkfs.xfs 7.x turns on by default.
So yes, TO_EOF is screwing up the file size even if it's actually
exchanging the data block correctly. The following is the intended
behavior of XFS_EXCHANGE_RANGE_TO_EOF, as described by its manpage:
"Ignore the length parameter. All bytes in file1_fd from file1_offset
to EOF are moved to file2_fd, and file2's size is set to
(file2_offset+(file1_length-file1_offset)). Meanwhile, all bytes in
file2 from file2_offset to EOF are moved to file1 and file1's size is
set to (file1_offset+(file2_length-file2_offset))."
In other words, if we only exchange a portion of the file data, then we
should only exchange the portion of the file size that relates to the
exchanged data. Unfortunately, the code swaps the ondisk file size and
the incore file sizes, which is incorrect if the caller passes in a
nonzero offset.
Having confused the file sizes, the PoC uses a second exchange-range
call on what looks like a non-reflinked single-block file. Because the
file size is set incorrectly, exchange-range thinks it's exchanging the
full contents of two files and clears the reflink flag on the broken
file. That enables an extending write of the broken file to rewrite the
shared block that's just past EOF. This has become known colloquially
as refluxfs.
Once the file sizes are set correctly, the second exchange-range no
longer thinks that it's doing a full-contents swap, so it won't clear
the reflink flag on either of its file arguments.
Norbert followed up to an earlier version of this patch with a second
complaint about us not handling the i_disk_size state correctly:
> The offsets are bounded against the incore i_size but subtracted from
>
---truncated---
In the Linux kernel, the following vulnerability has been resolved:
drm/bridge: Fix unlocked list_del in drm_bridge_add()
When re-adding a bridge that was previously removed, drm_bridge_add()
drops it from bridge_lingering_list without holding bridge_lock.
Both bridge_list and bridge_lingering_list are protected by bridge_lock,
as they are concurrently modified by drm_bridge_remove() and
__drm_bridge_free(), and walked by the debugfs 'bridges' file. Running
the list_empty() test and the list_del_init() outside of the lock may
therefore corrupt either list.
Perform both under bridge_lock.
In the Linux kernel, the following vulnerability has been resolved:
cgroup/cpuset: Don't access cpuset_cgrp_subsys.root in is_in_v2_mode()
After seeing the patch [1] to guard is_in_v2_mode() with RCU, it makes
me realize that is_in_v2_mode() may be called in a context where a new
cgroup filesystem is being rebound with stale cpuset_cgrp_subsys.root
pointer. Avoid this potential UaF situation by adding a new cpuset_v2_mode
flag which is set when the cpuset_v2_mode mount option is used. This
flag is written into only when cpuset_bind() is being called with a
stable cpuset_cgrp_subsys.root value. The is_in_v2_mode() helper is
modified to read the new cpuset_v2_mode flag instead of accessing
cpuset_cgrp_subsys.root directly.
[1] https://lore.kernel.org/lkml/20260929084124.626693-2-arighi@nvidia.com
In the Linux kernel, the following vulnerability has been resolved:
gpio: exar: initialize the ID before registering its cleanup
devm_add_action_or_reset() invokes the cleanup action immediately when
registering the action fails. gpio_exar_probe() currently stores the
allocated ID in exar_gpio only after registering the action, so that
failure makes exar_devm_ida_free() release ID 0. If the allocated ID is
nonzero, this leaks that ID and can free ID 0 belonging to another device.
Store the ID before registering the cleanup action so both the immediate
failure path and normal device teardown release the correct allocation.
In the Linux kernel, the following vulnerability has been resolved:
drm/vc4: Free the BO cache size list on teardown
vc4_bo_cache_destroy() purges the cached BOs but never frees the size
list, leaking it on every unbind.
In the Linux kernel, the following vulnerability has been resolved:
net/sched: em_text: reject unterminated algo before autoload
This is a follow-up to commit 9c572a83037a ("net/sched: fix potential
stack infoleak in em_text_dump()"), which zero-initialised the dump-side
struct. Sashiko pointed at a pre-existing bug that exposed the change-side
input-validation gap open.
em_text_change() casts the raw netlink attribute payload to struct
tcf_em_text and passes conf->algo, a 16-byte array with no enforced NUL
terminator, to textsearch_prepare(). On the TS_AUTOLOAD retry
textsearch_prepare() calls request_module("ts_%s", algo); vsnprintf()
walks %s until it finds a NUL, so an algo[] with no NUL reads past the
array into the adjacent struct fields and payload and feeds those bytes
to the usermode helper command line. This is an out-of-bounds read and
a kernel memory disclosure.
Reject the attribute when no NUL exists within TC_EM_TEXT_ALGOSIZ bytes,
matching the check xt_string has had since 3ab720881b6e.
Conditions to recreate the bug:
- CAP_NET_ADMIN on the target netns (namespace-local via unshare -Urn
is enough)
- install clsact on an interface, then add a basic filter with one text
ematch whose algo[] is 16 bytes with no NUL (e.g. "A"*16) and
pattern_len at least 1
- the add fails with -ENOENT and the kernel invokes modprobe with a
module name made of "ts_" plus the 16 bytes and the adjacent payload
In the Linux kernel, the following vulnerability has been resolved:
amt: send the relay's General Query directly from the receive path
An skb queued in a qdisc can outlive the tunnel it references
through a raw pointer in skb->cb. For example, with igmp_qrv set
to 1 on the relay a tunnel lives for 135s, so a netem delay of
180s on the amt device outlives it; when the tunnel expires and is
freed, the subsequent dequeue triggers a use-after-free in
amt_dev_xmit().
BUG: KASAN: slab-use-after-free in amt_dev_xmit+0x2763/0x2e20
Call Trace:
amt_dev_xmit+0x2763/0x2e20 [drivers/net/amt.c:1262]
dev_hard_start_xmit+0x22f/0x620
sch_direct_xmit+0x12e/0xac0
netem_dequeue+0x333/0xc50
net_tx_action+0x35c/0xa60
amt_send_igmp_gq() and amt_send_mld_gq() are only called from
amt_request_handler(), inside the rcu_read_lock_bh() section of
amt_rcv(). amt_request_handler() already has the tunnel the query is
for: it found or created it inside that section. Queuing the query with
dev_queue_xmit() only leads back into amt_dev_xmit(), which strips the
Ethernet header and calls amt_send_membership_query() for that tunnel.
Make that call directly from the two senders instead, the same way
amt_send_advertisement() transmits from the receive path. The query
never waits in a qdisc, the tunnel is only dereferenced inside the
RCU section that found or created it, and nothing is stored in
skb->cb, so no lookup or refcount is needed. Remove the query branch
of amt_dev_xmit(), amt_skb_cb() and struct amt_skb_cb, which have no
users left.
Behaviour changes:
- The relay's own General Queries no longer pass through the amt
device's egress path: its qdisc, tc egress (clsact/tcx), the
netfilter egress hook and packet taps. They are still visible as
UDP on the underlay.
- A query that is sent successfully is no longer counted as
tx_dropped. The old query branch left through the unlock label,
which counted every sent query as dropped.
- A query that reaches amt_dev_xmit() on a relay from elsewhere,
such as a userspace querier, is now dropped at the IGMP/MLD type
switch. Before, it trusted whatever skb->cb held, and a NULL
tunnel hit the WARN_ON(1).
In the Linux kernel, the following vulnerability has been resolved:
cxgb4/ch_ktls: disable softirqs around tid_list erases
chcr_ktls_cpl_act_open_rpl() inserts into tid_list from the receive path
using xa_insert_bh(). The array is shared by the adapter's connections
and initialized with XA_FLAGS_LOCK_BH. However, chcr_ktls_dev_del() and
the chcr_ktls_dev_add() error path use xa_erase(), which takes the XArray
lock without disabling softirqs.
When cleanup runs with softirqs enabled, a receive softirq for another
connection on the same CPU can interrupt the erase and spin on the lock
held by the interrupted task. Use xa_erase_bh() at both sites to match
the insertion path.
In the Linux kernel, the following vulnerability has been resolved:
bonding: fix the program leak when XDP is replaced
We need to release the reference on the old BPF prog, whether
we are going to no-prog, or replacing the prog with another.
Spotted by AI while running tests for the propagation series
(leaked programs were accumulating). Selftest will come in
the net-next series to avoid conflicts.
In the Linux kernel, the following vulnerability has been resolved:
ice: Restore Ordered MMIO Writes for Tx Doorbells
The DQL accounting state which tracks the number of bytes queued for the
NIC to transmit, namely dql->num_queued, is updated by the ICE driver
when it invokes __netdev_tx_sent_queue(). Subsequently the driver
updates the hardware queue tail allowing the NIC to begin processing the
work queue. The queue accounting update should be ordered before the NIC
begins transmitting descriptors.
The transmit path currently updates the hardware doorbell using
writel_relaxed(), which (on arm64) does not provide the same ordering
guarantees between writes to normal memory and writes to MMIO registers
that writel() does. This introduces a potential race where the NIC
begins transmitting descriptors before the dql->num_queued update is
globally visible. If the NIC finishes before the update is observed by
dql_completed(), it detects an invalid state where more bytes have been
completed than have been queued. When this happens the following BUG_ON
is triggered.
BUG_ON(count > num_queued - dql->num_completed);
This has been observed and manifests as the following crash (note that
the trace has been trimmed) in an environment with sustained network
load that uses an Intel Corporation Ethernet Controller E810-XXV for SFP
(rev 02) on an arm64 machine. Replacing writel_relaxed() with writel()
in a test kernel eliminated the crash in the user's workload, which
previously reproduced the issue reliably.
kernel BUG at lib/dynamic_queue_limits.c:99
Internal error: Oops - BUG: 00000000f2000800 [#1] SMP
pc : dql_completed+0x268/0x2a0
lr : ice_clean_tx_irq+0x1d4/0x620 [ice]
Call trace:
dql_completed+0x268/0x2a0 (P)
ice_napi_poll+0x94/0x520 [ice]
__napi_poll+0x48/0x3f0
net_rx_action+0x194/0x420
This restores the behaviour prior to commit ccde82e90946 ("ice: add E830
Earliest TxTime First Offload support"), which changed the notification
mechanism from writel() to writel_relaxed().
In the Linux kernel, the following vulnerability has been resolved:
nvmet: use the local P2P device on metadata allocation failure
nvmet_req_alloc_p2pmem_sgls() assigns req->p2p_dev only after the data
and metadata SGL allocations succeed. If the data allocation succeeds
but the metadata allocation fails, out_free_sg passes req->p2p_dev to
pci_p2pmem_free_sgl() before that assignment has taken place.
Pass the local p2p_dev argument instead, so cleanup uses the same PCI
device as the successful data allocation.
In the Linux kernel, the following vulnerability has been resolved:
btrfs: always return -EIOCBQUEUED after btrfs_uring_read_extent_endio
If all bios finish before btrfs_encoded_read_regular_fill_pages()
returns, it calls btrfs_uring_read_extent_endio() and previously
returned the I/O status. A negative errno then made
btrfs_uring_read_extent() unlock and free while
btrfs_uring_read_finished() did the same again.
Return -EIOCBQUEUED so only the deferred path cleans up.
In the Linux kernel, the following vulnerability has been resolved:
btrfs: free iov when btrfs_uring_read_extent() fails
After btrfs_uring_read_extent(), the caller always jumped to out_acct.
That skips kfree(data->iov), which is only correct for -EIOCBQUEUED
where the deferred path owns the iov. On failure, fall through to
out_free instead.
In the Linux kernel, the following vulnerability has been resolved:
btrfs: unlock inode and extent in caller when io_uring read extent fails
btrfs_uring_read_extent() runs only after btrfs_encoded_read() has taken
the inode shared lock and the extent lock. On failure it used to unlock
in out_fail, and a pages-array allocation failure returned -ENOMEM
without unlocking at all.
Unlock in the caller instead on all failure returns, matching the
copy_to_user() error path. The deferred -EIOCBQUEUED path still unlocks
in btrfs_uring_read_finished().
In the Linux kernel, the following vulnerability has been resolved:
btrfs: don't stash io_uring encoded data across -EAGAIN
Returning -EAGAIN while leaving btrfs_uring_encoded_data in the cmd PDU
leaks if the request is cancelled or the ring exits before reissue.
io_uring does not free driver PDU allocations on cleanup.
Write: io_queue_sqe() always issues with IO_URING_F_NONBLOCK first, so
return -EAGAIN before allocating and free data on every exit.
Read: free on nowait -EAGAIN too; only -EIOCBQUEUED keeps the allocation
for btrfs_uring_read_finished().