Re: [PATCH 0/2] iommu/io-pgfault: reject partial faults after device removal

From: Baolu Lu

Date: Fri Nov 20 2026 - 20:25:40 EST


On 11/20/2026 5:26 PM, Qihang wrote:
Hi Baolu,

Thanks. I agree that the driver must drain pending faults before
calling iopf_queue_remove_device(). My selftest removed the mock
driver's serialization, so it doesn't establish that fault reporting
can overlap removal in an unmodified driver. I'm withdrawing this
series and won't resend it.

I still have two questions about where the drivers establish that guarantee.

On VT-d, blocking_domain_attach_dev() calls
iopf_for_domain_remove(old, dev) before device_block_translation(dev).
When that drops the last IOPF reference, it calls
iopf_queue_remove_device(). On an SM-capable IOMMU, for a device that
isn't a real DMA subdevice, the only PRQ drain on this path is reached
through device_block_translation() → intel_pasid_tear_down_entry() →
intel_iommu_drain_pasid_prq(). By then iopf_refcount is zero, so the
drain returns early.

In the same file, identity_domain_attach_dev() calls
device_block_translation(dev) first and releases the old domain's IOPF
reference afterwards. The comment there explains that putting the
device in the blocking state leaves the IOPF reference count
unchanged. Should blocking_domain_attach_dev() use the same order? If
the current order is intentional, what ensures that pending PRQ
reports for this device have drained before the last reference is
dropped?

Good catch. This is a bug in the vt-d driver. I already have a patch
that changes the order in blocking_domain_attach_dev(). I will submit it
for discussion after the necessary assessment.

On AMD, detach_device() calls iopf_queue_flush_dev() immediately
before amd_iommu_iopf_remove_device(). The flush helper waits for
queued work and documents that the caller must have flushed the
low-level queue first. The PPR interrupt thread calls
iommu_report_device_fault() directly, so waiting for queued work alone
doesn't establish that the thread has finished reporting faults for
this device. What in the detach path ensures that PPR reporting for
this device has quiesced before the device is removed from the IOPF
queue?

I may be missing a step in both cases. A pointer would be enough.

Thanks,
Qihang

On Fri, Nov 20, 2026 at 3:56 PM Baolu Lu <baolu.lu@xxxxxxxxxxxxxxx> wrote:

On 11/17/2026 9:54 AM, Qihang wrote:
iopf_queue_remove_device() kfree()s the entries on
fault_param->partial without unlinking them, leaving the list head
pointing into freed memory. Since a74c077b9021 the fault parameter is
refcounted and can outlive the drain, a report that took its reference
earlier can still list_add() through those pointers. Unlinking
instead turns that into a leak: the late node lands on a parameter
that is no longer on queue->devices, whose final put does not free
partial faults.

This needs a device removed from its iopf queue while a non-last
(partial) request is in flight, with the caller not blocking new reports
for the duration.

Patch 1 unlinks each partial before freeing it, matching
iopf_queue_discard_partial().

Patch 2 adds fault_param->removing, set and tested under
fault_param->lock: a late partial is either drained, or rejected and
freed by report_partial_fault() (-ENODEV). It covers partial-list
admission only; callers must still quiesce reports for the remaining
lifetimes.

Both patches are scoped to fault_param->partial, not the group path,
the attach handle, or paths that never call iopf_queue_remove_device().

Verified on v7.4-rc3 with KASAN + DEBUG_LIST and the iommufd mock
driver, its caller-side IOPF serialization deliberately removed and
test-only delays added; this shows the mechanism, not production
reachability. The runs cover queue-before-removal,
removal-before-insertion, concurrent late reporters and re-enable.
Each round queues one partial before the removal and then races the
late reporters against the drain, so a clean run ends with every node
accounted for: 1800 = 900 + 900 and 3600 = 900 + 2700 (allocated =
drained + rejected), counted at a final removal with no reporter in
flight. Nothing was left on fault_param->partial and KASAN stayed
quiet.

With patch 1 reverted, the report is:

BUG: KASAN: slab-use-after-free in __list_add_valid_or_report
Freed by task 1: kfree <- iopf_queue_remove_device
56 bytes inside a freed 64-byte region (struct iopf_fault, list.prev)

With patch 1 only, the same reproducer runs 1460 rounds without a KASAN
report and leaks one late node per round, which patch 2 addresses.

Reachability on an unmodified production kernel, and any security
impact, remain unconfirmed: the test shows memory corruption and a
KASAN panic on the modified mock driver, not a user-triggerable bug on
a stock kernel.

An AI coding assistant helped find this defect and build the
reproducer and instrumentation, which are available on request along
with the KASAN logs.

+Jason

The issues addressed by both patches are false positives. They rely on a
fault being reported for a device after iopf_queue_remove_device() has
been called for that device. The iommu driver must not allow that.
Pending i/o page faults must be drained before the device is removed
from its iommu's io page fault queue.

Thanks,
baolu

Thanks,
baolu