[PATCH] iommufd: reject mmap() of not-yet-finalized objects with iommufd_try_inc_users()

From: YanXin Li

Date: Sat Nov 21 2026 - 22:24:47 EST


iommufd_fops_mmap() takes a users reference on the object that owns the
mmapped region with refcount_inc_not_zero(&immap->owner->users). The
reference is meant to pin the owner across the lifetime of the VMA, but
refcount_inc_not_zero() cannot tell an already-created object from an
object that is still being constructed.

Objects with an MMIO region are published in ictx->mt_mmap by their
creation ioctl before they are finalized: IOMMU_VIOMMU_ALLOC inserts
the immap in mock_viommu_init() (or in a driver's viommu_init())
through iommufd_viommu_alloc_mmap(), and at that point the viommu is
reserved as XA_ZERO_ENTRY in ictx->objects and obj->users is still the
initial 1. A concurrent mmap() on the same /dev/iommu fd therefore
succeeds in taking a users reference on the half-constructed object,
because refcount_inc_not_zero() only fails once the refcount has
dropped to zero. When the ioctl later fails - for IOMMU_VIOMMU_ALLOC
this is reachable by simply placing the ioctl
command buffer on a
read-only page so that iommufd_ucmd_respond() faults - the error path
runs iommufd_object_abort_and_destroy(), which calls the driver destroy
op that erases the immap from mt_mmap and kfree()s it, and then
iommufd_object_abort() WARNs because obj->users never returns to zero.
The next munmap() (or fork/exit) reaches iommufd_fops_vma_close(), which
dereferences the freed immap. On v7.4-rc3 with CONFIG_IOMMUFD_TEST,
racing a loop of mmap()/munmap() against aborted IOMMU_VIOMMU_ALLOC
yields:

WARNING: drivers/iommu/iommufd/main.c:126 at iommufd_object_abort+0x11d/0x1a0
BUG: KASAN: slab-use-after-free in iommufd_fops_vma_close+0x87/0x90
Read of size 8 at addr ffff888010f4a5c0 by task test_iommufd_mm
...
Oops: 0002 [#1] SMP KASAN

The reproducer uses CONFIG_IOMMUFD_TEST and needs no hardware IOMMU.
On production kernels, the bug requires a driver that publishes an
mmap region before its creation ioctl can still fail, such as
tegra241-cmdqv o
n real hardware.

The kernel already has the correct helper for this invariant: main.c
documents that "the users refcount cannot be increased on an already
created object unless the object is installed in the xarray, otherwise
things are racing with a parallel destruction", and iommufd_try_inc_users()
implements exactly that check under the xarray lock. Use it in
iommufd_fops_mmap() instead of refcount_inc_not_zero(). Before
finalize, xa_load() returns XA_ZERO_ENTRY instead of the object, the
helper returns -EBUSY, and the mmap() is rejected with -ENXIO while the
ioctl is still constructing the object; after finalize it behaves like
the open-coded version. The mmap path now nests the objects xarray
lock inside the mt_mmap lock. Object teardown releases the xarray
lock before the driver destroy op erases the mmap entry, so it does
not acquire these locks in reverse order.

With this change a successful (finalized) viommu can still be mmap()ed
and unmapped normally,
while 3000 aborted allocations racing two
mmap()/munmap() loops produced zero kernel splats, versus the use-
after-free above on the unmodified kernel. A test case and full
validation logs are available on request.

The vulnerable code is present in mainline commit c44104f3642d
(post-v7.4-rc3). Runtime verification and fix validation were
performed in an x86-64 QEMU/KVM guest built from that baseline with
Generic KASAN, UBSAN, lockdep, and DEBUG_OBJECTS.

Fixes: 56e9a0d8e53f ("iommufd: Add mmap interface")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: YanXin Li <fadouse@xxxxx>
---
drivers/iommu/iommufd/main.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/iommu/iommufd/main.c b/drivers/iommu/iommufd/main.c
index 9a921b15316..f0b4308fe34 100644
--- a/drivers/iommu/iommufd/main.c
+++ b/drivers/iommu/iommufd/main.c
@@ -601,7 +601,7 @@ static int iommufd_fops_mmap(struct file *filp, struct vm_area_struct *vma)
mtree_lock(&ictx->mt_mmap);

/* vma->vm_pgoff carries a page-shifted start position to an immap */
immap = mtree_load(&ictx->mt_mmap, vma->vm_pgoff << PAGE_SHIFT);
- if (!immap || !refcount_inc_not_zero(&immap->owner->users)) {
+ if (!immap || iommufd_try_inc_users(ictx, immap->owner)) {
mtree_unlock(&ictx->mt_mmap);
return -ENXIO;
}
--
2.43.0

Attachment: publickey - fadouse@pm.me - 0x99485AA3.asc
Description: application/pgp-keys

Attachment: signature.asc
Description: OpenPGP digital signature