Re: [PATCH v2] iommufd: only hold references to files memfd_pin_folios() can pin
From: Samiullah Khawaja
Date: Mon Nov 16 2026 - 20:58:38 EST
On Tue, Nov 17, 2026 at 09:32:44AM +0800, Jiahao Xu wrote:
iopt_map_file_pages() stores a reference to the file passed to
IOMMU_IOAS_MAP_FILE in the iopt_pages of the new mapping:
iopt_map_file_pages()
iopt_alloc_file_pages()
pages->file = get_file(file)
iopt_release_pages() drops that reference only when the iopt_pages is
destroyed, which for a file mapping happens while the iommufd file is
being torn down. Nothing restricts which file can be passed, so
userspace can map the iommufd file itself, or a vfio device fd bound to
the same iommufd, both of which own an iommufd_ctx reference. The
reference graph then contains a cycle and the file, the iopt_pages, the
IOAS and the iommufd_ctx are all leaked; kmemleak reports it as
"memory leak in path_openat" with the holder taken through
iopt_map_file_pages().
Reverse the test instead of trying to enumerate the files that own an
iommufd_ctx: accept only the files that memfd_pin_folios() is able to
pin, which is exactly the set of files the mapping can be built from.
memfd_pin_folios() already enforces
if (!shmem_file(memfd) && !is_file_hugepages(memfd))
return -EINVAL;
and both predicates are inline helpers, so the same check can run here
before the reference is handed to iopt_pages. dma-bufs are unaffected:
they are taken by the dma_buf_get() branch above and never reach the
file branch, and iopt_pages holds the dma_buf there, not a struct file.
Fixes: f4986a72d6e4 ("iommufd: Add IOMMU_IOAS_MAP_FILE")
Reported-by: syzbot+7666ed2c3d42196bf7e3@xxxxxxxxxxxxxxxxxxxxxxxxx
Closes: https://syzkaller.appspot.com/bug?extid=7666ed2c3d42196bf7e3
Link: https://lore.kernel.org/all/20261112105945.285526-1-xujiahao@xxxxxxxxxxxxxxxxxxxx/
Signed-off-by: Jiahao Xu <xujiahao@xxxxxxxxxxxxxxxxxxxx>
---
v2:
- Replace the iommufd_fops blocklist with the allowlist of files that
memfd_pin_folios() can pin, as Jason asked for. This covers a vfio
device fd bound to the same iommufd, and any other fd that owns an
iommufd_ctx, instead of only the iommufd file itself.
- dma-buf mappings are untouched: they are taken by the dma_buf_get()
branch and iopt_pages holds the dma_buf there, not a struct file.
drivers/iommu/iommufd/io_pagetable.c | 14 ++++++++++++++
1 file changed, 14 insertions(+)
diff --git a/drivers/iommu/iommufd/io_pagetable.c b/drivers/iommu/iommufd/io_pagetable.c
index 5ac736cf5..4d10b2253 100644
--- a/drivers/iommu/iommufd/io_pagetable.c
+++ b/drivers/iommu/iommufd/io_pagetable.c
@@ -15,7 +15,9 @@
#include <linux/iommu.h>
#include <linux/iommufd.h>
#include <linux/lockdep.h>
+#include <linux/hugetlb.h>
#include <linux/sched/mm.h>
+#include <linux/shmem_fs.h>
#include <linux/slab.h>
#include <uapi/linux/iommufd.h>
@@ -515,6 +517,18 @@ int iopt_map_file_pages(struct iommufd_ctx *ictx, struct io_pagetable *iopt,
if (!file)
return -EBADF;
+ /*
+ * iopt_pages keeps a reference to the file until the mapping is
+ * destroyed, so only accept the files memfd_pin_folios() is able
+ * to pin. Any other fd, in particular one that itself owns an
+ * iommufd_ctx (iommufd, vfio, ...), would put the file into a
+ * reference cycle with this iommufd and never be released.
+ */
+ if (!shmem_file(file) && !is_file_hugepages(file)) {
+ fput(file);
+ return -EINVAL;
+ }
+
pages = iopt_alloc_file_pages(file, start_byte, start, length,
iommu_prot & IOMMU_WRITE);
fput(file);
--
2.43.0
Reviewed-by: Samiullah Khawaja <skhawaja@xxxxxxxxxx>
Thanks,
Sami