fs: aio: reject partial mremap to avoid Null-pointer-dereference error

[BUG]
Recently, our internal syzkaller testing uncovered a null pointer
dereference issue:
BUG: kernel NULL pointer dereference, address: 0000000000000000
...
[   51.111664]  filemap_read_folio+0x25/0xe0
[   51.112410]  filemap_fault+0xad7/0x1250
[   51.113112]  __do_fault+0x4b/0x460
[   51.113699]  do_pte_missing+0x5bc/0x1db0
[   51.114250]  ? __pte_offset_map+0x23/0x170
[   51.114822]  __handle_mm_fault+0x9f8/0x1680
...
Crash analysis showed the file involved was an AIO ring file. The
phenomenon triggered is the same as the issue described in [1].

[CAUSE]
Consider the following scenario: userspace sets up an AIO context via
io_setup(), which creates a VMA covering the entire ring buffer. Then
userspace calls mremap() with the AIO ring address as the source, a smaller
old_len (less than the full ring size), MREMAP_MAYMOVE set, and without
MREMAP_DONTUNMAP. The kernel will relocate the requested portion to a new
destination address.

During this move, __split_vma() splits the original AIO ring VMA. The
requested portion is unmapped from the source and re-established at the
destination, while the remainder stays at the original source address as
an orphan VMA. The aio_ring_mremap() callback fires on the new destination
VMA, updating ctx->mmap_base to the destination address. But the callback
is unaware that only a partial region was moved and that an orphan VMA
still exists at the source:

  source(AIO):
  +-------------------+---------------------+
  |  moved to dest    |  orphan VMA (AIO)   |
  +-------------------+---------------------+
  A                 A+partial_len        A+ctx->mmap_size

  dest:
  +-------------------+
  |  moved VMA (AIO)  |
  +-------------------+
  B                 B+partial_len

Later, io_destroy() calls vm_munmap(ctx->mmap_base, ctx->mmap_size), which
unmaps the destination. This not only fails to unmap the orphan VMA at the
source, but also overshoots the destination VMA and may unmap unrelated
mappings adjacent to it! After put_aio_ring_file() calls truncate_setsize()
to remove all pages from the pagecache, any subsequent access to the orphan
VMA triggers filemap_fault(), which calls a_ops->read_folio(). Since aio
does not implement read_folio, this results in a NULL pointer dereference.

[FIX]
Note that expanding mremap (new_len > old_len) is already rejected because
AIO ring VMAs are created with VM_DONTEXPAND. The only problematic case is
a partial move where "old_len == new_len" but both are smaller than the
full ring size.

Fix this by checking in aio_ring_mremap() that the new VMA covers the
entire ring. This ensures the AIO ring is always moved as a whole,
preventing orphan VMAs and the subsequent crash.

[1]: https://lore.kernel.org/all/20260413010814.548568-1-wozizhi@huawei.com/

Signed-off-by: Zizhi Wo <wozizhi@huaweicloud.com>
Link: https://patch.msgid.link/20260418060634.3713620-1-wozizhi@huaweicloud.com
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Christian Brauner <brauner@kernel.org>

diff --git a/fs/aio.c b/fs/aio.c
index d7910c7c93a621103c60c74988bc07d669473005..72247656084808217346402eeb24d16adc5a0215 100644 (file)
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -422,7 +422,8 @@ static int aio_ring_mremap(struct vm_area_struct *vma)
 
                ctx = rcu_dereference(table->table[i]);
                if (ctx && ctx->aio_ring_file == file) {
-                       if (!atomic_read(&ctx->dead)) {
+                       if (!atomic_read(&ctx->dead) &&
+                           (ctx->mmap_size == (vma->vm_end - vma->vm_start))) {
                                ctx->user_id = ctx->mmap_base = vma->vm_start;
                                res = 0;
                        }