cifs: Fix memory leak in direct I/O

When __cifs_readv() and __cifs_writev() extract pages from a user-backed
iterator into a BVEC-type iterator, they set ->bv_need_unpin to note
whether they need to unpin the pages later.  However, in both cases they
examine the BVEC-type iterator and not the source iterator - and so
bv_need_unpin doesn't get set and the pages are leaked.

I think this may be responsible for the generic/208 xfstest failing
occasionally with:

	WARNING: CPU: 0 PID: 3064 at mm/gup.c:218 try_grab_page+0x65/0x100
	RIP: 0010:try_grab_page+0x65/0x100
	? __kmalloc_large_node+0xb1/0x120
	? __kmalloc_node+0xbe/0x130
	netfs_extract_user_iter+0xbf/0x200 [netfs]
	__cifs_writev+0x150/0x330 [cifs]

with the page refcount going negative.  This is less unlikely than it seems
because the page is being pinned, not simply got, and so the refcount
increased by 1024 each time, and so only needs to be called around ~2097152
for the refcount to go negative.

Further, the test program (aio-dio-invalidate-failure) uses a 32MiB static
buffer and all the PTEs covering it refer to the same page because it's
never written to.

The warning in try_grab_page():

	if (WARN_ON_ONCE(folio_ref_count(folio) <= 0))
		return -ENOMEM;

then trips and prevents us ever using the page again for DIO at least.

Fixes: d08089f649a0 ("cifs: Change the I/O paths to use an iterator rather than a page list")
Reported-by: Murphy Zhou <>
Signed-off-by: David Howells <>
Reviewed-by: Paulo Alcantara (SUSE) <>
cc: Shyam Prasad N <>
cc: Rohith Surabattula <>
cc: Jeff Layton <>
Signed-off-by: Steve French <>
1 file changed