cifs: Fix memory leak in direct I/O 6.3-rc-smb3-client-fixes-part2
authorDavid Howells <dhowells@redhat.com>
Tue, 28 Feb 2023 22:38:38 +0000 (22:38 +0000)
committerSteve French <stfrench@microsoft.com>
Thu, 2 Mar 2023 00:18:25 +0000 (18:18 -0600)
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
follow_page_pte+0x1a7/0x570
__get_user_pages+0x1a2/0x650
__gup_longterm_locked+0xdc/0xb50
internal_get_user_pages_fast+0x17f/0x310
pin_user_pages_fast+0x46/0x60
iov_iter_extract_pages+0xc9/0x510
? __kmalloc_large_node+0xb1/0x120
? __kmalloc_node+0xbe/0x130
netfs_extract_user_iter+0xbf/0x200 [netfs]
__cifs_writev+0x150/0x330 [cifs]
vfs_write+0x2a8/0x3c0
ksys_pwrite64+0x65/0xa0

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 <jencce.kernel@gmail.com>
Link: https://lore.kernel.org/r/CAH2r5mvaTsJ---n=265a4zqRA7pP+o4MJ36WCQUS6oPrOij8cw@mail.gmail.com
Signed-off-by: David Howells <dhowells@redhat.com>
Reviewed-by: Paulo Alcantara (SUSE) <pc@manguebit.com>
cc: Shyam Prasad N <nspmangalore@gmail.com>
cc: Rohith Surabattula <rohiths.msft@gmail.com>
cc: Jeff Layton <jlayton@kernel.org>
cc: linux-cifs@vger.kernel.org
Signed-off-by: Steve French <stfrench@microsoft.com>
fs/cifs/file.c

index ec0694a65c7b6ce592704158df3e0d29b5546396..4d4a2d82636d2a359f785c41c5efca5ba5aa8504 100644 (file)
@@ -3612,7 +3612,7 @@ static ssize_t __cifs_writev(
 
                ctx->nr_pinned_pages = rc;
                ctx->bv = (void *)ctx->iter.bvec;
-               ctx->bv_need_unpin = iov_iter_extract_will_pin(&ctx->iter);
+               ctx->bv_need_unpin = iov_iter_extract_will_pin(from);
        } else if ((iov_iter_is_bvec(from) || iov_iter_is_kvec(from)) &&
                   !is_sync_kiocb(iocb)) {
                /*
@@ -4148,7 +4148,7 @@ static ssize_t __cifs_readv(
 
                ctx->nr_pinned_pages = rc;
                ctx->bv = (void *)ctx->iter.bvec;
-               ctx->bv_need_unpin = iov_iter_extract_will_pin(&ctx->iter);
+               ctx->bv_need_unpin = iov_iter_extract_will_pin(to);
                ctx->should_dirty = true;
        } else if ((iov_iter_is_bvec(to) || iov_iter_is_kvec(to)) &&
                   !is_sync_kiocb(iocb)) {