ext4: fix deadlock allocating crypto bounce page from mempool

[ Upstream commit 547c556f4db7c09447ecf5f833ab6aaae0c5ab58 ]

ext4_writepages() on an encrypted file has to encrypt the data, but it
can't modify the pagecache pages in-place, so it encrypts the data into
bounce pages and writes those instead.  All bounce pages are allocated
from a mempool using GFP_NOFS.

This is not correct use of a mempool, and it can deadlock.  This is
because GFP_NOFS includes __GFP_DIRECT_RECLAIM, which enables the "never
fail" mode for mempool_alloc() where a failed allocation will fall back
to waiting for one of the preallocated elements in the pool.

But since this mode is used for all a bio's pages and not just the
first, it can deadlock waiting for pages already in the bio to be freed.

This deadlock can be reproduced by patching mempool_alloc() to pretend
that pool->alloc() always fails (so that it always falls back to the
preallocations), and then creating an encrypted file of size > 128 KiB.

Fix it by only using GFP_NOFS for the first page in the bio.  For
subsequent pages just use GFP_NOWAIT, and if any of those fail, just
submit the bio and start a new one.

This will need to be fixed in f2fs too, but that's less straightforward.

Fixes: c9af28fdd449 ("ext4 crypto: don't let data integrity writebacks fail with ENOMEM")
Cc: stable@vger.kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
Link: https://lore.kernel.org/r/20191231181149.47619-1-ebiggers@kernel.org
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Sasha Levin <sashal@kernel.org>

diff --git a/fs/ext4/page-io.c b/fs/ext4/page-io.c
index 0094923e5ebf56a6a65198cf5a7040078b73240a..94f60f9d57fd11b1dbd27839c1ad33d3631f1ec6 100644 (file)
--- a/fs/ext4/page-io.c
+++ b/fs/ext4/page-io.c
@@ -469,16 +469,25 @@ int ext4_bio_write_page(struct ext4_io_submit *io,
            nr_to_submit) {
                gfp_t gfp_flags = GFP_NOFS;
 
+               /*
+                * Since bounce page allocation uses a mempool, we can only use
+                * a waiting mask (i.e. request guaranteed allocation) on the
+                * first page of the bio.  Otherwise it can deadlock.
+                */
+               if (io->io_bio)
+                       gfp_flags = GFP_NOWAIT | __GFP_NOWARN;
        retry_encrypt:
                data_page = fscrypt_encrypt_page(inode, page, gfp_flags);
                if (IS_ERR(data_page)) {
                        ret = PTR_ERR(data_page);
-                       if (ret == -ENOMEM && wbc->sync_mode == WB_SYNC_ALL) {
-                               if (io->io_bio) {
+                       if (ret == -ENOMEM &&
+                           (io->io_bio || wbc->sync_mode == WB_SYNC_ALL)) {
+                               gfp_flags = GFP_NOFS;
+                               if (io->io_bio)
                                        ext4_io_submit(io);
-                                       congestion_wait(BLK_RW_ASYNC, HZ/50);
-                               }
-                               gfp_flags |= __GFP_NOFAIL;
+                               else
+                                       gfp_flags |= __GFP_NOFAIL;
+                               congestion_wait(BLK_RW_ASYNC, HZ/50);
                                goto retry_encrypt;
                        }
                        data_page = NULL;