[PATCH] ext4: Prevent successful remount as r/w after filesystem abort

[Fanqi Yu <fanqi.yu@columbia.edu>]

While the order of writing to the shutdown and read-only flags has been
enforced by a write memory barrier, the read side in remount does not
have a pairing read barrier.

In the event of a fs forced shutdown remounting as read-only,
remounting it again as read-write can succeed when the flag reads are
reordered such that sb_rdonly() returns true and
ext4_forced_shutdown() returns false.

Commit 4418e14112e3 ("ext4: Fix fsync error handling after filesystem
abort") has added the other barriers related to these two flags but
seems to have missed this one.

Signed-off-by: Fanqi Yu <fanqi.yu@columbia.edu>
---
 fs/ext4/super.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/fs/ext4/super.c b/fs/ext4/super.c
index 044135796f2b..c5d3f8969dec 100644
--- a/fs/ext4/super.c
+++ b/fs/ext4/super.c
@@ -6542,6 +6542,13 @@ static int __ext4_remount(struct fs_context *fc, struct super_block *sb)
 	flush_work(&sbi->s_sb_upd_work);
 
 	if ((bool)(fc->sb_flags & SB_RDONLY) != sb_rdonly(sb)) {
+
+		/*
+		 * Make sure we read the updated s_ext4_flags.
+		 * Pairs with smp_wmb() in ext4_handle_error().
+		 */
+		smp_rmb();
+
 		if (ext4_forced_shutdown(sb)) {
 			err = -EROFS;
 			goto restore_opts;
-- 
2.34.1