ext4: fix superblock checksum calculation race

The race condition could cause the persisted superblock checksum
to not match the contents of the superblock, causing the
superblock to be considered corrupt.

An example of the race follows.  A first thread is interrupted in the
middle of a checksum calculation. Then, another thread changes the
superblock, calculates a new checksum, and sets it. Then, the first
thread resumes and sets the checksum based on the older superblock.

To fix, serialize the superblock checksum calculation using the buffer
header lock. While a spinlock is sufficient, the buffer header is
already there and there is precedent for locking it (e.g. in
ext4_commit_super).

Tested the patch by booting up a kernel with the patch, creating
a filesystem and some files (including some orphans), and then
unmounting and remounting the file system.

Cc: stable@kernel.org
Signed-off-by: Constantine Sapuntzakis <costa@purestorage.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Suggested-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20200914161014.22275-1-costa@purestorage.comSigned-off-by: Theodore Ts'o <tytso@mit.edu>

fs/ext4/super.c

@@ -201,7 +201,18 @@ void ext4_superblock_csum_set(struct super_block *sb)
 	if (!ext4_has_metadata_csum(sb))
 		return;
 
+	/*
+	 * Locking the superblock prevents the scenario
+	 * where:
+	 *  1) a first thread pauses during checksum calculation.
+	 *  2) a second thread updates the superblock, recalculates
+	 *     the checksum, and updates s_checksum
+	 *  3) the first thread resumes and finishes its checksum calculation
+	 *     and updates s_checksum with a potentially stale or torn value.
+	 */
+	lock_buffer(EXT4_SB(sb)->s_sbh);
 	es->s_checksum = ext4_superblock_csum(sb, es);
+	unlock_buffer(EXT4_SB(sb)->s_sbh);
 }
 
 ext4_fsblk_t ext4_block_bitmap(struct super_block *sb,