jbd2: fix deadlock while checkpoint thread waits commit thread to finish

This issue was found when I tried to put checkpoint work in a separate thread,
the deadlock below happened:
Thread1 | Thread2
__jbd2_log_wait_for_space |
jbd2_log_do_checkpoint (hold j_checkpoint_mutex)|
if (jh->b_transaction != NULL) |
... |
jbd2_log_start_commit(journal, tid); |jbd2_update_log_tail
| will lock j_checkpoint_mutex,
| but will be blocked here.
|
jbd2_log_wait_commit(journal, tid); |
wait_event(journal->j_wait_done_commit, |
!tid_gt(tid, journal->j_commit_sequence)); |
... |wake_up(j_wait_done_commit)
} |

then deadlock occurs, Thread1 will never be waken up.

To fix this issue, drop j_checkpoint_mutex in jbd2_log_do_checkpoint()
when we are going to wait for transaction commit.

Reviewed-by: Jan Kara <[e-mail]>
Signed-off-by: Xiaoguang Wang <[e-mail]>
Signed-off-by: Theodore Ts'o <[e-mail]>

fs/jbd2/checkpoint.c

···
 	nblocks = jbd2_space_needed(journal);
 	while (jbd2_log_space_left(journal) < nblocks) {
 		write_unlock(&journal->j_state_lock);
-		mutex_lock(&journal->j_checkpoint_mutex);
+		mutex_lock_io(&journal->j_checkpoint_mutex);
 
 		/*
 		 * Test again, another process may have checkpointed while we
···
 		"JBD2: %s: Waiting for Godot: block %llu\n",
 		journal->j_devname, (unsigned long long) bh->b_blocknr);
 
+			if (batch_count)
+				__flush_batch(journal, &batch_count);
 			jbd2_log_start_commit(journal, tid);
+			/*
+			 * jbd2_journal_commit_transaction() may want
+			 * to take the checkpoint_mutex if JBD2_FLUSHED
+			 * is set, jbd2_update_log_tail() called by
+			 * jbd2_journal_commit_transaction() may also take
+			 * checkpoint_mutex.  So we need to temporarily
+			 * drop it.
+			 */
+			mutex_unlock(&journal->j_checkpoint_mutex);
 			jbd2_log_wait_commit(journal, tid);
-			goto retry;
+			mutex_lock_io(&journal->j_checkpoint_mutex);
+			spin_lock(&journal->j_list_lock);
+			goto restart;
 		}
 		if (!buffer_dirty(bh)) {
 			if (unlikely(buffer_write_io_error(bh)) && !result)

fs/jbd2/journal.c

···
 	err = jbd2_journal_skip_recovery(journal);
 	if (write) {
 		/* Lock to make assertions happy... */
-		mutex_lock(&journal->j_checkpoint_mutex);
+		mutex_lock_io(&journal->j_checkpoint_mutex);
 		jbd2_mark_journal_empty(journal, REQ_SYNC | REQ_FUA);
 		mutex_unlock(&journal->j_checkpoint_mutex);
 	}