jbd2: ensure abort the journal if detect IO error when writing original buffer back

Although we merged c044f3d8360 ("jbd2: abort journal if free a async
write error metadata buffer"), there is a race between
jbd2_journal_try_to_free_buffers() and jbd2_journal_destroy(), so the
jbd2_log_do_checkpoint() may still fail to detect the buffer write
io error flag which may lead to filesystem inconsistency.

jbd2_journal_try_to_free_buffers()     ext4_put_super()
                                        jbd2_journal_destroy()
  __jbd2_journal_remove_checkpoint()
  detect buffer write error              jbd2_log_do_checkpoint()
                                         jbd2_cleanup_journal_tail()
                                           <--- lead to inconsistency
  jbd2_journal_abort()

Fix this issue by introducing a new atomic flag which only have one
JBD2_CHECKPOINT_IO_ERROR bit now, and set it in
__jbd2_journal_remove_checkpoint() when freeing a checkpoint buffer
which has write_io_error flag. Then jbd2_journal_destroy() will detect
this mark and abort the journal to prevent updating log tail.

Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20210610112440.3438139-3-yi.zhang@huawei.com
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

2 files changed, 26 insertions, 0 deletions

diff --git a/fs/jbd2/checkpoint.c b/fs/jbd2/checkpoint.c
index bf5511d19ac56..d27c10f4502f6 100644
--- a/fs/jbd2/checkpoint.c
+++ b/fs/jbd2/checkpoint.c
@@ -564,6 +564,7 @@ int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
 	struct transaction_chp_stats_s *stats;
 	transaction_t *transaction;
 	journal_t *journal;
+	struct buffer_head *bh = jh2bh(jh);
 
 	JBUFFER_TRACE(jh, "entry");
 
@@ -575,6 +576,17 @@ int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
 	journal = transaction->t_journal;
 
 	JBUFFER_TRACE(jh, "removing from transaction");
+
+	/*
+	 * If we have failed to write the buffer out to disk, the filesystem
+	 * may become inconsistent. We cannot abort the journal here since
+	 * we hold j_list_lock and we have to be careful about races with
+	 * jbd2_journal_destroy(). So mark the writeback IO error in the
+	 * journal here and we abort the journal later from a better context.
+	 */
+	if (buffer_write_io_error(bh))
+		set_bit(JBD2_CHECKPOINT_IO_ERROR, &journal->j_atomic_flags);
+
 	__buffer_unlink(jh);
 	jh->b_cp_transaction = NULL;
 	jbd2_journal_put_journal_head(jh);
diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c
index f88895b4920cd..8b3f5bbd65f95 100644
--- a/fs/jbd2/journal.c
+++ b/fs/jbd2/journal.c
@@ -1610,6 +1610,10 @@ int jbd2_journal_update_sb_log_tail(journal_t *journal, tid_t tail_tid,
 
 	if (is_journal_aborted(journal))
 		return -EIO;
+	if (test_bit(JBD2_CHECKPOINT_IO_ERROR, &journal->j_atomic_flags)) {
+		jbd2_journal_abort(journal, -EIO);
+		return -EIO;
+	}
 
 	BUG_ON(!mutex_is_locked(&journal->j_checkpoint_mutex));
 	jbd_debug(1, "JBD2: updating superblock (start %lu, seq %u)\n",
@@ -2091,6 +2095,16 @@ int jbd2_journal_destroy(journal_t *journal)
 	J_ASSERT(journal->j_checkpoint_transactions == NULL);
 	spin_unlock(&journal->j_list_lock);
 
+	/*
+	 * OK, all checkpoint transactions have been checked, now check the
+	 * write out io error flag and abort the journal if some buffer failed
+	 * to write back to the original location, otherwise the filesystem
+	 * may become inconsistent.
+	 */
+	if (!is_journal_aborted(journal) &&
+	    test_bit(JBD2_CHECKPOINT_IO_ERROR, &journal->j_atomic_flags))
+		jbd2_journal_abort(journal, -EIO);
+
 	if (journal->j_sb_buffer) {
 		if (!is_journal_aborted(journal)) {
 			mutex_lock_io(&journal->j_checkpoint_mutex);