jbd: don't wait (forever) for stale tid caused by wraparound
Jan Kara [Wed, 27 Mar 2013 16:30:59 +0000 (17:30 +0100)]
In the case where an inode has a very stale transaction id (tid) in
i_datasync_tid or i_sync_tid, it's possible that after a very large
(2**31) number of transactions, that the tid number space might wrap,
causing tid_geq()'s calculations to fail.

Commit d9b0193 "jbd: fix fsync() tid wraparound bug" attempted to fix
this problem, but it only avoided kjournald spinning forever by fixing
the logic in jbd_log_start_commit().

Signed-off-by: Jan Kara <jack@suse.cz>

fs/jbd/journal.c

index 81cc7ea..81880c6 100644 (file)
@@ -564,6 +564,16 @@ int log_wait_commit(journal_t *journal, tid_t tid)
        spin_unlock(&journal->j_state_lock);
 #endif
        spin_lock(&journal->j_state_lock);
+       /*
+        * Not running or committing trans? Must be already committed. This
+        * saves us from waiting for a *long* time when tid overflows.
+        */
+       if (!((journal->j_running_transaction &&
+              journal->j_running_transaction->t_tid == tid) ||
+             (journal->j_committing_transaction &&
+              journal->j_committing_transaction->t_tid == tid)))
+               goto out_unlock;
+
        if (!tid_geq(journal->j_commit_waited, tid))
                journal->j_commit_waited = tid;
        while (tid_gt(tid, journal->j_commit_sequence)) {
@@ -575,6 +585,7 @@ int log_wait_commit(journal_t *journal, tid_t tid)
                                !tid_gt(tid, journal->j_commit_sequence));
                spin_lock(&journal->j_state_lock);
        }
+out_unlock:
        spin_unlock(&journal->j_state_lock);
 
        if (unlikely(is_journal_aborted(journal))) {