jbd2: fix assertion failure in jbd2_journal_flush()
authorJan Kara <jack@suse.cz>
Fri, 21 Dec 2012 05:15:51 +0000 (00:15 -0500)
committerTheodore Ts'o <tytso@mit.edu>
Fri, 21 Dec 2012 05:15:51 +0000 (00:15 -0500)
commitd7961c7fa4d2e3c3f12be67e21ba8799b5a7238a
tree3341ce3bdff7f2283e18f0cb02c438f421d746b0
parent261cb20cb2f0737a247aaf08dff7eb065e3e5b66
jbd2: fix assertion failure in jbd2_journal_flush()

The following race is possible between start_this_handle() and someone
calling jbd2_journal_flush().

Process A                              Process B
start_this_handle().
  if (journal->j_barrier_count) # false
  if (!journal->j_running_transaction) { #true
    read_unlock(&journal->j_state_lock);
                                       jbd2_journal_lock_updates()
                                       jbd2_journal_flush()
                                         write_lock(&journal->j_state_lock);
                                         if (journal->j_running_transaction) {
                                           # false
                                         ... wait for committing trans ...
                                         write_unlock(&journal->j_state_lock);
    ...
    write_lock(&journal->j_state_lock);
    if (!journal->j_running_transaction) { # true
      jbd2_get_transaction(journal, new_transaction);
    write_unlock(&journal->j_state_lock);
    goto repeat; # eventually blocks on j_barrier_count > 0
                                         ...
                                         J_ASSERT(!journal->j_running_transaction);
                                           # fails

We fix the race by rechecking j_barrier_count after reacquiring j_state_lock
in exclusive mode.

Reported-by: yjwsignal@empal.com
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@vger.kernel.org
fs/jbd2/transaction.c