xfs: Fix a deadlock in xfs_log_commit_cil() code path
authorChandra Seetharaman <sekharan@us.ibm.com>
Fri, 19 Jul 2013 22:31:38 +0000 (17:31 -0500)
committerBen Myers <bpm@sgi.com>
Mon, 22 Jul 2013 18:58:10 +0000 (13:58 -0500)
While testing and rearranging pquota/gquota code, I stumbled
on a xfs_shutdown() during a mount. But the mount just hung.

Debugged and found that there is a deadlock involving
&log->l_cilp->xc_ctx_lock.

It is in a code path where &log->l_cilp->xc_ctx_lock is first
acquired in read mode and some levels down the same semaphore
is being acquired in write mode causing a deadlock.

This is the stack:
xfs_log_commit_cil -> acquires &log->l_cilp->xc_ctx_lock in read mode
  xlog_print_tic_res
    xfs_force_shutdown
      xfs_log_force_umount
        xlog_cil_force
          xlog_cil_force_lsn
            xlog_cil_push_foreground
              xlog_cil_push - tries to acquire same semaphore in write mode

This patch fixes the deadlock by changing the reason code for
xfs_force_shutdown in xlog_print_tic_res() to SHUTDOWN_LOG_IO_ERROR.

SHUTDOWN_LOG_IO_ERROR is the right reason code to be set since
we are in the log path.

Thanks to Dave for suggesting this solution.

Signed-off-by: Chandra Seetharaman <sekharan@us.ibm.com>
Reviewed-by: Ben Myers <bpm@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>
fs/xfs/xfs_log.c

index d852a2b3e1fdfae0c4fb5bf18452ec79fd03ab01..bf89eb97fefdcf24cf95a3fc3eb57c4b8ecb31cc 100644 (file)
@@ -1941,7 +1941,7 @@ xlog_print_tic_res(
 
        xfs_alert_tag(mp, XFS_PTAG_LOGRES,
                "xlog_write: reservation ran out. Need to up reservation");
-       xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
+       xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 }
 
 /*