xfs: don't sleep in xlog_cil_force_lsn on shutdown
authorDave Chinner <dchinner@redhat.com>
Tue, 6 May 2014 22:05:50 +0000 (08:05 +1000)
committerDave Chinner <david@fromorbit.com>
Tue, 6 May 2014 22:05:50 +0000 (08:05 +1000)
Reports of a shutdown hang when fsyncing a directory have surfaced,
such as this:

[ 3663.394472] Call Trace:
[ 3663.397199]  [<ffffffff815f1889>] schedule+0x29/0x70
[ 3663.402743]  [<ffffffffa01feda5>] xlog_cil_force_lsn+0x185/0x1a0 [xfs]
[ 3663.416249]  [<ffffffffa01fd3af>] _xfs_log_force_lsn+0x6f/0x2f0 [xfs]
[ 3663.429271]  [<ffffffffa01a339d>] xfs_dir_fsync+0x7d/0xe0 [xfs]
[ 3663.435873]  [<ffffffff811df8c5>] do_fsync+0x65/0xa0
[ 3663.441408]  [<ffffffff811dfbc0>] SyS_fsync+0x10/0x20
[ 3663.447043]  [<ffffffff815fc7d9>] system_call_fastpath+0x16/0x1b

If we trigger a shutdown in xlog_cil_push() from xlog_write(), we
will never wake waiters on the current push sequence number, so
anything waiting in xlog_cil_force_lsn() for that push sequence
number to come up will not get woken and hence stall the shutdown.

Fix this by ensuring we call wake_up_all(&cil->xc_commit_wait) in
the push abort handling, in the log shutdown code when waking all
waiters, and adding a shutdown check in the sequence completion wait
loops to ensure they abort when a wakeup due to a shutdown occurs.

Reported-by: Boris Ranto <branto@redhat.com>
Reported-by: Eric Sandeen <esandeen@redhat.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
fs/xfs/xfs_log.c
fs/xfs/xfs_log_cil.c

index 8497a00..82f5602 100644 (file)
@@ -3915,11 +3915,14 @@ xfs_log_force_umount(
                retval = xlog_state_ioerror(log);
                spin_unlock(&log->l_icloglock);
        }
+
        /*
-        * Wake up everybody waiting on xfs_log_force.
-        * Callback all log item committed functions as if the
-        * log writes were completed.
+        * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
+        * as if the log writes were completed. The abort handling in the log
+        * item committed callback functions will do this again under lock to
+        * avoid races.
         */
+       wake_up_all(&log->l_cilp->xc_commit_wait);
        xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
 
 #ifdef XFSERRORDEBUG
index 7e54553..039c873 100644 (file)
@@ -385,7 +385,15 @@ xlog_cil_committed(
        xfs_extent_busy_clear(mp, &ctx->busy_extents,
                             (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
 
+       /*
+        * If we are aborting the commit, wake up anyone waiting on the
+        * committing list.  If we don't, then a shutdown we can leave processes
+        * waiting in xlog_cil_force_lsn() waiting on a sequence commit that
+        * will never happen because we aborted it.
+        */
        spin_lock(&ctx->cil->xc_push_lock);
+       if (abort)
+               wake_up_all(&ctx->cil->xc_commit_wait);
        list_del(&ctx->committing);
        spin_unlock(&ctx->cil->xc_push_lock);
 
@@ -564,8 +572,18 @@ restart:
        spin_lock(&cil->xc_push_lock);
        list_for_each_entry(new_ctx, &cil->xc_committing, committing) {
                /*
+                * Avoid getting stuck in this loop because we were woken by the
+                * shutdown, but then went back to sleep once already in the
+                * shutdown state.
+                */
+               if (XLOG_FORCED_SHUTDOWN(log)) {
+                       spin_unlock(&cil->xc_push_lock);
+                       goto out_abort_free_ticket;
+               }
+
+               /*
                 * Higher sequences will wait for this one so skip them.
-                * Don't wait for own own sequence, either.
+                * Don't wait for our own sequence, either.
                 */
                if (new_ctx->sequence >= ctx->sequence)
                        continue;
@@ -810,6 +828,13 @@ restart:
         */
        spin_lock(&cil->xc_push_lock);
        list_for_each_entry(ctx, &cil->xc_committing, committing) {
+               /*
+                * Avoid getting stuck in this loop because we were woken by the
+                * shutdown, but then went back to sleep once already in the
+                * shutdown state.
+                */
+               if (XLOG_FORCED_SHUTDOWN(log))
+                       goto out_shutdown;
                if (ctx->sequence > sequence)
                        continue;
                if (!ctx->commit_lsn) {
@@ -833,14 +858,12 @@ restart:
         * push sequence after the above wait loop and the CIL still contains
         * dirty objects.
         *
-        * When the push occurs, it will empty the CIL and
-        * atomically increment the currect sequence past the push sequence and
-        * move it into the committing list. Of course, if the CIL is clean at
-        * the time of the push, it won't have pushed the CIL at all, so in that
-        * case we should try the push for this sequence again from the start
-        * just in case.
+        * When the push occurs, it will empty the CIL and atomically increment
+        * the currect sequence past the push sequence and move it into the
+        * committing list. Of course, if the CIL is clean at the time of the
+        * push, it won't have pushed the CIL at all, so in that case we should
+        * try the push for this sequence again from the start just in case.
         */
-
        if (sequence == cil->xc_current_sequence &&
            !list_empty(&cil->xc_cil)) {
                spin_unlock(&cil->xc_push_lock);
@@ -849,6 +872,17 @@ restart:
 
        spin_unlock(&cil->xc_push_lock);
        return commit_lsn;
+
+       /*
+        * We detected a shutdown in progress. We need to trigger the log force
+        * to pass through it's iclog state machine error handling, even though
+        * we are already in a shutdown state. Hence we can't return
+        * NULLCOMMITLSN here as that has special meaning to log forces (i.e.
+        * LSN is already stable), so we return a zero LSN instead.
+        */
+out_shutdown:
+       spin_unlock(&cil->xc_push_lock);
+       return 0;
 }
 
 /*