1 files changed, 58 insertions, 51 deletions

diff --git a/fs/xfs/xfs_log.c b/fs/xfs/xfs_log.c
index a8034c0afdf2..499e15b24215 100644
--- a/fs/xfs/xfs_log.c
+++ b/fs/xfs/xfs_log.c
@@ -487,7 +487,10 @@ out_error:
  * Run all the pending iclog callbacks and wake log force waiters and iclog
  * space waiters so they can process the newly set shutdown state. We really
  * don't care what order we process callbacks here because the log is shut down
- * and so state cannot change on disk anymore.
+ * and so state cannot change on disk anymore. However, we cannot wake waiters
+ * until the callbacks have been processed because we may be in unmount and
+ * we must ensure that all AIL operations the callbacks perform have completed
+ * before we tear down the AIL.
  *
  * We avoid processing actively referenced iclogs so that we don't run callbacks
  * while the iclog owner might still be preparing the iclog for IO submssion.
@@ -501,7 +504,6 @@ xlog_state_shutdown_callbacks(
 	struct xlog_in_core	*iclog;
 	LIST_HEAD(cb_list);
 
-	spin_lock(&log->l_icloglock);
 	iclog = log->l_iclog;
 	do {
 		if (atomic_read(&iclog->ic_refcnt)) {
@@ -509,26 +511,22 @@ xlog_state_shutdown_callbacks(
 			continue;
 		}
 		list_splice_init(&iclog->ic_callbacks, &cb_list);
+		spin_unlock(&log->l_icloglock);
+
+		xlog_cil_process_committed(&cb_list);
+
+		spin_lock(&log->l_icloglock);
 		wake_up_all(&iclog->ic_write_wait);
 		wake_up_all(&iclog->ic_force_wait);
 	} while ((iclog = iclog->ic_next) != log->l_iclog);
 
 	wake_up_all(&log->l_flush_wait);
-	spin_unlock(&log->l_icloglock);
-
-	xlog_cil_process_committed(&cb_list);
 }
 
 /*
  * Flush iclog to disk if this is the last reference to the given iclog and the
  * it is in the WANT_SYNC state.
  *
- * If the caller passes in a non-zero @old_tail_lsn and the current log tail
- * does not match, there may be metadata on disk that must be persisted before
- * this iclog is written.  To satisfy that requirement, set the
- * XLOG_ICL_NEED_FLUSH flag as a condition for writing this iclog with the new
- * log tail value.
- *
  * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
  * log tail is updated correctly. NEED_FUA indicates that the iclog will be
  * written to stable storage, and implies that a commit record is contained
@@ -545,12 +543,10 @@ xlog_state_shutdown_callbacks(
  * always capture the tail lsn on the iclog on the first NEED_FUA release
  * regardless of the number of active reference counts on this iclog.
  */
-
 int
 xlog_state_release_iclog(
 	struct xlog		*log,
-	struct xlog_in_core	*iclog,
-	xfs_lsn_t		old_tail_lsn)
+	struct xlog_in_core	*iclog)
 {
 	xfs_lsn_t		tail_lsn;
 	bool			last_ref;
@@ -561,18 +557,14 @@ xlog_state_release_iclog(
 	/*
 	 * Grabbing the current log tail needs to be atomic w.r.t. the writing
 	 * of the tail LSN into the iclog so we guarantee that the log tail does
-	 * not move between deciding if a cache flush is required and writing
-	 * the LSN into the iclog below.
+	 * not move between the first time we know that the iclog needs to be
+	 * made stable and when we eventually submit it.
 	 */
-	if (old_tail_lsn || iclog->ic_state == XLOG_STATE_WANT_SYNC) {
+	if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
+	     (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
+	    !iclog->ic_header.h_tail_lsn) {
 		tail_lsn = xlog_assign_tail_lsn(log->l_mp);
-
-		if (old_tail_lsn && tail_lsn != old_tail_lsn)
-			iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
-
-		if ((iclog->ic_flags & XLOG_ICL_NEED_FUA) &&
-		    !iclog->ic_header.h_tail_lsn)
-			iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
+		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
 	}
 
 	last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
@@ -583,11 +575,8 @@ xlog_state_release_iclog(
 		 * pending iclog callbacks that were waiting on the release of
 		 * this iclog.
 		 */
-		if (last_ref) {
-			spin_unlock(&log->l_icloglock);
+		if (last_ref)
 			xlog_state_shutdown_callbacks(log);
-			spin_lock(&log->l_icloglock);
-		}
 		return -EIO;
 	}
 
@@ -600,8 +589,6 @@ xlog_state_release_iclog(
 	}
 
 	iclog->ic_state = XLOG_STATE_SYNCING;
-	if (!iclog->ic_header.h_tail_lsn)
-		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
 	xlog_verify_tail_lsn(log, iclog);
 	trace_xlog_iclog_syncing(iclog, _RET_IP_);
 
@@ -873,7 +860,7 @@ xlog_force_iclog(
 	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
 	if (iclog->ic_state == XLOG_STATE_ACTIVE)
 		xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
-	return xlog_state_release_iclog(iclog->ic_log, iclog, 0);
+	return xlog_state_release_iclog(iclog->ic_log, iclog);
 }
 
 /*
@@ -1373,7 +1360,7 @@ xlog_ioend_work(
 	 */
 	if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
 		xfs_alert(log->l_mp, "log I/O error %d", error);
-		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
+		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
 	}
 
 	xlog_state_done_syncing(iclog);
@@ -1912,7 +1899,7 @@ xlog_write_iclog(
 	iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
 
 	if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
-		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
+		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
 		return;
 	}
 	if (is_vmalloc_addr(iclog->ic_data))
@@ -2411,7 +2398,7 @@ xlog_write_copy_finish(
 		ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
 			xlog_is_shutdown(log));
 release_iclog:
-	error = xlog_state_release_iclog(log, iclog, 0);
+	error = xlog_state_release_iclog(log, iclog);
 	spin_unlock(&log->l_icloglock);
 	return error;
 }
@@ -2487,7 +2474,7 @@ xlog_write(
 		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
 		     "ctx ticket reservation ran out. Need to up reservation");
 		xlog_print_tic_res(log->l_mp, ticket);
-		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
+		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
 	}
 
 	len = xlog_write_calc_vec_length(ticket, log_vector, optype);
@@ -2628,7 +2615,7 @@ next_lv:
 
 	spin_lock(&log->l_icloglock);
 	xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
-	error = xlog_state_release_iclog(log, iclog, 0);
+	error = xlog_state_release_iclog(log, iclog);
 	spin_unlock(&log->l_icloglock);
 
 	return error;
@@ -3052,7 +3039,7 @@ restart:
 		 * reference to the iclog.
 		 */
 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
-			error = xlog_state_release_iclog(log, iclog, 0);
+			error = xlog_state_release_iclog(log, iclog);
 		spin_unlock(&log->l_icloglock);
 		if (error)
 			return error;
@@ -3821,9 +3808,10 @@ xlog_verify_iclog(
 #endif
 
 /*
- * Perform a forced shutdown on the log. This should be called once and once
- * only by the high level filesystem shutdown code to shut the log subsystem
- * down cleanly.
+ * Perform a forced shutdown on the log.
+ *
+ * This can be called from low level log code to trigger a shutdown, or from the
+ * high level mount shutdown code when the mount shuts down.
  *
  * Our main objectives here are to make sure that:
  *	a. if the shutdown was not due to a log IO error, flush the logs to
@@ -3832,6 +3820,8 @@ xlog_verify_iclog(
  *	   parties to find out. Nothing new gets queued after this is done.
  *	c. Tasks sleeping on log reservations, pinned objects and
  *	   other resources get woken up.
+ *	d. The mount is also marked as shut down so that log triggered shutdowns
+ *	   still behave the same as if they called xfs_forced_shutdown().
  *
  * Return true if the shutdown cause was a log IO error and we actually shut the
  * log down.
@@ -3843,25 +3833,25 @@ xlog_force_shutdown(
 {
 	bool		log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
 
-	/*
-	 * If this happens during log recovery then we aren't using the runtime
-	 * log mechanisms yet so there's nothing to shut down.
-	 */
-	if (!log || xlog_in_recovery(log))
+	if (!log)
 		return false;
 
-	ASSERT(!xlog_is_shutdown(log));
-
 	/*
 	 * Flush all the completed transactions to disk before marking the log
 	 * being shut down. We need to do this first as shutting down the log
 	 * before the force will prevent the log force from flushing the iclogs
 	 * to disk.
 	 *
-	 * Re-entry due to a log IO error shutdown during the log force is
-	 * prevented by the atomicity of higher level shutdown code.
+	 * When we are in recovery, there are no transactions to flush, and
+	 * we don't want to touch the log because we don't want to perturb the
+	 * current head/tail for future recovery attempts. Hence we need to
+	 * avoid a log force in this case.
+	 *
+	 * If we are shutting down due to a log IO error, then we must avoid
+	 * trying to write the log as that may just result in more IO errors and
+	 * an endless shutdown/force loop.
 	 */
-	if (!log_error)
+	if (!log_error && !xlog_in_recovery(log))
 		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
 
 	/*
@@ -3878,12 +3868,25 @@ xlog_force_shutdown(
 	spin_lock(&log->l_icloglock);
 	if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
 		spin_unlock(&log->l_icloglock);
-		ASSERT(0);
 		return false;
 	}
 	spin_unlock(&log->l_icloglock);
 
 	/*
+	 * If this log shutdown also sets the mount shutdown state, issue a
+	 * shutdown warning message.
+	 */
+	if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &log->l_mp->m_opstate)) {
+		xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
+"Filesystem has been shut down due to log error (0x%x).",
+				shutdown_flags);
+		xfs_alert(log->l_mp,
+"Please unmount the filesystem and rectify the problem(s).");
+		if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
+			xfs_stack_trace();
+	}
+
+	/*
 	 * We don't want anybody waiting for log reservations after this. That
 	 * means we have to wake up everybody queued up on reserveq as well as
 	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
@@ -3903,8 +3906,12 @@ xlog_force_shutdown(
 	wake_up_all(&log->l_cilp->xc_start_wait);
 	wake_up_all(&log->l_cilp->xc_commit_wait);
 	spin_unlock(&log->l_cilp->xc_push_lock);
+
+	spin_lock(&log->l_icloglock);
 	xlog_state_shutdown_callbacks(log);
+	spin_unlock(&log->l_icloglock);
 
+	wake_up_var(&log->l_opstate);
 	return log_error;
 }