GFS2: Fix recovery issues for spectators

This patch fixes a couple problems dealing with spectators who
remain with gfs2 mounts after the last non-spectator node fails.

Before this patch, spectator mounts would try to acquire the dlm's
mounted lock EX as part of its normal recovery sequence.
The mounted lock is only used to determine whether the node is
the first mounter, the first node to mount the file system, for
the purposes of file system recovery and journal replay.

It's not necessary for spectators: they should never do journal
recovery. If they acquire the lock it will prevent another "real"
first-mounter from acquiring the lock in EX mode, which means it
also cannot do journal recovery because it doesn't think it's the
first node to mount the file system.

This patch checks if the mounter is a spectator, and if so, avoids
grabbing the mounted lock. This allows a secondary mounter who is
really the first non-spectator mounter, to do journal recovery:
since the spectator doesn't acquire the lock, it can grab it in
EX mode, and therefore consider itself to be the first mounter
both as a "real" first mount, and as a first-real-after-spectator.

Note that the control lock still needs to be taken in PR mode
in order to fetch the lvb value so it has the current status of
all journal's recovery. This is used as it is today by a first
mounter to replay the journals. For spectators, it's merely
used to fetch the status bits. All recovery is bypassed and the
node waits until recovery is completed by a non-spectator node.

I also improved the cryptic message given by control_mount when
a spectator is waiting for a non-spectator to perform recovery.

It also fixes a problem in gfs2_recover_set whereby spectators
were never queueing recovery work for their own journal.
They cannot do recovery themselves, but they still need to queue
the work so they can check the recovery bits and clear the
DFL_BLOCK_LOCKS bit once the recovery happens on another node.

When the work queue runs on a spectator, it bypasses most of the
work so it won't print a bunch of annoying messages. All it will
print is a bunch of messages that look like this until recovery
completes on the non-spectator node:

GFS2: fsid=mycluster:scratch.s: recover generation 3 jid 0
GFS2: fsid=mycluster:scratch.s: recover jid 0 result busy

These continue every 1.5 seconds until the recovery is done by
the non-spectator, at which time it says:

GFS2: fsid=mycluster:scratch.s: recover generation 4 done

Then it proceeds with its mount.

If the file system is mounted in spectator node and the last
remaining non-spectator is fenced, any IO to the file system is
blocked by dlm and the spectator waits until recovery is
performed by a non-spectator.

If a spectator tries to mount the file system before any
non-spectators, it blocks and repeatedly gives this kernel
message:

GFS2: fsid=mycluster:scratch: Recovery is required. Waiting for a non-spectator to mount.
GFS2: fsid=mycluster:scratch: Recovery is required. Waiting for a non-spectator to mount.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>

fs/gfs2/lock_dlm.c

@@ -821,6 +821,13 @@ static int control_mount(struct gfs2_sbd *sdp)
 		goto fail;
 	}
 
+	/**
+	 * If we're a spectator, we don't want to take the lock in EX because
+	 * we cannot do the first-mount responsibility it implies: recovery.
+	 */
+	if (sdp->sd_args.ar_spectator)
+		goto locks_done;
+
 	error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
 	if (!error) {
 		mounted_mode = DLM_LOCK_EX;
@@ -896,9 +903,16 @@ static int control_mount(struct gfs2_sbd *sdp)
 	if (lvb_gen < mount_gen) {
 		/* wait for mounted nodes to update control_lock lvb to our
 		   generation, which might include new recovery bits set */
-		fs_info(sdp, "control_mount wait1 block %u start %u mount %u "
-			"lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
-			lvb_gen, ls->ls_recover_flags);
+		if (sdp->sd_args.ar_spectator) {
+			fs_info(sdp, "Recovery is required. Waiting for a "
+				"non-spectator to mount.\n");
+			msleep_interruptible(1000);
+		} else {
+			fs_info(sdp, "control_mount wait1 block %u start %u "
+				"mount %u lvb %u flags %lx\n", block_gen,
+				start_gen, mount_gen, lvb_gen,
+				ls->ls_recover_flags);
+		}
 		spin_unlock(&ls->ls_recover_spin);
 		goto restart;
 	}

fs/gfs2/recovery.c

@@ -413,12 +413,13 @@ void gfs2_recover_func(struct work_struct *work)
 	ktime_t t_start, t_jlck, t_jhd, t_tlck, t_rep;
 	int ro = 0;
 	unsigned int pass;
-	int error;
+	int error = 0;
 	int jlocked = 0;
 
 	t_start = ktime_get();
-	if (sdp->sd_args.ar_spectator ||
-	    (jd->jd_jid != sdp->sd_lockstruct.ls_jid)) {
+	if (sdp->sd_args.ar_spectator)
+		goto fail;
+	if (jd->jd_jid != sdp->sd_lockstruct.ls_jid) {
 		fs_info(sdp, "jid=%u: Trying to acquire journal lock...\n",
 			jd->jd_jid);
 		jlocked = 1;

fs/gfs2/sys.c

@@ -429,11 +429,18 @@ int gfs2_recover_set(struct gfs2_sbd *sdp, unsigned jid)
 
 	spin_lock(&sdp->sd_jindex_spin);
 	rv = -EBUSY;
-	if (sdp->sd_jdesc->jd_jid == jid)
+	/**
+	 * If we're a spectator, we use journal0, but it's not really ours.
+	 * So we need to wait for its recovery too. If we skip it we'd never
+	 * queue work to the recovery workqueue, and so its completion would
+	 * never clear the DFL_BLOCK_LOCKS flag, so all our locks would
+	 * permanently stop working.
+	 */
+	if (sdp->sd_jdesc->jd_jid == jid && !sdp->sd_args.ar_spectator)
 		goto out;
 	rv = -ENOENT;
 	list_for_each_entry(jd, &sdp->sd_jindex_list, jd_list) {
-		if (jd->jd_jid != jid)
+		if (jd->jd_jid != jid && !sdp->sd_args.ar_spectator)
 			continue;
 		rv = gfs2_recover_journal(jd, false);
 		break;