mdmon: periodically checkpoint recovery

The kernel updates and notifies md/sync_completed when it is time to
take a checkpoint.  When this occurs (at 1/16 array size intervals)
write 'idle' to md/sync_action to have the current recovery position
updated in recovery_start and resync_start.

Requires the metadata handler to reset ->last_checkpoint when it has
determined that recovery has ended.

Signed-off-by: Dan Williams <dan.j.williams@intel.com>

managemon.c

@@ -540,6 +540,7 @@ static void manage_new(struct mdstat_ent *mdstat,
 	new->info.state_fd = sysfs_open(new->devnum, NULL, "array_state");
 	new->resync_start_fd = sysfs_open(new->devnum, NULL, "resync_start");
 	new->metadata_fd = sysfs_open(new->devnum, NULL, "metadata_version");
+	new->sync_completed_fd = sysfs_open(new->devnum, NULL, "sync_completed");
 	dprintf("%s: inst: %d action: %d state: %d\n", __func__, atoi(inst),
 		new->action_fd, new->info.state_fd);
 

mdmon.h

@@ -32,6 +32,15 @@ struct active_array {
 	int action_fd;
 	int resync_start_fd;
 	int metadata_fd; /* for monitoring rw/ro status */
+	int sync_completed_fd; /* for checkpoint notification events */
+	unsigned long long last_checkpoint; /* sync_completed fires for many
+					     * reasons this field makes sure the
+					     * kernel has made progress before
+					     * moving the checkpoint.  It is
+					     * cleared by the metadata handler
+					     * when it determines recovery is
+					     * terminated.
+					     */
 
 	enum array_state prev_state, curr_state, next_state;
 	enum sync_action prev_action, curr_action, next_action;

monitor.c

@@ -80,6 +80,24 @@ static unsigned long long read_resync_start(int fd)
 		return strtoull(buf, NULL, 10);
 }
 
+static unsigned long long read_sync_completed(int fd)
+{
+	unsigned long long val;
+	char buf[50];
+	int n;
+	char *ep;
+
+	n = read_attr(buf, 50, fd);
+
+	if (n <= 0)
+		return 0;
+	buf[n] = 0;
+	val = strtoull(buf, &ep, 0);
+	if (ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))
+		return 0;
+	return val;
+}
+
 static enum array_state read_state(int fd)
 {
 	char buf[20];
@@ -195,6 +213,7 @@ static void signal_manager(void)
 
 static int read_and_act(struct active_array *a)
 {
+	unsigned long long sync_completed;
 	int check_degraded = 0;
 	int deactivate = 0;
 	struct mdinfo *mdi;
@@ -206,6 +225,7 @@ static int read_and_act(struct active_array *a)
 	a->curr_state = read_state(a->info.state_fd);
 	a->curr_action = read_action(a->action_fd);
 	a->info.resync_start = read_resync_start(a->resync_start_fd);
+	sync_completed = read_sync_completed(a->sync_completed_fd);
 	for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 		mdi->next_state = 0;
 		if (mdi->state_fd >= 0) {
@@ -307,6 +327,18 @@ static int read_and_act(struct active_array *a)
 		}
 	}
 
+	/* Check for recovery checkpoint notifications.  We need to be a
+	 * minimum distance away from the last checkpoint to prevent
+	 * over checkpointing.  Note reshape checkpointing is not
+	 * handled here.
+	 */
+	if (sync_completed > a->last_checkpoint &&
+	    sync_completed - a->last_checkpoint > a->info.component_size >> 4 &&
+	    a->curr_action > reshape && a->next_action == bad_action) {
+		a->last_checkpoint = sync_completed;
+		a->next_action = idle;
+	}
+
 	a->container->ss->sync_metadata(a->container);
 	dprintf("%s(%d): state:%s action:%s next(", __func__, a->info.container_member,
 		array_states[a->curr_state], sync_actions[a->curr_action]);
@@ -461,6 +493,7 @@ static int wait_and_act(struct supertype *container, int nowait)
 
 		add_fd(&rfds, &maxfd, a->info.state_fd);
 		add_fd(&rfds, &maxfd, a->action_fd);
+		add_fd(&rfds, &maxfd, a->sync_completed_fd);
 		for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 			add_fd(&rfds, &maxfd, mdi->state_fd);
 

super-intel.c

@@ -4384,6 +4384,7 @@ static int imsm_set_array_state(struct active_array *a, int consistent)
 			dprintf("imsm: mark resync done\n");
 			end_migration(dev, map_state);
 			super->updates_pending++;
+			a->last_checkpoint = 0;
 		}
 	} else if (!is_resyncing(dev) && !failed) {
 		/* mark the start of the init process if nothing is failed */
@@ -4476,17 +4477,20 @@ static void imsm_set_disk(struct active_array *a, int n, int state)
 		map = get_imsm_map(dev, 0);
 		map->failed_disk_num = ~0;
 		super->updates_pending++;
+		a->last_checkpoint = 0;
 	} else if (map_state == IMSM_T_STATE_DEGRADED &&
 		   map->map_state != map_state &&
 		   !dev->vol.migr_state) {
 		dprintf("imsm: mark degraded\n");
 		map->map_state = map_state;
 		super->updates_pending++;
+		a->last_checkpoint = 0;
 	} else if (map_state == IMSM_T_STATE_FAILED &&
 		   map->map_state != map_state) {
 		dprintf("imsm: mark failed\n");
 		end_migration(dev, map_state);
 		super->updates_pending++;
+		a->last_checkpoint = 0;
 	}
 }