ping_manager() to prevent 'add' before 'remove' completes

It is currently possible to remove a device and re-add it without the
manager noticing, i.e. without detecting a mdstat->devcnt
container->devcnt mismatch.  Introduce ping_manager() to arrange for
mdmon to run manage_container() prior to mdadm dropping the exclusive
open() on the container.  Despite these precautions sysfs_read() may
still fail.  If this happens invalidate container->devcnt to ensure
manage_container() runs at the next event.

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

4 files changed, 60 insertions, 8 deletions

diff --git a/Manage.c b/Manage.c
index 76447ed..5853515 100644
--- a/Manage.c
+++ b/Manage.c
@@ -712,6 +712,23 @@ int Manage_subdevs(char *devname, int fd,
 					close(lfd);
 				return 1;
 			}
+			if (tst->ss->external) {
+				/*
+				 * Before dropping our exclusive open we make an
+				 * attempt at preventing mdmon from seeing an
+				 * 'add' event before reconciling this 'remove'
+				 * event.
+				 */
+				char *name = devnum2devname(fd2devnum(fd));
+
+				if (!name) {
+					fprintf(stderr, Name ": unable to get container name\n");
+					return 1;
+				}
+
+				ping_manager(name);
+				free(name);
+			}
 			close(lfd);
 			if (verbose >= 0)
 				fprintf(stderr, Name ": hot removed %s\n",
diff --git a/managemon.c b/managemon.c
index fc34d77..fc5da71 100644
--- a/managemon.c
+++ b/managemon.c
@@ -265,8 +265,11 @@ static void manage_container(struct mdstat_ent *mdstat,
 		 * These need to be remove from, or added to, the array
 		 */
 		mdi = sysfs_read(-1, mdstat->devnum, GET_DEVS);
-		if (!mdi)
+		if (!mdi) {
+			/* invalidate the current count so we can try again */
+			container->devcnt = -1;
 			return;
+		}
 
 		/* check for removals */
 		for (cdp = &container->devs; *cdp; ) {
@@ -525,14 +528,15 @@ static void handle_message(struct supertype *container, struct metadata_update *
 
 	struct metadata_update *mu;
 
-	if (msg->len == 0) {
-		int cnt;
-		
+	if (msg->len <= 0)
 		while (update_queue_pending || update_queue) {
 			check_update_queue(container);
 			usleep(15*1000);
 		}
 
+	if (msg->len == 0) { /* ping_monitor */
+		int cnt;
+		
 		cnt = monitor_loop_cnt;
 		if (cnt & 1)
 			cnt += 2; /* wait until next pselect */
@@ -542,6 +546,11 @@ static void handle_message(struct supertype *container, struct metadata_update *
 
 		while (monitor_loop_cnt - cnt < 0)
 			usleep(10 * 1000);
+	} else if (msg->len == -1) { /* ping_manager */
+		struct mdstat_ent *mdstat = mdstat_read(1, 0);
+
+		manage(mdstat, container);
+		free_mdstat(mdstat);
 	} else {
 		mu = malloc(sizeof(*mu));
 		mu->len = msg->len;
diff --git a/msg.c b/msg.c
index 013bcb9..5a4839f 100644
--- a/msg.c
+++ b/msg.c
@@ -81,12 +81,12 @@ static int recv_buf(int fd, void* buf, int len, int tmo)
 
 int send_message(int fd, struct metadata_update *msg, int tmo)
 {
-	__u32 len = msg->len;
+	__s32 len = msg->len;
 	int rv;
 
 	rv = send_buf(fd, &start_magic, 4, tmo);
 	rv = rv ?: send_buf(fd, &len, 4, tmo);
-	if (len)
+	if (len > 0)
 		rv = rv ?: send_buf(fd, msg->buf, msg->len, tmo);
 	rv = send_buf(fd, &end_magic, 4, tmo);
 
@@ -96,7 +96,7 @@ int send_message(int fd, struct metadata_update *msg, int tmo)
 int receive_message(int fd, struct metadata_update *msg, int tmo)
 {
 	__u32 magic;
-	__u32 len;
+	__s32 len;
 	int rv;
 
 	rv = recv_buf(fd, &magic, 4, tmo);
@@ -105,7 +105,7 @@ int receive_message(int fd, struct metadata_update *msg, int tmo)
 	rv = recv_buf(fd, &len, 4, tmo);
 	if (rv < 0 || len > MSG_MAX_LEN)
 		return -1;
-	if (len) {
+	if (len > 0) {
 		msg->buf = malloc(len);
 		if (msg->buf == NULL)
 			return -1;
@@ -177,6 +177,7 @@ int connect_monitor(char *devname)
 	return sfd;
 }
 
+/* give the monitor a chance to update the metadata */
 int ping_monitor(char *devname)
 {
 	int sfd = connect_monitor(devname);
@@ -196,3 +197,27 @@ int ping_monitor(char *devname)
 	close(sfd);
 	return err;
 }
+
+/* give the manager a chance to view the updated container state.  This
+ * would naturally happen due to the manager noticing a change in
+ * /proc/mdstat; however, pinging encourages this detection to happen
+ * while an exclusive open() on the container is active
+ */
+int ping_manager(char *devname)
+{
+	int sfd = connect_monitor(devname);
+	struct metadata_update msg = { .len = -1 };
+	int err = 0;
+
+	if (sfd < 0)
+		return sfd;
+
+	err = send_message(sfd, &msg, 20);
+
+	/* check the reply */
+	if (!err && wait_reply(sfd, 20) != 0)
+		err = -1;
+
+	close(sfd);
+	return err;
+}
diff --git a/msg.h b/msg.h
index 4dc805e..b9bd205 100644
--- a/msg.h
+++ b/msg.h
@@ -27,5 +27,6 @@ extern int ack(int fd, int tmo);
 extern int wait_reply(int fd, int tmo);
 extern int connect_monitor(char *devname);
 extern int ping_monitor(char *devname);
+extern int ping_manager(char *devname);
 
 #define MSG_MAX_LEN (4*1024*1024)