[PATCH] device-mapper multipath: Avoid possible suspension deadlock

To avoid deadlock when suspending a multipath device after all its paths have
failed, stop queueing any I/O that is about to fail *before* calling
freeze_bdev instead of after.

Instead of setting a multipath 'suspended' flag which would have to be reset
if an error occurs during the process, save the previous queueing state and
leave userspace to restore if it wishes.

Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

2 files changed, 20 insertions, 17 deletions

diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c
index fa72f0153206..98da8eee2d2c 100644
--- a/drivers/md/dm-mpath.c
+++ b/drivers/md/dm-mpath.c
@@ -72,7 +72,7 @@ struct multipath {
 
         unsigned queue_io;              /* Must we queue all I/O? */
         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
-        unsigned suspended;             /* Has dm core suspended our I/O? */
+        unsigned saved_queue_if_no_path;/* Saved state during suspension */
 
         struct work_struct process_queued_ios;
         struct bio_list queued_ios;
@@ -304,7 +304,7 @@ static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
                 m->queue_size--;
 
         if ((pgpath && m->queue_io) ||
-            (!pgpath && m->queue_if_no_path && !m->suspended)) {
+            (!pgpath && m->queue_if_no_path)) {
                 /* Queue for the daemon to resubmit */
                 bio_list_add(&m->queued_ios, bio);
                 m->queue_size++;
@@ -333,6 +333,7 @@ static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path)
 
         spin_lock_irqsave(&m->lock, flags);
 
+        m->saved_queue_if_no_path = m->queue_if_no_path;
         m->queue_if_no_path = queue_if_no_path;
         if (!m->queue_if_no_path)
                 queue_work(kmultipathd, &m->process_queued_ios);
@@ -391,7 +392,7 @@ static void process_queued_ios(void *data)
         pgpath = m->current_pgpath;
 
         if ((pgpath && m->queue_io) ||
-            (!pgpath && m->queue_if_no_path && !m->suspended))
+            (!pgpath && m->queue_if_no_path))
                 must_queue = 1;
 
         init_required = m->pg_init_required;
@@ -998,7 +999,7 @@ static int do_end_io(struct multipath *m, struct bio *bio,
 
         spin_lock(&m->lock);
         if (!m->nr_valid_paths) {
-                if (!m->queue_if_no_path || m->suspended) {
+                if (!m->queue_if_no_path) {
                         spin_unlock(&m->lock);
                         return -EIO;
                 } else {
@@ -1059,27 +1060,27 @@ static int multipath_end_io(struct dm_target *ti, struct bio *bio,
 
 /*
  * Suspend can't complete until all the I/O is processed so if
- * the last path failed we will now error any queued I/O.
+ * the last path fails we must error any remaining I/O.
+ * Note that if the freeze_bdev fails while suspending, the
+ * queue_if_no_path state is lost - userspace should reset it.
  */
 static void multipath_presuspend(struct dm_target *ti)
 {
         struct multipath *m = (struct multipath *) ti->private;
-        unsigned long flags;
 
-        spin_lock_irqsave(&m->lock, flags);
-        m->suspended = 1;
-        if (m->queue_if_no_path)
-                queue_work(kmultipathd, &m->process_queued_ios);
-        spin_unlock_irqrestore(&m->lock, flags);
+        queue_if_no_path(m, 0);
 }
 
+/*
+ * Restore the queue_if_no_path setting.
+ */
 static void multipath_resume(struct dm_target *ti)
 {
         struct multipath *m = (struct multipath *) ti->private;
         unsigned long flags;
 
         spin_lock_irqsave(&m->lock, flags);
-        m->suspended = 0;
+        m->queue_if_no_path = m->saved_queue_if_no_path;
         spin_unlock_irqrestore(&m->lock, flags);
 }
 
diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index 5d40555b42ba..bb3ad79c14d7 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -1055,14 +1055,17 @@ int dm_suspend(struct mapped_device *md)
         if (test_bit(DMF_BLOCK_IO, &md->flags))
                 goto out_read_unlock;
 
-        error = __lock_fs(md);
-        if (error)
-                goto out_read_unlock;
-
         map = dm_get_table(md);
         if (map)
+                /* This does not get reverted if there's an error later. */
                 dm_table_presuspend_targets(map);
 
+        error = __lock_fs(md);
+        if (error) {
+                dm_table_put(map);
+                goto out_read_unlock;
+        }
+
         up_read(&md->lock);
 
         /*
@@ -1121,7 +1124,6 @@ int dm_suspend(struct mapped_device *md)
         return 0;
 
 out_unfreeze:
-        /* FIXME Undo dm_table_presuspend_targets */
         __unlock_fs(md);
         clear_bit(DMF_BLOCK_IO, &md->flags);
 out_write_unlock: