[PATCH] dm-multipath: fix stall on noflush suspend/resume

Allow noflush suspend/resume of device-mapper device only for the case
where the device size is unchanged.

Otherwise, dm-multipath devices can stall when resumed if noflush was used
when suspending them, all paths have failed and queue_if_no_path is set.

Explanation:
 1. Something is doing fsync() on the block dev,
    holding inode->i_sem
 2. The fsync write is blocked by all-paths-down and queue_if_no_path
 3. Someone requests to suspend the dm device with noflush.
    Pending writes are left in queue.
 4. In the middle of dm_resume(), __bind() tries to get
    inode->i_sem to do __set_size() and waits forever.

'noflush suspend' is a new device-mapper feature introduced in
early 2.6.20. So I hope the fix being included before 2.6.20 is
released.

Example of reproducer:
 1. Create a multipath device by dmsetup
 2. Fail all paths during mkfs
 3. Do dmsetup suspend --noflush and load new map with healthy paths
 4. Do dmsetup resume

Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Acked-by: Alasdair G Kergon <agk@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 19 insertions, 8 deletions

diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index fe7c56e10435..3668b170ea68 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -1116,7 +1116,8 @@ static int __bind(struct mapped_device *md, struct dm_table *t)
         if (size != get_capacity(md->disk))
                 memset(&md->geometry, 0, sizeof(md->geometry));
 
-        __set_size(md, size);
+        if (md->suspended_bdev)
+                __set_size(md, size);
         if (size == 0)
                 return 0;
 
@@ -1264,6 +1265,11 @@ int dm_swap_table(struct mapped_device *md, struct dm_table *table)
         if (!dm_suspended(md))
                 goto out;
 
+        /* without bdev, the device size cannot be changed */
+        if (!md->suspended_bdev)
+                if (get_capacity(md->disk) != dm_table_get_size(table))
+                        goto out;
+
         __unbind(md);
         r = __bind(md, table);
 
@@ -1341,11 +1347,14 @@ int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
         /* This does not get reverted if there's an error later. */
         dm_table_presuspend_targets(map);
 
-        md->suspended_bdev = bdget_disk(md->disk, 0);
-        if (!md->suspended_bdev) {
-                DMWARN("bdget failed in dm_suspend");
-                r = -ENOMEM;
-                goto flush_and_out;
+        /* bdget() can stall if the pending I/Os are not flushed */
+        if (!noflush) {
+                md->suspended_bdev = bdget_disk(md->disk, 0);
+                if (!md->suspended_bdev) {
+                        DMWARN("bdget failed in dm_suspend");
+                        r = -ENOMEM;
+                        goto flush_and_out;
+                }
         }
 
         /*
@@ -1473,8 +1482,10 @@ int dm_resume(struct mapped_device *md)
 
         unlock_fs(md);
 
-        bdput(md->suspended_bdev);
-        md->suspended_bdev = NULL;
+        if (md->suspended_bdev) {
+                bdput(md->suspended_bdev);
+                md->suspended_bdev = NULL;
+        }
 
         clear_bit(DMF_SUSPENDED, &md->flags);