umount: avoid racy block-size set

Setting the block size back to the original blocksize at umount time
is unnecessary, and can cause buffer errors and infinite loops in
__getblk_slow, and possibly other problems.

Just getting bdev_lock instead causes other issues, namely AB-BA deadlocks
with /proc/meminfo etc. Just drop it, and replace with a "sync_blockdev()"
to cause a flush (which was a side effect of changing the blocksize before).

Debugging and testing by Jeff Mahoney <jeffm@suse.com>

diff --git a/fs/super.c b/fs/super.c
index c5b48939c9a86b1ceec228471aab3910571e6bcd..3a1b8ca04ba601b37ffa5dadb8f9695272952e39 100644 (file)
--- a/fs/super.c
+++ b/fs/super.c
@@ -732,7 +732,7 @@ void kill_block_super(struct super_block *sb)
 
        bdev_uevent(bdev, KOBJ_UMOUNT);
        generic_shutdown_super(sb);
-       set_blocksize(bdev, sb->s_old_blocksize);
+       sync_blockdev(bdev);
        close_bdev_excl(bdev);
 }