block: detach bdev inode from its wb in __blkdev_put()

Since 52ebea749aae ("writeback: make backing_dev_info host
cgroup-specific bdi_writebacks") inode, at some point in its lifetime,
gets attached to a wb (struct bdi_writeback).  Detaching happens on
evict, in inode_detach_wb() called from __destroy_inode(), and involves
updating wb.

However, detaching an internal bdev inode from its wb in
__destroy_inode() is too late.  Its bdi and by extension root wb are
embedded into struct request_queue, which has different lifetime rules
and can be freed long before the final bdput() is called (can be from
__fput() of a corresponding /dev inode, through dput() - evict() -
bd_forget().  bdevs hold onto the underlying disk/queue pair only while
opened; as soon as bdev is closed all bets are off.  In fact,
disk/queue can be gone before __blkdev_put() even returns:

1499 static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part)
1500 {
...
1518         if (bdev->bd_contains == bdev) {
1519                 if (disk->fops->release)
1520                         disk->fops->release(disk, mode);

[ Driver puts its references to disk/queue ]

1521         }
1522         if (!bdev->bd_openers) {
1523                 struct module *owner = disk->fops->owner;
1524
1525                 disk_put_part(bdev->bd_part);
1526                 bdev->bd_part = NULL;
1527                 bdev->bd_disk = NULL;
1528                 if (bdev != bdev->bd_contains)
1529                         victim = bdev->bd_contains;
1530                 bdev->bd_contains = NULL;
1531
1532                 put_disk(disk);

[ We put ours, the queue is gone
  The last bdput() would result in a write to invalid memory ]

1533                 module_put(owner);
...
1539 }

Since bdev inodes are special anyway, detach them in __blkdev_put()
after clearing inode's dirty bits, turning the problematic
inode_detach_wb() in __destroy_inode() into a noop.

add_disk() grabs its disk->queue since 523e1d399ce0 ("block: make
gendisk hold a reference to its queue"), so the old ->release comment
is removed in favor of the new inode_detach_wb() comment.

Cc: stable@vger.kernel.org # 4.2+, needs backporting
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Acked-by: Tejun Heo <tj@kernel.org>
Tested-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Signed-off-by: Jens Axboe <axboe@fb.com>

diff --git a/fs/block_dev.c b/fs/block_dev.c
index c25639e907bd21a194aecedc6c60a93fb4994375..44d4a1e9244e74e923d7b26018d4b420aa2e1323 100644 (file)
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -1523,11 +1523,14 @@ static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part)
                WARN_ON_ONCE(bdev->bd_holders);
                sync_blockdev(bdev);
                kill_bdev(bdev);
+
+               bdev_write_inode(bdev);
                /*
-                * ->release can cause the queue to disappear, so flush all
-                * dirty data before.
+                * Detaching bdev inode from its wb in __destroy_inode()
+                * is too late: the queue which embeds its bdi (along with
+                * root wb) can be gone as soon as we put_disk() below.
                 */
-               bdev_write_inode(bdev);
+               inode_detach_wb(bdev->bd_inode);
        }
        if (bdev->bd_contains == bdev) {
                if (disk->fops->release)