On Sun 12-02-17 13:40:27, Tejun Heo wrote:

Hello, Jan.

On Thu, Feb 09, 2017 at 01:44:31PM +0100, Jan Kara wrote:

quoted

When block device is closed, we call inode_detach_wb() in __blkdev_put()
which sets inode->i_wb to NULL. That is contrary to expectations that
inode->i_wb stays valid once set during the whole inode's lifetime and
leads to oops in wb_get() in locked_inode_to_wb_and_lock_list() because
inode_to_wb() returned NULL.

The reason why we called inode_detach_wb() is not valid anymore though.
BDI is guaranteed to stay along until we call bdi_put() from
bdev_evict_inode() so we can postpone calling inode_detach_wb() to that
moment. A complication is that i_wb can point to non-root wb_writeback
structure and in that case we do need to clean it up as bdi_unregister()
blocks waiting for all non-root wb_writeback references to get dropped.
Thus this i_wb reference could block device removal e.g. from
__scsi_remove_device() (which indirectly ends up calling
bdi_unregister()). We cannot rely on block device inode to go away soon
(and thus i_wb reference to get dropped) as the device may got
hot-removed e.g. under a mounted filesystem. We deal with these issues
by switching block device inode from non-root wb_writeback structure to
bdi->wb when needed.  Since this is rather expensive (requires
synchronize_rcu()) we do the switching only in del_gendisk() when we
know the device is going away.

So, the only reason cgwb_bdi_destroy() is synchronous is because bdi
destruction was synchronous.  Now that bdi is properly reference
counted and can be decoupled from gendisk / q destruction, I can't
think of a reason to keep cgwb destruction synchronous.  Switching
wb's on destruction is kinda clumsy and it almost always hurts to
expose synchronize_rcu() in userland visible paths.

Wouldn't something like the following work?

* Remove bdi->usage_cnt and the synchronous waiting in
  cgwb_bdi_destroy().

* Instead, make cgwb's hold bdi->refcnt and put it from
  cgwb_release_workfn().

Then, we don't have to switch during shutdown and can just let things
drain.

At first sight this looks workable and would mean less special code so I
like it. I'll experiment with it and see how it works out.

								Honza
-- 
Jan Kara [off-list ref]
SUSE Labs, CR