Re: [RFC][PATCH 04/16] writeback: fix possible race when shutting down bdi

[Artem Bityutskiy <dedekind1@gmail.com>]

On Sun, 2010-07-18 at 02:47 -0400, Christoph Hellwig wrote:
> On Fri, Jul 16, 2010 at 03:45:00PM +0300, Artem Bityutskiy wrote:
> > From: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
> > 
> > Current bdi code has the following race between 'bdi_wb_shutdown()'
> > and 'bdi_forker_thread()'.
> > 
> > Initial condition: BDI_pending is cleaned, bdi has no writeback thread,
> > because it was inactive and exited, 'bdi_wb_shutdown()' and
> > 'bdi_forker_thread()' are executed concurrently.
> 
> Wouldn't it be better to have a per-bdi mutex to serialize thread
> creation and shutdown?

There are several parties which want to have some serialization with bdi
trheads creation and shutdown:

1. 'bdi_queue_work()' - this should not take any mutex and should be
   fast. It uses spinlock and this is should stay this way

2. I'm going to modify '__mark_inode_dirty()' to wake-up bdi thread -
   this is similar to 'bdi_queue_work()'

3. 'bdi_wb_shutdown()' - this uses the 'BDI_pending' for serialization
   now, but can use a mutex instead.

I guess you mean that for 1 and 2 things stay the same, but for 3 we can
use a mutex. Then the forker thread should also take this mutex. Right?

If yes, this looks fine for me. I am going to try this approach. Then
-->

>   And please also kill the bit wait in favour
> of a proper wait queue - the bit wait interface really is just a hack
> for structures that are very size sensitive, which the backing device
> is not.

--> the bit should go away and so no wait queue will be needed as well. 

-- 
Best Regards,
Artem Bityutskiy (Артём Битюцкий)