lkml.org 
[lkml]   [2014]   [Mar]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [RFC PATCH] cifs: Fix possible deadlock with cifs and work queues
From
2014-03-21 6:23 GMT+04:00 Steven Rostedt <rostedt@goodmis.org>:
> On Thu, 20 Mar 2014 17:02:39 -0400
> Jeff Layton <jlayton@redhat.com> wrote:
>
>> Eventually the server should just allow the read to complete even if
>> the client doesn't respond to the oplock break. It has to since clients
>> can suddenly drop off the net while holding an oplock. That should
>> allow everything to unwedge eventually (though it may take a while).
>>
>> If that's not happening then I'd be curious as to why...
>
> The problem is that the data is being filled in the page and the reader
> is waiting for the page lock to be released. The kworker for the reader
> will issue the complete() and unlock the page to wake up the reader.
>
> But because the other workqueue callback calls down_read(), and there
> can be a down_write() waiting for the reader to finish, this
> down_read() will block on the lock as well (rwsems are fair locks).
> This blocks the other workqueue callback from issuing the complete and
> page_unlock() that will wake up the reader that is holding the rwsem
> with down_read().
>
> DEADLOCK.

Thank you for reporting and clarifying the issue!

Read and write codepaths both obtain lock_sem for read and then wait
for cifsiod_wq to complete and release lock_sem. They don't do any
lock_sem operations inside their work task queued to cifsiod_wq. But
oplock code can obtain/release lock_sem in its work task. So, that's
why I agree with Jeff and suggest to move the oplock code to a
different work queue (cifsioopd_wq?) but leave read and write
codepaths use cifsiod_wq.

--
Best regards,
Pavel Shilovsky.


\
 
 \ /
  Last update: 2014-03-21 10:21    [W:0.091 / U:0.124 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site