[patch] cciss dequeue race fix (was Re: preempt-patch cleared of blame)

[Jens Axboe <axboe@suse.de>]

[-- Attachment #1: Type: text/plain, Size: 630 bytes --]

On Thu, Nov 08 2001, Jens Axboe wrote:
> On Thu, Nov 08 2001, Ingo Molnar wrote:
> > 
> > On Thu, 8 Nov 2001, Jens Axboe wrote:
> > 
> > > On Wed, Nov 07 2001, J Sloan wrote:
> > > > I think there may be a problem with the
> > > > compaq smart/2p raid drivers, since
> > > > the "do_ida_intr" code keeps showing
> > > > up in the oops, and I have not seen a
> > > > problem with 2.4.14 on any other system.
> > >
> > > Does this fix it?
> > 
> > it did the trick on my system, which used to oops in/around do_ida_intr as
> > well.
> 
> Good, so it was a dequeue race. Thanks Ingo.

BTW, cciss has a similar race.

-- 
Jens Axboe


[-- Attachment #2: cciss-dequeue-1 --]
[-- Type: text/plain, Size: 625 bytes --]

--- linux/drivers/block/cciss.c~	Thu Nov  8 11:36:24 2001
+++ linux/drivers/block/cciss.c	Thu Nov  8 11:37:03 2001
@@ -1307,6 +1307,8 @@
 	if (( c = cmd_alloc(h, 1)) == NULL)
 		goto startio;
 
+	blkdev_dequeue_request(creq);
+
 	spin_unlock_irq(&io_request_lock);
 
 	c->cmd_type = CMD_RWREQ;      
@@ -1386,12 +1388,6 @@
 
 	spin_lock_irq(&io_request_lock);
 
-	blkdev_dequeue_request(creq);
-
-        /*
-         * ehh, we can't really end the request here since it's not
-         * even started yet. for now it shouldn't hurt though
-         */
 	addQ(&(h->reqQ),c);
 	h->Qdepth++;
 	if(h->Qdepth > h->maxQsinceinit)