Re: [OOPS] 2.6.11 - NMI lockup with CFQ scheduler

[Jens Axboe]

On Wed, Apr 06 2005, James Bottomley wrote:
> On Wed, 2005-04-06 at 19:58 +0200, Jens Axboe wrote:
> > I rather like the queue lock being a pointer, so you can share at
> > whatever level you want. Lets not grow the request_queue a full lock
> > just to work around a bug elsewhere.
> 
> I'm not proposing that it not be a pointer, merely that it could be
> intialiased to point to a lock structure within the request queue.

Sorry that's what I meant, you are adding a lock inside the queue. I
didn't mean removing q->queue_lock or making it a non-pointer.

> Doing this looks much simpler than your current patch ... one of the
> problems with which looks to be that removing the scsi_driver module is
> in trouble because we currently have the queue_release in the sdev
> release (which won't get called while the queue holds a reference).

It is simpler, but it only solves this particular problem of the lock
going away.

> I think the correct model for all of this is that the block driver
> shouldn't care (or be tied to) the scsi one.  Thus, as long as SCSI can
> reject requests from a queue whose device has been released (without
> checking the device) then everything is fine as long as we sort out the
> lock lifetime problem.

But you are tying it completely to the SCSI problem, since we have no
locking problems of this sort elsewhere. At least the notified could
potentially be used for something else. The lock is just hack number two
to work around the problem.

-- 
Jens Axboe

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/