Re: [patch-mm] Workaround for RAID breakage

[Jens Axboe <jens.axboe@oracle.com>]

On Tue, Jan 16 2007, Thomas Gleixner wrote:
> On Tue, 2007-01-16 at 11:41 +1100, Jens Axboe wrote:
> > > AFAICS this is intentional to avoid checks all over the place, but the
> > > underflow check is missing. All we need to do is make sure, that in case
> > > of ioc->plugged == 0 we return early and bug, if there is either a queue
> > > plugged in or the plugged_list is not empty.
> > > 
> > > Jens ?
> > 
> > It should not go negative, that would be a bug elsewhere. So it's
> > interesting if it does, we should definitely put a WARN_ON() check in
> > there for that.
> 
> Well. All offenders come via queue_sync_plugs(). queue_sync_plugs()
> calls blk_unplug_current().

Ah, again a problem because the #plug branch wasn't pushed to #for-akpm
in due time. That problem was fixed already :/

> One path which triggers is blk_sync_queue(). This happens e.g. in the
> cleanup of the floppy check. There are other call pathes too.
> 
> The other is raid md_super_write(). It is not plugged and calls with the
> barrier bit set, which executes the unlikely path in __make_request():
> 
>         if (unlikely(bio_barrier(bio))) {
>                 queue_sync_plugs(q);
>                 spin_lock_irq(q->queue_lock);
>                 goto get_rq;
>         }
> 
> So you either need checks all over the place to avoid calling
> blk_unplug_current(), or you prevent the unplug below 0 like I did. The
> BUG_ON()s I added should catch any real invalid callers. But it's up to
> you.

blk_replug_current_nested() should do it. I'll make sure the branch is
updated for next -mm.

The BUG_ON()'s are indeed correct, they should just be moved further
down the function.

-- 
Jens Axboe