Re: Panic when cpu hot-remove

[Alex Williamson <alex.williamson@redhat.com>]

On Wed, 2015-06-17 at 13:52 +0200, Joerg Roedeljoro wrote:
> On Wed, Jun 17, 2015 at 10:42:49AM +0000, 范冬冬 wrote:
> > Hi maintainer,
> > 
> > We found a problem that a panic happen when cpu was hot-removed. We also trace the problem according to the calltrace information.
> > An endless loop happen because value head is not equal to value tail forever in the function qi_check_fault( ).
> > The location code is as follows:
> > 
> > 
> > do {
> >         if (qi->desc_status[head] == QI_IN_USE)
> >         qi->desc_status[head] = QI_ABORT;
> >         head = (head - 2 + QI_LENGTH) % QI_LENGTH;
> >     } while (head != tail);
> 
> Hmm, this code interates only over every second QI descriptor, and tail
> probably points to a descriptor that is not iterated over.
> 
> Jiang, can you please have a look?

I think that part is normal, the way we use the queue is to always
submit a work operation followed by a wait operation so that we can
determine the work operation is complete.  That's done via
qi_submit_sync().  We have had spurious reports of the queue getting
impossibly out of sync though.  I saw one that was somehow linked to the
I/O AT DMA engine.  Roland Dreier saw something similar[1].  I'm not
sure if they're related to this, but maybe worth comparing.  Thanks,

Alex

[1] http://lists.linuxfoundation.org/pipermail/iommu/2015-January/011502.html