Re: [PATCH v10 2/2] kvm: Add resampling irqfds for level triggered interrupts

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

On Wed, 2012-09-19 at 21:48 +0300, Michael S. Tsirkin wrote:
> On Wed, Sep 19, 2012 at 12:23:13PM -0600, Alex Williamson wrote:
> > On Wed, 2012-09-19 at 11:59 +0300, Avi Kivity wrote:
> > > On 09/18/2012 06:16 AM, Alex Williamson wrote:
> > > > @@ -92,6 +156,43 @@ irqfd_shutdown(struct work_struct *work)
> > > >  	 */
> > > >  	flush_work_sync(&irqfd->inject);
> > > >  
> > > > +	if (irqfd->resampler) {
> > > > +		struct _irqfd_resampler *resampler = irqfd->resampler;
> > > > +		struct kvm *kvm = resampler->kvm;
> > > > +
> > > > +		mutex_lock(&kvm->irq_lock);
> > > > +		spin_lock_irq(&irqfd->kvm->irqfds.lock);
> > > > +
> > > > +		list_del_rcu(&irqfd->resampler_list);
> > > > +
> > > > +		/*
> > > > +		 * On removal of the last irqfd in the resampler list,
> > > > +		 * remove the resampler and unregister the irq ack
> > > > +		 * notifier.  It's possible to race the ack of the final
> > > > +		 * injection here, so manually de-assert the gsi to avoid
> > > > +		 * leaving an unmanaged, asserted interrupt line.
> > > > +		 */
> > > > +		if (list_empty(&resampler->irqfds)) {
> > > > +			list_del(&resampler->list);
> > > > +			__kvm_unregister_irq_ack_notifier(kvm,
> > > > +							  &resampler->notifier);
> > > > +			kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
> > > > +				    resampler->notifier.gsi, 0);
> > > > +			kfree(resampler);
> > > 
> > > Is this rcu safe?
> > 
> > No it's not and unfortunately this also points out another race in
> > trying to use a single source ID...
> > 
> > > > +		}
> > > > +
> > > > +		spin_unlock_irq(&irqfd->kvm->irqfds.lock);
> > > > +		mutex_unlock(&kvm->irq_lock);
> > > > +
> > > > +		/*
> > > > +		 * Both list_del_rcu & __kvm_unregister_irq_ack_notifier
> > > > +		 * require an rcu grace period/
> > > > +		 */
> > > > +		synchronize_rcu();
> > 
> > The kfree can't be done until here and we also have to assume that ack
> > notifies are firing until here.  That means that between the
> > mutex_unlock and the end of synchronize_rcu another resampling irqfd can
> > be registered, post an interrupt, and have it de-asserted by the wrong
> > resampler.  Maybe the conversion wasn't as clean as I first thought :(
> > > Quite ugly to expose the internals this way.
> > 
> > Yep.  I don't know how to clean it up though; between all the different
> > rcu operations and locks, it's a mess.  Thanks,
> > 
> > Alex
> 
> Add another mutex for the resamplers, keep it during the whole
> operation? This also removes the need for exposing the internals.
> If you do pls document lock nesting rules.

How does that hide the internals?  Seems like we'd just wrap this in yet
another mutex, but be largely the same.  Thanks,

Alex