Re: [PATCH] s390/vfio-ap: do not open code locks for VFIO_GROUP_NOTIFY_SET_KVM notification

[Tony Krowiak]

On 7/15/21 9:44 AM, Halil Pasic wrote:
> On Wed,  7 Jul 2021 11:41:56 -0400
> Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote:
>
> First sorry for being this late with having a more serious look at the
> code.
>
>
> > @@ -270,6 +270,9 @@ static struct ap_queue_status vfio_ap_irq_enable(struct vfio_ap_queue *q,
> >    * We take the matrix_dev lock to ensure serialization on queues and
> >    * mediated device access.
> >    *
> > + * Note: This function must be called with a read lock held on
> > + *	 vcpu->kvm->arch.crypto.pqap_hook_rwsem.
> > + *
>
> That is a fine synchronization for the pqap_hook, but I don't think it
> is sufficient for everything.
>
>
> >    * Return 0 if we could handle the request inside KVM.
> >    * otherwise, returns -EOPNOTSUPP to let QEMU handle the fault.
> >    */
> > @@ -287,22 +290,12 @@ static int handle_pqap(struct kvm_vcpu *vcpu)
> >   		return -EOPNOTSUPP;
> >
> >   	apqn = vcpu->run->s.regs.gprs[0] & 0xffff;
> > -	mutex_lock(&matrix_dev->lock);
> Here you drop a matrix_dev->lock critical section. And then
> you do all the interesting stuff. E.g.
> q = vfio_ap_get_queue(matrix_mdev, apqn);
> and
> vfio_ap_irq_enable(q, status & 0x07, vcpu->run->s.regs.gprs[2]);.
> Since in vfio_ap_get_queue() we do the check if the queue belongs
> to the given guest, and examine the matrix (apm, aqm) I suppose
> that needs to be done holding a lock that protects the matrix,
> and to my best knowledge this is still matrix_dev->lock. It would
> probably make sense to convert matrix_dev->lock into an rw_semaphore,
> or to introduce a some new rwlock which protects less state in the
> future, but right now AFAICT it is still matrix_dev->lock.
>
> So I don't think this patch should pass review.

Good catch. In an earlier patch review, Jason G suggested locking the 
kvm->lock
mutex outside of the kvm_arch_crypto_set_masks() and 
kvm_arch_crypto_clear_masks()
functions to resolve the lockdep splat resulting from locking the 
matrix_dev->lock
mutex prior to the kvm->lock mutex. I believe this will allow me to remove
the kvm_busy/wait queue scenario without introducing a new rwsem.

> Regards,
> Halil
>
> >   	if (!vcpu->kvm->arch.crypto.pqap_hook)
> >   		goto out_unlock;
> >   	matrix_mdev = container_of(vcpu->kvm->arch.crypto.pqap_hook,
> >   				   struct ap_matrix_mdev, pqap_hook);
> >
> > -	/*
> > -	 * If the KVM pointer is in the process of being set, wait until the
> > -	 * process has completed.
> > -	 */
> > -	wait_event_cmd(matrix_mdev->wait_for_kvm,
> > -		       !matrix_mdev->kvm_busy,
> > -		       mutex_unlock(&matrix_dev->lock),
> > -		       mutex_lock(&matrix_dev->lock));
> > -
> >   	/* If the there is no guest using the mdev, there is nothing to do */
> >   	if (!matrix_mdev->kvm)
> >   		goto out_unlock;