Re: [virtio-dev] Re: [PATCH V2 2/2] virtio-pci: add PASID configuration extended capability

[Jean-Philippe Brucker <jean-philippe@linaro.org>]

On Fri, Jan 14, 2022 at 11:15:35AM +0800, Jason Wang wrote:
> 
> å 2022/1/13 äå11:17, Michael S. Tsirkin åé:
> > On Thu, Jan 13, 2022 at 02:53:36PM +0000, Stefan Hajnoczi wrote:
> > > On Thu, Jan 13, 2022 at 05:45:11AM -0500, Michael S. Tsirkin wrote:
> > > > On Thu, Jan 13, 2022 at 10:32:11AM +0000, Stefan Hajnoczi wrote:
> > > > > > So my understanding is that instead of delaying the response to read
> > > > > > it's better to simply present the previous PASID value if there're any
> > > > > > pending transactions of previous PASID value. The driver can then
> > > > > > choose to poll or using timers etc.
> > > > Don't we limit these changes to before DRIVER_OK?
> > > > If we do then no previous transactions can exist.
> > > Not sure if that's possible for the vDPA virtqueue assignment use case
> > > where one large VIRTIO device provides virtqueues for many smaller vDPA
> > > devices?
> 
> 
> That could be the case but usually a smaller vDPA device requires a bunch of
> hardware resources more than just virtqueues.
> 
> Another example is assign a queue directly to userspace with PASID for SVA.
> 
> 
> > > 
> > > Stefan
> > Then I guess queue must be stopped?
> 
> 
> Yes, since we have virtqueue reset, I can limit the assigning before
> DRIVER_OK or queue(s) are reset. Does this work?

Yes, I was wondering how this would work for the SVA use-case, where
processes can request queues from the virtio driver at runtime:

* Process opens a fd to the queue
* virtio driver selects a disabled queue, gets a PASID for that
  task and assigns it to the queue's group 
* process mmaps its queue and works with it. When it is done, it closes
  the fd.
* virtio driver now makes sure that the virtqueue does not issue any
  more DMA with this PASID, it disables all the queues in the group by
  writing 1 to their queue_reset. After reset completes, it clears the
  group_pasid field.
* The driver releases the PASID (iommu_sva_unbind_device()).
  Both PASID and queue can now safely be assigned to other processes.

So making group_pasid field writeable only if all queues in the group are
disabled should work

Thanks,
Jean