Re: [RFC 0/4] Virtio uses DMA API for all devices

[Michael S. Tsirkin]

On Tue, Aug 07, 2018 at 08:13:56AM +1000, Benjamin Herrenschmidt wrote:
> On Tue, 2018-08-07 at 00:46 +0300, Michael S. Tsirkin wrote:
> > On Tue, Aug 07, 2018 at 07:26:35AM +1000, Benjamin Herrenschmidt wrote:
> > > On Mon, 2018-08-06 at 23:35 +0300, Michael S. Tsirkin wrote:
> > > > > As I said replying to Christoph, we are "leaking" into the interface
> > > > > something here that is really what's the VM is doing to itself, which
> > > > > is to stash its memory away in an inaccessible place.
> > > > > 
> > > > > Cheers,
> > > > > Ben.
> > > > 
> > > > I think Christoph merely objects to the specific implementation.  If
> > > > instead you do something like tweak dev->bus_dma_mask for the virtio
> > > > device I think he won't object.
> > > 
> > > Well, we don't have "bus_dma_mask" yet ..or you mean dma_mask ?
> > > 
> > > So, something like that would be a possibility, but the problem is that
> > > the current virtio (guest side) implementation doesn't honor this when
> > > not using dma ops and will not use dma ops if not using iommu, so back
> > > to square one.
> > 
> > Well we have the RFC for that - the switch to using DMA ops unconditionally isn't
> > problematic itself IMHO, for now that RFC is blocked
> > by its perfromance overhead for now but Christoph says
> > he's trying to remove that for direct mappings,
> > so we should hopefully be able to get there in X weeks.
> 
> That would be good yes.
> 
>  ../..
> 
> > > --- a/drivers/virtio/virtio_ring.c
> > > +++ b/drivers/virtio/virtio_ring.c
> > > @@ -155,7 +155,7 @@ static bool vring_use_dma_api(struct virtio_device
> > > *vdev)
> > >          * the DMA API if we're a Xen guest, which at least allows
> > >          * all of the sensible Xen configurations to work correctly.
> > >          */
> > > -       if (xen_domain())
> > > +       if (xen_domain() || arch_virtio_direct_dma_ops(&vdev->dev))
> > >                 return true;
> > >  
> > >         return false;
> > 
> > Right but can't we fix the retpoline overhead such that
> > vring_use_dma_api will not be called on data path any longer, making
> > this a setup time check?
> 
> Yes it needs to be a setup time check regardless actually !
> 
> The above is broken, sorry I was a bit quick here (too early in the
> morning... ugh). We don't want the arch to go override the dma ops
> every time that is callled.
> 
> But yes, if we can fix the overhead, it becomes just a matter of
> setting up the "right" ops automatically.
> 
> > > (Passing the dev allows the arch to know this is a virtio device in
> > > "direct" mode or whatever we want to call the !iommu case, and
> > > construct appropriate DMA ops for it, which aren't the same as the DMA
> > > ops of any other PCI device who *do* use the iommu).
> > 
> > I think that's where Christoph might have specific ideas about it.
> 
> OK well, assuming Christoph can solve the direct case in a way that
> also work for the virtio !iommu case, we still want some bit of logic
> somewhere that will "switch" to swiotlb based ops if the DMA mask is
> limited.
> 
> You mentioned an RFC for that ? Do you happen to have a link ?

No but Christoph did I think.

> It would be indeed ideal if all we had to do was setup some kind of
> bus_dma_mask on all PCI devices and have virtio automagically insert
> swiotlb when necessary.
> 
> Cheers,
> Ben.
>