[Spice-devel] [PATCH v2 1/1] video_video: Add the Virtio Video V4L2 driver

[Dmitry Sepp]

Hi Tomasz,

On Freitag, 13. März 2020 12:11:51 CET Tomasz Figa wrote:
> On Fri, Mar 13, 2020 at 11:27 AM Dmitry Sepp
> 
> <dmitry.sepp at opensynergy.com> wrote:
> > Hi Tomasz,
> > 
> > On Freitag, 13. März 2020 11:05:35 CET Tomasz Figa wrote:
> > > On Thu, Mar 12, 2020 at 12:48 PM Dmitry Sepp
> > > 
> > > <dmitry.sepp at opensynergy.com> wrote:
> > > > Hi Hans,
> > > > 
> > > > One more thing:
> > > > > GFP_DMA? That's unusual. I'd expect GFP_DMA32. All V4L2 drivers use
> > > > > that.
> > > > 
> > > > GFP_DMA32 had no effect for me on arm64. Probably I need to recheck.
> > > 
> > > What's the reason to use any specific GFP flags at all? GFP_DMA(32)
> > > memory in the guest would typically correspond to host pages without
> > > any specific location guarantee.
> > 
> > Typically, but not always, especially for non x86. Say, some platforms
> > don't have IOMMUs for codec devices and those devices require physically
> > contig low memory. We had to find a way to handle that.
> 
> So basically your hypervisor guarantees that the guest pages inside
> the GFP_DMA zone are contiguous and DMA-able on the host as well?
> Given the Linux-specific aspect of GFP flags and differences in the
> implementation across architectures, perhaps it would be a better idea
> to use the DMA mask instead? That wouldn't currently affect vb2_dma_sg
> allocations, but in that case the host decoder would have some IOMMU
> anyway, right?
> 

DMA mask has no effect for vb2_dma_sg, but GFP has. Unfortunately we need to 
support both of the two: low mem phys contig and low mem sg. So DMA mask 
cannot be an option. No, there are use-cases with obsolutely no iommus.

Best regards,
Dmitry.

> > Best regards,
> > Dmitry.
> > 
> > > Best regards,
> > > Tomasz
> > > 
> > > > Best regards,
> > > > Dmitry.
> > > > 
> > > > On Donnerstag, 12. März 2020 11:18:26 CET Hans Verkuil wrote:
> > > > > On 3/12/20 11:15 AM, Dmitry Sepp wrote:
> > > > > > Hi Hans,
> > > > > > 
> > > > > > Thank you for your great detailed review!
> > > > > > 
> > > > > > I won't provide inline answers as your comments totally make
> > > > > > sense.
> > > > > > There
> > > > > > is>
> > > > > > 
> > > > > > only one thing I want to mention:
> > > > > >>> + struct video_plane_format
> > > > > >>> plane_format[VIRTIO_VIDEO_MAX_PLANES];
> > > > > >> 
> > > > > >> Why is this virtio specific? Any reason for not using
> > > > > >> VIDEO_MAX_PLANES?
> > > > > > 
> > > > > > I'd say this is because VIDEO_MAX_PLANES does not exist outside of
> > > > > > the
> > > > > > Linux OS, so for whatever other system we need a virtio specific
> > > > > > definition.
> > > > > 
> > > > > OK, good reason :-)
> > > > > 
> > > > > It's probably a good thing to add a comment where
> > > > > VIRTIO_VIDEO_MAX_PLANES is defined that explains this.
> > > > > 
> > > > > Regards,
> > > > > 
> > > > >       Hans