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@xxxxxxxxxxxxxxx> 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@xxxxxxxxxxxxxxx> 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 _______________________________________________ Spice-devel mailing list Spice-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/spice-devel