Re: [PATCH v5 2/2] media: usb: pwc: Don't use coherent DMA buffers for ISO transfer

[Tomasz Figa]

On Tue, Dec 18, 2018 at 4:38 PM Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
>
> On Tue, Dec 18, 2018 at 04:22:43PM +0900, Tomasz Figa wrote:
> > It kind of limits the usability of this API, since it enforces
> > contiguous allocations even for big sizes even for devices behind
> > IOMMU (contrary to the case when DMA_ATTR_NON_CONSISTENT is not set),
> > but given that it's just a temporary solution for devices like these
> > USB cameras, I guess that's fine.
>
> The problem is that you can't have flexibility and simplicity at the
> same time.  Once you use kernel virtual address remapping you need to
> be prepared to have multiple segments.
>
> So as I said you can call dma_alloc_attrs with DMA_ATTR_NON_CONSISTENT
> in a loop with a suitably small chunk size, then stuff the results into
> a scatterlist and map that again for the device share with if you don't
> want a single contigous region.  You just have to either deal with
> non-contigous access from the kernel or use vmap and the right vmap
> cache flushing helpers.

The point is that you didn't have to do this small chunk loop without
DMA_ATTR_NON_CONSISTENT, so it's at least inconsistent now and not
sure why it could be better than just a loop of alloc_page().

>
> > Note that in V4L2 we use the DMA API extensively, so that we don't
> > need to embed any device-specific or integration-specific knowledge in
> > the framework. Right now we're using dma_alloc_attrs() with
> > driver-provided attrs [1], but current driver never request
> > non-consistent memory. We're however thinking about making it possible
> > to allocate non-consistent memory. What would you suggest for this?
> >
> > [1] https://elixir.bootlin.com/linux/v4.20-rc7/source/drivers/media/common/videobuf2/videobuf2-dma-contig.c#L139
>
> I would advice against new non-consistent users until this series
> goes through, mostly because dma_cache_sync is such an amazing bad
> API.  Otherwise things will just work at the allocation side, you'll
> just need to be careful to transfer ownership between the cpu and
> the device(s) carefully using the dma_sync_* APIs.

Just to clarify, the actual code isn't very likely to surface any time
soon. so I assume it would be after this series lands.

We will however need an API that can transparently handle both cases
of contiguous (without IOMMU) and page-by-page allocations (with
IOMMU) behind the scenes, like the current dma_alloc_attrs() without
DMA_ATTR_NON_CONSISTENT.

Best regards,
Tomasz