Re: [RFC 2/3] USB: core: Add non-coherent buffer allocation helpers

[Tomasz Figa]

Hi Christoph and everyone,

On Fri, Aug 31, 2018 at 3:51 PM Tomasz Figa <tfiga@xxxxxxxxxxxx> wrote:
>
> On Fri, Aug 31, 2018 at 2:50 PM Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
> >
> > On Thu, Aug 30, 2018 at 07:11:35PM -0300, Ezequiel Garcia wrote:
> > > On Thu, 2018-08-30 at 10:58 -0700, Christoph Hellwig wrote:
> > > > Please don't introduce new DMA_ATTR_NON_CONSISTENT users, it is
> > > > a rather horrible interface, and I plan to kill it off rather sooner
> > > > than later.  I plan to post some patches for a better interface
> > > > that can reuse the normal dma_sync_single_* interfaces for ownership
> > > > transfers.  I can happily include usb in that initial patch set based
> > > > on your work here if that helps.
> > >
> > > Please do. Until we have proper allocators that go thru the DMA API,
> > > drivers will have to kmalloc the USB transfer buffers, and have
> > > streaming mappings. Which in turns mean not using IOMMU or CMA.
> >
> > dma_map_page will of course use the iommu.
>
> Sure, dma_map*() will, but using kmalloc() defeats (half of) the
> purpose of it, since contiguous memory would be allocated
> unnecessarily, risking failures due to fragmentation.

Have we reached a conclusion here?

It sounds like it's a quite significant problem, at least for some of
the camera (media) devices over there and there are people interested
in solving it, so all we need here is a conclusion on how to do it. :)

Best regards,
Tomasz