Re: [PATCH] swiotlb: sync buffer when mapping FROM_DEVICE

[Robin Murphy]

On 2019-05-22 2:09 pm, Christoph Hellwig wrote:
> On Wed, May 22, 2019 at 01:50:47PM +0100, Robin Murphy wrote:
> > Would that work out any different from the existing DMA_ATTR_SKIP_CPU_SYNC?
> >
> > If drivers are prepared to handle this issue from their end, they can
> > already do so for single mappings by using that attr along with explicit
> > partial syncs via dma_sync_single(). For page/sg mappings we'd still have
> > the problem of identifying what part of "partial" actually matters, and
> > probably having to add some additional new sync operations to cope.
>
> Except that the same optimization we are tripping over here is also
> present in dma_sync_* - dma_sync_*_to_device with DMA_FROM_DEVICE is a
> no-op in swiotlb.

Sure, but that should be irrelevant since the effective problem here is 
in the sync_*_for_cpu direction, and it's the unmap which nobbles the 
buffer. If the driver does this:

	dma_map_single(whole buffer);
	<device writes to part of buffer>
	dma_unmap_single(whole buffer);
	<contents of rest of buffer now undefined>

then it could instead do this and be happy:

	dma_map_single(whole buffer, SKIP_CPU_SYNC);
	<device writes to part of buffer>
	dma_sync_single_for_cpu(updated part of buffer);
	dma_unmap_single(whole buffer, SKIP_CPU_SYNC);
	<contents of rest of buffer still valid>

Robin.