On Thu, 9 Aug 2018, Laurent Pinchart wrote:
> > >> There is no need to wonder. "Frequent DMA mapping/Cached memory" is
> > >> always faster than "No DMA mapping/Uncached memory".
> > >
> > > Is it really, doesn't it depend on the CPU access pattern ?
> >
> > Well, if your access pattern involves transferring data in from the
> > device and then throwing it away without reading it, you might get a
> > different result. :-) But assuming you plan to read the data after
> > transferring it, using uncached memory slows things down so much that
> > the overhead of DMA mapping/unmapping is negligible by comparison.
>
> :-) I suppose it would also depend on the access pattern, if I only need to
> access part of the buffer, performance figures may vary. In this case however
> the whole buffer needs to be copied.
>
> > The only exception might be if you were talking about very small
> > amounts of data. I don't know exactly where the crossover occurs, but
> > bear in mind that Matwey's tests required ~50 us for mapping/unmapping
> > and 3000 us for accessing uncached memory. He didn't say how large the
> > transfers were, but that's still a pretty big difference.
>
> For UVC devices using bulk endpoints data buffers are typically tens of kBs.
> For devices using isochronous endpoints, that goes down to possibly hundreds
> of bytes for some buffers. Devices can send less data than the maximum packet
> size, and mapping/unmapping would still invalidate the cache for the whole
> buffer. If we keep the mappings around and use the DMA sync API, we could
> possibly restrict the cache invalidation to the portion of the buffer actually
> written to.
Furthermore, invalidating a cache is likely to require less overhead
than using non-cacheable memory. After the cache has been invalidated,
it can be repopulated relatively quickly (an entire cache line at a
time), whereas reading uncached memory requires a slow transaction for
each individual read operation.
I think adding support to the USB core for
dma_sync_single_for_{cpu|device} would be a good approach. In fact, I
wonder whether using coherent mappings provides any benefit at all.
Alan Stern
