Hi Christoph

On Wed, Jan 27, 2021 at 4:56 PM . Christoph Hellwig [off-list ref] wrote:

On Wed, Jan 27, 2021 at 12:29:08AM +0100, Ricardo Ribalda wrote:

quoted

- Is there any platform where dma_alloc_noncontiguos can fail?
This is, !ops->alloc_noncontiguous and !dev->coherent_dma_mask
If yes then we need to add a function to let the driver know in
advance that it has to use the coherent allocator (usb_alloc_coherent
for uvc)

dev->coherent_dma_mask is set by the driver.  So the only reason why
dma_alloc_noncontiguos will fail is because is because it can't
allocate any memory.

quoted

- In dma_alloc_noncontiguos, on the dma_alloc_pages fallback. If we
have a device where the dma happens in only one direction, could not
get more performance with DMA_FROM/TO_DEVICE instead of
DMA_BIDIRECTIONAL ?

Yes, we could probably do that.

quoted

Then I have tried to use the API, and I have encountered a problem: on
uvcvideo the device passed to the memory allocator is different for
DMA_PAGES and NON_CONTIGUOUS:
https://github.com/ribalda/linux/blob/042cd497739f71c8d4a83a67ee970369e2baca4a/drivers/media/usb/uvc/uvc_video.c#L1236

I need to dig a bit tomorrow to figure out why this is, I have
hardware to test both paths, so it should not be too difficult.

I always found the USB dma alloc API a little weird, but we might have
to follow the scheme of the usb coherent wrappers there.

I have used the current API here:

https://git.kernel.org/pub/scm/linux/kernel/git/ribalda/linux.git/log/?h=uvc-noncontiguous

And I think the result is very clean. Great work!

I have tested it in X86 and in arm64, with similar performance as the
previous patchset.

Maybe you want to cherry pick that commit into your series I can also
send the patch to the list for review if you prefer so.

At least in 5.11 rc5 I the same dmadev worked in arm64 and x86.

Best regards!


-- 
Ricardo Ribalda