lkml.org 
[lkml]   [2018]   [Aug]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v4 2/2] media: usb: pwc: Don't use coherent DMA buffers for ISO transfer
пт, 10 авг. 2018 г. в 17:27, Alan Stern <stern@rowland.harvard.edu>:
>
> On Fri, 10 Aug 2018, Laurent Pinchart wrote:
>
> > > > Aren't you're missing a dma_sync_single_for_device() call before
> > > > submitting the URB ? IIRC that's required for correct operation of the DMA
> > > > mapping API on some platforms, depending on the cache architecture. The
> > > > additional sync can affect performances, so it would be useful to re-run
> > > > the perf test.
> > >
> > > This was already discussed:
> > >
> > > https://lkml.org/lkml/2018/7/23/1051
> > >
> > > I rely on Alan's reply:
> > >
> > > > According to Documentation/DMA-API-HOWTO.txt, the CPU should not write
> > > > to a DMA_FROM_DEVICE-mapped area, so dma_sync_single_for_device() is
> > > > not needed.
> >
> > I fully agree that the CPU should not write to the buffer. However, I think
> > the sync call is still needed. It's been a long time since I touched this
> > area, but IIRC, some cache architectures (VIVT ?) require both cache clean
> > before the transfer and cache invalidation after the transfer. On platforms
> > where no cache management operation is needed before the transfer in the
> > DMA_FROM_DEVICE direction, the dma_sync_*_for_device() calls should be no-ops
> > (and if they're not, it's a bug of the DMA mapping implementation).
>
> In general, I agree that the cache has to be clean before a transfer
> starts. This means some sort of mapping operation (like
> dma_sync_*_for-device) is indeed required at some point between the
> allocation and the first transfer.
>
> For subsequent transfers, however, the cache is already clean and it
> will remain clean because the CPU will not do any writes to the buffer.
> (Note: clean != empty. Rather, clean == !dirty.) Therefore transfers
> following the first should not need any dma_sync_*_for_device.
>
> If you don't accept this reasoning then you should ask the people who
> wrote DMA-API-HOWTO.txt. They certainly will know more about this
> issue than I do.

Laurent,

I have not seen any data corruption or glitches without
dma_sync_single_for_device() on ARM and x86.
It takes additional ~20usec for dma_sync_single_for_device() to run on
ARM. So It would be great to ensure that we can reliable save another
15% running time.

>
> Alan Stern
>


--
With best regards,
Matwey V. Kornilov

\
 
 \ /
  Last update: 2018-08-17 19:46    [W:0.093 / U:2.872 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site