Re: [RFC PATCH v1] media: uvcvideo: Cache URB header data before processing

[Laurent Pinchart]

Hi Alan,

On Wednesday, 8 August 2018 20:04:45 EEST Alan Stern wrote:
> On Wed, 8 Aug 2018, Laurent Pinchart wrote:
> > On Wednesday, 8 August 2018 17:20:21 EEST Alan Stern wrote:
> >> On Wed, 8 Aug 2018, Keiichi Watanabe wrote:
> >>> Hi Laurent, Kieran, Tomasz,
> >>> 
> >>> Thank you for reviews and suggestions.
> >>> I want to do additional measurements for improving the performance.
> >>> 
> >>> Let me clarify my understanding:
> >>> Currently, if the platform doesn't support coherent-DMA (e.g. ARM),
> >>> urb_buffer is allocated by usb_alloc_coherent with
> >>> URB_NO_TRANSFER_DMA_MAP flag instead of using kmalloc.
> >> 
> >> Not exactly.  You are mixing up allocation with mapping.  The speed of
> >> the allocation doesn't matter; all that matters is whether the memory
> >> is cached and when it gets mapped/unmapped.
> >> 
> >>> This is because we want to avoid frequent DMA mappings, which are
> >>> generally expensive. However, memories allocated in this way are not
> >>> cached.
> >>> 
> >>> So, we wonder if using usb_alloc_coherent is really fast.
> >>> In other words, we want to know which is better:
> >>> "No DMA mapping/Uncached memory" v.s. "Frequent DMA mapping/Cached
> >>> memory".
> > 
> > The second option should also be split in two:
> > 
> > - cached memory with DMA mapping/unmapping around each transfer
> > - cached memory with DMA mapping/unmapping at allocation/free time, and
> > DMA sync around each transfer
> > 
> > The second option should in theory lead to at least slightly better
> > performances, but tests with the pwc driver have reported contradictory
> > results. I'd like to know whether that's also the case with the uvcvideo
> > driver, and if so, why.
> > 
> >> 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.

> >> The only issue is that on some platform (such as x86) but not others,
> >> there is a third option: "No DMA mapping/Cached memory".  On platforms
> >> which support it, this is the fastest option.

-- 
Regards,

Laurent Pinchart