[lkml]   [2018]   [Aug]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
SubjectRe: [RFC PATCH v1] media: uvcvideo: Cache URB header data before processing
Hi All,

On 08/08/18 13:45, 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.
> 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".
> For this purpose, I'm planning to measure performance on ARM
> Chromebooks in the following conditions:
> 1. Current implementation with Kieran's patches
> 2. 1. + my patch
> 3. Use kmalloc instead
> 1 and 2 are the same conditions I reported in the first mail on this thread.
> For condition 3, I only have to add "#define CONFIG_DMA_NONCOHERENT"
> at the beginning of uvc_video.c.

I'd be interested in numbers/performances both with and without my async
if possible too.

The async path can be easily disabled temporarily with the following
change: (perhaps this should be a module option?)

diff --git a/drivers/media/usb/uvc/uvc_video.c
index 8bb6e90f3483..f9fbdc9bfa4b 100644
--- a/drivers/media/usb/uvc/uvc_video.c
+++ b/drivers/media/usb/uvc/uvc_video.c
@@ -1505,7 +1505,8 @@ static void uvc_video_complete(struct urb *urb)

INIT_WORK(&uvc_urb->work, uvc_video_copy_data_work);
- queue_work(stream->async_wq, &uvc_urb->work);
+// queue_work(stream->async_wq, &uvc_urb->work);
+ uvc_video_copy_data_work(&uvc_urb->work);


I do suspect that even with cached buffers, it's probably likely we
should still consider the async patches to move the memcopy out of
interrupt context.



> Does this plan sound reasonable?
> Best regards,
> Keiichi
> On Wed, Aug 8, 2018 at 5:42 PM Laurent Pinchart
> <> wrote:
>> Hi Tomasz,
>> On Wednesday, 8 August 2018 07:08:59 EEST Tomasz Figa wrote:
>>> On Tue, Jul 31, 2018 at 1:00 AM Laurent Pinchart wrote:
>>>> On Wednesday, 27 June 2018 13:34:08 EEST Keiichi Watanabe wrote:
>>>>> On some platforms with non-coherent DMA (e.g. ARM), USB drivers use
>>>>> uncached memory allocation methods. In such situations, it sometimes
>>>>> takes a long time to access URB buffers. This can be a cause of video
>>>>> flickering problems if a resolution is high and a USB controller has
>>>>> a very tight time limit. (e.g. dwc2) To avoid this problem, we copy
>>>>> header data from (uncached) URB buffer into (cached) local buffer.
>>>>> This change should make the elapsed time of the interrupt handler
>>>>> shorter on platforms with non-coherent DMA. We measured the elapsed
>>>>> time of each callback of uvc_video_complete without/with this patch
>>>>> while capturing Full HD video in
>>>>> I tested it on the top of Kieran Bingham's Asynchronous UVC series
>>>>> The test device was Jerry Chromebook (RK3288) with Logitech Brio 4K.
>>>>> I collected data for 5 seconds. (There were around 480 callbacks in
>>>>> this case.) The following result shows that this patch makes
>>>>> uvc_video_complete about 2x faster.
>>>>> | average | median | min | max | standard deviation
>>>>> w/o caching| 45319ns | 40250ns | 33834ns | 142625ns| 16611ns
>>>>> w/ caching| 20620ns | 19250ns | 12250ns | 56583ns | 6285ns
>>>>> In addition, we confirmed that this patch doesn't make it worse on
>>>>> coherent DMA architecture by performing the same measurements on a
>>>>> Broadwell Chromebox with the same camera.
>>>>> | average | median | min | max | standard deviation
>>>>> w/o caching| 21026ns | 21424ns | 12263ns | 23956ns | 1932ns
>>>>> w/ caching| 20728ns | 20398ns | 8922ns | 45120ns | 3368ns
>>>> This is very interesting, and it seems related to https://
>>>> You might have seen that discussion
>>>> as you got CC'ed at some point.
>>>> I wonder whether performances couldn't be further improved by allocating
>>>> the URB buffers cached, as that would speed up the memcpy() as well. Have
>>>> you tested that by any chance ?
>>> We haven't measure it, but the issue being solved here was indeed
>>> significantly reduced by using cached URB buffers, even without
>>> Kieran's async series. After we discovered the latter, we just
>>> backported it and decided to further tweak the last remaining bit, to
>>> avoid playing too much with the DMA API in code used in production on
>>> several different platforms (including both ARM and x86).
>>> If you think we could change the driver to use cached buffers instead
>>> (as the pwc driver mentioned in another thread), I wouldn't have
>>> anything against it obviously.
>> I think there's a chance that performances could be further improved.
>> Furthermore, it would lean to simpler code as we wouldn't need to deal with
>> caching headers manually. I would however like to see numbers before making a
>> decision.
>> --
>> Regards,
>> Laurent Pinchart


 \ /
  Last update: 2018-08-08 15:08    [W:0.060 / U:2.776 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site