Re: [PATCH v4 6/6] media: uvcvideo: Move decode processing to process context

[Kieran Bingham]

Hi Guennadi,

Thank you for reviewing / taking a look through the series.

On 04/06/18 13:09, Guennadi Liakhovetski wrote:
> Hi Kieran,
> 
> I've got a question:
> 
> On Tue, 27 Mar 2018, Kieran Bingham wrote:
> 
>> Newer high definition cameras, and cameras with multiple lenses such as
>> the range of stereo-vision cameras now available have ever increasing
>> data rates.
>>
>> The inclusion of a variable length packet header in URB packets mean
>> that we must memcpy the frame data out to our destination 'manually'.
>> This can result in data rates of up to 2 gigabits per second being
>> processed.
>>
>> To improve efficiency, and maximise throughput, handle the URB decode
>> processing through a work queue to move it from interrupt context, and
>> allow multiple processors to work on URBs in parallel.
>>
>> Signed-off-by: Kieran Bingham <kieran.bingham@xxxxxxxxxxxxxxxx>
>>
>> ---
>> v2:
>>  - Lock full critical section of usb_submit_urb()
>>
>> v3:
>>  - Fix race on submitting uvc_video_decode_data_work() to work queue.
>>  - Rename uvc_decode_op -> uvc_copy_op (Generic to encode/decode)
>>  - Rename decodes -> copy_operations
>>  - Don't queue work if there is no async task
>>  - obtain copy op structure directly in uvc_video_decode_data()
>>  - uvc_video_decode_data_work() -> uvc_video_copy_data_work()
>>
>> v4:
>>  - Provide for_each_uvc_urb()
>>  - Simplify fix for shutdown race to flush queue before freeing URBs
>>  - Rebase to v4.16-rc4 (linux-media/master) adjusting for metadata
>>    conflicts.
>>
>>  drivers/media/usb/uvc/uvc_video.c | 107 ++++++++++++++++++++++++-------
>>  drivers/media/usb/uvc/uvcvideo.h  |  28 ++++++++-
>>  2 files changed, 111 insertions(+), 24 deletions(-)
>>
>> diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/usb/uvc/uvc_video.c
>> index 7dd0dcb457f3..a62e8caf367c 100644
>> --- a/drivers/media/usb/uvc/uvc_video.c
>> +++ b/drivers/media/usb/uvc/uvc_video.c
>> @@ -1042,21 +1042,54 @@ static int uvc_video_decode_start(struct uvc_streaming *stream,
>>  	return data[0];
>>  }
>>  
>> -static void uvc_video_decode_data(struct uvc_streaming *stream,
>> +/*
>> + * uvc_video_decode_data_work: Asynchronous memcpy processing
>> + *
>> + * Perform memcpy tasks in process context, with completion handlers
>> + * to return the URB, and buffer handles.
>> + */
>> +static void uvc_video_copy_data_work(struct work_struct *work)
>> +{
>> +	struct uvc_urb *uvc_urb = container_of(work, struct uvc_urb, work);
>> +	unsigned int i;
>> +	int ret;
>> +
>> +	for (i = 0; i < uvc_urb->async_operations; i++) {
>> +		struct uvc_copy_op *op = &uvc_urb->copy_operations[i];
>> +
>> +		memcpy(op->dst, op->src, op->len);
>> +
>> +		/* Release reference taken on this buffer */
>> +		uvc_queue_buffer_release(op->buf);
>> +	}
>> +
>> +	ret = usb_submit_urb(uvc_urb->urb, GFP_ATOMIC);
> 
> Does this still have to be ATOMIC now that it's called from a work queue 
> context?

I think you're right.
This could very likely be changed to GFP_KERNEL.

Does this series impact anything on your async-controls series ?

--
Kieran


> 
>> +	if (ret < 0)
>> +		uvc_printk(KERN_ERR, "Failed to resubmit video URB (%d).\n",
>> +			   ret);
>> +}
> 
> [snip]
> 
> Thannks
> Guennadi
>