On 01/16/2018 10:39 AM, Alexandre Courbot wrote:
> On Mon, Jan 15, 2018 at 6:07 PM, Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
>> On 01/15/2018 09:24 AM, Alexandre Courbot wrote:
>>> On Fri, Jan 12, 2018 at 7:49 PM, Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
>>>> On 12/15/17 08:56, Alexandre Courbot wrote:
>>>>> Add throttling support for buffers when requests are in use on a given
>>>>> queue. Buffers associated to a request are kept into the vb2 queue until
>>>>> the request becomes active, at which point all the buffers are passed to
>>>>> the driver. The queue can also signal that is has processed all of a
>>>>> request's buffers.
>>>>>
>>>>> Also add support for the request parameter when handling the QBUF ioctl.
>>>>>
>>>>> Signed-off-by: Alexandre Courbot <acourbot@xxxxxxxxxxxx>
>>>>> ---
>>>>> drivers/media/v4l2-core/videobuf2-core.c | 59 ++++++++++++++++++++++++++++----
>>>>> drivers/media/v4l2-core/videobuf2-v4l2.c | 29 +++++++++++++++-
>>>>> include/media/videobuf2-core.h | 25 +++++++++++++-
>>>>> 3 files changed, 104 insertions(+), 9 deletions(-)
>>>>>
>>>>> diff --git a/drivers/media/v4l2-core/videobuf2-core.c b/drivers/media/v4l2-core/videobuf2-core.c
>>>>> index cb115ba6a1d2..c01038b7962a 100644
>>>>> --- a/drivers/media/v4l2-core/videobuf2-core.c
>>>>> +++ b/drivers/media/v4l2-core/videobuf2-core.c
>>>>> @@ -898,6 +898,8 @@ void vb2_buffer_done(struct vb2_buffer *vb, enum vb2_buffer_state state)
>>>>> state != VB2_BUF_STATE_REQUEUEING))
>>>>> state = VB2_BUF_STATE_ERROR;
>>>>>
>>>>> + WARN_ON(vb->request != q->cur_req);
>>>>
>>>> What's the reason for this WARN_ON? It's not immediately obvious to me.
>>>
>>> This is a safeguard against driver bugs: a buffer should not complete
>>> unless it is part of the request being currently processed.
>>>
>>>>
>>>>> +
>>>>> #ifdef CONFIG_VIDEO_ADV_DEBUG
>>>>> /*
>>>>> * Although this is not a callback, it still does have to balance
>>>>> @@ -920,6 +922,13 @@ void vb2_buffer_done(struct vb2_buffer *vb, enum vb2_buffer_state state)
>>>>> /* Add the buffer to the done buffers list */
>>>>> list_add_tail(&vb->done_entry, &q->done_list);
>>>>> vb->state = state;
>>>>> +
>>>>> + if (q->cur_req) {
>>>>> + WARN_ON(q->req_buf_cnt < 1);
>>>>> +
>>>>> + if (--q->req_buf_cnt == 0)
>>>>> + q->cur_req = NULL;
>>>>> + }
>>>>> }
>>>>> atomic_dec(&q->owned_by_drv_count);
>>>>> spin_unlock_irqrestore(&q->done_lock, flags);
>>>>> @@ -1298,6 +1307,16 @@ int vb2_core_prepare_buf(struct vb2_queue *q, unsigned int index, void *pb)
>>>>> }
>>>>> EXPORT_SYMBOL_GPL(vb2_core_prepare_buf);
>>>>>
>>>>> +static void vb2_queue_enqueue_current_buffers(struct vb2_queue *q)
>>>>> +{
>>>>> + struct vb2_buffer *vb;
>>>>> +
>>>>> + list_for_each_entry(vb, &q->queued_list, queued_entry) {
>>>>> + if (vb->request == q->cur_req)
>>>>> + __enqueue_in_driver(vb);
>>>>> + }
>>>>> +}
>>>>
>>>> I think this will clash big time with the v4l2 fence patch series...
>>>
>>> Indeed, but on the other hand I was not a big fan of going through the
>>> whole list. :) So I welcome the extra throttling introduced by the
>>> fence series.
>>
>> There is only throttling if fences are used by userspace. Otherwise there
>> is no change.
>>
>>>
>>>>
>>>>> +
>>>>> /**
>>>>> * vb2_start_streaming() - Attempt to start streaming.
>>>>> * @q: videobuf2 queue
>>>>> @@ -1318,8 +1337,7 @@ static int vb2_start_streaming(struct vb2_queue *q)
>>>>> * If any buffers were queued before streamon,
>>>>> * we can now pass them to driver for processing.
>>>>> */
>>>>> - list_for_each_entry(vb, &q->queued_list, queued_entry)
>>>>> - __enqueue_in_driver(vb);
>>>>> + vb2_queue_enqueue_current_buffers(q);
>>>>>
>>>>> /* Tell the driver to start streaming */
>>>>> q->start_streaming_called = 1;
>>>>> @@ -1361,7 +1379,8 @@ static int vb2_start_streaming(struct vb2_queue *q)
>>>>> return ret;
>>>>> }
>>>>>
>>>>> -int vb2_core_qbuf(struct vb2_queue *q, unsigned int index, void *pb)
>>>>> +int vb2_core_qbuf(struct vb2_queue *q, unsigned int index,
>>>>> + struct media_request *req, void *pb)
>>>>> {
>>>>> struct vb2_buffer *vb;
>>>>> int ret;
>>>>> @@ -1392,6 +1411,7 @@ int vb2_core_qbuf(struct vb2_queue *q, unsigned int index, void *pb)
>>>>> q->queued_count++;
>>>>> q->waiting_for_buffers = false;
>>>>> vb->state = VB2_BUF_STATE_QUEUED;
>>>>> + vb->request = req;
>>>>>
>>>>> if (pb)
>>>>> call_void_bufop(q, copy_timestamp, vb, pb);
>>>>> @@ -1401,8 +1421,11 @@ int vb2_core_qbuf(struct vb2_queue *q, unsigned int index, void *pb)
>>>>> /*
>>>>> * If already streaming, give the buffer to driver for processing.
>>>>> * If not, the buffer will be given to driver on next streamon.
>>>>> + *
>>>>> + * If using the request API, the buffer will be given to the driver
>>>>> + * when the request becomes active.
>>>>> */
>>>>> - if (q->start_streaming_called)
>>>>> + if (q->start_streaming_called && !req)
>>>>> __enqueue_in_driver(vb);
>>>>>
>>>>> /* Fill buffer information for the userspace */
>>>>> @@ -1427,6 +1450,28 @@ int vb2_core_qbuf(struct vb2_queue *q, unsigned int index, void *pb)
>>>>> }
>>>>> EXPORT_SYMBOL_GPL(vb2_core_qbuf);
>>>>>
>>>>> +void vb2_queue_start_request(struct vb2_queue *q, struct media_request *req)
>>>>> +{
>>>>> + struct vb2_buffer *vb;
>>>>> +
>>>>> + q->req_buf_cnt = 0;
>>>>> + list_for_each_entry(vb, &q->queued_list, queued_entry) {
>>>>> + if (vb->request == req)
>>>>> + ++q->req_buf_cnt;
>>>>> + }
>>>>> +
>>>>> + /* only consider the request if we actually have buffers for it */
>>>>> + if (q->req_buf_cnt == 0)
>>>>> + return;
>>>>> +
>>>>> + q->cur_req = req;
>>>>> +
>>>>> + /* If not streaming yet, we will enqueue the buffers later */
>>>>> + if (q->start_streaming_called)
>>>>> + vb2_queue_enqueue_current_buffers(q);
>>>>
>>>> If I understand all this correctly, then you are queuing one request at a
>>>> time to the vb2_queue. I.e. all the buffers queued to the driver belong to
>>>> the same request (q->cur_req).
>>>
>>> That is correct.
>>>
>>>> But that might work for codecs, but not
>>>> for camera drivers: you will typically have multiple requests queued up in
>>>> the driver.
>>>
>>> Aren't requests supposed to be performed sequentially, even in the
>>> camera case? Passing a buffer to the driver means that we allow it to
>>> process it using its current settings ; if another request is
>>> currently active, wouldn't that become an issue?
>>
>> Drivers often need multiple buffers queued before they can start the DMA
>> engine (usually at least two buffers have to be queued, just do a
>> git grep min_buffers_needed drivers/media).
>>
>> In addition, sensors often have to be programmed one or two frames earlier
>> for a new setting to take effect for the required frame.
>>
>> In other words: drivers need to be able to look ahead and vb2 should just
>> queue buffers/requests as soon as they are ready.
>
> Cannot drivers simply peek into their vb2_queue if they need to look
> ahead? My main concern here is that I would like to avoid having to
> make individual drivers aware of requests as much as possible. With
> the current design, drivers just need to care about unconditionally
> processing all the buffers that are passed to them by vb2, and not
> keeping it that way would complicate things.
I'm not sure what the problem is. Once buffers are ready (i.e. not waiting
for fences or unfinished requests) then they should be queued to the driver.
At that time the driver can look at whatever associated request data the
newly queued buffer has and program the hardware.
Once a buffer is ready, just queue it up to the driver. Drivers should never
look into buffers not explicitly queued to them. They don't need to either,
since those buffers are not ready anyway.
It's also a matter of who owns a buffer: that can be either userspace, vb2
or the driver. Each 'owner' can only touch the buffers it owns and should
never access buffers from other owners.
> The issue of programming sensors ahead-of-time is more complex. As
> always, it would probably be preferable to manage as much of this at
> the framework level. I am not sure that a bulletproof solution exists
> here, especially given the extra synchronization introduced by fences:
> you may want to program a sensor two frames ahead, but what if the
> frame in between has a non-yet signaled input fence? You could
> potentially block and miss the timing. The solution here would be to
> delay the processing of frame F-2 until the fence is signaled, or
> introduce an empty "latency frame" once the fence signals to get the
> timing right. Unless I am missing something, the ability to look into
> the vb2_queue should be enough to achieve this.
It is the responsibility of the application to make sure buffers aren't
held up waiting for fences. That's no different to what happens today
if the application cannot process buffers fast enough and the DMA engine
is starved of buffers.
Handling such complex scenarios is out-of-scope to the request API. The
request framework should just hand over the buffers + request data to
the driver as soon as it can, and that's it.
We will probably want to provide helper functions for drivers to use
to simplify such complex scenarios, but that's something for the future.
Right now we just need to get core functionality right.
>
>>
>>>
>>>>
>>>> In any case, I don't think the req_buf_cnt and cur_req fields belong in
>>>> vb2_queue.
>>>>
>>>> Another weird thing here is that it appears that you allow for multiple
>>>> buffers for the same device in the same request. I'm not sure that is
>>>> useful. For one, it will postpone the completion of the request until
>>>> all buffers are dequeued.
>>>
>>> s/dequeued/completed. A request is marked as completed as soon as all
>>> its buffers are marked as done, and can be polled before its buffers
>>> are dequeued by user-space.
>>>
>>> Do you suggest that we enforce a "one buffer per queue per request"
>>> rule? I cannot thing of any case where this would be a hard limiting
>>> factor (and it would certainly simplify the code), on the other hand
>>> it may slow things down a bit in the case where we want to e.g. take
>>> several shots in fast succession with the same parameters. But I have
>>> no evidence that the extra lag would be noticeable.
>>
>> This is something that should go into an RFC for discussion. My opinion
>> at this time is that you should stick with one buffer per queue per request.
>> This can always be relaxed in the future. For now it is better to keep
>> it simple.
>
> The more I think about this idea, the more it looks like a good one.
> It also answers the question "can we queue buffers not associated to a
> request after we started using requests?". The answer would be yes:
> such buffers would be processed immediately after the preceding
> request completes, almost emulating the behavior of having several
> buffers per request (with the exceptions that the request completes
> earlier and with a different ending state, but this can be fixed by
> associated another request without any control set to the last frame).
>
It sounds reasonable, but, as always, the devil is in the details.
Regards,
Hans
