by Dafna Hirschfeld

[permalink] [raw]

Subject: Re: [PATCH v3 3/9] media: vimc: Add usage count to subdevices

Am 19.08.20 um 20:04 schrieb Kaaira Gupta:
> From: Niklas Söderlund <[email protected]>
>
> Prepare for multiple video streams from the same sensor by adding a use
> counter to vimc_ent_device. The counter is increased for every s_stream(1)
> and decremented for every s_stream(0) call.
>
> The subdevice stream is not started or stopped unless the usage count go
> from 0 to 1 (started) or from 1 to 0 (stopped). This allows for multiple
> s_stream() calls to try to either start or stop the device while only
> the first/last call will actually effect the state of the device.
>
> Initialise and increment use_count for capture as well, as use_count
> will be used in subsequent patches for starting process_frame as well.
>
> [Kaaira: moved use_count to vimc entity device instead of declaring it
> for each subdevice, used use_count for capture as well and rebased
> the patch on current HEAD of master to help with the current series]
>
> Signed-off-by: Niklas Söderlund <[email protected]>
> Signed-off-by: Kaaira Gupta <[email protected]>

Hi,
maybe incrementing/decrementing the usage count can be
done from the streamer code instead of
the s_stream of each entity?

Thanks,
Dafna

> ---
> drivers/media/test-drivers/vimc/vimc-capture.c | 3 +++
> drivers/media/test-drivers/vimc/vimc-common.h | 2 ++
> drivers/media/test-drivers/vimc/vimc-debayer.c | 7 +++++++
> drivers/media/test-drivers/vimc/vimc-scaler.c | 7 +++++++
> drivers/media/test-drivers/vimc/vimc-sensor.c | 6 ++++++
> 5 files changed, 25 insertions(+)
>
> diff --git a/drivers/media/test-drivers/vimc/vimc-capture.c b/drivers/media/test-drivers/vimc/vimc-capture.c
> index a8cbb8e4d5ba..93418cb5a139 100644
> --- a/drivers/media/test-drivers/vimc/vimc-capture.c
> +++ b/drivers/media/test-drivers/vimc/vimc-capture.c
> @@ -243,6 +243,7 @@ static int vimc_cap_start_streaming(struct vb2_queue *vq, unsigned int count)
> struct media_entity *entity = &vcap->vdev.entity;
> int ret;
>
> + atomic_inc(&vcap->ved.use_count);
> vcap->sequence = 0;
>
> /* Start the media pipeline */
> @@ -270,6 +271,7 @@ static void vimc_cap_stop_streaming(struct vb2_queue *vq)
> {
> struct vimc_cap_device *vcap = vb2_get_drv_priv(vq);
>
> + atomic_dec(&vcap->ved.use_count);
> vimc_streamer_s_stream(&vcap->stream, &vcap->ved, 0);
>
> /* Stop the media pipeline */
> @@ -424,6 +426,7 @@ static struct vimc_ent_device *vimc_cap_add(struct vimc_device *vimc,
> vcap->vdev.entity.name = vcfg_name;
> vcap->vdev.entity.function = MEDIA_ENT_F_IO_V4L;
> vcap->pad.flags = MEDIA_PAD_FL_SINK;
> + atomic_set(&vcap->ved.use_count, 0);
> ret = media_entity_pads_init(&vcap->vdev.entity,
> 1, &vcap->pad);
> if (ret)
> diff --git a/drivers/media/test-drivers/vimc/vimc-common.h b/drivers/media/test-drivers/vimc/vimc-common.h
> index 287d66edff49..c214f5ec7818 100644
> --- a/drivers/media/test-drivers/vimc/vimc-common.h
> +++ b/drivers/media/test-drivers/vimc/vimc-common.h
> @@ -85,6 +85,7 @@ struct vimc_pix_map {
> *
> * @dev: a pointer of the device struct of the driver
> * @ent: the pointer to struct media_entity for the node
> + * @use_count: a count to show the number of streams entity is part of
> * @get_frame: callback that sends a frame processed by the entity
> * @process_frame: callback that processes a frame
> * @vdev_get_format: callback that returns the current format a pad, used
> @@ -102,6 +103,7 @@ struct vimc_pix_map {
> struct vimc_ent_device {
> struct device *dev;
> struct media_entity *ent;
> + atomic_t use_count;
> void * (*get_frame)(struct vimc_ent_device *ved);
> int (*process_frame)(struct vimc_ent_device *ved);
> void (*vdev_get_format)(struct vimc_ent_device *ved,
> diff --git a/drivers/media/test-drivers/vimc/vimc-debayer.c b/drivers/media/test-drivers/vimc/vimc-debayer.c
> index f61e6e8899ac..60c4c0ec2030 100644
> --- a/drivers/media/test-drivers/vimc/vimc-debayer.c
> +++ b/drivers/media/test-drivers/vimc/vimc-debayer.c
> @@ -343,6 +343,9 @@ static int vimc_deb_s_stream(struct v4l2_subdev *sd, int enable)
> const struct vimc_pix_map *vpix;
> unsigned int frame_size;
>
> + if (atomic_inc_return(&vdeb->ved.use_count) != 1)
> + return 0;
> +
> if (vdeb->src_frame)
> return 0;
>
> @@ -368,6 +371,9 @@ static int vimc_deb_s_stream(struct v4l2_subdev *sd, int enable)
> return -ENOMEM;
>
> } else {
> + if (atomic_dec_return(&vdeb->ved.use_count) != 0)
> + return 0;
> +
> if (!vdeb->src_frame)
> return 0;
>
> @@ -608,6 +614,7 @@ static struct vimc_ent_device *vimc_deb_add(struct vimc_device *vimc,
> vdeb->ved.get_frame = vimc_deb_get_frame;
> vdeb->ved.dev = vimc->mdev.dev;
> vdeb->mean_win_size = vimc_deb_ctrl_mean_win_size.def;
> + atomic_set(&vdeb->ved.use_count, 0);
>
> /* Initialize the frame format */
> vdeb->sink_fmt = sink_fmt_default;
> diff --git a/drivers/media/test-drivers/vimc/vimc-scaler.c b/drivers/media/test-drivers/vimc/vimc-scaler.c
> index 347f9cd4a168..d511e1f2152d 100644
> --- a/drivers/media/test-drivers/vimc/vimc-scaler.c
> +++ b/drivers/media/test-drivers/vimc/vimc-scaler.c
> @@ -340,6 +340,9 @@ static int vimc_sca_s_stream(struct v4l2_subdev *sd, int enable)
> const struct vimc_pix_map *vpix;
> unsigned int frame_size;
>
> + if (atomic_inc_return(&vsca->ved.use_count) != 1)
> + return 0;
> +
> if (vsca->src_frame)
> return 0;
>
> @@ -363,6 +366,9 @@ static int vimc_sca_s_stream(struct v4l2_subdev *sd, int enable)
> return -ENOMEM;
>
> } else {
> + if (atomic_dec_return(&vsca->ved.use_count) != 0)
> + return 0;
> +
> if (!vsca->src_frame)
> return 0;
>
> @@ -518,6 +524,7 @@ static struct vimc_ent_device *vimc_sca_add(struct vimc_device *vimc,
> vsca->ved.process_frame = vimc_sca_process_frame;
> vsca->ved.get_frame = vimc_sca_get_frame;
> vsca->ved.dev = vimc->mdev.dev;
> + atomic_set(&vsca->ved.use_count, 0);
>
> /* Initialize the frame format */
> vsca->sink_fmt = sink_fmt_default;
> diff --git a/drivers/media/test-drivers/vimc/vimc-sensor.c b/drivers/media/test-drivers/vimc/vimc-sensor.c
> index 32a2c39de2cd..ced8ef06b01e 100644
> --- a/drivers/media/test-drivers/vimc/vimc-sensor.c
> +++ b/drivers/media/test-drivers/vimc/vimc-sensor.c
> @@ -256,6 +256,9 @@ static int vimc_sen_s_stream(struct v4l2_subdev *sd, int enable)
> const struct vimc_pix_map *vpix;
> unsigned int frame_size;
>
> + if (atomic_inc_return(&vsen->ved.use_count) != 1)
> + return 0;
> +
> vsen->start_stream_ts = ktime_get_ns();
>
> /* Calculate the frame size */
> @@ -275,6 +278,8 @@ static int vimc_sen_s_stream(struct v4l2_subdev *sd, int enable)
> vimc_sen_tpg_s_format(vsen);
>
> } else {
> + if (atomic_dec_return(&vsen->ved.use_count) != 0)
> + return 0;
>
> vfree(vsen->frame);
> vsen->frame = NULL;
> @@ -437,6 +442,7 @@ static struct vimc_ent_device *vimc_sen_add(struct vimc_device *vimc,
> vsen->ved.process_frame = vimc_sen_process_frame;
> vsen->ved.get_frame = vimc_sen_get_frame;
> vsen->ved.dev = vimc->mdev.dev;
> + atomic_set(&vsen->ved.use_count, 0);
>
> /* Initialize the frame format */
> vsen->mbus_format = fmt_default;
>

2020-10-02 12:20:29

by Dafna Hirschfeld

[permalink] [raw]

Subject: Re: [PATCH v3 4/9] media: vimc: Separate starting stream from pipeline initialisation

Am 12.09.20 um 12:21 schrieb Kaaira Gupta:
> On Wed, Sep 02, 2020 at 10:56:46AM +0100, Kieran Bingham wrote:
>> Hi Kaaira, Dafna,
>>
>> On 28/08/2020 21:37, Dafna Hirschfeld wrote:
>>> Hi,
>>>
>>> Am 21.08.20 um 23:01 schrieb Kaaira Gupta:
>>>> Hi,
>>>>
>>>> On Fri, Aug 21, 2020 at 05:11:23PM +0200, Dafna Hirschfeld wrote:
>>>>>
>>>>>
>>>>> Am 19.08.20 um 20:04 schrieb Kaaira Gupta:
>>>>>> Separate the process of initialising pipeline array from starting
>>>>>> streaming for all entities in path of a stream. This is needed because
>>>>>> multiple streams can stream, but only one pipeline object is needed.
>>>>>>
>>>>>> Process frames only for those entities in a pipeline which are
>>>>>> streaming. This is known through their use counts.
>>>>>>
>>>>>> Signed-off-by: Kaaira Gupta <[email protected]>
>>>>>> ---
>>>>>>    .../media/test-drivers/vimc/vimc-streamer.c   | 95
>>>>>> ++++++++++++++++---
>>>>>>    1 file changed, 83 insertions(+), 12 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/media/test-drivers/vimc/vimc-streamer.c
>>>>>> b/drivers/media/test-drivers/vimc/vimc-streamer.c
>>>>>> index c1644d69686d..cc40ecabe95a 100644
>>>>>> --- a/drivers/media/test-drivers/vimc/vimc-streamer.c
>>>>>> +++ b/drivers/media/test-drivers/vimc/vimc-streamer.c
>>>>>> @@ -40,33 +40,30 @@ static void
>>>>>> vimc_streamer_pipeline_terminate(struct vimc_stream *stream)
>>>>>>    }
>>>>>>    /**
>>>>>> - * vimc_streamer_pipeline_init - Initializes the stream structure
>>>>>> + * vimc_streamer_stream_start - Starts streaming for all entities
>>>>>> + * in a stream
>>>>>>     *
>>>>>> - * @stream: the pointer to the stream structure to be initialized
>>>>>>     * @ved:    the pointer to the vimc entity initializing the stream
>>>>>>     *
>>>>>> - * Initializes the stream structure. Walks through the entity graph to
>>>>>> - * construct the pipeline used later on the streamer thread.
>>>>>> - * Calls vimc_streamer_s_stream() to enable stream in all entities of
>>>>>> - * the pipeline.
>>>>>> + * Walks through the entity graph to call vimc_streamer_s_stream()
>>>>>> + * to enable stream in all entities in path of a stream.
>>>>>>     *
>>>>>>     * Return: 0 if success, error code otherwise.
>>>>>>     */
>>>>>> -static int vimc_streamer_pipeline_init(struct vimc_stream *stream,
>>>>>> -                       struct vimc_ent_device *ved)
>>>>>> +static int vimc_streamer_stream_start(struct vimc_stream *stream,
>>>>>> +                      struct vimc_ent_device *ved)
>>>>>>    {
>>>>>>        struct media_entity *entity;
>>>>>>        struct video_device *vdev;
>>>>>>        struct v4l2_subdev *sd;
>>>>>> +    int stream_size = 0;
>>>>>>        int ret = 0;
>>>>>> -    stream->pipe_size = 0;
>>>>>> -    while (stream->pipe_size < VIMC_STREAMER_PIPELINE_MAX_SIZE) {
>>>>>> +    while (stream_size < VIMC_STREAMER_PIPELINE_MAX_SIZE) {
>>>>>>            if (!ved) {
>>>>>>                vimc_streamer_pipeline_terminate(stream);
>>>>>>                return -EINVAL;
>>>>>>            }
>>>>>> -        stream->ved_pipeline[stream->pipe_size++] = ved;
>>>>>>            if (is_media_entity_v4l2_subdev(ved->ent)) {
>>>>>>                sd = media_entity_to_v4l2_subdev(ved->ent);
>>>>>> @@ -104,6 +101,73 @@ static int vimc_streamer_pipeline_init(struct
>>>>>> vimc_stream *stream,
>>>>>>                            entity);
>>>>>>                ved = video_get_drvdata(vdev);
>>>>>>            }
>>>>>> +        stream_size++;
>>>>>> +    }
>>>>>> +
>>>>>> +    vimc_streamer_pipeline_terminate(stream);
>>>>>> +    return -EINVAL;
>>>>>> +}
>>>>>> +
>>>>>> +/**
>>>>>> + * vimc_streamer_pipeline_init - Initialises pipeline and pipe size
>>>>>> + *
>>>>>> + * @stream: the pointer to the stream structure
>>>>>> + * @ved:    the pointer to the vimc entity initializing the stream
>>>>>> pipeline
>>
>> Which entity is this? Is it the start, or the end of the pipeline? I.e.
>> the sensor? or the capture ?
>
> It is the capture, I will add it to the documentation..thanks
>
>>
>>>>>> + *
>>>>>> + * Walks through the entity graph to initialise ved_pipeline and
>>>>>> updates
>>>>>> + * pipe_size too.
>>>>>> + *
>>>>>> + * Return: 0 if success, error code otherwise.
>>>>>> + */
>>>>>> +static int vimc_streamer_pipeline_init(struct vimc_stream *stream,
>>>>>> +                       struct vimc_ent_device *ved)
>>>>>> +{
>>>>>> +    struct media_entity *entity;
>>>>>> +    struct media_device *mdev;
>>>>>> +    struct media_graph graph;
>>>>>> +    struct video_device *vdev;
>>>>>> +    struct v4l2_subdev *sd;
>>>>>> +    int ret;
>>>>>> +
>>>>>> +    entity = ved->ent;
>>>>>> +    mdev = entity->graph_obj.mdev;
>>>>>> +
>>>>>> +    ret = media_graph_walk_init(&graph, mdev);
>>>>>> +    if (ret)
>>>>>> +        return ret;
>>>>>> +
>>>>>> +    media_graph_walk_start(&graph, entity);
>>>>>> +
>>>>>> +    /*
>>>>>> +     * Start pipeline array initialisation from RAW Capture only to
>>>>>> get
>>>>>> +     * entities in the correct order of their frame processing.
>>>>>> +     */
>>>>>> +    if (!strncmp(entity->name, "RGB", 3)) {
>>>>>
>>>>> I don't understand this condition, way is it good for?
>>>>
>>>> I have explained that later in the reply
>>
>> Matching on entity names is a bit awkward, as they could be changed I
>> guess quite easily, and easily missed in this string matching.
>
> Agreed, I need to think of a better way to prevent this
>
>>
>> I haven't fully understood this code block yet to work out if there's
>> another way we could do this though, but reading ahead I see there might
>> be a way to 'walk the graph' on a per-stream basis which might be a good
>> way of factoring this path.
>>
>> Although there still needs to be a construction of the paths available
>> to a stream which splits from the sensor.
>>
>>
>>>>
>>>>>
>>>>> I think the function should be generic and not assume names of entities
>>>>> or specific topology.
>>>>
>>>> It doesn't assume the topology, rather it is in place just to make sure
>>>> that the entities in ved_pipeline are in correct order.
>>>>
>>>>>
>>>>>
>>>>>> +        entity = media_graph_walk_next(&graph);
>>>>>> +        mdev = entity->graph_obj.mdev;
>>>>>> +        media_graph_walk_cleanup(&graph);
>>>>>> +
>>>>>> +        ret = media_graph_walk_init(&graph, mdev);
>>>>>> +        if (ret)
>>>>>> +            return ret;
>>>>>> +        media_graph_walk_start(&graph, entity);
>>>
>>> Hi, can you explain what this code does?
>>> Why does it start the search in the next entity?
>>>
>>>>>> +    }
>>>>>> +
>>>>>> +    while (stream->pipe_size < VIMC_STREAMER_PIPELINE_MAX_SIZE) {
>>>>>> +        if (is_media_entity_v4l2_subdev(entity)) {
>>>>>> +            sd = media_entity_to_v4l2_subdev(entity);
>>>>>> +            ved = v4l2_get_subdevdata(sd);
>>>>>> +        } else {
>>>>>> +            vdev = container_of(entity, struct video_device, entity);
>>>>>> +            ved = video_get_drvdata(vdev);
>>>>>> +        }
>>>>>> +        stream->ved_pipeline[stream->pipe_size++] = ved;
>>>>>> +        entity = media_graph_walk_next(&graph);
>>>>>> +
>>>>>> +        if (!strcmp(entity->name,
>>>>>> stream->ved_pipeline[0]->ent->name)) {
>>>>>
>>>>> I also don't understand this condition
>>>>>
>>>>>> +            media_graph_walk_cleanup(&graph);
>>>>>> +            return 0;
>>>>>> +        }
>>>>>>        }
>>>>>
>>>>> It is not clear what this function does, it looks like it adds to
>>>>> 'ved_pipeline'
>>>>> all entities that are connected to the video node, in addition to the
>>>>> entities
>>>>> that where there from previous calls, so some entities appear several
>>>>> times.
>>>>
>>>> This function returns all the entities in a pipe, weather they are
>>>> streaming or not. For example, if only the RAW Capture 1 streams, or
>>>> RGB/YUB capture streams, or both stream..in all three cases ved_pipeline
>>>> will have the same entities, in exactly same order, and all will occur
>>>> just once.
>>>> Since media_graph_walk goes to each node of the graph, it returns back
>>>> to the first one (as its a graph), hence the last condition, ie,
>>>>
>>>>     if (!strcmp(entity->name, stream->ved_pipeline[0]->ent->name)) {
>>>>
>>>> makes sure that the first entity is not added again to the array.
>>>>
>>>> First condition, ie
>>>>
>>>>     if (!strncmp(entity->name, "RGB", 3)) {
>>>>
>>>> Just makes sure that the search starts at RGB capture only. This is
>>>> because, in case of any other video node, the order of entities, as you
>>>> have mentioned later in the mail, will not be desirable, ie it won't
>>>> start at sensor and end at captures. So this condition just takes care
>>>> that all the entities in ved_pipeline array are in correct order
>>>> (starting at sensor, ending at captures).
>>>
>>> It is better to compare to the whole name of the entity to make it more
>>> clear.
>>> Also I think it is better to document that this function is called only
>>> upon the
>>> first streaming node.
>>
>> If this doesn't end up refactored with other helpers, then indeed a few
>> comments might help the readabilty here. The distinctions of each
>> re-initialisation of the graph walk are hard to interpret the purpose.
>>
>>
>>
>>>
>>> I still think this function should be independent of the topology.
>>> Maybe you can use Helen's patch that allow walking a graph only opposite
>>> to the link direction: https://patchwork.kernel.org/patch/11564899/
>>> This ensures that the Sensor will be first in the graph walk. Then the
>>> streaming thread
>>> iterates the ved_pipeline from 0 upwards and not from 'pipe_size'
>>> downwards.
>>
>> Being able to use a direct helper to walk the pipeline cleanly sounds
>> promising indeed.
>>
>> I suspect at some point int he next patches though - I'm going to see
>> something that needs to have full visibility of all paths enabled from
>> the sensor, as I think I recall that the thread will then process all
>> (enabled) entities in a single pass.
>
> Yes, that exactly is the problem :( This helper (that dafna has shared)
> can walk through one path, while the thread (which processes the frames)
> needs to view all the entites in all connected paths

Hi, what I think would be a good solution is to first to change the type of ved_pipeline
from an array to a list:
""
- struct vimc_ent_device *ved_pipeline[VIMC_STREAMER_PIPELINE_MAX_SIZE];
+ struct list_head ved_pipeline;
""

then you can use Helen's patch to iterate the topology in a depth first form only
from sink to source and if an entity is already streaming then you don't add it to the
ved_pipeline list (since it is already there). Adding to the ved_pipeline list
is always to the end of the list. This ensures that for each entity, its source entity
already processed its frame.

This way, you iterate not the entities of the whole pipe but only the ones that are streaming.
Let's look at two scenarios:

1) "RGB/YUV Capture" start streaming and afterwards "Raw Capture 0" start streaming:
After the 'start_stream" of the "RGB/YUV Capture" , the ved_pipeline list is:

Sensor A => Debayer A => Scaler => "RGB/YUV Capture"

After the start_stream of the "Raw Capture 0" the ved_pipeline list is:

Sensor A => Debayer A => Scaler => "RGB/YUV Capture" => "Raw Capture 0"

2) "Raw Capture 0" starts streaming and afterwards "RGB/YUV Capture" start streaming:
After the 'start_stream" of the "Raw Capture 0" , the ved_pipeline list is:

Sensor A => "Raw Capture 0"

After the start_stream of the "RGB/YUV Capture" the ved_pipeline list is:

Sensor A => "Raw Capture 0" => Debayer A => Scaler => "RGB/YUV Capture"

Thanks,
Dafna

>
>> --
>> Kieran
>>
>>
>>>
>>> Thanks,
>>> Dafna
>>>
>>>
>>>
>>>>
>>>> Thanks,
>>>> Kaaira
>>>>
>>>>>
>>>>> I think there is no need to use the graph walk here but to access the
>>>>> source entity
>>>>> in each iteration, the way done in vimc_streamer_stream_start
>>>>> also.
>>>>> I think the code should iterate here until it reaches an entity that
>>>>> is already streaming,
>>>>> this means that the entity is already in the `ved_pipeline`, also you
>>>>> should make sure
>>>>> that the sensor is the first entity that process a frame, therefore
>>>>> the sensor should be
>>>>> at the end/start of the list of entities. Generally each entity
>>>>> should appear exactly once
>>>>> in the 'ved_pipeline' array and the entities should be ordered such
>>>>> that when calling 'process_frame'
>>>>> on one entity should be after calling 'process_frame' on its source
>>>>> entity.
>>>>> maybe it is easyer to implement if 'ved_pipeline' is a linked list.
>>>>>
>>>>> Thanks,
>>>>> Dafna
>>>>>
>>>>>>        vimc_streamer_pipeline_terminate(stream);
>>>>>> @@ -138,8 +202,11 @@ static int vimc_streamer_thread(void *data)
>>>>>>            for (i = stream->pipe_size - 1; i >= 0; i--) {
>>>>>>                ved = stream->ved_pipeline[i];
>>>>>> -            ret = ved->process_frame(ved);
>>>>>> +            if (atomic_read(&ved->use_count) == 0)
>>>>>> +                continue;
>>>>>> +
>>>>>> +            ret = ved->process_frame(ved);
>>>>>>                if (ret)
>>>>>>                    break;
>>>>>>            }
>>>>>> @@ -179,6 +246,10 @@ int vimc_streamer_s_stream(struct vimc_stream
>>>>>> *stream,
>>>>>>            if (stream->kthread)
>>>>>>                return 0;
>>>>>> +        ret = vimc_streamer_stream_start(stream, ved);
>>>>>> +        if (ret)
>>>>>> +            return ret;
>>>>>> +
>>>>>>            ret = vimc_streamer_pipeline_init(stream, ved);
>>>>>>            if (ret)
>>>>>>                return ret;
>>>>>>
>>
>> --
>> Regards
>> --
>> Kieran