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

From: Dafna Hirschfeld
Date: Fri Oct 02 2020 - 08:18:08 EST




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 <kgupta@xxxxxxxxxxxxx>
---
   .../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