Memory allocation

Memory allocation and management are very important topics in multimedia. High definition video uses many megabytes to store one single image frame. It is important to reuse memory when possible instead of constantly allocating and freeing it.

Multimedia systems usually use special-purpose chips, such as DSPs or GPUs to perform the heavy lifting (especially for video). These special-purpose chips usually have strict requirements for the memory they operate on and how it is accessed.

This chapter talks about the memory-management features available to GStreamer plugins. We will first talk about the lowlevel GstMemory object that manages access to a piece of memory and then continue with one of it's main users, the GstBuffer, which is used to exchange data between plugins and with the application. We will also discuss the GstMeta. This object can be placed on buffers to provide extra info about it and its memory. We will also discuss the GstBufferPool, which allows to more-efficiently manage buffers of the same size.

To conclude this chapter we will take a look at the GST_QUERY_ALLOCATION query, which is used to negotiate memory management options between elements.

GstMemory

GstMemory is an object that manages a region of memory. This memory object points to a region of memory of “maxsize”. The area in this memory starting at “offset” and size “size” bytes is the accessible memory region. After a GstMemory is created its maxsize can no longer be changed, however, its "offset" and "size" can.

GstAllocator

GstMemory objects are created by a GstAllocator object. Most allocators implement the default gst_allocator_alloc() method but some might implement different ones, for example, when additional parameters are needed to allocate the specific memory.

Different allocators exist for system memory, shared memory and memory backed by a DMAbuf file descriptor. To implement support for a new kind of memory type, you must implement a new allocator object.

GstMemory API example

Data access to the memory wrapped by the GstMemory object is always protected with a gst_memory_map() and gst_memory_unmap() pair. An access mode (read/write) must be given when mapping memory. The map function returns a pointer to the valid memory region that can then be accessed according to the requested access mode.

Below is an example on creating a GstMemory object and using the gst_memory_map() to access the memory region.

[...]

  GstMemory *mem;
  GstMapInfo info;
  gint i;

  /* allocate 100 bytes */
  mem = gst_allocator_alloc (NULL, 100, NULL);

  /* get access to the memory in write mode */
  gst_memory_map (mem, &info, GST_MAP_WRITE);

  /* fill with pattern */
  for (i = 0; i < info.size; i++)
    info.data[i] = i;

  /* release memory */
  gst_memory_unmap (mem, &info);

[...]


Implementing a GstAllocator

WRITEME

GstBuffer

A GstBuffer is a lightweight object that is passed from an upstream to a downstream element and contains memory and metadata. It represents the multimedia content that is pushed to or pulled by downstream elements.

A GstBuffer contains one or more GstMemory objects. These objects hold the buffer's data.

Metadata in the buffer consists of:

  • DTS and PTS timestamps. These represent the decoding and presentation timestamps of the buffer content and are used by synchronizing elements to schedule buffers. These timestamps can be GST_CLOCK_TIME_NONE when unknown/undefined.

  • The duration of the buffer contents. This duration can be GST_CLOCK_TIME_NONE when unknown/undefined.

  • Media-specific offset and offset_end values. For video this is the frame number in the stream, for audio, the sample number. Other media might use different definitions.

  • Arbitrary structures via GstMeta, see below.

Writability

A GstBuffer is writable when the refcount of the object is exactly 1, meaning that only one object is holding a ref to the buffer. You can only modify the buffer when it is writable. This means that you need to call gst_buffer_make_writable() before changing the timestamps, offsets, metadata or adding and removing memory blocks.

API examples

You can create a GstBuffer with gst_buffer_new () and then you can add memory objects to it. You can alternatively use the convenience function gst_buffer_new_allocate () to perform both operations at once. It's also possible to wrap existing memory with gst_buffer_new_wrapped_full () and specify the function to call when the memory should be freed.

You can access the memory of a GstBuffer by getting and mapping the GstMemory objects individually or by using gst_buffer_map (). The latter merges all the memory into one big block and then gives you a pointer to it.

Below is an example of how to create a buffer and access its memory.

[...]
  GstBuffer *buffer;
  GstMemory *mem;
  GstMapInfo info;

  /* make empty buffer */
  buffer = gst_buffer_new ();

  /* make memory holding 100 bytes */
  mem = gst_allocator_alloc (NULL, 100, NULL);

  /* add the buffer */
  gst_buffer_append_memory (buffer, mem);

[...]

  /* get WRITE access to the memory and fill with 0xff */
  gst_buffer_map (buffer, &info, GST_MAP_WRITE);
  memset (info.data, 0xff, info.size);
  gst_buffer_unmap (buffer, &info);

[...]

  /* free the buffer */
  gst_buffer_unref (buffer);

[...]

GstMeta

With the GstMeta system you can add arbitrary structures to buffers. These structures describe extra properties of the buffer such as cropping, stride, region of interest, etc.

The metadata system separates API specification (what the metadata and its API look like) and its implementation (how it works). This makes it possible to have different implementations of the same API, for example, depending on the hardware you are running on.

API example

After allocating a new GstBuffer, you can add metadata to it with the metadata-specific API. This means that you will need to link to the header file where the metadata is defined to use its API.

By convention, a metadata API with name FooBar should provide two methods, a gst_buffer_add_foo_bar_meta () and a gst_buffer_get_foo_bar_meta (). Both functions should return a pointer to a FooBarMeta structure that contains the metadata fields. Some of the _add_*_meta () can have extra parameters that will usually be used to configure the metadata structure for you.

Let's have a look at the metadata that is used to specify a cropping region for video frames.

#include <gst/video/gstvideometa.h>

[...]
  GstVideoCropMeta *meta;

  /* buffer points to a video frame, add some cropping metadata */
  meta = gst_buffer_add_video_crop_meta (buffer);

  /* configure the cropping metadata */
  meta->x = 8;
  meta->y = 8;
  meta->width = 120;
  meta->height = 80;
[...]

An element can then use the metadata on the buffer when rendering the frame like this:

#include <gst/video/gstvideometa.h>

[...]
  GstVideoCropMeta *meta;

  /* buffer points to a video frame, get the cropping metadata */
  meta = gst_buffer_get_video_crop_meta (buffer);

  if (meta) {
    /* render frame with cropping */
    _render_frame_cropped (buffer, meta->x, meta->y, meta->width, meta->height);
  } else {
    /* render frame */
    _render_frame (buffer);
  }
[...]

Implementing new GstMeta

In the next sections we show how you can add new metadata to the system and use it on buffers.

Define the metadata API

First we need to define what our API will look like and we have to register this API to the system. This is important because the API definition will be used when elements negotiate what kind of metadata they will exchange. The API definition also contains arbitrary tags that give hints about what the metadata contains. This is important when we see how metadata is preserved as buffers pass through the pipeline.

If you are making a new implementation of an existing API, you can skip this step and move directly to the implementation.

First we start with making the my-example-meta.h header file that will contain the definition of the API and structure for our metadata.

#include <gst/gst.h>

typedef struct _MyExampleMeta MyExampleMeta;

struct _MyExampleMeta {
  GstMeta       meta;

  gint          age;
  gchar        *name;
};

GType my_example_meta_api_get_type (void);
#define MY_EXAMPLE_META_API_TYPE (my_example_meta_api_get_type())

#define gst_buffer_get_my_example_meta(b) \
  ((MyExampleMeta*)gst_buffer_get_meta((b),MY_EXAMPLE_META_API_TYPE))

The metadata API definition consists of the definition of the structure that holds a gint and a string. The first field in the structure must be a GstMeta.

We also define a my_example_meta_api_get_type () function that will register our metadata API definition and a convenience gst_buffer_get_my_example_meta () macro that simply finds and returns the metadata with our new API.

Let's have a look at how the my_example_meta_api_get_type () function is implemented in the my-example-meta.c file:

#include "my-example-meta.h"

GType
my_example_meta_api_get_type (void)
{
  static volatile GType type;
  static const gchar *tags[] = { "foo", "bar", NULL };

  if (g_once_init_enter (&type)) {
    GType _type = gst_meta_api_type_register ("MyExampleMetaAPI", tags);
    g_once_init_leave (&type, _type);
  }
  return type;
}

As you can see, it simply uses the gst_meta_api_type_register () function to register a name and some tags for the API. The result is a new GType pointer that defines the newly registered API.

Implementing a metadata API

Next we can make an implementation for a registered metadata API GType.

The implementation details of a metadata API are kept in a GstMetaInfo structure that you make available to the users of your metadata API implementation with a my_example_meta_get_info () function and a convenience MY_EXAMPLE_META_INFO macro. You also provide a method to add your metadata implementation to a GstBuffer. Your my-example-meta.h header file will need these additions:

[...]

/* implementation */
const GstMetaInfo *my_example_meta_get_info (void);
#define MY_EXAMPLE_META_INFO (my_example_meta_get_info())

MyExampleMeta * gst_buffer_add_my_example_meta (GstBuffer      *buffer,
                                                gint            age,
                                                const gchar    *name);

Let's have a look at how these functions are implemented in the my-example-meta.c file.

[...]

static gboolean
my_example_meta_init (GstMeta * meta, gpointer params, GstBuffer * buffer)
{
  MyExampleMeta *emeta = (MyExampleMeta *) meta;

  emeta->age = 0;
  emeta->name = NULL;

  return TRUE;
}

static gboolean
my_example_meta_transform (GstBuffer * transbuf, GstMeta * meta,
    GstBuffer * buffer, GQuark type, gpointer data)
{
  MyExampleMeta *emeta = (MyExampleMeta *) meta;

  /* we always copy no matter what transform */
  gst_buffer_add_my_example_meta (transbuf, emeta->age, emeta->name);

  return TRUE;
}

static void
my_example_meta_free (GstMeta * meta, GstBuffer * buffer)
{
  MyExampleMeta *emeta = (MyExampleMeta *) meta;

  g_free (emeta->name);
  emeta->name = NULL;
}

const GstMetaInfo *
my_example_meta_get_info (void)
{
  static const GstMetaInfo *meta_info = NULL;

  if (g_once_init_enter (&meta_info)) {
    const GstMetaInfo *mi = gst_meta_register (MY_EXAMPLE_META_API_TYPE,
        "MyExampleMeta",
        sizeof (MyExampleMeta),
        my_example_meta_init,
        my_example_meta_free,
        my_example_meta_transform);
    g_once_init_leave (&meta_info, mi);
  }
  return meta_info;
}

MyExampleMeta *
gst_buffer_add_my_example_meta (GstBuffer   *buffer,
                                gint         age,
                                const gchar *name)
{
  MyExampleMeta *meta;

  g_return_val_if_fail (GST_IS_BUFFER (buffer), NULL);

  meta = (MyExampleMeta *) gst_buffer_add_meta (buffer,
      MY_EXAMPLE_META_INFO, NULL);

  meta->age = age;
  meta->name = g_strdup (name);

  return meta;
}

gst_meta_register () registers the implementation details, like the API that you implement and the size of the metadata structure, alongside methods to initialize and free the memory area. You can also implement a transform function that will be called when a certain transformation (identified by the quark and quark specific data) is performed on a buffer.

Lastly, you implement a gst_buffer_add_*_meta() that adds the metadata implementation to a buffer and sets the values of the metadata.

GstBufferPool

The GstBufferPool object provides a convenient base class for managing lists of reusable buffers. Essential for this object is that all the buffers have the same properties such as size, padding, metadata and alignment.

A GstBufferPool can be configured to manage a minimum and maximum amount of buffers of a specific size. It can also be configured to use a specific GstAllocator for the memory of the buffers. There is also support in the bufferpool to enable bufferpool specific options, such as adding GstMeta to the pool's buffers or enabling specific padding on the buffers' memory.

A GstBufferPool can be either inactivate or active. In the inactive state, you can configure the pool. In the active state, you can't change the configuration anymore but you can acquire and release buffers from/to the pool.

In the following sections we take a look at how you can use a GstBufferPool.

API example

There can be many different GstBufferPool implementations; they are all subclasses of the GstBufferPool base class. For this example, we will assume we somehow have access to a buffer pool, either because we created it ourselves or because we were given one as a result of the ALLOCATION query, as we will see below.

The GstBufferPool is initially in the inactive state so that we can configure it. Trying to configure a GstBufferPool that is not in the inactive state will fail. Likewise, trying to activate a bufferpool that is not configured will also fail.

  GstStructure *config;

[...]

  /* get config structure */
  config = gst_buffer_pool_get_config (pool);

  /* set caps, size, minimum and maximum buffers in the pool */
  gst_buffer_pool_config_set_params (config, caps, size, min, max);

  /* configure allocator and parameters */
  gst_buffer_pool_config_set_allocator (config, allocator, &params);

  /* store the updated configuration again */
  gst_buffer_pool_set_config (pool, config);

[...]

The configuration of a GstBufferPool is maintained in a generic GstStructure that can be obtained with gst_buffer_pool_get_config(). Convenience methods exist to get and set the configuration options in this structure. After updating the structure, it is set as the current configuration in the GstBufferPool again with gst_buffer_pool_set_config().

The following options can be configured on a GstBufferPool:

  • The caps of the buffers to allocate.

  • The size of the buffers. This is the suggested size of the buffers in the pool. The pool might decide to allocate larger buffers to add padding.

  • The minimum and maximum amount of buffers in the pool. When minimum is set to \> 0, the bufferpool will pre-allocate this amount of buffers. When maximum is not 0, the bufferpool will allocate up to maximum amount of buffers.

  • The allocator and parameters to use. Some bufferpools might ignore the allocator and use its internal one.

  • Other arbitrary bufferpool options identified with a string. a bufferpool lists the supported options with gst_buffer_pool_get_options() and you can ask if an option is supported with gst_buffer_pool_has_option(). The option can be enabled by adding it to the configuration structure with gst_buffer_pool_config_add_option (). These options are used to enable things like letting the pool set metadata on the buffers or to add extra configuration options for padding, for example.

After the configuration is set on the bufferpool, the pool can be activated with gst_buffer_pool_set_active (pool, TRUE). From that point on you can use gst_buffer_pool_acquire_buffer () to retrieve a buffer from the pool, like this:

  [...]

  GstFlowReturn ret;
  GstBuffer *buffer;

  ret = gst_buffer_pool_acquire_buffer (pool, &buffer, NULL);
  if (G_UNLIKELY (ret != GST_FLOW_OK))
    goto pool_failed;

  [...]

It is important to check the return value of the acquire function because it is possible that it fails: When your element shuts down, it will deactivate the bufferpool and then all calls to acquire will return GST_FLOW_FLUSHING.

All buffers that are acquired from the pool will have their pool member set to the original pool. When the last ref is decremented on the buffer, GStreamer will automatically call gst_buffer_pool_release_buffer() to release the buffer back to the pool. You (or any other downstream element) don't need to know if a buffer came from a pool, you can just unref it.

Implementing a new GstBufferPool

WRITEME

GST_QUERY_ALLOCATION

The ALLOCATION query is used to negotiate GstMeta, GstBufferPool and GstAllocator between elements. Negotiation of the allocation strategy is always initiated and decided by a srcpad after it has negotiated a format and before it decides to push buffers. A sinkpad can suggest an allocation strategy but it is ultimately the source pad that will decide based on the suggestions of the downstream sink pad.

The source pad will do a GST_QUERY_ALLOCATION with the negotiated caps as a parameter. This is needed so that the downstream element knows what media type is being handled. A downstream sink pad can answer the allocation query with the following results:

  • An array of possible GstBufferPool suggestions with suggested size, minimum and maximum amount of buffers.

  • An array of GstAllocator objects along with suggested allocation parameters such as flags, prefix, alignment and padding. These allocators can also be configured in a bufferpool when this is supported by the bufferpool.

  • An array of supported GstMeta implementations along with metadata specific parameters. It is important that the upstream element knows what kind of metadata is supported downstream before it places that metadata on buffers.

When the GST_QUERY_ALLOCATION returns, the source pad will select from the available bufferpools, allocators and metadata how it will allocate buffers.

ALLOCATION query example

Below is an example of the ALLOCATION query.

#include <gst/video/video.h>
#include <gst/video/gstvideometa.h>
#include <gst/video/gstvideopool.h>

  GstCaps *caps;
  GstQuery *query;
  GstStructure *structure;
  GstBufferPool *pool;
  GstStructure *config;
  guint size, min, max;

[...]

  /* find a pool for the negotiated caps now */
  query = gst_query_new_allocation (caps, TRUE);

  if (!gst_pad_peer_query (scope->srcpad, query)) {
    /* query failed, not a problem, we use the query defaults */
  }

  if (gst_query_get_n_allocation_pools (query) > 0) {
    /* we got configuration from our peer, parse them */
    gst_query_parse_nth_allocation_pool (query, 0, &pool, &size, &min, &max);
  } else {
    pool = NULL;
    size = 0;
    min = max = 0;
  }

  if (pool == NULL) {
    /* we did not get a pool, make one ourselves then */
    pool = gst_video_buffer_pool_new ();
  }

  config = gst_buffer_pool_get_config (pool);
  gst_buffer_pool_config_add_option (config, GST_BUFFER_POOL_OPTION_VIDEO_META);
  gst_buffer_pool_config_set_params (config, caps, size, min, max);
  gst_buffer_pool_set_config (pool, config);

  /* and activate */
  gst_buffer_pool_set_active (pool, TRUE);

[...]

This particular implementation will make a custom GstVideoBufferPool object that is specialized in allocating video buffers. You can also enable the pool to put GstVideoMeta metadata on the buffers from the pool doing:

gst_buffer_pool_config_add_option (config, GST_BUFFER_POOL_OPTION_VIDEO_META)

The ALLOCATION query in base classes

In many base classes you will see the following virtual methods for influencing the allocation strategy:

  • propose_allocation () should suggest allocation parameters for the upstream element.

  • decide_allocation () should decide the allocation parameters from the suggestions received from downstream.

Implementors of these methods should modify the given GstQuery object by updating the pool options and allocation options.

Negotiating the exact layout of video buffers

Hardware elements may have specific constraints on the layout of their input buffers, requiring to add vertical and/or horizontal paddings to their planes. If the producer is able to create buffers fulfilling these requirements, we can ensure zero-copy by configuring its driver accordingly before starting to produce buffers.

In such setup on Linux we'll generally use dmabuf to exchange buffers in order to reduce memory copies. The producer can either export its buffers to the consumer (dmabuf export) or import them from it (dmabuf import).

In this section we'll outline the steps for how the consumer can inform the producer of its expected buffer layout for import and export use cases. Let's consider v4l2src (the producer) feeding buffers to omxvideoenc (the consumer) for encoding.

v4l2src importing buffers from omxvideoenc

  1. omxvideoenc: query the hardware for its requirements and create a GstVideoAlignment accordingly.
  2. omxvideoenc: in its buffer pool alloc_buffer implementation, call gst_buffer_add_video_meta_full() and then gst_video_meta_set_alignment() on the returned meta with the requested alignment. The alignment will be added to the meta, allowing v4l2src to configure its driver before trying to import buffers.
      meta = gst_buffer_add_video_meta_full (buf, GST_VIDEO_FRAME_FLAG_NONE,
          GST_VIDEO_INFO_FORMAT (&pool->video_info),
          GST_VIDEO_INFO_WIDTH (&pool->video_info),
          GST_VIDEO_INFO_HEIGHT (&pool->video_info),
          GST_VIDEO_INFO_N_PLANES (&pool->video_info), offset, stride);

      if (gst_omx_video_get_port_padding (pool->port, &pool->video_info,
              &align))
        gst_video_meta_set_alignment (meta, align);
  1. omxvideoenc: propose its pool to the producer when replying to the ALLOCATION query (propose_allocation()).
  2. v4l2src: when receiving the reply from the ALLOCATION query (decide_allocation()) acquire a single buffer from the suggested pool and retrieve its layout using GstVideoMeta.stride and gst_video_meta_get_plane_height().
  3. v4l2src: configure its driver to produce data matching those requirements, if possible, then try to import the buffer. If not, v4l2src won't be able to import from omxvideoenc and so will fallback to sending its own buffers to omxvideoenc which will have to copy each input buffer to fit its requirements.

v4l2src exporting buffers to omxvideoenc

  1. omxvideoenc: query the hardware for its requirements and create a GstVideoAlignment accordingly.
  2. omxvideoenc: create a GstStructure named video-meta serializing the alignment:
params = gst_structure_new ("video-meta",
    "padding-top", G_TYPE_UINT, align.padding_top,
    "padding-bottom", G_TYPE_UINT, align.padding_bottom,
    "padding-left", G_TYPE_UINT, align.padding_left,
    "padding-right", G_TYPE_UINT, align.padding_right,
    NULL);
  1. omxvideoenc: when handling the ALLOCATION query (propose_allocation()), pass this structure as parameter when adding the GST_VIDEO_META_API_TYPE meta:
gst_query_add_allocation_meta (query, GST_VIDEO_META_API_TYPE, params);
  1. v4l2src: when receiving the reply from the ALLOCATION query (decide_allocation()) retrieve the GST_VIDEO_META_API_TYPE parameters to compute the expected buffers layout:
guint video_idx;
GstStructure *params;

if (gst_query_find_allocation_meta (query, GST_VIDEO_META_API_TYPE, &video_idx)) {
  gst_query_parse_nth_allocation_meta (query, video_idx, &params);

  if (params) {
    GstVideoAlignment align;
    GstVideoInfo info;
    gsize plane_size[GST_VIDEO_MAX_PLANES];

    gst_video_alignment_reset (&align);

    gst_structure_get_uint (s, "padding-top", &align.padding_top);
    gst_structure_get_uint (s, "padding-bottom", &align.padding_bottom);
    gst_structure_get_uint (s, "padding-left", &align.padding_left);
    gst_structure_get_uint (s, "padding-right", &align.padding_right);

    gst_video_info_from_caps (&info, caps);

    gst_video_info_align_full (&info, align, plane_size);
  }
}
  1. v4l2src: retrieve the requested buffers layout using GstVideoInfo.stride and GST_VIDEO_INFO_PLANE_HEIGHT().
  2. v4l2src: configure its driver to produce data matching those requirements, if possible. If not, driver will produce buffers using its own layout but omxvideoenc will have to copy each input buffer to fit its requirements.

The results of the search are