ffmpeg 内存池AVBufferPool

1、首先参考：FFmpeg5.0源码阅读——内存池AVBufferPool
2、链表是怎么串起来的？
看下面三个核心函数：

static void pool_release_buffer(void *opaque, uint8_t *data)
{
    BufferPoolEntry *buf = opaque;
    AVBufferPool *pool = buf->pool;

    if(CONFIG_MEMORY_POISONING)
        memset(buf->data, FF_MEMORY_POISON, pool->size);

    ff_mutex_lock(&pool->mutex);
    buf->next = pool->pool;
    pool->pool = buf;
    ff_mutex_unlock(&pool->mutex);

    if (atomic_fetch_sub_explicit(&pool->refcount, 1, memory_order_acq_rel) == 1)
        buffer_pool_free(pool);
}

/* allocate a new buffer and override its free() callback so that
 * it is returned to the pool on free */
static AVBufferRef *pool_alloc_buffer(AVBufferPool *pool)
{
    BufferPoolEntry *buf;
    AVBufferRef     *ret;

    av_assert0(pool->alloc || pool->alloc2);

    ret = pool->alloc2 ? pool->alloc2(pool->opaque, pool->size) :
                         pool->alloc(pool->size);
    if (!ret)
        return NULL;

    buf = av_mallocz(sizeof(*buf));
    if (!buf) {
        av_buffer_unref(&ret);
        return NULL;
    }

    buf->data   = ret->buffer->data;
    buf->opaque = ret->buffer->opaque;
    buf->free   = ret->buffer->free;
    buf->pool   = pool;

    ret->buffer->opaque = buf;
    ret->buffer->free   = pool_release_buffer;

    return ret;
}

AVBufferRef *av_buffer_pool_get(AVBufferPool *pool)
{
    AVBufferRef *ret;
    BufferPoolEntry *buf;

    ff_mutex_lock(&pool->mutex);
    buf = pool->pool;
    if (buf) {
        ret = av_buffer_create(buf->data, pool->size, pool_release_buffer,
                               buf, 0);
        if (ret) {
            pool->pool = buf->next;
            buf->next = NULL;
        }
    } else {
        ret = pool_alloc_buffer(pool);
    }
    ff_mutex_unlock(&pool->mutex);

    if (ret)
        atomic_fetch_add_explicit(&pool->refcount, 1, memory_order_relaxed);

    return ret;
}

其中用户调用av_buffer_pool_get() -> pool_alloc_buffer()
如果是第一次创建的时候：

int av_hwframe_ctx_init(AVBufferRef *ref)
{
    AVHWFramesContext *ctx = (AVHWFramesContext*)ref->data;
    const enum AVPixelFormat *pix_fmt;
    int ret;

    if (ctx->internal->source_frames) {
        /* A derived frame context is already initialised. */
        return 0;
    }

    /* validate the pixel format */
    for (pix_fmt = ctx->internal->hw_type->pix_fmts; *pix_fmt != AV_PIX_FMT_NONE; pix_fmt++) {
        if (*pix_fmt == ctx->format)
            break;
    }
    if (*pix_fmt == AV_PIX_FMT_NONE) {
        av_log(ctx, AV_LOG_ERROR,
               "The hardware pixel format '%s' is not supported by the device type '%s'\n",
               av_get_pix_fmt_name(ctx->format), ctx->internal->hw_type->name);
        return AVERROR(ENOSYS);
    }

    /* validate the dimensions */
    ret = av_image_check_size(ctx->width, ctx->height, 0, ctx);
    if (ret < 0)
        return ret;

    /* format-specific init */
    if (ctx->internal->hw_type->frames_init) {
        ret = ctx->internal->hw_type->frames_init(ctx);
        if (ret < 0)
            goto fail;
    }

    if (ctx->internal->pool_internal && !ctx->pool)
        ctx->pool = ctx->internal->pool_internal;

    /* preallocate the frames in the pool, if requested */
    if (ctx->initial_pool_size > 0) {
        ret = hwframe_pool_prealloc(ref);
        if (ret < 0)
            goto fail;
    }

    return 0;
fail:
    if (ctx->internal->hw_type->frames_uninit)
        ctx->internal->hw_type->frames_uninit(ctx);
    return ret;
}

hwframe_pool_prealloc（）是第一次创建内存池。

static int hwframe_pool_prealloc(AVBufferRef *ref)
{
    AVHWFramesContext *ctx = (AVHWFramesContext*)ref->data;
    AVFrame **frames;
    int i, ret = 0;

    frames = av_mallocz_array(ctx->initial_pool_size, sizeof(*frames));
    if (!frames)
        return AVERROR(ENOMEM);

    for (i = 0; i < ctx->initial_pool_size; i++) {
        frames[i] = av_frame_alloc();
        if (!frames[i])
            goto fail;

        ret = av_hwframe_get_buffer(ref, frames[i], 0);
        if (ret < 0)
            goto fail;
    }

fail:
    for (i = 0; i < ctx->initial_pool_size; i++)
        av_frame_free(&frames[i]);
    av_freep(&frames);

    return ret;
}

可以看到for循环创建的过程。av_hwframe_get_buffer()

int av_hwframe_get_buffer(AVBufferRef *hwframe_ref, AVFrame *frame, int flags)
{
    AVHWFramesContext *ctx = (AVHWFramesContext*)hwframe_ref->data;
    int ret;

    if (ctx->internal->source_frames) {
        // This is a derived frame context, so we allocate in the source
        // and map the frame immediately.
        AVFrame *src_frame;

        frame->format = ctx->format;
        frame->hw_frames_ctx = av_buffer_ref(hwframe_ref);
        if (!frame->hw_frames_ctx)
            return AVERROR(ENOMEM);

        src_frame = av_frame_alloc();
        if (!src_frame)
            return AVERROR(ENOMEM);

        ret = av_hwframe_get_buffer(ctx->internal->source_frames,
                                    src_frame, 0);
        if (ret < 0) {
            av_frame_free(&src_frame);
            return ret;
        }

        ret = av_hwframe_map(frame, src_frame,
                             ctx->internal->source_allocation_map_flags);
        if (ret) {
            av_log(ctx, AV_LOG_ERROR, "Failed to map frame into derived "
                   "frame context: %d.\n", ret);
            av_frame_free(&src_frame);
            return ret;
        }

        // Free the source frame immediately - the mapped frame still
        // contains a reference to it.
        av_frame_free(&src_frame);

        return 0;
    }

    if (!ctx->internal->hw_type->frames_get_buffer)
        return AVERROR(ENOSYS);

    if (!ctx->pool)
        return AVERROR(EINVAL);

    frame->hw_frames_ctx = av_buffer_ref(hwframe_ref);
    if (!frame->hw_frames_ctx)
        return AVERROR(ENOMEM);

    ret = ctx->internal->hw_type->frames_get_buffer(ctx, frame);
    if (ret < 0) {
        av_buffer_unref(&frame->hw_frames_ctx);
        return ret;
    }

    frame->extended_data = frame->data;

    return 0;
}

其中会调用用户设置的一个回调函数：
hw_type->frames_get_buffe(),这里我们看英伟达的解码器：

static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
{
    CUDAFramesContext *priv = ctx->internal->priv;
    int res;

    frame->buf[0] = av_buffer_pool_get(ctx->pool);
    if (!frame->buf[0])
        return AVERROR(ENOMEM);

    res = av_image_fill_arrays(frame->data, frame->linesize, frame->buf[0]->data,
                               ctx->sw_format, ctx->width, ctx->height, priv->tex_alignment);
    if (res < 0)
        return res;

    // YUV420P is a special case.
    // Nvenc expects the U/V planes in swapped order from how ffmpeg expects them, also chroma is half-aligned
    if (ctx->sw_format == AV_PIX_FMT_YUV420P) {
        frame->linesize[1] = frame->linesize[2] = frame->linesize[0] / 2;
        frame->data[2]     = frame->data[1];
        frame->data[1]     = frame->data[2] + frame->linesize[2] * (ctx->height / 2);
    }

    frame->format = AV_PIX_FMT_CUDA;
    frame->width  = ctx->width;
    frame->height = ctx->height;

    return 0;
}

所以真正调用的是av_buffer_pool_get()，这里就和我们最开始的代码一致了。
av_buffer_pool_get（）在第一次创建了一个BufferPoolEntry，在pool_release_buffer()的时候串起来，注意，这里是个单链表：

static void pool_release_buffer(void *opaque, uint8_t *data)
{
    BufferPoolEntry *buf = opaque;
    AVBufferPool *pool = buf->pool;

    if(CONFIG_MEMORY_POISONING)
        memset(buf->data, FF_MEMORY_POISON, pool->size);

    ff_mutex_lock(&pool->mutex);
    buf->next = pool->pool;
    //这里是链表头，也就是最后一个release的是头
    pool->pool = buf;
    ff_mutex_unlock(&pool->mutex);

    if (atomic_fetch_sub_explicit(&pool->refcount, 1, memory_order_acq_rel) == 1)
        buffer_pool_free(pool);
}