h264-v4l2m2m: Remove bogus cleanup

include/h264-encoder.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 - 2022 Andri Yngvason
+ * Copyright (c) 2021 - 2024 Andri Yngvason
  *
  * Permission to use, copy, modify, and/or distribute this software for any
  * purpose with or without fee is hereby granted, provided that the above
@@ -17,13 +17,28 @@
 
 #include <stdint.h>
 #include <unistd.h>
+#include <stdbool.h>
 
-struct h264_encoder;
 struct nvnc_fb;
+struct h264_encoder;
 
 typedef void (*h264_encoder_packet_handler_fn)(const void* payload, size_t size,
 		uint64_t pts, void* userdata);
 
+struct h264_encoder_impl {
+	struct h264_encoder* (*create)(uint32_t width, uint32_t height,
+			uint32_t format, int quality);
+	void (*destroy)(struct h264_encoder*);
+	void (*feed)(struct h264_encoder*, struct nvnc_fb*);
+};
+
+struct h264_encoder {
+	struct h264_encoder_impl *impl;
+	h264_encoder_packet_handler_fn on_packet_ready;
+	void* userdata;
+	bool next_frame_should_be_keyframe;
+};
+
 struct h264_encoder* h264_encoder_create(uint32_t width, uint32_t height,
 		uint32_t format, int quality);
 

meson.build

@@ -138,13 +138,26 @@ if gbm.found()
 	config.set('HAVE_GBM', true)
 endif
 
-if gbm.found() and libdrm.found() and libavcodec.found() and libavfilter.found() and libavutil.found()
-	sources += [ 'src/h264-encoder.c', 'src/open-h264.c' ]
+have_ffmpeg = gbm.found() and libdrm.found() and libavcodec.found() and libavfilter.found() and libavutil.found()
+have_v4l2 = gbm.found() and libdrm.found() and cc.check_header('linux/videodev2.h')
+
+if have_ffmpeg
+	sources += [ 'src/h264-encoder-ffmpeg-impl.c' ]
 	dependencies += [libdrm, libavcodec, libavfilter, libavutil]
-	config.set('ENABLE_OPEN_H264', true)
+	config.set('HAVE_FFMPEG', true)
 	config.set('HAVE_LIBAVUTIL', true)
 endif
 
+if have_v4l2
+	sources += [ 'src/h264-encoder-v4l2m2m-impl.c' ]
+	config.set('HAVE_V4L2', true)
+endif
+
+if have_ffmpeg or have_v4l2
+	sources += [ 'src/h264-encoder.c', 'src/open-h264.c' ]
+	config.set('ENABLE_OPEN_H264', true)
+endif
+
 if enable_websocket
 	sources += [
 		'src/ws-handshake.c',

src/h264-encoder-ffmpeg-impl.c

@@ -0,0 +1,625 @@
+/*
+ * Copyright (c) 2021 - 2024 Andri Yngvason
+ *
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+ * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+ * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+ * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+ * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
+ * OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#include "h264-encoder.h"
+#include "neatvnc.h"
+#include "fb.h"
+#include "sys/queue.h"
+#include "vec.h"
+#include "usdt.h"
+
+#include <stdlib.h>
+#include <stdint.h>
+#include <stdbool.h>
+#include <unistd.h>
+#include <assert.h>
+#include <gbm.h>
+#include <xf86drm.h>
+#include <aml.h>
+
+#include <libavcodec/avcodec.h>
+#include <libavutil/hwcontext.h>
+#include <libavutil/hwcontext_drm.h>
+#include <libavutil/pixdesc.h>
+#include <libavutil/dict.h>
+#include <libavfilter/avfilter.h>
+#include <libavfilter/buffersink.h>
+#include <libavfilter/buffersrc.h>
+
+#include <libdrm/drm_fourcc.h>
+
+struct h264_encoder;
+
+struct fb_queue_entry {
+	struct nvnc_fb* fb;
+	TAILQ_ENTRY(fb_queue_entry) link;
+};
+
+TAILQ_HEAD(fb_queue, fb_queue_entry);
+
+struct h264_encoder_ffmpeg {
+	struct h264_encoder base;
+
+	uint32_t width;
+	uint32_t height;
+	uint32_t format;
+
+	AVRational timebase;
+	AVRational sample_aspect_ratio;
+	enum AVPixelFormat av_pixel_format;
+
+	/* type: AVHWDeviceContext */
+	AVBufferRef* hw_device_ctx;
+
+	/* type: AVHWFramesContext */
+	AVBufferRef* hw_frames_ctx;
+
+	AVCodecContext* codec_ctx;
+
+	AVFilterGraph* filter_graph;
+	AVFilterContext* filter_in;
+	AVFilterContext* filter_out;
+
+	struct fb_queue fb_queue;
+
+	struct aml_work* work;
+	struct nvnc_fb* current_fb;
+	struct vec current_packet;
+	bool current_frame_is_keyframe;
+
+	bool please_destroy;
+};
+
+struct h264_encoder_impl h264_encoder_ffmpeg_impl;
+
+static enum AVPixelFormat drm_to_av_pixel_format(uint32_t format)
+{
+	switch (format) {
+	case DRM_FORMAT_XRGB8888:
+	case DRM_FORMAT_ARGB8888:
+		return AV_PIX_FMT_BGR0;
+	case DRM_FORMAT_XBGR8888:
+	case DRM_FORMAT_ABGR8888:
+		return AV_PIX_FMT_RGB0;
+	case DRM_FORMAT_RGBX8888:
+	case DRM_FORMAT_RGBA8888:
+		return AV_PIX_FMT_0BGR;
+	case DRM_FORMAT_BGRX8888:
+	case DRM_FORMAT_BGRA8888:
+		return AV_PIX_FMT_0RGB;
+	}
+
+	return AV_PIX_FMT_NONE;
+}
+
+static void hw_frame_desc_free(void* opaque, uint8_t* data)
+{
+	struct AVDRMFrameDescriptor* desc = (void*)data;
+	assert(desc);
+
+	for (int i = 0; i < desc->nb_objects; ++i)
+		close(desc->objects[i].fd);
+
+	free(desc);
+}
+
+// TODO: Maybe do this once per frame inside nvnc_fb?
+static AVFrame* fb_to_avframe(struct nvnc_fb* fb)
+{
+	struct gbm_bo* bo = fb->bo;
+
+	int n_planes = gbm_bo_get_plane_count(bo);
+
+	AVDRMFrameDescriptor* desc = calloc(1, sizeof(*desc));
+	desc->nb_objects = n_planes;
+
+	desc->nb_layers = 1;
+	desc->layers[0].format = gbm_bo_get_format(bo);
+	desc->layers[0].nb_planes = n_planes;
+
+	for (int i = 0; i < n_planes; ++i) {
+		uint32_t stride = gbm_bo_get_stride_for_plane(bo, i);
+
+		desc->objects[i].fd = gbm_bo_get_fd_for_plane(bo, i);
+		desc->objects[i].size = stride * fb->height;
+		desc->objects[i].format_modifier = gbm_bo_get_modifier(bo);
+
+		desc->layers[0].format = gbm_bo_get_format(bo);
+		desc->layers[0].planes[i].object_index = i;
+		desc->layers[0].planes[i].offset = gbm_bo_get_offset(bo, i);
+		desc->layers[0].planes[i].pitch = stride;
+	}
+
+	AVFrame* frame = av_frame_alloc();
+	if (!frame) {
+		hw_frame_desc_free(NULL, (void*)desc);
+		return NULL;
+	}
+
+	frame->opaque = fb;
+	frame->width = fb->width;
+	frame->height = fb->height;
+	frame->format = AV_PIX_FMT_DRM_PRIME;
+	frame->sample_aspect_ratio = (AVRational){1, 1};
+
+	AVBufferRef* desc_ref = av_buffer_create((void*)desc, sizeof(*desc),
+			hw_frame_desc_free, NULL, 0);
+	if (!desc_ref) {
+		hw_frame_desc_free(NULL, (void*)desc);
+		av_frame_free(&frame);
+		return NULL;
+	}
+
+	frame->buf[0] = desc_ref;
+	frame->data[0] = (void*)desc_ref->data;
+
+	// TODO: Set colorspace?
+
+	return frame;
+}
+
+static struct nvnc_fb* fb_queue_dequeue(struct fb_queue* queue)
+{
+	if (TAILQ_EMPTY(queue))
+		return NULL;
+
+	struct fb_queue_entry* entry = TAILQ_FIRST(queue);
+	TAILQ_REMOVE(queue, entry, link);
+	struct nvnc_fb* fb = entry->fb;
+	free(entry);
+
+	return fb;
+}
+
+static int fb_queue_enqueue(struct fb_queue* queue, struct nvnc_fb* fb)
+{
+	struct fb_queue_entry* entry = calloc(1, sizeof(*entry));
+	if (!entry)
+		return -1;
+
+	entry->fb = fb;
+	nvnc_fb_ref(fb);
+	TAILQ_INSERT_TAIL(queue, entry, link);
+
+	return 0;
+}
+
+static int h264_encoder__init_buffersrc(struct h264_encoder_ffmpeg* self)
+{
+	int rc;
+
+	/* Placeholder values are used to pacify input checking and the real
+	 * values are set below.
+	 */
+	rc = avfilter_graph_create_filter(&self->filter_in,
+			avfilter_get_by_name("buffer"), "in",
+			"width=1:height=1:pix_fmt=drm_prime:time_base=1/1", NULL,
+			self->filter_graph);
+	if (rc != 0)
+		return -1;
+
+	AVBufferSrcParameters *params = av_buffersrc_parameters_alloc();
+	if (!params)
+		return -1;
+
+	params->format = AV_PIX_FMT_DRM_PRIME;
+	params->width = self->width;
+	params->height = self->height;
+	params->sample_aspect_ratio = self->sample_aspect_ratio;
+	params->time_base = self->timebase;
+	params->hw_frames_ctx = self->hw_frames_ctx;
+
+	rc = av_buffersrc_parameters_set(self->filter_in, params);
+	assert(rc == 0);
+
+	av_free(params);
+	return 0;
+}
+
+static int h264_encoder__init_filters(struct h264_encoder_ffmpeg* self)
+{
+	int rc;
+
+	self->filter_graph = avfilter_graph_alloc();
+	if (!self->filter_graph)
+		return -1;
+
+	rc = h264_encoder__init_buffersrc(self);
+	if (rc != 0)
+		goto failure;
+
+	rc = avfilter_graph_create_filter(&self->filter_out,
+			avfilter_get_by_name("buffersink"), "out", NULL,
+			NULL, self->filter_graph);
+	if (rc != 0)
+		goto failure;
+
+	AVFilterInOut* inputs = avfilter_inout_alloc();
+	if (!inputs)
+		goto failure;
+
+	inputs->name = av_strdup("in");
+	inputs->filter_ctx = self->filter_in;
+	inputs->pad_idx = 0;
+	inputs->next = NULL;
+
+	AVFilterInOut* outputs = avfilter_inout_alloc();
+	if (!outputs) {
+		avfilter_inout_free(&inputs);
+		goto failure;
+	}
+
+	outputs->name = av_strdup("out");
+	outputs->filter_ctx = self->filter_out;
+	outputs->pad_idx = 0;
+	outputs->next = NULL;
+
+	rc = avfilter_graph_parse(self->filter_graph,
+			"hwmap=mode=direct:derive_device=vaapi"
+			",scale_vaapi=format=nv12:mode=fast",
+			outputs, inputs, NULL);
+	if (rc != 0)
+		goto failure;
+
+	assert(self->hw_device_ctx);
+
+	for (unsigned int i = 0; i < self->filter_graph->nb_filters; ++i) {
+		self->filter_graph->filters[i]->hw_device_ctx =
+			av_buffer_ref(self->hw_device_ctx);
+	}
+
+	rc = avfilter_graph_config(self->filter_graph, NULL);
+	if (rc != 0)
+		goto failure;
+
+	return 0;
+
+failure:
+	avfilter_graph_free(&self->filter_graph);
+	return -1;
+}
+
+static int h264_encoder__init_codec_context(struct h264_encoder_ffmpeg* self,
+		const AVCodec* codec, int quality)
+{
+	self->codec_ctx = avcodec_alloc_context3(codec);
+	if (!self->codec_ctx)
+		return -1;
+
+	struct AVCodecContext* c = self->codec_ctx;
+	c->width = self->width;
+	c->height = self->height;
+	c->time_base = self->timebase;
+	c->sample_aspect_ratio = self->sample_aspect_ratio;
+	c->pix_fmt = AV_PIX_FMT_VAAPI;
+	c->gop_size = INT32_MAX; /* We'll select key frames manually */
+	c->max_b_frames = 0; /* B-frames are bad for latency */
+	c->global_quality = quality;
+
+	/* open-h264 requires baseline profile, so we use constrained
+	 * baseline.
+	 */
+	c->profile = 578;
+
+	return 0;
+}
+
+static int h264_encoder__init_hw_frames_context(struct h264_encoder_ffmpeg* self)
+{
+	self->hw_frames_ctx = av_hwframe_ctx_alloc(self->hw_device_ctx);
+	if (!self->hw_frames_ctx)
+		return -1;
+
+	AVHWFramesContext* c = (AVHWFramesContext*)self->hw_frames_ctx->data;
+	c->format = AV_PIX_FMT_DRM_PRIME;
+	c->sw_format = drm_to_av_pixel_format(self->format);
+	c->width = self->width;
+	c->height = self->height;
+
+	if (av_hwframe_ctx_init(self->hw_frames_ctx) < 0)
+		av_buffer_unref(&self->hw_frames_ctx);
+
+	return 0;
+}
+
+static int h264_encoder__schedule_work(struct h264_encoder_ffmpeg* self)
+{
+	if (self->current_fb)
+		return 0;
+
+	self->current_fb = fb_queue_dequeue(&self->fb_queue);
+	if (!self->current_fb)
+		return 0;
+
+	DTRACE_PROBE1(neatvnc, h264_encode_frame_begin, self->current_fb->pts);
+
+	self->current_frame_is_keyframe = self->base.next_frame_should_be_keyframe;
+	self->base.next_frame_should_be_keyframe = false;
+
+	return aml_start(aml_get_default(), self->work);
+}
+
+static int h264_encoder__encode(struct h264_encoder_ffmpeg* self,
+		AVFrame* frame_in)
+{
+	int rc;
+
+	rc = av_buffersrc_add_frame_flags(self->filter_in, frame_in,
+			AV_BUFFERSRC_FLAG_KEEP_REF);
+	if (rc != 0)
+		return -1;
+
+	AVFrame* filtered_frame = av_frame_alloc();
+	if (!filtered_frame)
+		return -1;
+
+	rc = av_buffersink_get_frame(self->filter_out, filtered_frame);
+	if (rc != 0)
+		goto get_frame_failure;
+
+	rc = avcodec_send_frame(self->codec_ctx, filtered_frame);
+	if (rc != 0)
+		goto send_frame_failure;
+
+	AVPacket* packet = av_packet_alloc();
+	assert(packet); // TODO
+
+	while (1) {
+		rc = avcodec_receive_packet(self->codec_ctx, packet);
+		if (rc != 0)
+			break;
+
+		vec_append(&self->current_packet, packet->data, packet->size);
+
+		packet->stream_index = 0;
+		av_packet_unref(packet);
+	}
+
+	// Frame should always start with a zero:
+	assert(self->current_packet.len == 0 ||
+			((char*)self->current_packet.data)[0] == 0);
+
+	av_packet_free(&packet);
+send_frame_failure:
+	av_frame_unref(filtered_frame);
+get_frame_failure:
+	av_frame_free(&filtered_frame);
+	return rc == AVERROR(EAGAIN) ? 0 : rc;
+}
+
+static void h264_encoder__do_work(void* handle)
+{
+	struct h264_encoder_ffmpeg* self = aml_get_userdata(handle);
+
+	AVFrame* frame = fb_to_avframe(self->current_fb);
+	assert(frame); // TODO
+
+	frame->hw_frames_ctx = av_buffer_ref(self->hw_frames_ctx);
+
+	if (self->current_frame_is_keyframe) {
+#if LIBAVUTIL_VERSION_INT >= AV_VERSION_INT(58, 7, 100)
+		frame->flags |= AV_FRAME_FLAG_KEY;
+#else
+		frame->key_frame = 1;
+#endif
+		frame->pict_type = AV_PICTURE_TYPE_I;
+	} else {
+#if LIBAVUTIL_VERSION_INT >= AV_VERSION_INT(58, 7, 100)
+		frame->flags &= ~AV_FRAME_FLAG_KEY;
+#else
+		frame->key_frame = 0;
+#endif
+		frame->pict_type = AV_PICTURE_TYPE_P;
+	}
+
+	int rc = h264_encoder__encode(self, frame);
+	if (rc != 0) {
+		char err[256];
+		av_strerror(rc, err, sizeof(err));
+		nvnc_log(NVNC_LOG_ERROR, "Failed to encode packet: %s", err);
+		goto failure;
+	}
+
+failure:
+	av_frame_unref(frame);
+	av_frame_free(&frame);
+}
+
+static void h264_encoder__on_work_done(void* handle)
+{
+	struct h264_encoder_ffmpeg* self = aml_get_userdata(handle);
+
+	uint64_t pts = nvnc_fb_get_pts(self->current_fb);
+	nvnc_fb_release(self->current_fb);
+	nvnc_fb_unref(self->current_fb);
+	self->current_fb = NULL;
+
+	DTRACE_PROBE1(neatvnc, h264_encode_frame_end, pts);
+
+	if (self->please_destroy) {
+		vec_destroy(&self->current_packet);
+		h264_encoder_destroy(&self->base);
+		return;
+	}
+
+	if (self->current_packet.len == 0) {
+		nvnc_log(NVNC_LOG_WARNING, "Whoops, encoded packet length is 0");
+		return;
+	}
+
+	void* userdata = self->base.userdata;
+
+	// Must make a copy of packet because the callback might destroy the
+	// encoder object.
+	struct vec packet;
+	vec_init(&packet, self->current_packet.len);
+	vec_append(&packet, self->current_packet.data,
+			self->current_packet.len);
+
+	vec_clear(&self->current_packet);
+	h264_encoder__schedule_work(self);
+
+	self->base.on_packet_ready(packet.data, packet.len, pts, userdata);
+	vec_destroy(&packet);
+}
+
+static int find_render_node(char *node, size_t maxlen) {
+	bool r = -1;
+	drmDevice *devices[64];
+
+	int n = drmGetDevices2(0, devices, sizeof(devices) / sizeof(devices[0]));
+	for (int i = 0; i < n; ++i) {
+		drmDevice *dev = devices[i];
+		if (!(dev->available_nodes & (1 << DRM_NODE_RENDER)))
+			continue;
+
+		strncpy(node, dev->nodes[DRM_NODE_RENDER], maxlen);
+		node[maxlen - 1] = '\0';
+		r = 0;
+		break;
+	}
+
+	drmFreeDevices(devices, n);
+	return r;
+}
+
+static struct h264_encoder* h264_encoder_ffmpeg_create(uint32_t width,
+		uint32_t height, uint32_t format, int quality)
+{
+	int rc;
+
+	struct h264_encoder_ffmpeg* self = calloc(1, sizeof(*self));
+	if (!self)
+		return NULL;
+
+	self->base.impl = &h264_encoder_ffmpeg_impl;
+
+	if (vec_init(&self->current_packet, 65536) < 0)
+		goto packet_failure;
+
+	self->work = aml_work_new(h264_encoder__do_work,
+			h264_encoder__on_work_done, self, NULL);
+	if (!self->work)
+		goto worker_failure;
+
+	char render_node[64];
+	if (find_render_node(render_node, sizeof(render_node)) < 0)
+		goto render_node_failure;
+
+	rc = av_hwdevice_ctx_create(&self->hw_device_ctx,
+			AV_HWDEVICE_TYPE_DRM, render_node, NULL, 0);
+	if (rc != 0)
+		goto hwdevice_ctx_failure;
+
+	self->base.next_frame_should_be_keyframe = true;
+	TAILQ_INIT(&self->fb_queue);
+
+	self->width = width;
+	self->height = height;
+	self->format = format;
+	self->timebase = (AVRational){1, 1000000};
+	self->sample_aspect_ratio = (AVRational){1, 1};
+	self->av_pixel_format = drm_to_av_pixel_format(format);
+	if (self->av_pixel_format == AV_PIX_FMT_NONE)
+		goto pix_fmt_failure;
+
+	const AVCodec* codec = avcodec_find_encoder_by_name("h264_vaapi");
+	if (!codec)
+		goto codec_failure;
+
+	if (h264_encoder__init_hw_frames_context(self) < 0)
+		goto hw_frames_context_failure;
+
+	if (h264_encoder__init_filters(self) < 0)
+		goto filter_failure;
+
+	if (h264_encoder__init_codec_context(self, codec, quality) < 0)
+		goto codec_context_failure;
+
+	self->codec_ctx->hw_frames_ctx =
+		av_buffer_ref(self->filter_out->inputs[0]->hw_frames_ctx);
+
+	AVDictionary *opts = NULL;
+	av_dict_set_int(&opts, "async_depth", 1, 0);
+
+	rc = avcodec_open2(self->codec_ctx, codec, &opts);
+	av_dict_free(&opts);
+
+	if (rc != 0)
+		goto avcodec_open_failure;
+
+	return &self->base;
+
+avcodec_open_failure:
+	avcodec_free_context(&self->codec_ctx);
+codec_context_failure:
+filter_failure:
+	av_buffer_unref(&self->hw_frames_ctx);
+hw_frames_context_failure:
+codec_failure:
+pix_fmt_failure:
+	av_buffer_unref(&self->hw_device_ctx);
+hwdevice_ctx_failure:
+render_node_failure:
+	aml_unref(self->work);
+worker_failure:
+	vec_destroy(&self->current_packet);
+packet_failure:
+	free(self);
+	return NULL;
+}
+
+static void h264_encoder_ffmpeg_destroy(struct h264_encoder* base)
+{
+	struct h264_encoder_ffmpeg* self = (struct h264_encoder_ffmpeg*)base;
+
+	if (self->current_fb) {
+		self->please_destroy = true;
+		return;
+	}
+
+	vec_destroy(&self->current_packet);
+	av_buffer_unref(&self->hw_frames_ctx);
+	avcodec_free_context(&self->codec_ctx);
+	av_buffer_unref(&self->hw_device_ctx);
+	avfilter_graph_free(&self->filter_graph);
+	aml_unref(self->work);
+	free(self);
+}
+
+static void h264_encoder_ffmpeg_feed(struct h264_encoder* base,
+		struct nvnc_fb* fb)
+{
+	struct h264_encoder_ffmpeg* self = (struct h264_encoder_ffmpeg*)base;
+	assert(fb->type == NVNC_FB_GBM_BO);
+
+	// TODO: Add transform filter
+	assert(fb->transform == NVNC_TRANSFORM_NORMAL);
+
+	int rc = fb_queue_enqueue(&self->fb_queue, fb);
+	assert(rc == 0); // TODO
+
+	nvnc_fb_hold(fb);
+
+	rc = h264_encoder__schedule_work(self);
+	assert(rc == 0); // TODO
+}
+
+struct h264_encoder_impl h264_encoder_ffmpeg_impl = {
+	.create = h264_encoder_ffmpeg_create,
+	.destroy = h264_encoder_ffmpeg_destroy,
+	.feed = h264_encoder_ffmpeg_feed,
+};

src/h264-encoder-v4l2m2m-impl.c

@@ -0,0 +1,725 @@
+#include "h264-encoder.h"
+#include "neatvnc.h"
+#include "fb.h"
+#include "pixels.h"
+
+#include <assert.h>
+#include <string.h>
+#include <stdio.h>
+#include <inttypes.h>
+#include <fcntl.h>
+#include <sys/mman.h>
+#include <sys/ioctl.h>
+#include <linux/videodev2.h>
+#include <drm_fourcc.h>
+#include <gbm.h>
+#include <aml.h>
+#include <dirent.h>
+
+#define UDIV_UP(a, b) (((a) + (b) - 1) / (b))
+#define ALIGN_UP(a, b) ((b) * UDIV_UP((a), (b)))
+#define ARRAY_LENGTH(a) (sizeof(a) / sizeof((a)[0]))
+
+#define N_SRC_BUFS 3
+#define N_DST_BUFS 3
+
+struct h264_encoder_v4l2m2m_dst_buf {
+	struct v4l2_buffer buffer;
+	struct v4l2_plane plane;
+	void* payload;
+};
+
+struct h264_encoder_v4l2m2m_src_buf {
+	struct v4l2_buffer buffer;
+	struct v4l2_plane planes[4];
+	int fd;
+	bool is_taken;
+	struct nvnc_fb* fb;
+};
+
+struct h264_encoder_v4l2m2m {
+	struct h264_encoder base;
+
+	uint32_t width;
+	uint32_t height;
+	uint32_t format;
+	int quality; // TODO: Can we affect the quality?
+
+	char driver[16];
+
+	int fd;
+	struct aml_handler* handler;
+
+	struct h264_encoder_v4l2m2m_src_buf src_bufs[N_SRC_BUFS];
+	int src_buf_index;
+
+	struct h264_encoder_v4l2m2m_dst_buf dst_bufs[N_DST_BUFS];
+};
+
+struct h264_encoder_impl h264_encoder_v4l2m2m_impl;
+
+static int v4l2_qbuf(int fd, const struct v4l2_buffer* inbuf)
+{
+	assert(inbuf->length <= 4);
+	struct v4l2_plane planes[4];
+	struct v4l2_buffer outbuf;
+	outbuf = *inbuf;
+	memcpy(&planes, inbuf->m.planes, inbuf->length * sizeof(planes[0]));
+	outbuf.m.planes = planes;
+	return ioctl(fd, VIDIOC_QBUF, &outbuf);
+}
+
+static inline int v4l2_dqbuf(int fd, struct v4l2_buffer* buf)
+{
+	return ioctl(fd, VIDIOC_DQBUF, buf);
+}
+
+static struct h264_encoder_v4l2m2m_src_buf* take_src_buffer(
+		struct h264_encoder_v4l2m2m* self)
+{
+	unsigned int count = 0;
+	int i = self->src_buf_index;
+
+	struct h264_encoder_v4l2m2m_src_buf* buffer;
+	do {
+		buffer = &self->src_bufs[i++];
+		i %= ARRAY_LENGTH(self->src_bufs);
+	} while (++count < ARRAY_LENGTH(self->src_bufs) && buffer->is_taken);
+
+	if (buffer->is_taken)
+		return NULL;
+
+	self->src_buf_index = i;
+	buffer->is_taken = true;
+
+	return buffer;
+}
+
+static bool any_src_buf_is_taken(struct h264_encoder_v4l2m2m* self)
+{
+	bool result = false;
+	for (unsigned int i = 0; i < ARRAY_LENGTH(self->src_bufs); ++i)
+		if (self->src_bufs[i].is_taken)
+			result = true;
+	return result;
+}
+
+static int u32_cmp(const void* pa, const void* pb)
+{
+	const uint32_t *a = pa;
+	const uint32_t *b = pb;
+	return *a < *b ? -1 : *a > *b;
+}
+
+static size_t get_supported_formats(struct h264_encoder_v4l2m2m* self,
+		uint32_t* formats, size_t max_len)
+{
+	size_t i = 0;
+	for (;; ++i) {
+		struct v4l2_fmtdesc desc = {
+			.index = i,
+			.type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE,
+		};
+		int rc = ioctl(self->fd, VIDIOC_ENUM_FMT, &desc);
+		if (rc < 0)
+			break;
+
+		nvnc_trace("Got pixel format: %s", desc.description);
+
+		formats[i] = desc.pixelformat;
+	}
+
+	qsort(formats, i, sizeof(*formats), u32_cmp);
+
+	return i;
+}
+
+static bool have_v4l2_format(const uint32_t* formats, size_t n_formats,
+		uint32_t format)
+{
+	return bsearch(&format, formats, n_formats, sizeof(format), u32_cmp);
+}
+
+static uint32_t v4l2_format_from_drm(const uint32_t* formats,
+		size_t n_formats, uint32_t drm_format)
+{
+#define TRY_FORMAT(f) \
+	if (have_v4l2_format(formats, n_formats, f)) \
+		return f
+
+	switch (drm_format) {
+	case DRM_FORMAT_RGBX8888:
+	case DRM_FORMAT_RGBA8888:
+		TRY_FORMAT(V4L2_PIX_FMT_RGBX32);
+		TRY_FORMAT(V4L2_PIX_FMT_RGBA32);
+		break;
+	case DRM_FORMAT_XRGB8888:
+	case DRM_FORMAT_ARGB8888:
+		TRY_FORMAT(V4L2_PIX_FMT_XRGB32);
+		TRY_FORMAT(V4L2_PIX_FMT_ARGB32);
+		TRY_FORMAT(V4L2_PIX_FMT_RGB32);
+		break;
+	case DRM_FORMAT_BGRX8888:
+	case DRM_FORMAT_BGRA8888:
+		TRY_FORMAT(V4L2_PIX_FMT_XBGR32);
+		TRY_FORMAT(V4L2_PIX_FMT_ABGR32);
+		TRY_FORMAT(V4L2_PIX_FMT_BGR32);
+		break;
+	case DRM_FORMAT_XBGR8888:
+	case DRM_FORMAT_ABGR8888:
+		TRY_FORMAT(V4L2_PIX_FMT_BGRX32);
+		TRY_FORMAT(V4L2_PIX_FMT_BGRA32);
+		break;
+	// TODO: More formats
+	}
+
+	return 0;
+#undef TRY_FORMAT
+}
+
+// This driver mixes up pixel formats...
+static uint32_t v4l2_format_from_drm_bcm2835(const uint32_t* formats,
+		size_t n_formats, uint32_t drm_format)
+{
+	switch (drm_format) {
+	case DRM_FORMAT_XRGB8888:
+	case DRM_FORMAT_ARGB8888:
+		return V4L2_PIX_FMT_RGBA32;
+	case DRM_FORMAT_BGRX8888:
+	case DRM_FORMAT_BGRA8888:
+		// TODO: This could also be ABGR, based on how this driver
+		// behaves
+		return V4L2_PIX_FMT_BGR32;
+	}
+	return 0;
+}
+
+static int set_src_fmt(struct h264_encoder_v4l2m2m* self)
+{
+	int rc;
+
+	uint32_t supported_formats[256];
+	size_t n_formats = get_supported_formats(self, supported_formats,
+			ARRAY_LENGTH(supported_formats));
+
+	uint32_t format;
+	if (strcmp(self->driver, "bcm2835-codec") == 0)
+		format = v4l2_format_from_drm_bcm2835(supported_formats,
+				n_formats, self->format);
+	else
+		format = v4l2_format_from_drm(supported_formats, n_formats,
+				self->format);
+	if (!format) {
+		nvnc_log(NVNC_LOG_DEBUG, "Failed to find a proper pixel format");
+		return -1;
+	}
+
+	struct v4l2_format fmt = {
+		.type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE,
+	};
+	rc = ioctl(self->fd, VIDIOC_G_FMT, &fmt);
+	if (rc < 0) {
+		return -1;
+	}
+
+	struct v4l2_pix_format_mplane* pix_fmt = &fmt.fmt.pix_mp;
+	pix_fmt->pixelformat = format;
+	pix_fmt->width = ALIGN_UP(self->width, 16);
+	pix_fmt->height = ALIGN_UP(self->height, 16);
+
+	rc = ioctl(self->fd, VIDIOC_S_FMT, &fmt);
+	if (rc < 0) {
+		return -1;
+	}
+
+	return 0;
+}
+
+static int set_dst_fmt(struct h264_encoder_v4l2m2m* self)
+{
+	int rc;
+
+	struct v4l2_format fmt = {
+		.type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE,
+	};
+	rc = ioctl(self->fd, VIDIOC_G_FMT, &fmt);
+	if (rc < 0) {
+		return -1;
+	}
+
+	struct v4l2_pix_format_mplane* pix_fmt = &fmt.fmt.pix_mp;
+	pix_fmt->pixelformat = V4L2_PIX_FMT_H264;
+	pix_fmt->width = self->width;
+	pix_fmt->height = self->height;
+
+	rc = ioctl(self->fd, VIDIOC_S_FMT, &fmt);
+	if (rc < 0) {
+		return -1;
+	}
+
+	return 0;
+}
+
+static int alloc_dst_buffers(struct h264_encoder_v4l2m2m* self)
+{
+	int n_bufs = ARRAY_LENGTH(self->dst_bufs);
+	int rc;
+
+	struct v4l2_requestbuffers req = {
+		.memory = V4L2_MEMORY_MMAP,
+		.count = n_bufs,
+		.type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE,
+	};
+	rc = ioctl(self->fd, VIDIOC_REQBUFS, &req);
+	if (rc < 0)
+		return -1;
+
+	for (unsigned int i = 0; i < req.count; ++i) {
+		struct h264_encoder_v4l2m2m_dst_buf* buffer = &self->dst_bufs[i];
+		struct v4l2_buffer* buf = &buffer->buffer;
+
+		buf->index = i;
+		buf->type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE;
+		buf->memory = V4L2_MEMORY_MMAP;
+		buf->length = 1;
+		buf->m.planes = &buffer->plane;
+
+		rc = ioctl(self->fd, VIDIOC_QUERYBUF, buf);
+		if (rc < 0)
+			return -1;
+
+		buffer->payload = mmap(0, buffer->plane.length,
+				PROT_READ | PROT_WRITE, MAP_SHARED, self->fd,
+				buffer->plane.m.mem_offset);
+		if (buffer->payload == MAP_FAILED) {
+			nvnc_log(NVNC_LOG_ERROR, "Whoops, mapping failed: %m");
+			return -1;
+		}
+	}
+
+	return 0;
+}
+
+static void enqueue_dst_buffers(struct h264_encoder_v4l2m2m* self)
+{
+	for (unsigned int i = 0; i < ARRAY_LENGTH(self->dst_bufs); ++i) {
+		int rc = v4l2_qbuf(self->fd, &self->dst_bufs[i].buffer);
+		assert(rc >= 0);
+	}
+}
+
+static void process_dst_bufs(struct h264_encoder_v4l2m2m* self)
+{
+	int rc;
+	struct v4l2_plane plane = { 0 };
+	struct v4l2_buffer buf = {
+		.type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE,
+		.memory = V4L2_MEMORY_MMAP,
+		.length = 1,
+		.m.planes = &plane,
+	};
+
+	while (true) {
+		rc = v4l2_dqbuf(self->fd, &buf);
+		if (rc < 0)
+			break;
+
+		uint64_t pts = buf.timestamp.tv_sec * UINT64_C(1000000) +
+			buf.timestamp.tv_usec;
+		struct h264_encoder_v4l2m2m_dst_buf* dstbuf =
+			&self->dst_bufs[buf.index];
+		size_t size = buf.m.planes[0].bytesused;
+
+		static uint64_t last_pts;
+		if (last_pts && last_pts > pts) {
+			nvnc_log(NVNC_LOG_ERROR, "pts - last_pts = %"PRIi64,
+					(int64_t)pts - (int64_t)last_pts);
+		}
+		last_pts = pts;
+
+		nvnc_trace("Encoded frame (index %d) at %"PRIu64" µs with size: %zu",
+				buf.index, pts, size);
+
+		self->base.on_packet_ready(dstbuf->payload, size, pts,
+				self->base.userdata);
+
+		v4l2_qbuf(self->fd, &buf);
+	}
+}
+
+static void process_src_bufs(struct h264_encoder_v4l2m2m* self)
+{
+	int rc;
+	struct v4l2_plane planes[4] = { 0 };
+	struct v4l2_buffer buf = {
+		.type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE,
+		.memory = V4L2_MEMORY_DMABUF,
+		.length = 1,
+		.m.planes = planes,
+	};
+
+	while (true) {
+		rc = v4l2_dqbuf(self->fd, &buf);
+		if (rc < 0)
+			break;
+
+		struct h264_encoder_v4l2m2m_src_buf* srcbuf =
+			&self->src_bufs[buf.index];
+		srcbuf->is_taken = false;
+
+		// TODO: This assumes that there's only one fd
+		close(srcbuf->planes[0].m.fd);
+
+		nvnc_fb_unmap(srcbuf->fb);
+		nvnc_fb_release(srcbuf->fb);
+		nvnc_fb_unref(srcbuf->fb);
+		srcbuf->fb = NULL;
+	}
+}
+
+static void stream_off(struct h264_encoder_v4l2m2m* self)
+{
+	int type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE;
+	ioctl(self->fd, VIDIOC_STREAMOFF, &type);
+
+	type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE;
+	ioctl(self->fd, VIDIOC_STREAMOFF, &type);
+}
+
+static void free_dst_buffers(struct h264_encoder_v4l2m2m* self)
+{
+	for (unsigned int i = 0; i < ARRAY_LENGTH(self->dst_bufs); ++i) {
+		struct h264_encoder_v4l2m2m_dst_buf* buf = &self->dst_bufs[i];
+		munmap(buf->payload, buf->plane.length);
+	}
+}
+
+static int stream_on(struct h264_encoder_v4l2m2m* self)
+{
+	int type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE;
+	ioctl(self->fd, VIDIOC_STREAMON, &type);
+
+	type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE;
+	return ioctl(self->fd, VIDIOC_STREAMON, &type);
+}
+
+static int alloc_src_buffers(struct h264_encoder_v4l2m2m* self)
+{
+	int rc;
+
+	struct v4l2_requestbuffers req = {
+		.memory = V4L2_MEMORY_DMABUF,
+		.count = N_SRC_BUFS,
+		.type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE,
+	};
+	rc = ioctl(self->fd, VIDIOC_REQBUFS, &req);
+	if (rc < 0)
+		return -1;
+
+	for (int i = 0; i < N_SRC_BUFS; ++i) {
+		struct h264_encoder_v4l2m2m_src_buf* buffer = &self->src_bufs[i];
+		struct v4l2_buffer* buf = &buffer->buffer;
+
+		buf->index = i;
+		buf->type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE;
+		buf->memory = V4L2_MEMORY_DMABUF;
+		buf->length = 1;
+		buf->m.planes = buffer->planes;
+
+		rc = ioctl(self->fd, VIDIOC_QUERYBUF, buf);
+		if (rc < 0)
+			return -1;
+	}
+
+	return 0;
+}
+
+static void force_key_frame(struct h264_encoder_v4l2m2m* self)
+{
+	struct v4l2_control ctrl = { 0 };
+	ctrl.id = V4L2_CID_MPEG_VIDEO_FORCE_KEY_FRAME;
+	ctrl.value = 0;
+	ioctl(self->fd, VIDIOC_S_CTRL, &ctrl);
+}
+
+static void encode_buffer(struct h264_encoder_v4l2m2m* self,
+		struct nvnc_fb* fb)
+{
+	struct h264_encoder_v4l2m2m_src_buf* srcbuf = take_src_buffer(self);
+	if (!srcbuf) {
+		nvnc_log(NVNC_LOG_ERROR, "Out of source buffers. Dropping frame...");
+		return;
+	}
+
+	assert(!srcbuf->fb);
+
+	nvnc_fb_ref(fb);
+	nvnc_fb_hold(fb);
+
+	/* For some reason the v4l2m2m h264 encoder in the Rapberry Pi 4 gets
+	 * really glitchy unless the buffer is mapped first.
+	 * This should probably be handled by the driver, but it's not.
+	 */
+	nvnc_fb_map(fb);
+
+	srcbuf->fb = fb;
+
+	struct gbm_bo* bo = nvnc_fb_get_gbm_bo(fb);
+
+	int n_planes = gbm_bo_get_plane_count(bo);
+	int fd = gbm_bo_get_fd(bo);
+	uint32_t height = ALIGN_UP(gbm_bo_get_height(bo), 16);
+
+	for (int i = 0; i < n_planes; ++i) {
+		uint32_t stride = gbm_bo_get_stride_for_plane(bo, i);
+		uint32_t offset = gbm_bo_get_offset(bo, i);
+		uint32_t size = stride * height;
+
+		srcbuf->buffer.m.planes[i].m.fd = fd;
+		srcbuf->buffer.m.planes[i].bytesused = size;
+		srcbuf->buffer.m.planes[i].length = size;
+		srcbuf->buffer.m.planes[i].data_offset = offset;
+	}
+
+	srcbuf->buffer.timestamp.tv_sec = fb->pts / UINT64_C(1000000);
+	srcbuf->buffer.timestamp.tv_usec = fb->pts % UINT64_C(1000000);
+
+	if (self->base.next_frame_should_be_keyframe)
+		force_key_frame(self);
+	self->base.next_frame_should_be_keyframe = false;
+
+	int rc = v4l2_qbuf(self->fd, &srcbuf->buffer);
+	if (rc < 0) {
+		nvnc_log(NVNC_LOG_PANIC, "Failed to enqueue buffer: %m");
+	}
+}
+
+static void process_fd_events(void* handle)
+{
+	struct h264_encoder_v4l2m2m* self = aml_get_userdata(handle);
+	process_dst_bufs(self);
+}
+
+static void h264_encoder_v4l2m2m_configure(struct h264_encoder_v4l2m2m* self)
+{
+	struct v4l2_control ctrl = { 0 };
+
+	ctrl.id = V4L2_CID_MPEG_VIDEO_H264_PROFILE;
+	ctrl.value = V4L2_MPEG_VIDEO_H264_PROFILE_CONSTRAINED_BASELINE;
+	ioctl(self->fd, VIDIOC_S_CTRL, &ctrl);
+
+	ctrl.id = V4L2_CID_MPEG_VIDEO_H264_I_PERIOD;
+	ctrl.value = INT_MAX;
+	ioctl(self->fd, VIDIOC_S_CTRL, &ctrl);
+
+	ctrl.id = V4L2_CID_MPEG_VIDEO_BITRATE_MODE;
+	ctrl.value = V4L2_MPEG_VIDEO_BITRATE_MODE_CQ;
+	ioctl(self->fd, VIDIOC_S_CTRL, &ctrl);
+
+	ctrl.id = V4L2_CID_MPEG_VIDEO_CONSTANT_QUALITY;
+	ctrl.value = self->quality;
+	ioctl(self->fd, VIDIOC_S_CTRL, &ctrl);
+}
+
+static bool can_encode_to_h264(int fd)
+{
+	size_t i = 0;
+	for (;; ++i) {
+		struct v4l2_fmtdesc desc = {
+			.index = i,
+			.type = V4L2_BUF_TYPE_VIDEO_CAPTURE_MPLANE,
+		};
+		int rc = ioctl(fd, VIDIOC_ENUM_FMT, &desc);
+		if (rc < 0)
+			break;
+
+		if (desc.pixelformat == V4L2_PIX_FMT_H264)
+			return true;
+	}
+	return false;
+}
+
+static bool can_handle_frame_size(int fd, uint32_t width, uint32_t height)
+{
+	size_t i = 0;
+	for (;; ++i) {
+		struct v4l2_frmsizeenum size = {
+			.index = i,
+			.pixel_format = V4L2_PIX_FMT_H264,
+		};
+		int rc = ioctl(fd, VIDIOC_ENUM_FRAMESIZES, &size);
+		if (rc < 0)
+			break;
+
+		switch (size.type) {
+		case V4L2_FRMSIZE_TYPE_DISCRETE:
+			if (size.discrete.width == width &&
+					size.discrete.height == height)
+				return true;
+			break;
+		case V4L2_FRMSIZE_TYPE_CONTINUOUS:
+		case V4L2_FRMSIZE_TYPE_STEPWISE:
+			if (size.stepwise.min_width <= width &&
+					width <= size.stepwise.max_width &&
+					size.stepwise.min_height <= height &&
+					height <= size.stepwise.max_height &&
+					(16 % size.stepwise.step_width) == 0 &&
+					(16 % size.stepwise.step_height) == 0)
+				return true;
+			break;
+		}
+	}
+	return false;
+}
+
+static bool is_device_capable(int fd, uint32_t width, uint32_t height)
+{
+	struct v4l2_capability cap = { 0 };
+	int rc = ioctl(fd, VIDIOC_QUERYCAP, &cap);
+	if (rc < 0)
+		return false;
+
+	uint32_t required_caps = V4L2_CAP_VIDEO_M2M_MPLANE | V4L2_CAP_STREAMING;
+	if ((cap.capabilities & required_caps) != required_caps)
+		return false;
+
+	if (!can_encode_to_h264(fd))
+		return false;
+
+	if (!can_handle_frame_size(fd, width, height))
+		return false;
+
+	return true;
+}
+
+static int find_capable_device(uint32_t width, uint32_t height)
+{
+	int fd = -1;
+	DIR *dir = opendir("/dev");
+	assert(dir);
+
+	for (;;) {
+		struct dirent* entry = readdir(dir);
+		if (!entry)
+			break;
+
+		if (strncmp(entry->d_name, "video", 5) != 0)
+			continue;
+
+		char path[256];
+		snprintf(path, sizeof(path), "/dev/%s", entry->d_name);
+		fd = open(path, O_RDWR | O_CLOEXEC);
+		if (fd < 0) {
+			continue;
+		}
+
+		if (is_device_capable(fd, width, height)) {
+			nvnc_log(NVNC_LOG_DEBUG, "Using v4l2m2m device: %s",
+					path);
+			break;
+		}
+		close(fd);
+		fd = -1;
+	}
+
+	closedir(dir);
+	return fd;
+}
+
+static struct h264_encoder* h264_encoder_v4l2m2m_create(uint32_t width,
+		uint32_t height, uint32_t format, int quality)
+{
+	struct h264_encoder_v4l2m2m* self = calloc(1, sizeof(*self));
+	if (!self)
+		return NULL;
+
+	self->base.impl = &h264_encoder_v4l2m2m_impl;
+	self->fd = -1;
+	self->width = width;
+	self->height = height;
+	self->format = format;
+	self->quality = quality;
+
+	self->fd = find_capable_device(width, height);
+	if (self->fd < 0)
+		goto failure;
+
+	struct v4l2_capability cap = { 0 };
+	ioctl(self->fd, VIDIOC_QUERYCAP, &cap);
+	strncpy(self->driver, (const char*)cap.driver, sizeof(self->driver));
+
+	if (set_src_fmt(self) < 0)
+		goto failure;
+
+	if (set_dst_fmt(self) < 0)
+		goto failure;
+
+	h264_encoder_v4l2m2m_configure(self);
+
+	if (alloc_dst_buffers(self) < 0)
+		goto failure;
+
+	if (alloc_src_buffers(self) < 0)
+		goto failure;
+
+	enqueue_dst_buffers(self);
+
+	if (stream_on(self) < 0)
+		goto failure;
+
+	int flags = fcntl(self->fd, F_GETFL);
+	fcntl(self->fd, F_SETFL, flags | O_NONBLOCK);
+
+	self->handler = aml_handler_new(self->fd, process_fd_events, self, NULL);
+	aml_set_event_mask(self->handler, AML_EVENT_READ);
+
+	if (aml_start(aml_get_default(), self->handler) < 0) {
+		aml_unref(self->handler);
+		goto failure;
+	}
+
+	return &self->base;
+
+failure:
+	if (self->fd >= 0)
+		close(self->fd);
+	return NULL;
+}
+
+static void claim_all_src_bufs(
+		struct h264_encoder_v4l2m2m* self)
+{
+	for (;;) {
+		 process_src_bufs(self);
+		 if (!any_src_buf_is_taken(self))
+			 break;
+		 usleep(10000);
+	}
+}
+
+static void h264_encoder_v4l2m2m_destroy(struct h264_encoder* base)
+{
+	struct h264_encoder_v4l2m2m* self = (struct h264_encoder_v4l2m2m*)base;
+	claim_all_src_bufs(self);
+	aml_stop(aml_get_default(), self->handler);
+	aml_unref(self->handler);
+	stream_off(self);
+	free_dst_buffers(self);
+	if (self->fd >= 0)
+		close(self->fd);
+	free(self);
+}
+
+static void h264_encoder_v4l2m2m_feed(struct h264_encoder* base,
+		struct nvnc_fb* fb)
+{
+	struct h264_encoder_v4l2m2m* self = (struct h264_encoder_v4l2m2m*)base;
+	process_src_bufs(self);
+	encode_buffer(self, fb);
+}
+
+struct h264_encoder_impl h264_encoder_v4l2m2m_impl = {
+	.create = h264_encoder_v4l2m2m_create,
+	.destroy = h264_encoder_v4l2m2m_destroy,
+	.feed = h264_encoder_v4l2m2m_feed,
+};

src/h264-encoder.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 - 2022 Andri Yngvason
+ * Copyright (c) 2024 Andri Yngvason
  *
  * Permission to use, copy, modify, and/or distribute this software for any
  * purpose with or without fee is hereby granted, provided that the above
@@ -15,614 +15,60 @@
  */
 
 #include "h264-encoder.h"
-#include "neatvnc.h"
-#include "fb.h"
-#include "sys/queue.h"
-#include "vec.h"
-#include "usdt.h"
+#include "config.h"
 
-#include <stdlib.h>
-#include <stdint.h>
-#include <stdbool.h>
-#include <unistd.h>
-#include <assert.h>
-#include <gbm.h>
-#include <xf86drm.h>
-#include <aml.h>
-
-#include <libavcodec/avcodec.h>
-#include <libavutil/hwcontext.h>
-#include <libavutil/hwcontext_drm.h>
-#include <libavutil/pixdesc.h>
-#include <libavutil/dict.h>
-#include <libavfilter/avfilter.h>
-#include <libavfilter/buffersink.h>
-#include <libavfilter/buffersrc.h>
-
-#include <libdrm/drm_fourcc.h>
-
-struct h264_encoder;
-
-struct fb_queue_entry {
-	struct nvnc_fb* fb;
-	TAILQ_ENTRY(fb_queue_entry) link;
-};
-
-TAILQ_HEAD(fb_queue, fb_queue_entry);
-
-struct h264_encoder {
-	h264_encoder_packet_handler_fn on_packet_ready;
-	void* userdata;
-
-	uint32_t width;
-	uint32_t height;
-	uint32_t format;
-
-	AVRational timebase;
-	AVRational sample_aspect_ratio;
-	enum AVPixelFormat av_pixel_format;
-
-	/* type: AVHWDeviceContext */
-	AVBufferRef* hw_device_ctx;
-
-	/* type: AVHWFramesContext */
-	AVBufferRef* hw_frames_ctx;
-
-	AVCodecContext* codec_ctx;
-
-	AVFilterGraph* filter_graph;
-	AVFilterContext* filter_in;
-	AVFilterContext* filter_out;
-
-	bool next_frame_should_be_keyframe;
-	struct fb_queue fb_queue;
-
-	struct aml_work* work;
-	struct nvnc_fb* current_fb;
-	struct vec current_packet;
-	bool current_frame_is_keyframe;
-
-	bool please_destroy;
-};
-
-static enum AVPixelFormat drm_to_av_pixel_format(uint32_t format)
-{
-	switch (format) {
-	case DRM_FORMAT_XRGB8888:
-	case DRM_FORMAT_ARGB8888:
-		return AV_PIX_FMT_BGR0;
-	case DRM_FORMAT_XBGR8888:
-	case DRM_FORMAT_ABGR8888:
-		return AV_PIX_FMT_RGB0;
-	case DRM_FORMAT_RGBX8888:
-	case DRM_FORMAT_RGBA8888:
-		return AV_PIX_FMT_0BGR;
-	case DRM_FORMAT_BGRX8888:
-	case DRM_FORMAT_BGRA8888:
-		return AV_PIX_FMT_0RGB;
-	}
-
-	return AV_PIX_FMT_NONE;
-}
-
-static void hw_frame_desc_free(void* opaque, uint8_t* data)
-{
-	struct AVDRMFrameDescriptor* desc = (void*)data;
-	assert(desc);
-
-	for (int i = 0; i < desc->nb_objects; ++i)
-		close(desc->objects[i].fd);
-
-	free(desc);
-}
-
-// TODO: Maybe do this once per frame inside nvnc_fb?
-static AVFrame* fb_to_avframe(struct nvnc_fb* fb)
-{
-	struct gbm_bo* bo = fb->bo;
-
-	int n_planes = gbm_bo_get_plane_count(bo);
-
-	AVDRMFrameDescriptor* desc = calloc(1, sizeof(*desc));
-	desc->nb_objects = n_planes;
-
-	desc->nb_layers = 1;
-	desc->layers[0].format = gbm_bo_get_format(bo);
-	desc->layers[0].nb_planes = n_planes;
-
-	for (int i = 0; i < n_planes; ++i) {
-		uint32_t stride = gbm_bo_get_stride_for_plane(bo, i);
-
-		desc->objects[i].fd = gbm_bo_get_fd_for_plane(bo, i);
-		desc->objects[i].size = stride * fb->height;
-		desc->objects[i].format_modifier = gbm_bo_get_modifier(bo);
-
-		desc->layers[0].format = gbm_bo_get_format(bo);
-		desc->layers[0].planes[i].object_index = i;
-		desc->layers[0].planes[i].offset = gbm_bo_get_offset(bo, i);
-		desc->layers[0].planes[i].pitch = stride;
-	}
-
-	AVFrame* frame = av_frame_alloc();
-	if (!frame) {
-		hw_frame_desc_free(NULL, (void*)desc);
-		return NULL;
-	}
-
-	frame->opaque = fb;
-	frame->width = fb->width;
-	frame->height = fb->height;
-	frame->format = AV_PIX_FMT_DRM_PRIME;
-	frame->sample_aspect_ratio = (AVRational){1, 1};
-
-	AVBufferRef* desc_ref = av_buffer_create((void*)desc, sizeof(*desc),
-			hw_frame_desc_free, NULL, 0);
-	if (!desc_ref) {
-		hw_frame_desc_free(NULL, (void*)desc);
-		av_frame_free(&frame);
-		return NULL;
-	}
-
-	frame->buf[0] = desc_ref;
-	frame->data[0] = (void*)desc_ref->data;
-
-	// TODO: Set colorspace?
-
-	return frame;
-}
-
-static struct nvnc_fb* fb_queue_dequeue(struct fb_queue* queue)
-{
-	if (TAILQ_EMPTY(queue))
-		return NULL;
-
-	struct fb_queue_entry* entry = TAILQ_FIRST(queue);
-	TAILQ_REMOVE(queue, entry, link);
-	struct nvnc_fb* fb = entry->fb;
-	free(entry);
-
-	return fb;
-}
-
-static int fb_queue_enqueue(struct fb_queue* queue, struct nvnc_fb* fb)
-{
-	struct fb_queue_entry* entry = calloc(1, sizeof(*entry));
-	if (!entry)
-		return -1;
-
-	entry->fb = fb;
-	nvnc_fb_ref(fb);
-	TAILQ_INSERT_TAIL(queue, entry, link);
-
-	return 0;
-}
-
-static int h264_encoder__init_buffersrc(struct h264_encoder* self)
-{
-	int rc;
-
-	/* Placeholder values are used to pacify input checking and the real
-	 * values are set below.
-	 */
-	rc = avfilter_graph_create_filter(&self->filter_in,
-			avfilter_get_by_name("buffer"), "in",
-			"width=1:height=1:pix_fmt=drm_prime:time_base=1/1", NULL,
-			self->filter_graph);
-	if (rc != 0)
-		return -1;
-
-	AVBufferSrcParameters *params = av_buffersrc_parameters_alloc();
-	if (!params)
-		return -1;
-
-	params->format = AV_PIX_FMT_DRM_PRIME;
-	params->width = self->width;
-	params->height = self->height;
-	params->sample_aspect_ratio = self->sample_aspect_ratio;
-	params->time_base = self->timebase;
-	params->hw_frames_ctx = self->hw_frames_ctx;
-
-	rc = av_buffersrc_parameters_set(self->filter_in, params);
-	assert(rc == 0);
-
-	av_free(params);
-	return 0;
-}
-
-static int h264_encoder__init_filters(struct h264_encoder* self)
-{
-	int rc;
-
-	self->filter_graph = avfilter_graph_alloc();
-	if (!self->filter_graph)
-		return -1;
-
-	rc = h264_encoder__init_buffersrc(self);
-	if (rc != 0)
-		goto failure;
-
-	rc = avfilter_graph_create_filter(&self->filter_out,
-			avfilter_get_by_name("buffersink"), "out", NULL,
-			NULL, self->filter_graph);
-	if (rc != 0)
-		goto failure;
-
-	AVFilterInOut* inputs = avfilter_inout_alloc();
-	if (!inputs)
-		goto failure;
-
-	inputs->name = av_strdup("in");
-	inputs->filter_ctx = self->filter_in;
-	inputs->pad_idx = 0;
-	inputs->next = NULL;
-
-	AVFilterInOut* outputs = avfilter_inout_alloc();
-	if (!outputs) {
-		avfilter_inout_free(&inputs);
-		goto failure;
-	}
-
-	outputs->name = av_strdup("out");
-	outputs->filter_ctx = self->filter_out;
-	outputs->pad_idx = 0;
-	outputs->next = NULL;
-
-	rc = avfilter_graph_parse(self->filter_graph,
-			"hwmap=mode=direct:derive_device=vaapi"
-			",scale_vaapi=format=nv12:mode=fast",
-			outputs, inputs, NULL);
-	if (rc != 0)
-		goto failure;
-
-	assert(self->hw_device_ctx);
-
-	for (unsigned int i = 0; i < self->filter_graph->nb_filters; ++i) {
-		self->filter_graph->filters[i]->hw_device_ctx =
-			av_buffer_ref(self->hw_device_ctx);
-	}
-
-	rc = avfilter_graph_config(self->filter_graph, NULL);
-	if (rc != 0)
-		goto failure;
-
-	return 0;
-
-failure:
-	avfilter_graph_free(&self->filter_graph);
-	return -1;
-}
-
-static int h264_encoder__init_codec_context(struct h264_encoder* self,
-		const AVCodec* codec, int quality)
-{
-	self->codec_ctx = avcodec_alloc_context3(codec);
-	if (!self->codec_ctx)
-		return -1;
-
-	struct AVCodecContext* c = self->codec_ctx;
-	c->width = self->width;
-	c->height = self->height;
-	c->time_base = self->timebase;
-	c->sample_aspect_ratio = self->sample_aspect_ratio;
-	c->pix_fmt = AV_PIX_FMT_VAAPI;
-	c->gop_size = INT32_MAX; /* We'll select key frames manually */
-	c->max_b_frames = 0; /* B-frames are bad for latency */
-	c->global_quality = quality;
-
-	/* open-h264 requires baseline profile, so we use constrained
-	 * baseline.
-	 */
-	c->profile = 578;
-
-	return 0;
-}
-
-static int h264_encoder__init_hw_frames_context(struct h264_encoder* self)
-{
-	self->hw_frames_ctx = av_hwframe_ctx_alloc(self->hw_device_ctx);
-	if (!self->hw_frames_ctx)
-		return -1;
-
-	AVHWFramesContext* c = (AVHWFramesContext*)self->hw_frames_ctx->data;
-	c->format = AV_PIX_FMT_DRM_PRIME;
-	c->sw_format = drm_to_av_pixel_format(self->format);
-	c->width = self->width;
-	c->height = self->height;
-
-	if (av_hwframe_ctx_init(self->hw_frames_ctx) < 0)
-		av_buffer_unref(&self->hw_frames_ctx);
-
-	return 0;
-}
-
-static int h264_encoder__schedule_work(struct h264_encoder* self)
-{
-	if (self->current_fb)
-		return 0;
-
-	self->current_fb = fb_queue_dequeue(&self->fb_queue);
-	if (!self->current_fb)
-		return 0;
-
-	DTRACE_PROBE1(neatvnc, h264_encode_frame_begin, self->current_fb->pts);
-
-	self->current_frame_is_keyframe = self->next_frame_should_be_keyframe;
-	self->next_frame_should_be_keyframe = false;
-
-	return aml_start(aml_get_default(), self->work);
-}
-
-static int h264_encoder__encode(struct h264_encoder* self, AVFrame* frame_in)
-{
-	int rc;
-
-	rc = av_buffersrc_add_frame_flags(self->filter_in, frame_in,
-			AV_BUFFERSRC_FLAG_KEEP_REF);
-	if (rc != 0)
-		return -1;
-
-	AVFrame* filtered_frame = av_frame_alloc();
-	if (!filtered_frame)
-		return -1;
-
-	rc = av_buffersink_get_frame(self->filter_out, filtered_frame);
-	if (rc != 0)
-		goto get_frame_failure;
-
-	rc = avcodec_send_frame(self->codec_ctx, filtered_frame);
-	if (rc != 0)
-		goto send_frame_failure;
-
-	AVPacket* packet = av_packet_alloc();
-	assert(packet); // TODO
-
-	while (1) {
-		rc = avcodec_receive_packet(self->codec_ctx, packet);
-		if (rc != 0)
-			break;
-
-		vec_append(&self->current_packet, packet->data, packet->size);
-
-		packet->stream_index = 0;
-		av_packet_unref(packet);
-	}
-
-	// Frame should always start with a zero:
-	assert(self->current_packet.len == 0 ||
-			((char*)self->current_packet.data)[0] == 0);
-
-	av_packet_free(&packet);
-send_frame_failure:
-	av_frame_unref(filtered_frame);
-get_frame_failure:
-	av_frame_free(&filtered_frame);
-	return rc == AVERROR(EAGAIN) ? 0 : rc;
-}
-
-static void h264_encoder__do_work(void* handle)
-{
-	struct h264_encoder* self = aml_get_userdata(handle);
-
-	AVFrame* frame = fb_to_avframe(self->current_fb);
-	assert(frame); // TODO
-
-	frame->hw_frames_ctx = av_buffer_ref(self->hw_frames_ctx);
-
-	if (self->current_frame_is_keyframe) {
-#if LIBAVUTIL_VERSION_INT >= AV_VERSION_INT(58, 7, 100)
-		frame->flags |= AV_FRAME_FLAG_KEY;
-#else
-		frame->key_frame = 1;
+#ifdef HAVE_FFMPEG
+extern struct h264_encoder_impl h264_encoder_ffmpeg_impl;
 #endif
-		frame->pict_type = AV_PICTURE_TYPE_I;
-	} else {
-#if LIBAVUTIL_VERSION_INT >= AV_VERSION_INT(58, 7, 100)
-		frame->flags &= ~AV_FRAME_FLAG_KEY;
-#else
-		frame->key_frame = 0;
+
+#ifdef HAVE_V4L2
+extern struct h264_encoder_impl h264_encoder_v4l2m2m_impl;
 #endif
-		frame->pict_type = AV_PICTURE_TYPE_P;
-	}
-
-	int rc = h264_encoder__encode(self, frame);
-	if (rc != 0) {
-		char err[256];
-		av_strerror(rc, err, sizeof(err));
-		nvnc_log(NVNC_LOG_ERROR, "Failed to encode packet: %s", err);
-		goto failure;
-	}
-
-failure:
-	av_frame_unref(frame);
-	av_frame_free(&frame);
-}
-
-static void h264_encoder__on_work_done(void* handle)
-{
-	struct h264_encoder* self = aml_get_userdata(handle);
-
-	uint64_t pts = nvnc_fb_get_pts(self->current_fb);
-	nvnc_fb_release(self->current_fb);
-	nvnc_fb_unref(self->current_fb);
-	self->current_fb = NULL;
-
-	DTRACE_PROBE1(neatvnc, h264_encode_frame_end, pts);
-
-	if (self->please_destroy) {
-		vec_destroy(&self->current_packet);
-		h264_encoder_destroy(self);
-		return;
-	}
-
-	if (self->current_packet.len == 0) {
-		nvnc_log(NVNC_LOG_WARNING, "Whoops, encoded packet length is 0");
-		return;
-	}
-
-	void* userdata = self->userdata;
-
-	// Must make a copy of packet because the callback might destroy the
-	// encoder object.
-	struct vec packet;
-	vec_init(&packet, self->current_packet.len);
-	vec_append(&packet, self->current_packet.data,
-			self->current_packet.len);
-
-	vec_clear(&self->current_packet);
-	h264_encoder__schedule_work(self);
-
-	self->on_packet_ready(packet.data, packet.len, pts, userdata);
-	vec_destroy(&packet);
-}
-
-static int find_render_node(char *node, size_t maxlen) {
-	bool r = -1;
-	drmDevice *devices[64];
-
-	int n = drmGetDevices2(0, devices, sizeof(devices) / sizeof(devices[0]));
-	for (int i = 0; i < n; ++i) {
-		drmDevice *dev = devices[i];
-		if (!(dev->available_nodes & (1 << DRM_NODE_RENDER)))
-			continue;
-
-		strncpy(node, dev->nodes[DRM_NODE_RENDER], maxlen);
-		node[maxlen - 1] = '\0';
-		r = 0;
-		break;
-	}
-
-	drmFreeDevices(devices, n);
-	return r;
-}
 
 struct h264_encoder* h264_encoder_create(uint32_t width, uint32_t height,
 		uint32_t format, int quality)
 {
-	int rc;
+	struct h264_encoder* encoder = NULL;
 
-	struct h264_encoder* self = calloc(1, sizeof(*self));
-	if (!self)
-		return NULL;
+#ifdef HAVE_V4L2
+	encoder = h264_encoder_v4l2m2m_impl.create(width, height, format, quality);
+	if (encoder) {
+		return encoder;
+	}
+#endif
 
-	if (vec_init(&self->current_packet, 65536) < 0)
-		goto packet_failure;
+#ifdef HAVE_FFMPEG
+	encoder = h264_encoder_ffmpeg_impl.create(width, height, format, quality);
+	if (encoder) {
+		return encoder;
+	}
+#endif
 
-	self->work = aml_work_new(h264_encoder__do_work,
-			h264_encoder__on_work_done, self, NULL);
-	if (!self->work)
-		goto worker_failure;
-
-	char render_node[64];
-	if (find_render_node(render_node, sizeof(render_node)) < 0)
-		goto render_node_failure;
-
-	rc = av_hwdevice_ctx_create(&self->hw_device_ctx,
-			AV_HWDEVICE_TYPE_DRM, render_node, NULL, 0);
-	if (rc != 0)
-		goto hwdevice_ctx_failure;
-
-	self->next_frame_should_be_keyframe = true;
-	TAILQ_INIT(&self->fb_queue);
-
-	self->width = width;
-	self->height = height;
-	self->format = format;
-	self->timebase = (AVRational){1, 1000000};
-	self->sample_aspect_ratio = (AVRational){1, 1};
-	self->av_pixel_format = drm_to_av_pixel_format(format);
-	if (self->av_pixel_format == AV_PIX_FMT_NONE)
-		goto pix_fmt_failure;
-
-	const AVCodec* codec = avcodec_find_encoder_by_name("h264_vaapi");
-	if (!codec)
-		goto codec_failure;
-
-	if (h264_encoder__init_hw_frames_context(self) < 0)
-		goto hw_frames_context_failure;
-
-	if (h264_encoder__init_filters(self) < 0)
-		goto filter_failure;
-
-	if (h264_encoder__init_codec_context(self, codec, quality) < 0)
-		goto codec_context_failure;
-
-	self->codec_ctx->hw_frames_ctx =
-		av_buffer_ref(self->filter_out->inputs[0]->hw_frames_ctx);
-
-	AVDictionary *opts = NULL;
-	av_dict_set_int(&opts, "async_depth", 1, 0);
-
-	rc = avcodec_open2(self->codec_ctx, codec, &opts);
-	av_dict_free(&opts);
-
-	if (rc != 0)
-		goto avcodec_open_failure;
-
-	return self;
-
-avcodec_open_failure:
-	avcodec_free_context(&self->codec_ctx);
-codec_context_failure:
-filter_failure:
-	av_buffer_unref(&self->hw_frames_ctx);
-hw_frames_context_failure:
-codec_failure:
-pix_fmt_failure:
-	av_buffer_unref(&self->hw_device_ctx);
-hwdevice_ctx_failure:
-render_node_failure:
-	aml_unref(self->work);
-worker_failure:
-	vec_destroy(&self->current_packet);
-packet_failure:
-	free(self);
-	return NULL;
+	return encoder;
 }
 
 void h264_encoder_destroy(struct h264_encoder* self)
 {
-	if (self->current_fb) {
-		self->please_destroy = true;
-		return;
-	}
-
-	vec_destroy(&self->current_packet);
-	av_buffer_unref(&self->hw_frames_ctx);
-	avcodec_free_context(&self->codec_ctx);
-	av_buffer_unref(&self->hw_device_ctx);
-	avfilter_graph_free(&self->filter_graph);
-	aml_unref(self->work);
-	free(self);
+	self->impl->destroy(self);
 }
 
 void h264_encoder_set_packet_handler_fn(struct h264_encoder* self,
-		h264_encoder_packet_handler_fn value)
+		h264_encoder_packet_handler_fn fn)
 {
-	self->on_packet_ready = value;
+	self->on_packet_ready = fn;
 }
 
-void h264_encoder_set_userdata(struct h264_encoder* self, void* value)
+void h264_encoder_set_userdata(struct h264_encoder* self, void* userdata)
 {
-	self->userdata = value;
+	self->userdata = userdata;
+}
+
+void h264_encoder_feed(struct h264_encoder* self, struct nvnc_fb* fb)
+{
+	self->impl->feed(self, fb);
 }
 
 void h264_encoder_request_keyframe(struct h264_encoder* self)
 {
 	self->next_frame_should_be_keyframe = true;
 }
-
-void h264_encoder_feed(struct h264_encoder* self, struct nvnc_fb* fb)
-{
-	assert(fb->type == NVNC_FB_GBM_BO);
-
-	// TODO: Add transform filter
-	assert(fb->transform == NVNC_TRANSFORM_NORMAL);
-
-	int rc = fb_queue_enqueue(&self->fb_queue, fb);
-	assert(rc == 0); // TODO
-
-	nvnc_fb_hold(fb);
-
-	rc = h264_encoder__schedule_work(self);
-	assert(rc == 0); // TODO
-}