]> git.sesse.net Git - nageru/blob - nageru/video_encoder.cpp
Fix a dangling reference (found by GCC 14).
[nageru] / nageru / video_encoder.cpp
1 #include <algorithm>
2 #include <assert.h>
3 #include <chrono>
4 #include <epoxy/gl.h>
5 #include <memory>
6 #include <movit/image_format.h>
7 #include <mutex>
8 #include <netdb.h>
9 #include <stdint.h>
10 #include <stdio.h>
11 #include <string.h>
12 #include <string>
13 #include <sys/socket.h>
14 #include <thread>
15 #include <time.h>
16 #include <unistd.h>
17 #include <vector>
18
19 extern "C" {
20 #include <libavutil/mem.h>
21 #include <libavformat/avformat.h>
22 #include <libavformat/avio.h>
23 #include <libavutil/avutil.h>
24 #include <libavutil/rational.h>
25 }
26
27 #include "audio_encoder.h"
28 #ifdef HAVE_AV1
29 #include "av1_encoder.h"
30 #endif
31 #include "defs.h"
32 #include "shared/ffmpeg_raii.h"
33 #include "flags.h"
34 #include "shared/httpd.h"
35 #include "shared/mux.h"
36 #include "quicksync_encoder.h"
37 #include "shared/timebase.h"
38 #include "x264_encoder.h"
39 #include "video_encoder.h"
40 #include "shared/metrics.h"
41 #include "shared/ref_counted_gl_sync.h"
42 #include "shared/shared_defs.h"
43
44 class RefCountedFrame;
45
46 using namespace std;
47 using namespace std::chrono;
48 using namespace movit;
49
50 namespace {
51
52 string generate_local_dump_filename(int frame)
53 {
54         time_t now = time(NULL);
55         tm now_tm;
56         localtime_r(&now, &now_tm);
57
58         char timestamp[64];
59         strftime(timestamp, sizeof(timestamp), "%F-%H%M%S%z", &now_tm);
60
61         // Use the frame number to disambiguate between two cuts starting
62         // on the same second.
63         char filename[256];
64         snprintf(filename, sizeof(filename), "%s/%s%s-f%02d%s",
65                 global_flags.recording_dir.c_str(),
66                 LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX);
67         return filename;
68 }
69
70 }  // namespace
71
72 VideoEncoder::VideoEncoder(ResourcePool *resource_pool, QSurface *surface, const std::string &va_display, int width, int height, HTTPD *httpd, DiskSpaceEstimator *disk_space_estimator)
73         : resource_pool(resource_pool), surface(surface), va_display(va_display), width(width), height(height), httpd(httpd), disk_space_estimator(disk_space_estimator)
74 {
75         // TODO: If we're outputting AV1, we can't use MPEG-TS currently.
76         srt_oformat = av_guess_format("mpegts", nullptr, nullptr);
77         assert(srt_oformat != nullptr);
78
79         oformat = av_guess_format(global_flags.stream_mux_name.c_str(), nullptr, nullptr);
80         assert(oformat != nullptr);
81         if (global_flags.stream_audio_codec_name.empty()) {
82                 stream_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
83         } else {
84                 stream_audio_encoder.reset(new AudioEncoder(global_flags.stream_audio_codec_name, global_flags.stream_audio_codec_bitrate, oformat));
85         }
86         if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
87                 x264_encoder.reset(new X264Encoder(oformat, /*use_separate_disk_params=*/false));
88         }
89         VideoCodecInterface *http_encoder = x264_encoder.get();
90         VideoCodecInterface *disk_encoder = x264_encoder.get();
91 #ifdef HAVE_AV1
92         if (global_flags.av1_video_to_http) {
93                 av1_encoder.reset(new AV1Encoder(oformat));
94                 http_encoder = av1_encoder.get();
95         }
96 #endif
97         if (global_flags.x264_separate_disk_encode) {
98                 x264_disk_encoder.reset(new X264Encoder(oformat, /*use_separate_disk_params=*/true));
99                 disk_encoder = x264_disk_encoder.get();
100         }
101
102         string filename = generate_local_dump_filename(/*frame=*/0);
103         quicksync_encoder.reset(new QuickSyncEncoder(filename, resource_pool, surface, va_display, width, height, oformat, http_encoder, disk_encoder, disk_space_estimator));
104
105         open_output_streams();
106         stream_audio_encoder->add_mux(http_mux.get());
107         if (srt_mux != nullptr) {
108                 stream_audio_encoder->add_mux(srt_mux.get());
109         }
110         quicksync_encoder->set_http_mux(http_mux.get());
111         if (srt_mux != nullptr) {
112                 quicksync_encoder->set_srt_mux(srt_mux.get());
113         }
114         if (global_flags.x264_video_to_http) {
115                 x264_encoder->add_mux(http_mux.get());
116                 if (srt_mux != nullptr) {
117                         x264_encoder->add_mux(srt_mux.get());
118                 }
119         }
120 #ifdef HAVE_AV1
121         if (global_flags.av1_video_to_http) {
122                 av1_encoder->add_mux(http_mux.get());
123                 if (srt_mux != nullptr) {
124                         av1_encoder->add_mux(srt_mux.get());
125                 }
126         }
127 #endif
128 }
129
130 VideoEncoder::~VideoEncoder()
131 {
132         should_quit = true;
133         quicksync_encoder->shutdown();
134         x264_encoder.reset(nullptr);
135         x264_disk_encoder.reset(nullptr);
136         quicksync_encoder->close_file();
137         quicksync_encoder.reset(nullptr);
138         while (quicksync_encoders_in_shutdown.load() > 0) {
139                 usleep(10000);
140         }
141 }
142
143 void VideoEncoder::do_cut(int frame)
144 {
145         string filename = generate_local_dump_filename(frame);
146         printf("Starting new recording: %s\n", filename.c_str());
147
148         // Do the shutdown of the old encoder in a separate thread, since it can
149         // take some time (it needs to wait for all the frames in the queue to be
150         // done encoding, for one) and we are running on the main mixer thread.
151         // However, since this means both encoders could be sending packets at
152         // the same time, it means pts could come out of order to the stream mux,
153         // and we need to plug it until the shutdown is complete.
154         http_mux->plug();
155         lock(qs_mu, qs_audio_mu);
156         lock_guard<mutex> lock1(qs_mu, adopt_lock), lock2(qs_audio_mu, adopt_lock);
157         QuickSyncEncoder *old_encoder = quicksync_encoder.release();  // When we go C++14, we can use move capture instead.
158         X264Encoder *old_x264_encoder = nullptr;
159         X264Encoder *old_x264_disk_encoder = nullptr;
160         if (global_flags.x264_video_to_disk) {
161                 old_x264_encoder = x264_encoder.release();
162         }
163         if (global_flags.x264_separate_disk_encode) {
164                 old_x264_disk_encoder = x264_disk_encoder.release();
165         }
166         thread([old_encoder, old_x264_encoder, old_x264_disk_encoder, this]{
167                 old_encoder->shutdown();
168                 delete old_x264_encoder;
169                 delete old_x264_disk_encoder;
170                 old_encoder->close_file();
171                 http_mux->unplug();
172
173                 // We cannot delete the encoder here, as this thread has no OpenGL context.
174                 // We'll deal with it in begin_frame().
175                 lock_guard<mutex> lock(qs_mu);
176                 qs_needing_cleanup.emplace_back(old_encoder);
177         }).detach();
178
179         if (global_flags.x264_video_to_disk) {
180                 x264_encoder.reset(new X264Encoder(oformat, /*use_separate_disk_params=*/false));
181                 assert(global_flags.x264_video_to_http);
182                 if (global_flags.x264_video_to_http) {
183                         x264_encoder->add_mux(http_mux.get());
184                 }
185                 if (overriding_bitrate != 0) {
186                         x264_encoder->change_bitrate(overriding_bitrate);
187                 }
188         }
189         X264Encoder *http_encoder = x264_encoder.get();
190         X264Encoder *disk_encoder = x264_encoder.get();
191         if (global_flags.x264_separate_disk_encode) {
192                 x264_disk_encoder.reset(new X264Encoder(oformat, /*use_separate_disk_params=*/true));
193                 disk_encoder = x264_disk_encoder.get();
194         }
195
196         quicksync_encoder.reset(new QuickSyncEncoder(filename, resource_pool, surface, va_display, width, height, oformat, http_encoder, disk_encoder, disk_space_estimator));
197         quicksync_encoder->set_http_mux(http_mux.get());
198 }
199
200 void VideoEncoder::change_x264_bitrate(unsigned rate_kbit)
201 {
202         overriding_bitrate = rate_kbit;
203         x264_encoder->change_bitrate(rate_kbit);
204 }
205
206 void VideoEncoder::add_audio(int64_t pts, std::vector<float> audio)
207 {
208         // Take only qs_audio_mu, since add_audio() is thread safe
209         // (we can only conflict with do_cut(), which takes qs_audio_mu)
210         // and we don't want to contend with begin_frame().
211         {
212                 lock_guard<mutex> lock(qs_audio_mu);
213                 quicksync_encoder->add_audio(pts, audio);
214         }
215         stream_audio_encoder->encode_audio(audio, pts + quicksync_encoder->global_delay());
216 }
217
218 bool VideoEncoder::is_zerocopy() const
219 {
220         // Explicitly do _not_ take qs_mu; this is called from the mixer,
221         // and qs_mu might be contended. is_zerocopy() is thread safe
222         // and never called in parallel with do_cut() (both happen only
223         // from the mixer thread).
224         return quicksync_encoder->is_zerocopy();
225 }
226
227 bool VideoEncoder::begin_frame(int64_t pts, int64_t duration, movit::YCbCrLumaCoefficients ycbcr_coefficients, const std::vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
228 {
229         lock_guard<mutex> lock(qs_mu);
230         qs_needing_cleanup.clear();  // Since we have an OpenGL context here, and are called regularly.
231         return quicksync_encoder->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
232 }
233
234 RefCountedGLsync VideoEncoder::end_frame()
235 {
236         want_srt_metric_update = true;
237         lock_guard<mutex> lock(qs_mu);
238         return quicksync_encoder->end_frame();
239 }
240
241 void VideoEncoder::open_output_streams()
242 {
243         for (bool is_srt : {false, true}) {
244                 if (is_srt && global_flags.srt_destination_host.empty()) {
245                         continue;
246                 }
247
248                 AVFormatContext *avctx = avformat_alloc_context();
249                 avctx->oformat = is_srt ? srt_oformat : oformat;
250
251                 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
252                 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
253                 if (is_srt) {
254                         avctx->pb->write_packet = &VideoEncoder::write_srt_packet_thunk;
255                 } else {
256                         avctx->pb->write_data_type = &VideoEncoder::write_packet2_thunk;
257                         avctx->pb->ignore_boundary_point = 1;
258                 }
259
260                 Mux::Codec video_codec;
261                 if (global_flags.av1_video_to_http) {
262                         video_codec = Mux::CODEC_AV1;
263                 } else {
264                         video_codec = Mux::CODEC_H264;
265                 }
266
267                 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
268
269                 string video_extradata;
270                 if (global_flags.x264_video_to_http) {
271                         video_extradata = x264_encoder->get_global_headers();
272 #ifdef HAVE_AV1
273                 } else if (global_flags.av1_video_to_http) {
274                         video_extradata = av1_encoder->get_global_headers();
275 #endif
276                 }
277
278                 Mux *mux = new Mux(avctx, width, height, video_codec, video_extradata, stream_audio_encoder->get_codec_parameters().get(),
279                         get_color_space(global_flags.ycbcr_rec709_coefficients), COARSE_TIMEBASE,
280                         /*write_callback=*/nullptr, is_srt ? Mux::WRITE_BACKGROUND : Mux::WRITE_FOREGROUND, { is_srt ? &srt_mux_metrics : &http_mux_metrics });
281                 if (is_srt) {
282                         srt_mux.reset(mux);
283                         srt_mux_metrics.init({{ "destination", "srt" }});
284                         srt_metrics.init({{ "cardtype", "output" }});
285                         global_metrics.add("srt_num_connection_attempts", {{ "cardtype", "output" }}, &metric_srt_num_connection_attempts);
286                 } else {
287                         http_mux.reset(mux);
288                         http_mux_metrics.init({{ "destination", "http" }});
289                 }
290         }
291 }
292
293 int VideoEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
294 {
295         VideoEncoder *video_encoder = (VideoEncoder *)opaque;
296         return video_encoder->write_packet2(buf, buf_size, type, time);
297 }
298
299 int VideoEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
300 {
301         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
302                 seen_sync_markers = true;
303         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
304                 // We don't know if this is a keyframe or not (the muxer could
305                 // avoid marking it), so we just have to make the best of it.
306                 type = AVIO_DATA_MARKER_SYNC_POINT;
307         }
308
309         if (type == AVIO_DATA_MARKER_HEADER) {
310                 http_mux_header.append((char *)buf, buf_size);
311                 httpd->set_header(HTTPD::StreamID{ HTTPD::MAIN_STREAM, 0 }, http_mux_header);
312         } else {
313                 httpd->add_data(HTTPD::StreamID{ HTTPD::MAIN_STREAM, 0 }, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
314         }
315         return buf_size;
316 }
317
318 int VideoEncoder::write_srt_packet_thunk(void *opaque, uint8_t *buf, int buf_size)
319 {
320         VideoEncoder *video_encoder = (VideoEncoder *)opaque;
321         return video_encoder->write_srt_packet(buf, buf_size);
322 }
323
324 static string print_addrinfo(const addrinfo *ai)
325 {
326         char hoststr[NI_MAXHOST], portstr[NI_MAXSERV];
327         if (getnameinfo(ai->ai_addr, ai->ai_addrlen, hoststr, sizeof(hoststr), portstr, sizeof(portstr), NI_DGRAM | NI_NUMERICHOST | NI_NUMERICSERV) != 0) {
328                 return "<unknown address>";  // Should basically never happen, since we're not doing DNS lookups.
329         }
330
331         if (ai->ai_family == AF_INET6) {
332                 return string("[") + hoststr + "]:" + portstr;
333         } else {
334                 return string(hoststr) + ":" + portstr;
335         }
336 }
337
338 int VideoEncoder::open_srt_socket()
339 {
340         int sock = srt_create_socket();
341         if (sock == -1) {
342                 fprintf(stderr, "srt_create_socket(): %s\n", srt_getlasterror_str());
343                 return -1;
344         }
345
346         SRT_TRANSTYPE live = SRTT_LIVE;
347         if (srt_setsockopt(sock, 0, SRTO_TRANSTYPE, &live, sizeof(live)) < 0) {
348                 fprintf(stderr, "srt_setsockopt(SRTO_TRANSTYPE): %s\n", srt_getlasterror_str());
349                 srt_close(sock);
350                 return -1;
351         }
352
353         if (srt_setsockopt(sock, 0, SRTO_LATENCY, &global_flags.srt_output_latency_ms, sizeof(global_flags.srt_output_latency_ms)) < 0) {
354                 fprintf(stderr, "srt_setsockopt(SRTO_LATENCY): %s\n", srt_getlasterror_str());
355                 srt_close(sock);
356                 return -1;
357         }
358
359         if (!global_flags.srt_streamid.empty()) {
360                 if (srt_setsockopt(sock, 0, SRTO_STREAMID, global_flags.srt_streamid.data(), global_flags.srt_streamid.size()) < 0) {
361                         fprintf(stderr, "srt_setsockopt(SRTO_STREAMID): %s\n", srt_getlasterror_str());
362                         srt_close(sock);
363                         return -1;
364                 }
365         }
366
367         if (!global_flags.srt_passphrase.empty()) {
368                 if (srt_setsockopt(sock, 0, SRTO_PASSPHRASE, global_flags.srt_passphrase.data(), global_flags.srt_passphrase.size()) < 0) {
369                         fprintf(stderr, "srt_setsockopt(SRTO_PASSPHRASE): %s\n", srt_getlasterror_str());
370                         srt_close(sock);
371                         return -1;
372                 }
373         }
374
375         return sock;
376 }
377
378 int VideoEncoder::connect_to_srt()
379 {
380         // We need to specify SOCK_DGRAM as a hint, or we'll get all addresses
381         // three times (for each of TCP, UDP, raw).
382         addrinfo hints;
383         memset(&hints, 0, sizeof(hints));
384         hints.ai_flags = AI_ADDRCONFIG;
385         hints.ai_socktype = SOCK_DGRAM;
386
387         addrinfo *ai;
388         int ret = getaddrinfo(global_flags.srt_destination_host.c_str(), global_flags.srt_destination_port.c_str(), &hints, &ai);
389         if (ret != 0) {
390                 fprintf(stderr, "getaddrinfo(%s:%s): %s\n", global_flags.srt_destination_host.c_str(), global_flags.srt_destination_port.c_str(), gai_strerror(ret));
391                 return -1;
392         }
393
394         unique_ptr<addrinfo, decltype(freeaddrinfo) *> ai_cleanup(ai, &freeaddrinfo);
395
396         for (const addrinfo *cur = ai; cur != nullptr; cur = cur->ai_next) {
397                 // Seemingly, srt_create_socket() isn't universal; once we try to connect,
398                 // it gets locked to either IPv4 or IPv6. So we need to create a new one
399                 // for every address we try.
400                 int sock = open_srt_socket();
401                 if (sock == -1) {
402                         // Die immediately.
403                         return sock;
404                 }
405                 ++metric_srt_num_connection_attempts;
406
407                 // We do a non-blocking connect, so that we can check should_quit
408                 // every now and then.
409                 int blocking = 0;
410                 if (srt_setsockopt(sock, 0, SRTO_RCVSYN, &blocking, sizeof(blocking)) < 0) {
411                         fprintf(stderr, "srt_setsockopt(SRTO_SNDSYN=0): %s\n", srt_getlasterror_str());
412                         srt_close(sock);
413                         continue;
414                 }
415                 if (srt_connect(sock, cur->ai_addr, cur->ai_addrlen) < 0) {
416                         fprintf(stderr, "srt_connect(%s): %s\n", print_addrinfo(cur).c_str(), srt_getlasterror_str());
417                         srt_close(sock);
418                         continue;
419                 }
420                 int eid = srt_epoll_create();
421                 if (eid < 0) {
422                         fprintf(stderr, "srt_epoll_create(): %s\n", srt_getlasterror_str());
423                         srt_close(sock);
424                         continue;
425                 }
426                 int modes = SRT_EPOLL_ERR | SRT_EPOLL_OUT;
427                 if (srt_epoll_add_usock(eid, sock, &modes) < 0) {
428                         fprintf(stderr, "srt_epoll_usock(): %s\n", srt_getlasterror_str());
429                         srt_close(sock);
430                         srt_epoll_release(eid);
431                         continue;
432                 }
433                 bool ok;
434                 while (!should_quit.load()) {
435                         SRTSOCKET errfds[1], writefds[1];
436                         int num_errfds = 1, num_writefds = 1;
437                         int poll_time_ms = 100;
438                         int ret = srt_epoll_wait(eid, errfds, &num_errfds, writefds, &num_writefds, poll_time_ms, 0, 0, 0, 0);
439                         if (ret < 0) {
440                                 if (srt_getlasterror(nullptr) == SRT_ETIMEOUT) {
441                                         continue;
442                                 } else {
443                                         fprintf(stderr, "srt_epoll_wait(): %s\n", srt_getlasterror_str());
444                                         srt_close(sock);
445                                         srt_epoll_release(eid);
446                                         return -1;
447                                 }
448                         } else if (ret > 0) {
449                                 // The SRT epoll framework is pretty odd, but seemingly,
450                                 // this is the way. Getting the same error code as srt_connect()
451                                 // would normally return seems to be impossible, though.
452                                 ok = (num_errfds == 0);
453                                 break;
454                                 fprintf(stderr, "num_errfds=%d num_writefds=%d last_err=%s\n", num_errfds, num_writefds, srt_getlasterror_str());
455                                 break;
456                         }
457                 }
458                 srt_epoll_release(eid);
459                 if (should_quit.load()) {
460                         srt_close(sock);
461                         return -1;
462                 }
463                 if (ok) {
464                         fprintf(stderr, "Connected to destination SRT endpoint at %s.\n", print_addrinfo(cur).c_str());
465                         return sock;
466                 } else {
467                         fprintf(stderr, "srt_connect(%s): %s\n", print_addrinfo(cur).c_str(), srt_getlasterror_str());
468                         srt_close(sock);
469                 }
470         }
471
472         // Out of candidates, so give up.
473         return -1;
474 }
475
476 int VideoEncoder::write_srt_packet(uint8_t *buf, int buf_size)
477 {
478         if (want_srt_metric_update.exchange(false) && srt_sock != -1) {
479                 srt_metrics.update_srt_stats(srt_sock);
480         }
481
482         bool has_drained = false;
483         bool trying_reconnect = false;
484         steady_clock::time_point first_connect_start;
485
486         while (buf_size > 0 && !should_quit.load()) {
487                 if (srt_sock == -1) {
488                         if (!trying_reconnect) {
489                                 first_connect_start = steady_clock::now();
490                                 trying_reconnect = true;
491                         }
492                         srt_sock = connect_to_srt();
493                         if (srt_sock == -1) {
494                                 usleep(100000);
495                                 if (!has_drained && duration<double>(steady_clock::now() - first_connect_start).count() >= global_flags.srt_output_latency_ms * 1e-3) {
496                                         // The entire concept for SRT is to have fixed, low latency.
497                                         // If we've been out for more than a latency period, we shouldn't
498                                         // try to send the entire backlog. (But we should be tolerant
499                                         // of a quick disconnect and reconnect.) Maybe it would be better
500                                         // to have a sliding window of how much we remove, but it quickly
501                                         // starts getting esoteric, so juts drop it all.
502                                         fprintf(stderr, "WARNING: No SRT connection for more than %d ms, dropping data.\n",
503                                                 global_flags.srt_output_latency_ms);
504                                         srt_mux->drain();
505                                         has_drained = true;
506                                 }
507                                 continue;
508                         }
509                         srt_metrics.update_srt_stats(srt_sock);
510                 }
511                 if (has_drained) {
512                         // Now that we're reconnected, we can start accepting data again,
513                         // but discard the rest of this write (it is very old by now).
514                         srt_mux->undrain();
515                         break;
516                 }
517                 int to_send = min(buf_size, SRT_LIVE_DEF_PLSIZE);
518                 int ret = srt_send(srt_sock, (char *)buf, to_send);
519                 if (ret < 0)  {
520                         fprintf(stderr, "srt_send(): %s\n", srt_getlasterror_str());
521                         srt_close(srt_sock);
522                         srt_metrics.metric_srt_uptime_seconds = 0.0 / 0.0;
523                         if (!trying_reconnect) {
524                                 first_connect_start = steady_clock::now();
525                                 trying_reconnect = true;
526                         }
527                         srt_sock = connect_to_srt();
528                         continue;
529                 }
530                 buf += ret;
531                 buf_size -= ret;
532         }
533         return buf_size;
534 }
535