]> git.sesse.net Git - cubemap/blob - stream.cpp
Fix infinite CPU usage on waitpid().
[cubemap] / stream.cpp
1 #include <assert.h>
2 #include <errno.h>
3 #include <inttypes.h>
4 #include <limits.h>
5 #include <math.h>
6 #include <netinet/in.h>
7 #include <stdio.h>
8 #include <stdlib.h>
9 #include <string.h>
10 #include <sys/types.h>
11 #include <algorithm>
12 #include <string>
13 #include <queue>
14 #include <vector>
15
16 #include "log.h"
17 #include "metacube2.h"
18 #include "state.pb.h"
19 #include "stream.h"
20 #include "util.h"
21
22 using namespace std;
23
24 Stream::Stream(const string &url,
25                size_t backlog_size,
26                uint64_t prebuffering_bytes,
27                Encoding encoding,
28                Encoding src_encoding,
29                unsigned hls_frag_duration,
30                size_t hls_backlog_margin,
31                const std::string &allow_origin)
32         : url(url),
33           encoding(encoding),
34           src_encoding(src_encoding),
35           allow_origin(allow_origin),
36           data_fd(make_tempfile("")),
37           backlog_size(backlog_size),
38           prebuffering_bytes(prebuffering_bytes),
39           hls_frag_duration(hls_frag_duration),
40           hls_backlog_margin(hls_backlog_margin)
41 {
42         if (data_fd == -1) {
43                 exit(1);
44         }
45 }
46
47 Stream::~Stream()
48 {
49         if (data_fd != -1) {
50                 safe_close(data_fd);
51         }
52 }
53
54 Stream::Stream(const StreamProto &serialized, int data_fd)
55         : url(serialized.url()),
56           unavailable(serialized.unavailable()),
57           http_header(serialized.http_header()),
58           stream_header(serialized.stream_header()),
59           encoding(Stream::STREAM_ENCODING_RAW),  // Will be changed later.
60           data_fd(data_fd),
61           backlog_size(serialized.backlog_size()),
62           bytes_received(serialized.bytes_received()),
63           first_fragment_index(serialized.first_fragment_index()),
64           discontinuity_counter(serialized.discontinuity_counter())
65 {
66         if (data_fd == -1) {
67                 exit(1);
68         }
69
70         for (ssize_t point : serialized.suitable_starting_point()) {
71                 if (point == -1) {
72                         // Can happen when upgrading from before 1.1.3,
73                         // where this was an optional field with -1 signifying
74                         // "no such point".
75                         continue;
76                 }
77                 suitable_starting_points.push_back(point);
78         }
79
80         for (const FragmentStartProto &fragment : serialized.fragment()) {
81                 fragments.push_back(FragmentStart { size_t(fragment.byte_position()), fragment.pts(), fragment.begins_header() });
82         }
83 }
84
85 StreamProto Stream::serialize()
86 {
87         StreamProto serialized;
88         serialized.set_unavailable(unavailable);
89         serialized.set_http_header(http_header);
90         serialized.set_stream_header(stream_header);
91         serialized.add_data_fds(data_fd);
92         serialized.set_backlog_size(backlog_size);
93         serialized.set_bytes_received(bytes_received);
94         for (size_t point : suitable_starting_points) {
95                 serialized.add_suitable_starting_point(point);
96         }
97         for (const FragmentStart &fragment : fragments) {
98                 FragmentStartProto *proto = serialized.add_fragment();
99                 proto->set_byte_position(fragment.byte_position);
100                 proto->set_pts(fragment.pts);
101                 proto->set_begins_header(fragment.begins_header);
102         }
103         serialized.set_first_fragment_index(first_fragment_index);
104         serialized.set_discontinuity_counter(discontinuity_counter);
105
106         serialized.set_url(url);
107         data_fd = -1;
108         return serialized;
109 }
110         
111 void Stream::set_backlog_size(size_t new_size)
112 {
113         if (backlog_size == new_size) {
114                 return;
115         }
116
117         string existing_data;
118         if (!read_tempfile_and_close(data_fd, &existing_data)) {
119                 exit(1);
120         }
121
122         // Unwrap the data so it's no longer circular.
123         if (bytes_received <= backlog_size) {
124                 existing_data.resize(bytes_received);
125         } else {
126                 size_t pos = bytes_received % backlog_size;
127                 existing_data = existing_data.substr(pos, string::npos) +
128                         existing_data.substr(0, pos);
129         }
130
131         // See if we need to discard data.
132         if (new_size < existing_data.size()) {
133                 size_t to_discard = existing_data.size() - new_size;
134                 existing_data = existing_data.substr(to_discard, string::npos);
135         }
136
137         // Create a new, empty data file.
138         data_fd = make_tempfile("");
139         if (data_fd == -1) {
140                 exit(1);
141         }
142         backlog_size = new_size;
143
144         // Now cheat a bit by rewinding, and adding all the old data back.
145         bytes_received -= existing_data.size();
146         DataElement data_element;
147         data_element.data.iov_base = const_cast<char *>(existing_data.data());
148         data_element.data.iov_len = existing_data.size();
149         data_element.metacube_flags = 0;  // Ignored by add_data_raw().
150
151         vector<DataElement> data_elements;
152         data_elements.push_back(data_element);
153         add_data_raw(data_elements);
154         remove_obsolete_starting_points();
155 }
156
157 void Stream::set_header(const std::string &new_http_header, const std::string &new_stream_header)
158 {
159         unavailable = false;
160         http_header = new_http_header;
161         if (new_stream_header == stream_header) {
162                 return;
163         }
164
165         // We cannot start at any of the older starting points anymore,
166         // since they'd get the wrong header for the stream (not to mention
167         // that a changed header probably means the stream restarted,
168         // which means any client starting on the old one would probably
169         // stop playing properly at the change point). Next block
170         // should be a suitable starting point (if not, something is
171         // pretty strange), so it will fill up again soon enough.
172         suitable_starting_points.clear();
173
174         // HLS, on the other hand, can deal with discontinuities and multiple
175         // headers. At least in theory (client support varies wildly).
176         if (!fragments.empty()) {
177                 // Commit the old header to the backlog, so that we can serve it
178                 // for all the old fragments for as long as they exist.
179                 if (!stream_header.empty()) {
180                         // End the current fragment and make a new one for the header.
181                         fragments.push_back(Stream::FragmentStart { bytes_received, 0.0, true });
182                         process_queued_data();
183                         Stream::DataElement elem;
184                         elem.data.iov_base = (char *)stream_header.data();
185                         elem.data.iov_len = stream_header.size();
186                         add_data_raw({ elem });
187                         remove_obsolete_starting_points();
188
189                         // The discontinuity counter will be increased when
190                         // this header goes out of the backlog.
191                 }
192                 clear_hls_playlist_cache();
193         }
194         stream_header = new_stream_header;
195 }
196
197 void Stream::put_client_to_sleep(Client *client)
198 {
199         sleeping_clients.push_back(client);
200 }
201
202 // Return a new set of iovecs that contains only the first <bytes_wanted> bytes of <data>.
203 vector<iovec> collect_iovecs(const vector<Stream::DataElement> &data, size_t bytes_wanted)
204 {
205         vector<iovec> ret;
206         size_t max_iovecs = min<size_t>(data.size(), IOV_MAX);
207         for (size_t i = 0; i < max_iovecs && bytes_wanted > 0; ++i) {
208                 if (data[i].data.iov_len <= bytes_wanted) {
209                         // Consume the entire iovec.
210                         ret.push_back(data[i].data);
211                         bytes_wanted -= data[i].data.iov_len;
212                 } else {
213                         // Take only parts of this iovec.
214                         iovec iov;
215                         iov.iov_base = data[i].data.iov_base;
216                         iov.iov_len = bytes_wanted;
217                         ret.push_back(iov);
218                         bytes_wanted = 0;
219                 }
220         }
221         return ret;
222 }
223
224 // Return a new set of iovecs that contains all of <data> except the first <bytes_wanted> bytes.
225 vector<Stream::DataElement> remove_iovecs(const vector<Stream::DataElement> &data, size_t bytes_wanted)
226 {
227         vector<Stream::DataElement> ret;
228         size_t i;
229         for (i = 0; i < data.size() && bytes_wanted > 0; ++i) {
230                 if (data[i].data.iov_len <= bytes_wanted) {
231                         // Consume the entire iovec.
232                         bytes_wanted -= data[i].data.iov_len;
233                 } else {
234                         // Take only parts of this iovec.
235                         Stream::DataElement data_element;
236                         data_element.data.iov_base = reinterpret_cast<char *>(data[i].data.iov_base) + bytes_wanted;
237                         data_element.data.iov_len = data[i].data.iov_len - bytes_wanted;
238                         data_element.metacube_flags = METACUBE_FLAGS_NOT_SUITABLE_FOR_STREAM_START;
239                         data_element.pts = RationalPTS();
240                         ret.push_back(data_element);
241                         bytes_wanted = 0;
242                 }
243         }
244
245         // Add the rest of the iovecs unchanged.
246         ret.insert(ret.end(), data.begin() + i, data.end());
247         return ret;
248 }
249
250 void Stream::add_data_raw(const vector<DataElement> &orig_data)
251 {
252         vector<DataElement> data = orig_data;
253         while (!data.empty()) {
254                 size_t pos = bytes_received % backlog_size;
255
256                 // Collect as many iovecs as we can before we hit the point
257                 // where the circular buffer wraps around.
258                 vector<iovec> to_write = collect_iovecs(data, backlog_size - pos);
259                 ssize_t ret;
260                 do {
261                         ret = pwritev(data_fd, to_write.data(), to_write.size(), pos);
262                 } while (ret == -1 && errno == EINTR);
263
264                 if (ret == -1) {
265                         log_perror("pwritev");
266                         // Dazed and confused, but trying to continue...
267                         return;
268                 }
269                 bytes_received += ret;
270
271                 // Remove the data that was actually written from the set of iovecs.
272                 data = remove_iovecs(data, ret);
273         }
274 }
275
276 void Stream::remove_obsolete_starting_points()
277 {
278         // We could do a binary search here (std::lower_bound), but it seems
279         // overkill for removing what's probably only a few points.
280         while (!suitable_starting_points.empty() &&
281                bytes_received - suitable_starting_points[0] > backlog_size) {
282                 suitable_starting_points.pop_front();
283         }
284         assert(backlog_size >= hls_backlog_margin);
285         while (!fragments.empty() &&
286                bytes_received - fragments[0].byte_position > (backlog_size - hls_backlog_margin)) {
287                 if (fragments[0].begins_header) {
288                         ++discontinuity_counter;
289                 } else {
290                         ++first_fragment_index;
291                 }
292                 fragments.pop_front();
293                 clear_hls_playlist_cache();
294         }
295 }
296
297 void Stream::add_data_deferred(const char *data, size_t bytes, uint16_t metacube_flags, const RationalPTS &pts)
298 {
299         // For regular output, we don't want to send the client twice
300         // (it's already sent out together with the HTTP header).
301         // However, for Metacube output, we need to send it so that
302         // the Cubemap instance in the other end has a chance to update it.
303         // It may come twice in its stream, but Cubemap doesn't care.
304         if (encoding == Stream::STREAM_ENCODING_RAW &&
305             (metacube_flags & METACUBE_FLAGS_HEADER) != 0) {
306                 return;
307         }
308
309         lock_guard<mutex> lock(queued_data_mutex);
310
311         DataElement data_element;
312         data_element.metacube_flags = metacube_flags;
313         data_element.pts = pts;
314
315         if (encoding == Stream::STREAM_ENCODING_METACUBE) {
316                 // Construct a PTS metadata block. (We'll avoid sending it out
317                 // if we don't have a valid PTS.)
318                 metacube2_pts_packet pts_packet;
319                 pts_packet.type = htobe64(METACUBE_METADATA_TYPE_NEXT_BLOCK_PTS);
320                 pts_packet.pts = htobe64(pts.pts);
321                 pts_packet.timebase_num = htobe64(pts.timebase_num);
322                 pts_packet.timebase_den = htobe64(pts.timebase_den);
323
324                 metacube2_block_header pts_hdr;
325                 memcpy(pts_hdr.sync, METACUBE2_SYNC, sizeof(pts_hdr.sync));
326                 pts_hdr.size = htonl(sizeof(pts_packet));
327                 pts_hdr.flags = htons(METACUBE_FLAGS_METADATA);
328                 pts_hdr.csum = htons(metacube2_compute_crc(&pts_hdr));
329
330                 // Add a Metacube block header before the data.
331                 metacube2_block_header hdr;
332                 memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
333                 hdr.size = htonl(bytes);
334                 hdr.flags = htons(metacube_flags);
335                 hdr.csum = htons(metacube2_compute_crc(&hdr));
336
337                 data_element.data.iov_len = bytes + sizeof(hdr);
338                 if (pts.timebase_num != 0) {
339                         data_element.data.iov_len += sizeof(pts_hdr) + sizeof(pts_packet);
340                 }
341                 data_element.data.iov_base = new char[data_element.data.iov_len];
342
343                 char *ptr = reinterpret_cast<char *>(data_element.data.iov_base);
344                 if (pts.timebase_num != 0) {
345                         memcpy(ptr, &pts_hdr, sizeof(pts_hdr));
346                         ptr += sizeof(pts_hdr);
347                         memcpy(ptr, &pts_packet, sizeof(pts_packet));
348                         ptr += sizeof(pts_packet);
349                 }
350
351                 memcpy(ptr, &hdr, sizeof(hdr));
352                 ptr += sizeof(hdr);
353                 memcpy(ptr, data, bytes);
354
355                 queued_data.push_back(data_element);
356         } else if (encoding == Stream::STREAM_ENCODING_RAW) {
357                 // Just add the data itself.
358                 data_element.data.iov_base = new char[bytes];
359                 memcpy(data_element.data.iov_base, data, bytes);
360                 data_element.data.iov_len = bytes;
361
362                 queued_data.push_back(data_element);
363         } else {
364                 assert(false);
365         }
366 }
367
368 void Stream::process_queued_data()
369 {
370         vector<DataElement> queued_data_copy;
371
372         // Hold the lock for as short as possible, since add_data_raw() can possibly
373         // write to disk, which might disturb the input thread.
374         {
375                 lock_guard<mutex> lock(queued_data_mutex);
376                 if (queued_data.empty()) {
377                         return;
378                 }
379
380                 swap(queued_data, queued_data_copy);
381         }
382
383         // Add suitable starting points for the stream, if the queued data
384         // contains such starting points. Note that we drop starting points
385         // if they're less than 10 kB apart, so that we don't get a huge
386         // amount of them for e.g. each and every MPEG-TS 188-byte cell.
387         // The 10 kB value is somewhat arbitrary, but at least it should make
388         // the RAM cost of saving the position ~0.1% (or less) of the actual
389         // data, and 10 kB is a very fine granularity in most streams.
390         static const int minimum_start_point_distance = 10240;
391         size_t byte_position = bytes_received;
392         bool need_hls_clear = false;
393         for (const DataElement &elem : queued_data_copy) {
394                 if ((elem.metacube_flags & METACUBE_FLAGS_NOT_SUITABLE_FOR_STREAM_START) == 0) {
395                         size_t num_points = suitable_starting_points.size();
396                         if (num_points >= 2 &&
397                             suitable_starting_points[num_points - 1] - suitable_starting_points[num_points - 2] < minimum_start_point_distance) {
398                                 // p[n-1] - p[n-2] < 10 kB, so drop p[n-1].
399                                 suitable_starting_points.pop_back();
400                         }
401                         suitable_starting_points.push_back(byte_position);
402
403                         if (elem.pts.timebase_num != 0) {
404                                 need_hls_clear |= add_fragment_boundary(byte_position, elem.pts);
405                         }
406                 }
407                 byte_position += elem.data.iov_len;
408         }
409         if (need_hls_clear) {
410                 clear_hls_playlist_cache();
411         }
412
413         add_data_raw(queued_data_copy);
414         remove_obsolete_starting_points();
415         for (const DataElement &elem : queued_data_copy) {
416                 char *data = reinterpret_cast<char *>(elem.data.iov_base);
417                 delete[] data;
418         }
419
420         // We have more data, so wake up all clients.
421         if (to_process.empty()) {
422                 swap(sleeping_clients, to_process);
423         } else {
424                 to_process.insert(to_process.end(), sleeping_clients.begin(), sleeping_clients.end());
425                 sleeping_clients.clear();
426         }
427 }
428
429 bool Stream::add_fragment_boundary(size_t byte_position, const RationalPTS &pts)
430 {
431         double pts_double = double(pts.pts) * pts.timebase_den / pts.timebase_num;
432
433         if (fragments.size() <= 1 ||
434             fragments[fragments.size() - 1].begins_header ||
435             fragments[fragments.size() - 2].begins_header) {
436                 // Just starting up, so try to establish the first in-progress fragment.
437                 fragments.push_back(FragmentStart{ byte_position, pts_double, false });
438                 return false;
439         }
440
441         // Keep extending the in-progress fragment as long as we do not
442         // exceed the target duration by more than half a second
443         // (RFC 8216 4.3.3.1) and we get closer to the target by doing so.
444         // Note that in particular, this means we'll always extend
445         // as long as we don't exceed the target duration.
446         double current_duration = pts_double - fragments[fragments.size() - 1].pts;
447         double candidate_duration = pts_double - fragments[fragments.size() - 2].pts;
448         if (lrintf(candidate_duration) <= hls_frag_duration &&
449             fabs(candidate_duration - hls_frag_duration) < fabs(current_duration - hls_frag_duration)) {
450                 fragments.back() = FragmentStart{ byte_position, pts_double, false };
451                 return false;
452         } else {
453                 // Extending the in-progress fragment would make it too long,
454                 // so finalize it and start a new in-progress fragment.
455                 fragments.push_back(FragmentStart{ byte_position, pts_double, false });
456                 return true;
457         }
458 }
459
460 void Stream::clear_hls_playlist_cache()
461 {
462         hls_playlist_http10.reset();
463         hls_playlist_http11_close.reset();
464         hls_playlist_http11_persistent.reset();
465 }
466
467 shared_ptr<const string> Stream::generate_hls_playlist(bool http_11, bool close_after_response)
468 {
469         char buf[256];
470         snprintf(buf, sizeof(buf),
471                 "#EXTM3U\r\n"
472                 "#EXT-X-VERSION:7\r\n"
473                 "#EXT-X-TARGETDURATION:%u\r\n"
474                 "#EXT-X-MEDIA-SEQUENCE:%" PRIu64 "\r\n"
475                 "#EXT-X-DISCONTINUITY-SEQUENCE:%" PRIu64 "\r\n",
476                 hls_frag_duration,
477                 first_fragment_index,
478                 discontinuity_counter);
479
480         string playlist = buf;
481
482         if (fragments.size() >= 3) {
483                 bool printed_header_for_this_group = false;
484                 bool printed_first_header = false;
485                 for (size_t i = 0; i < fragments.size() - 2; ++i) {
486                         char buf[256];
487
488                         if (fragments[i].begins_header) {
489                                 // End of this group. (We've already printed the header
490                                 // as part of the previous group.)
491                                 printed_header_for_this_group = false;
492                                 continue;
493                         }
494                         if (!printed_header_for_this_group) {
495                                 // Look forward until we find the header for this group (if any).
496                                 for (size_t j = i + 1; j < fragments.size() - 1; ++j) {
497                                         if (fragments[j].begins_header) {
498                                                 if (printed_first_header) {
499                                                         playlist += "#EXT-X-DISCONTINUITY\r\n";
500                                                 }
501                                                 snprintf(buf, sizeof(buf),
502                                                         "#EXT-X-MAP:URI=\"%s?frag=%" PRIu64 "-%" PRIu64 "\"\r\n",
503                                                         url.c_str(), fragments[j].byte_position,
504                                                         fragments[j + 1].byte_position);
505                                                 playlist += buf;
506                                                 printed_first_header = true;
507                                                 printed_header_for_this_group = true;
508                                                 break;
509                                         }
510                                 }
511
512                                 if (!printed_header_for_this_group && !stream_header.empty()) {
513                                         if (printed_first_header) {
514                                                 playlist += "#EXT-X-DISCONTINUITY\r\n";
515                                         }
516                                         snprintf(buf, sizeof(buf), "#EXT-X-MAP:URI=\"%s?frag=header\"\r\n", url.c_str());
517                                         playlist += buf;
518                                 }
519
520                                 // Even if we didn't find anything, we don't want to search again for each fragment.
521                                 printed_first_header = true;
522                                 printed_header_for_this_group = true;
523                         }
524
525                         if (fragments[i + 1].begins_header) {
526                                 // Since we only have start pts for each block and not duration,
527                                 // we have no idea how long this fragment is; the encoder restarted
528                                 // before it got to output the next pts. However, it's likely
529                                 // to be very short, so instead of trying to guess, we just skip it.
530                                 continue;
531                         }
532
533                         snprintf(buf, sizeof(buf), "#EXTINF:%f,\r\n%s?frag=%" PRIu64 "-%" PRIu64 "\r\n",
534                                 fragments[i + 1].pts - fragments[i].pts,
535                                 url.c_str(),
536                                 fragments[i].byte_position,
537                                 fragments[i + 1].byte_position);
538                         playlist += buf;
539                 }
540         }
541
542         string response;
543         if (http_11) {
544                 response = "HTTP/1.1 200 OK\r\n";
545                 if (close_after_response) {
546                         response.append("Connection: close\r\n");
547                 }
548         } else {
549                 assert(close_after_response);
550                 response = "HTTP/1.0 200 OK\r\n";
551         }
552         snprintf(buf, sizeof(buf), "Content-Length: %zu\r\n", playlist.size());
553         response.append(buf);
554         response.append("Content-Type: application/x-mpegURL\r\n");
555         if (!allow_origin.empty()) {
556                 response.append("Access-Control-Allow-Origin: ");
557                 response.append(allow_origin);
558                 response.append("\r\n");
559         }
560         response.append("\r\n");
561         response.append(move(playlist));
562
563         return shared_ptr<const string>(new string(move(response)));
564 }