Fix a crash when a HTTP input connected to an UDP output goes unavailable.
[cubemap] / httpinput.cpp
1 #include <assert.h>
2 #include <errno.h>
3 #include <fcntl.h>
4 #include <math.h>
5 #include <netdb.h>
6 #include <netinet/in.h>
7 #include <poll.h>
8 #include <stdint.h>
9 #include <stdlib.h>
10 #include <string.h>
11 #include <sys/ioctl.h>
12 #include <sys/socket.h>
13 #include <sys/time.h>
14 #include <sys/types.h>
15 #include <sys/wait.h>
16 #include <time.h>
17 #include <unistd.h>
18 #include <math.h>
19 #include <spawn.h>
20 #include <map>
21 #include <string>
22 #include <utility>
23 #include <vector>
24
25 #include "httpinput.h"
26 #include "log.h"
27 #include "metacube2.h"
28 #include "parse.h"
29 #include "serverpool.h"
30 #include "state.pb.h"
31 #include "stream.h"
32 #include "timespec.h"
33 #include "util.h"
34 #include "version.h"
35
36 using namespace std;
37
38 namespace {
39
40 string host_header(const string &host, const string &port)
41 {
42         if (port == "http" || atoi(port.c_str()) == 80) {
43                 return host;
44         } else {
45                 return host + ":" + port;
46         }
47 }
48
49 }  // namespace
50
51 extern ServerPool *servers;
52
53 HTTPInput::HTTPInput(const string &url, Input::Encoding encoding)
54         : state(NOT_CONNECTED),
55           url(url),
56           encoding(encoding)
57 {
58         stats.url = url;
59         stats.bytes_received = 0;
60         stats.data_bytes_received = 0;
61         stats.metadata_bytes_received = 0;
62         stats.connect_time = -1;
63         stats.latency_sec = HUGE_VAL;
64 }
65
66 HTTPInput::HTTPInput(const InputProto &serialized)
67         : state(State(serialized.state())),
68           url(serialized.url()),
69           encoding(serialized.is_metacube_encoded() ?
70                    Input::INPUT_ENCODING_METACUBE :
71                    Input::INPUT_ENCODING_RAW),
72           request(serialized.request()),
73           request_bytes_sent(serialized.request_bytes_sent()),
74           response(serialized.response()),
75           http_header(serialized.http_header()),
76           stream_header(serialized.stream_header()),
77           has_metacube_header(serialized.has_metacube_header()),
78           sock(serialized.sock()),
79           child_pid(serialized.child_pid())
80 {
81         // Set back the close-on-exec flag for the socket.
82         // (This can't leak into a child, since we haven't been started yet.)
83         fcntl(sock, F_SETFD, O_CLOEXEC);
84
85         pending_data.resize(serialized.pending_data().size());
86         memcpy(&pending_data[0], serialized.pending_data().data(), serialized.pending_data().size());
87
88         string protocol, user;
89         parse_url(url, &protocol, &user, &host, &port, &path);  // Don't care if it fails.
90
91         stats.url = url;
92         stats.bytes_received = serialized.bytes_received();
93         stats.data_bytes_received = serialized.data_bytes_received();
94         stats.metadata_bytes_received = serialized.metadata_bytes_received();
95         if (serialized.has_connect_time()) {
96                 stats.connect_time = serialized.connect_time();
97         } else {
98                 stats.connect_time = time(nullptr);
99         }
100         if (serialized.has_latency_sec()) {
101                 stats.latency_sec = serialized.latency_sec();
102         } else {
103                 stats.latency_sec = HUGE_VAL;
104         }
105
106         last_verbose_connection.tv_sec = -3600;
107         last_verbose_connection.tv_nsec = 0;
108 }
109
110 void HTTPInput::close_socket()
111 {
112         if (sock != -1) {
113                 safe_close(sock);
114                 sock = -1;
115         }
116         if (child_pid != -1) {
117                 // Kill the child process group, forcibly.
118                 // TODO: Consider using a pidfd on newer kernels, so that we're guaranteed
119                 // never to kill the wrong process.
120                 kill(-child_pid, SIGKILL);
121         }
122         child_pid = -1;
123
124         lock_guard<mutex> lock(stats_mutex);
125         stats.connect_time = -1;
126 }
127
128 InputProto HTTPInput::serialize() const
129 {
130         // Unset the close-on-exec flag for the socket.
131         // (This can't leak into a child, since there's only one thread left.)
132         fcntl(sock, F_SETFD, 0);
133
134         InputProto serialized;
135         serialized.set_state(state);
136         serialized.set_url(url);
137         serialized.set_request(request);
138         serialized.set_request_bytes_sent(request_bytes_sent);
139         serialized.set_response(response);
140         serialized.set_http_header(http_header);
141         serialized.set_stream_header(stream_header);
142         serialized.set_pending_data(string(pending_data.begin(), pending_data.end()));
143         serialized.set_has_metacube_header(has_metacube_header);
144         serialized.set_sock(sock);
145         serialized.set_child_pid(child_pid);
146         serialized.set_bytes_received(stats.bytes_received);
147         serialized.set_data_bytes_received(stats.data_bytes_received);
148         if (isfinite(stats.latency_sec)) {
149                 serialized.set_latency_sec(stats.latency_sec);
150         }
151         serialized.set_connect_time(stats.connect_time);
152         if (encoding == Input::INPUT_ENCODING_METACUBE) {
153                 serialized.set_is_metacube_encoded(true);
154         } else {
155                 assert(encoding == Input::INPUT_ENCODING_RAW);
156                 serialized.set_is_metacube_encoded(false);
157         }
158         return serialized;
159 }
160
161 int HTTPInput::lookup_and_connect(const string &host, const string &port)
162 {
163         addrinfo *ai;
164         int err = getaddrinfo(host.c_str(), port.c_str(), nullptr, &ai);
165         if (err != 0) {
166                 if (!suppress_logging) {
167                         log(WARNING, "[%s] Lookup of '%s' failed (%s).",
168                                 url.c_str(), host.c_str(), gai_strerror(err));
169                 }
170                 return -1;
171         }
172
173         addrinfo *base_ai = ai;
174
175         // Connect to everything in turn until we have a socket.
176         for ( ; ai && !should_stop(); ai = ai->ai_next) {
177                 // Now do a non-blocking connect. This is important because we want to be able to be
178                 // woken up, even though it's rather cumbersome.
179                 int sock = socket(ai->ai_family, SOCK_STREAM | SOCK_NONBLOCK | SOCK_CLOEXEC, IPPROTO_TCP);
180                 if (sock == -1) {
181                         // Could be e.g. EPROTONOSUPPORT. The show must go on.
182                         continue;
183                 }
184
185                 // Do a non-blocking connect.
186                 do {
187                         err = connect(sock, ai->ai_addr, ai->ai_addrlen);
188                 } while (err == -1 && errno == EINTR);
189
190                 if (err == -1 && errno != EINPROGRESS) {
191                         log_perror("connect");
192                         safe_close(sock);
193                         continue;
194                 }
195
196                 // Wait for the connect to complete, or an error to happen.
197                 for ( ;; ) {
198                         bool complete = wait_for_activity(sock, POLLIN | POLLOUT, nullptr);
199                         if (should_stop()) {
200                                 safe_close(sock);
201                                 return -1;
202                         }
203                         if (complete) {
204                                 break;
205                         }
206                 }
207
208                 // Check whether it ended in an error or not.
209                 socklen_t err_size = sizeof(err);
210                 if (getsockopt(sock, SOL_SOCKET, SO_ERROR, &err, &err_size) == -1) {
211                         log_perror("getsockopt");
212                         safe_close(sock);
213                         continue;
214                 }
215
216                 errno = err;
217
218                 if (err == 0) {
219                         // Successful connect.
220                         freeaddrinfo(base_ai);
221                         return sock;
222                 }
223
224                 safe_close(sock);
225         }
226
227         // Give the last one as error.
228         if (!suppress_logging) {
229                 log(WARNING, "[%s] Connect to '%s' failed (%s)",
230                         url.c_str(), host.c_str(), strerror(errno));
231         }
232         freeaddrinfo(base_ai);
233         return -1;
234 }
235
236 int HTTPInput::open_child_process(const string &cmdline)
237 {
238         int devnullfd = open("/dev/null", O_RDONLY | O_CLOEXEC);
239         if (devnullfd == -1) {
240                 log_perror("/dev/null");
241                 return -1;
242         }
243
244         int pipefd[2];
245         if (pipe2(pipefd, O_CLOEXEC) == -1) {
246                 log_perror("pipe2()");
247                 close(devnullfd);
248                 return -1;
249         }
250
251         // Point stdout to us, stdin to /dev/null, and stderr remains where it is
252         // (probably the systemd log). All other file descriptors should be marked
253         // as close-on-exec, and should thus not leak into the child.
254         posix_spawn_file_actions_t actions;
255         posix_spawn_file_actions_init(&actions);
256         posix_spawn_file_actions_adddup2(&actions, devnullfd, 0);
257         posix_spawn_file_actions_adddup2(&actions, pipefd[1], 1);
258
259         // Make the process a leader of its own process group, so that we can easily
260         // kill it and any of its child processes (unless it's started new process
261         // groups itself, of course).
262         posix_spawnattr_t attr;
263         posix_spawnattr_init(&attr);
264         posix_spawnattr_setflags(&attr, POSIX_SPAWN_SETPGROUP);
265         posix_spawnattr_setpgroup(&attr, 0);
266
267         char * const argv[] = {
268                 strdup("/bin/sh"),
269                 strdup("-c"),
270                 strdup(path.c_str()),
271                 nullptr
272         };
273         int err = posix_spawn(&child_pid, "/bin/sh", &actions, &attr, argv, /*envp=*/nullptr);
274         posix_spawn_file_actions_destroy(&actions);
275         posix_spawnattr_destroy(&attr);
276         free(argv[0]);
277         free(argv[1]);
278         free(argv[2]);
279         close(devnullfd);
280         close(pipefd[1]);
281
282         if (err == 0) {
283                 return pipefd[0];
284         } else {
285                 child_pid = -1;
286                 log_perror(cmdline.c_str());
287                 close(pipefd[0]);
288                 return -1;
289         }
290 }
291         
292 bool HTTPInput::parse_response(const string &request)
293 {
294         vector<string> lines = split_lines(response);
295         if (lines.empty()) {
296                 if (!suppress_logging) {
297                         log(WARNING, "[%s] Empty HTTP response from input.", url.c_str());
298                 }
299                 return false;
300         }
301
302         vector<string> first_line_tokens = split_tokens(lines[0]);
303         if (first_line_tokens.size() < 2) {
304                 if (!suppress_logging) {
305                         log(WARNING, "[%s] Malformed response line '%s' from input.",
306                                 url.c_str(), lines[0].c_str());
307                 }
308                 return false;
309         }
310
311         int response = atoi(first_line_tokens[1].c_str());
312         if (response != 200) {
313                 if (!suppress_logging) {
314                         log(WARNING, "[%s] Non-200 response '%s' from input.",
315                                 url.c_str(), lines[0].c_str());
316                 }
317                 return false;
318         }
319
320         HTTPHeaderMultimap parameters = extract_headers(lines, url);
321
322         // Remove “Content-encoding: metacube”.
323         const auto encoding_it = parameters.find("Content-Encoding");
324         if (encoding_it != parameters.end() && encoding_it->second == "metacube") {
325                 parameters.erase(encoding_it);
326         }
327
328         // Change “Server: foo” to “Server: metacube/0.1 (reflecting: foo)”
329         // XXX: Use a Via: instead?
330         if (parameters.count("Server") == 0) {
331                 parameters.insert(make_pair("Server", SERVER_IDENTIFICATION));
332         } else {
333                 for (auto &key_and_value : parameters) {
334                         if (key_and_value.first != "Server") {
335                                 continue;
336                         }
337                         key_and_value.second = SERVER_IDENTIFICATION " (reflecting: " + key_and_value.second + ")";
338                 }
339         }
340
341         // Erase “Connection: close”; we'll set it on the sending side if needed.
342         parameters.erase("Connection");
343
344         // Construct the new HTTP header.
345         http_header = "HTTP/1.0 200 OK\r\n";
346         for (const auto &key_and_value : parameters) {
347                 http_header.append(key_and_value.first + ": " + key_and_value.second + "\r\n");
348         }
349
350         for (int stream_index : stream_indices) {
351                 servers->set_header(stream_index, http_header, stream_header);
352         }
353
354         return true;
355 }
356
357 void HTTPInput::do_work()
358 {
359         timespec last_activity;
360
361         // TODO: Make the timeout persist across restarts.
362         if (state == SENDING_REQUEST || state == RECEIVING_HEADER || state == RECEIVING_DATA) {
363                 int err = clock_gettime(CLOCK_MONOTONIC_COARSE, &last_activity);
364                 assert(err != -1);
365         }
366
367         while (!should_stop()) {
368                 if (state == SENDING_REQUEST || state == RECEIVING_HEADER || state == RECEIVING_DATA) {
369                         // Give the socket 30 seconds since last activity before we time out.
370                         static const int timeout_secs = 30;
371
372                         timespec now;
373                         int err = clock_gettime(CLOCK_MONOTONIC_COARSE, &now);
374                         assert(err != -1);
375
376                         timespec elapsed = clock_diff(last_activity, now);
377                         if (elapsed.tv_sec >= timeout_secs) {
378                                 // Timeout!
379                                 if (!suppress_logging) {
380                                         log(ERROR, "[%s] Timeout after %d seconds, closing.", url.c_str(), elapsed.tv_sec);
381                                 }
382                                 state = CLOSING_SOCKET;
383                                 continue;
384                         }
385
386                         // Basically calculate (30 - (now - last_activity)) = (30 + (last_activity - now)).
387                         // Add a second of slack to account for differences between clocks.
388                         timespec timeout = clock_diff(now, last_activity);
389                         timeout.tv_sec += timeout_secs + 1;
390                         assert(timeout.tv_sec > 0 || (timeout.tv_sec >= 0 && timeout.tv_nsec > 0));
391
392                         bool activity = wait_for_activity(sock, (state == SENDING_REQUEST) ? POLLOUT : POLLIN, &timeout);
393                         if (activity) {
394                                 err = clock_gettime(CLOCK_MONOTONIC_COARSE, &last_activity);
395                                 assert(err != -1);
396                         } else {
397                                 // OK. Most likely, should_stop was set, or we have timed out.
398                                 continue;
399                         }
400                 }
401
402                 switch (state) {
403                 case NOT_CONNECTED: {
404                         // Reap any exited children.
405                         int wstatus, err;
406                         do {
407                                 err = waitpid(-1, &wstatus, WNOHANG);
408                                 if (err == -1) {
409                                         if (errno == EINTR) {
410                                                 continue;
411                                         }
412                                         if (errno == ECHILD) {
413                                                 break;
414                                         }
415                                         log_perror("waitpid");
416                                         break;
417                                 }
418                         } while (err != 0);
419                         child_pid = -1;
420
421                         request.clear();
422                         request_bytes_sent = 0;
423                         response.clear();
424                         pending_data.clear();
425                         has_metacube_header = false;
426                         for (int stream_index : stream_indices) {
427                                 // Don't zero out the header; it might still be of use to HLS clients.
428                                 servers->set_unavailable(stream_index);
429                         }
430
431                         string protocol;
432                         {
433                                 string user;  // Thrown away.
434                                 if (!parse_url(url, &protocol, &user, &host, &port, &path)) {
435                                         if (!suppress_logging) {
436                                                 log(WARNING, "[%s] Failed to parse URL '%s'", url.c_str(), url.c_str());
437                                         }
438                                         break;
439                                 }
440
441                                 // Remove the brackets around IPv6 address literals.
442                                 // TODO: See if we can join this with the code in parse_ip_address(),
443                                 // or maybe even more it into parse_url().
444                                 if (!host.empty() && host[0] == '[' && host[host.size() - 1] == ']') {
445                                         host = host.substr(1, host.size() - 2);
446                                 }
447                         }
448
449                         if (suppress_logging) {
450                                 // See if there's more than one minute since last time we made a connection
451                                 // with logging enabled. If so, turn it on again.
452                                 timespec now;
453                                 int err = clock_gettime(CLOCK_MONOTONIC_COARSE, &now);
454                                 assert(err != -1);
455
456                                 double elapsed = now.tv_sec - last_verbose_connection.tv_sec +
457                                         1e-9 * (now.tv_nsec - last_verbose_connection.tv_nsec);
458                                 if (elapsed > 60.0) {
459                                         suppress_logging = false;
460                                 }
461                         }
462                         if (!suppress_logging) {
463                                 int err = clock_gettime(CLOCK_MONOTONIC_COARSE, &last_verbose_connection);
464                                 assert(err != -1);
465                         }
466                         ++num_connection_attempts;
467                         if (protocol == "pipe") {
468                                 sock = open_child_process(path.c_str());
469
470                                 if (sock != -1) {
471                                         // Construct a minimal HTTP header.
472                                         http_header = "HTTP/1.0 200 OK\r\n";
473                                         for (int stream_index : stream_indices) {
474                                                 servers->set_header(stream_index, http_header, stream_header);
475                                         }
476                                         state = RECEIVING_DATA;
477                                 }
478                         } else {
479                                 sock = lookup_and_connect(host, port);
480                                 if (sock != -1) {
481                                         // Yay, successful connect.
482                                         state = SENDING_REQUEST;
483                                         request = "GET " + path + " HTTP/1.0\r\nHost: " + host_header(host, port) + "\r\nUser-Agent: cubemap\r\n\r\n";
484                                         request_bytes_sent = 0;
485                                 }
486                         }
487                         if (sock != -1) {
488                                 lock_guard<mutex> lock(stats_mutex);
489                                 stats.connect_time = time(nullptr);
490                                 clock_gettime(CLOCK_MONOTONIC_COARSE, &last_activity);
491                         }
492                         break;
493                 }
494                 case SENDING_REQUEST: {
495                         size_t to_send = request.size() - request_bytes_sent;
496                         int ret;
497
498                         do {
499                                 ret = write(sock, request.data() + request_bytes_sent, to_send);
500                         } while (ret == -1 && errno == EINTR);
501
502                         if (ret == -1) {
503                                 log_perror("write");
504                                 state = CLOSING_SOCKET;
505                                 continue;
506                         }
507
508                         assert(ret >= 0);
509                         request_bytes_sent += ret;
510
511                         if (request_bytes_sent == request.size()) {
512                                 state = RECEIVING_HEADER;
513                         }
514                         break;
515                 }
516                 case RECEIVING_HEADER: {
517                         char buf[4096];
518                         int ret;
519
520                         do {
521                                 ret = read(sock, buf, sizeof(buf));
522                         } while (ret == -1 && errno == EINTR);
523
524                         if (ret == -1) {
525                                 log_perror("read");
526                                 state = CLOSING_SOCKET;
527                                 continue;
528                         }
529
530                         if (ret == 0) {
531                                 // This really shouldn't happen...
532                                 if (!suppress_logging) {
533                                         log(ERROR, "[%s] Socket unexpectedly closed while reading header",
534                                                    url.c_str());
535                                 }
536                                 state = CLOSING_SOCKET;
537                                 continue;
538                         }
539                         
540                         RequestParseStatus status = wait_for_double_newline(&response, buf, ret);
541                         
542                         if (status == RP_OUT_OF_SPACE) {
543                                 if (!suppress_logging) {
544                                         log(WARNING, "[%s] Server sent overlong HTTP response!", url.c_str());
545                                 }
546                                 state = CLOSING_SOCKET;
547                                 continue;
548                         } else if (status == RP_NOT_FINISHED_YET) {
549                                 continue;
550                         }
551         
552                         // OK, so we're fine, but there might be some of the actual data after the response.
553                         // We'll need to deal with that separately.
554                         string extra_data;
555                         if (status == RP_EXTRA_DATA) {
556                                 char *ptr = static_cast<char *>(
557                                         memmem(response.data(), response.size(), "\r\n\r\n", 4));
558                                 assert(ptr != nullptr);
559                                 extra_data = string(ptr + 4, &response[0] + response.size());
560                                 response.resize(ptr - response.data());
561                         }
562
563                         if (!parse_response(response)) {
564                                 state = CLOSING_SOCKET;
565                                 continue;
566                         }
567
568                         if (!extra_data.empty()) {
569                                 process_data(&extra_data[0], extra_data.size());
570                         }
571
572                         if (!suppress_logging) {
573                                 if (encoding == Input::INPUT_ENCODING_RAW) {
574                                         log(INFO, "[%s] Connected to '%s', receiving raw data.",
575                                                    url.c_str(), url.c_str());
576                                 } else {
577                                         assert(encoding == Input::INPUT_ENCODING_METACUBE);
578                                         log(INFO, "[%s] Connected to '%s', receiving data.",
579                                                    url.c_str(), url.c_str());
580                                 }
581                         }
582                         state = RECEIVING_DATA;
583                         break;
584                 }
585                 case RECEIVING_DATA: {
586                         char buf[4096];
587                         int ret;
588
589                         do {
590                                 ret = read(sock, buf, sizeof(buf));
591                         } while (ret == -1 && errno == EINTR);
592
593                         if (ret == -1) {
594                                 log_perror("read");
595                                 state = CLOSING_SOCKET;
596                                 continue;
597                         }
598
599                         if (ret == 0) {
600                                 // This really shouldn't happen...
601                                 if (!suppress_logging) {
602                                         log(ERROR, "[%s] Socket unexpectedly closed while reading data",
603                                                    url.c_str());
604                                 }
605                                 state = CLOSING_SOCKET;
606                                 continue;
607                         }
608
609                         num_connection_attempts = 0;  // Reset, since we have a successful read.
610                         if (suppress_logging) {
611                                 // This was suppressed earlier, so print it out now.
612                                 if (encoding == Input::INPUT_ENCODING_RAW) {
613                                         log(INFO, "[%s] Connected to '%s', receiving raw data.",
614                                                    url.c_str(), url.c_str());
615                                 } else {
616                                         assert(encoding == Input::INPUT_ENCODING_METACUBE);
617                                         log(INFO, "[%s] Connected to '%s', receiving data.",
618                                                    url.c_str(), url.c_str());
619                                 }
620                                 suppress_logging = false;
621                         }
622
623                         process_data(buf, ret);
624                         break;
625                 }
626                 case CLOSING_SOCKET: {
627                         close_socket();
628                         state = NOT_CONNECTED;
629                         break;
630                 }
631                 default:
632                         assert(false);
633                 }
634
635                 // If we are still in NOT_CONNECTED, either something went wrong,
636                 // or the connection just got closed.
637                 // The earlier steps have already given the error message, if any.
638                 if (state == NOT_CONNECTED && !should_stop()) {
639                         if (!suppress_logging) {
640                                 log(INFO, "[%s] Waiting 0.2 seconds and restarting...", url.c_str());
641                         }
642
643                         if (num_connection_attempts >= 3 && !suppress_logging) {
644                                 log(INFO, "[%s] %d failed connection attempts, suppressing logging for one minute.",
645                                         url.c_str(), num_connection_attempts);
646                                 suppress_logging = true;
647                         }
648                         timespec timeout_ts;
649                         timeout_ts.tv_sec = 0;
650                         timeout_ts.tv_nsec = 200000000;
651                         wait_for_wakeup(&timeout_ts);
652                 }
653         }
654 }
655
656 void HTTPInput::process_data(char *ptr, size_t bytes)
657 {
658         {
659                 lock_guard<mutex> lock(stats_mutex);
660                 stats.bytes_received += bytes;
661         }
662
663         if (encoding == Input::INPUT_ENCODING_RAW) {
664                 for (int stream_index : stream_indices) {
665                         servers->add_data(stream_index, ptr, bytes, /*metacube_flags=*/0, /*pts=*/RationalPTS());
666                 }
667                 return;
668         }
669
670         assert(encoding == Input::INPUT_ENCODING_METACUBE);
671         pending_data.insert(pending_data.end(), ptr, ptr + bytes);
672
673         for ( ;; ) {
674                 // If we don't have enough data (yet) for even the Metacube header, just return.
675                 if (pending_data.size() < sizeof(metacube2_block_header)) {
676                         return;
677                 }
678
679                 // Make sure we have the Metacube sync header at the start.
680                 // We may need to skip over junk data (it _should_ not happen, though).
681                 if (!has_metacube_header) {
682                         char *ptr = static_cast<char *>(
683                                 memmem(pending_data.data(), pending_data.size(),
684                                        METACUBE2_SYNC, strlen(METACUBE2_SYNC)));
685                         if (ptr == nullptr) {
686                                 // OK, so we didn't find the sync marker. We know then that
687                                 // we do not have the _full_ marker in the buffer, but we
688                                 // could have N-1 bytes. Drop everything before that,
689                                 // and then give up.
690                                 drop_pending_data(pending_data.size() - (strlen(METACUBE2_SYNC) - 1));
691                                 return;
692                         } else {
693                                 // Yay, we found the header. Drop everything (if anything) before it.
694                                 drop_pending_data(ptr - pending_data.data());
695                                 has_metacube_header = true;
696
697                                 // Re-check that we have the entire header; we could have dropped data.
698                                 if (pending_data.size() < sizeof(metacube2_block_header)) {
699                                         return;
700                                 }
701                         }
702                 }
703
704                 // Now it's safe to read the header.
705                 metacube2_block_header hdr;
706                 memcpy(&hdr, pending_data.data(), sizeof(hdr));
707                 assert(memcmp(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync)) == 0);
708                 uint32_t size = ntohl(hdr.size);
709                 uint16_t flags = ntohs(hdr.flags);
710                 uint16_t expected_csum = metacube2_compute_crc(&hdr);
711
712                 if (expected_csum != ntohs(hdr.csum)) {
713                         log(WARNING, "[%s] Metacube checksum failed (expected 0x%x, got 0x%x), "
714                                 "not reading block claiming to be %d bytes (flags=%x).",
715                                 url.c_str(), expected_csum, ntohs(hdr.csum),
716                                 size, flags);
717
718                         // Drop only the first byte, and let the rest of the code handle resync.
719                         pending_data.erase(pending_data.begin(), pending_data.begin() + 1);
720                         has_metacube_header = false;
721                         continue;
722                 }
723                 if (size > 10485760) {
724                         log(WARNING, "[%s] Metacube block of %d bytes (flags=%x); corrupted header??",
725                                 url.c_str(), size, flags);
726                 }
727
728                 // See if we have the entire block. If not, wait for more data.
729                 if (pending_data.size() < sizeof(metacube2_block_header) + size) {
730                         return;
731                 }
732
733                 // See if this is a metadata block. If so, we don't want to send it on,
734                 // but rather process it ourselves.
735                 // TODO: Keep metadata when sending on to other Metacube users.
736                 if (flags & METACUBE_FLAGS_METADATA) {
737                         {
738                                 lock_guard<mutex> lock(stats_mutex);
739                                 stats.metadata_bytes_received += size;
740                         }
741                         process_metacube_metadata_block(hdr, pending_data.data() + sizeof(hdr), size);
742                 } else {
743                         // Send this block on to the servers.
744                         {
745                                 lock_guard<mutex> lock(stats_mutex);
746                                 stats.data_bytes_received += size;
747                         }
748                         char *inner_data = pending_data.data() + sizeof(metacube2_block_header);
749                         if (flags & METACUBE_FLAGS_HEADER) {
750                                 stream_header = string(inner_data, inner_data + size);
751                                 for (int stream_index : stream_indices) {
752                                         servers->set_header(stream_index, http_header, stream_header);
753                                 }
754                         }
755                         for (int stream_index : stream_indices) {
756                                 servers->add_data(stream_index, inner_data, size, flags, next_block_pts);
757                         }
758                         next_block_pts = RationalPTS();
759                 }
760
761                 // Consume the block. This isn't the most efficient way of dealing with things
762                 // should we have many blocks, but these routines don't need to be too efficient
763                 // anyway.
764                 pending_data.erase(pending_data.begin(), pending_data.begin() + sizeof(metacube2_block_header) + size);
765                 has_metacube_header = false;
766         }
767 }
768
769 void HTTPInput::drop_pending_data(size_t num_bytes)
770 {
771         if (num_bytes == 0) {
772                 return;
773         }
774         log(WARNING, "[%s] Dropping %lld junk bytes; not a Metacube2 stream, or data was dropped from the middle of the stream.",
775                 url.c_str(), (long long)num_bytes);
776         assert(pending_data.size() >= num_bytes);
777         pending_data.erase(pending_data.begin(), pending_data.begin() + num_bytes);
778 }
779
780 void HTTPInput::add_destination(int stream_index)
781 {
782         stream_indices.push_back(stream_index);
783         servers->set_header(stream_index, http_header, stream_header);
784 }
785
786 InputStats HTTPInput::get_stats() const
787 {
788         lock_guard<mutex> lock(stats_mutex);
789         return stats;
790 }
791
792 void HTTPInput::process_metacube_metadata_block(const metacube2_block_header &hdr, const char *payload, uint32_t payload_size)
793 {
794         if (payload_size < sizeof(uint64_t)) {
795                 log(WARNING, "[%s] Undersized Metacube metadata block (%d bytes); corrupted header?",
796                         url.c_str(), payload_size);
797                 return;
798         }
799
800         uint64_t type = be64toh(*(const uint64_t *)payload);
801         if (type == METACUBE_METADATA_TYPE_ENCODER_TIMESTAMP) {
802                 timespec now;
803                 clock_gettime(CLOCK_REALTIME, &now);
804
805                 const metacube2_timestamp_packet *pkt = (const metacube2_timestamp_packet *)payload;
806                 if (payload_size != sizeof(*pkt)) {
807                         log(WARNING, "[%s] Metacube timestamp block of wrong size (%d bytes); ignoring.",
808                                 url.c_str(), payload_size);
809                         return;
810                 }
811
812                 double elapsed = now.tv_sec - be64toh(pkt->tv_sec) +
813                         1e-9 * (now.tv_nsec - long(be64toh(pkt->tv_nsec)));
814                 {
815                         lock_guard<mutex> lock(stats_mutex);
816                         stats.latency_sec = elapsed;
817                 }
818         } else if (type == METACUBE_METADATA_TYPE_NEXT_BLOCK_PTS) {
819                 const metacube2_pts_packet *pkt = (const metacube2_pts_packet *)payload;
820                 if (payload_size != sizeof(*pkt)) {
821                         log(WARNING, "[%s] Metacube pts block of wrong size (%d bytes); ignoring.",
822                                 url.c_str(), payload_size);
823                         return;
824                 }
825                 next_block_pts.pts = be64toh(pkt->pts);
826                 next_block_pts.timebase_num = be64toh(pkt->timebase_num);
827                 next_block_pts.timebase_den = be64toh(pkt->timebase_den);
828         } else {
829                 // Unknown metadata block, ignore
830                 log(INFO, "[%s] Metadata block %llu\n", url.c_str(), type);
831                 return;
832         }
833 }