6 #include <netinet/in.h>
11 #include <sys/ioctl.h>
12 #include <sys/socket.h>
14 #include <sys/types.h>
25 #include "httpinput.h"
27 #include "metacube2.h"
29 #include "serverpool.h"
40 string host_header(const string &host, const string &port)
42 if (port == "http" || atoi(port.c_str()) == 80) {
45 return host + ":" + port;
51 extern ServerPool *servers;
53 HTTPInput::HTTPInput(const string &url, Input::Encoding encoding)
54 : state(NOT_CONNECTED),
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;
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())
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.)
84 fcntl(sock, F_SETFD, FD_CLOEXEC);
87 pending_data.resize(serialized.pending_data().size());
88 memcpy(&pending_data[0], serialized.pending_data().data(), serialized.pending_data().size());
90 string protocol, user;
91 parse_url(url, &protocol, &user, &host, &port, &path); // Don't care if it fails.
94 stats.bytes_received = serialized.bytes_received();
95 stats.data_bytes_received = serialized.data_bytes_received();
96 stats.metadata_bytes_received = serialized.metadata_bytes_received();
97 if (serialized.has_connect_time()) {
98 stats.connect_time = serialized.connect_time();
100 stats.connect_time = time(nullptr);
102 if (serialized.has_latency_sec()) {
103 stats.latency_sec = serialized.latency_sec();
105 stats.latency_sec = HUGE_VAL;
108 last_verbose_connection.tv_sec = -3600;
109 last_verbose_connection.tv_nsec = 0;
112 void HTTPInput::close_socket()
118 if (child_pid != -1) {
119 // Kill the child process group, forcibly.
120 // TODO: Consider using a pidfd on newer kernels, so that we're guaranteed
121 // never to kill the wrong process.
122 kill(-child_pid, SIGKILL);
126 lock_guard<mutex> lock(stats_mutex);
127 stats.connect_time = -1;
130 InputProto HTTPInput::serialize() const
132 // Unset the close-on-exec flag for the socket.
133 // (This can't leak into a child, since there's only one thread left.)
134 fcntl(sock, F_SETFD, 0);
136 InputProto serialized;
137 serialized.set_state(state);
138 serialized.set_url(url);
139 serialized.set_request(request);
140 serialized.set_request_bytes_sent(request_bytes_sent);
141 serialized.set_response(response);
142 serialized.set_http_header(http_header);
143 serialized.set_stream_header(stream_header);
144 serialized.set_pending_data(string(pending_data.begin(), pending_data.end()));
145 serialized.set_has_metacube_header(has_metacube_header);
146 serialized.set_sock(sock);
147 serialized.set_child_pid(child_pid);
148 serialized.set_bytes_received(stats.bytes_received);
149 serialized.set_data_bytes_received(stats.data_bytes_received);
150 if (isfinite(stats.latency_sec)) {
151 serialized.set_latency_sec(stats.latency_sec);
153 serialized.set_connect_time(stats.connect_time);
154 if (encoding == Input::INPUT_ENCODING_METACUBE) {
155 serialized.set_is_metacube_encoded(true);
157 assert(encoding == Input::INPUT_ENCODING_RAW);
158 serialized.set_is_metacube_encoded(false);
163 int HTTPInput::lookup_and_connect(const string &host, const string &port)
166 int err = getaddrinfo(host.c_str(), port.c_str(), nullptr, &ai);
168 if (!suppress_logging) {
169 log(WARNING, "[%s] Lookup of '%s' failed (%s).",
170 url.c_str(), host.c_str(), gai_strerror(err));
176 addrinfo *base_ai = ai;
178 // Connect to everything in turn until we have a socket.
179 for ( ; ai && !should_stop(); ai = ai->ai_next) {
180 // Now do a non-blocking connect. This is important because we want to be able to be
181 // woken up, even though it's rather cumbersome.
182 int sock = socket(ai->ai_family, SOCK_STREAM | SOCK_NONBLOCK | SOCK_CLOEXEC, IPPROTO_TCP);
184 // Could be e.g. EPROTONOSUPPORT. The show must go on.
188 // Do a non-blocking connect.
190 err = connect(sock, ai->ai_addr, ai->ai_addrlen);
191 } while (err == -1 && errno == EINTR);
193 if (err == -1 && errno != EINPROGRESS) {
194 log_perror("connect");
199 // Wait for the connect to complete, or an error to happen.
201 bool complete = wait_for_activity(sock, POLLIN | POLLOUT, nullptr);
204 freeaddrinfo(base_ai);
212 // Check whether it ended in an error or not.
213 socklen_t err_size = sizeof(err);
214 if (getsockopt(sock, SOL_SOCKET, SO_ERROR, &err, &err_size) == -1) {
215 log_perror("getsockopt");
223 // Successful connect.
224 freeaddrinfo(base_ai);
231 // Give the last one as error.
232 if (!suppress_logging) {
233 log(WARNING, "[%s] Connect to '%s' failed (%s)",
234 url.c_str(), host.c_str(), strerror(errno));
236 freeaddrinfo(base_ai);
240 int HTTPInput::open_child_process(const string &cmdline)
242 int devnullfd = open("/dev/null", O_RDONLY | O_CLOEXEC);
243 if (devnullfd == -1) {
244 log_perror("/dev/null");
249 if (pipe2(pipefd, O_CLOEXEC) == -1) {
250 log_perror("pipe2()");
255 // Point stdout to us, stdin to /dev/null, and stderr remains where it is
256 // (probably the systemd log). All other file descriptors should be marked
257 // as close-on-exec, and should thus not leak into the child.
258 posix_spawn_file_actions_t actions;
259 posix_spawn_file_actions_init(&actions);
260 posix_spawn_file_actions_adddup2(&actions, devnullfd, 0);
261 posix_spawn_file_actions_adddup2(&actions, pipefd[1], 1);
263 // Make the process a leader of its own process group, so that we can easily
264 // kill it and any of its child processes (unless it's started new process
265 // groups itself, of course).
266 posix_spawnattr_t attr;
267 posix_spawnattr_init(&attr);
268 posix_spawnattr_setflags(&attr, POSIX_SPAWN_SETPGROUP);
269 posix_spawnattr_setpgroup(&attr, 0);
271 char * const argv[] = {
274 strdup(path.c_str()),
277 int err = posix_spawn(&child_pid, "/bin/sh", &actions, &attr, argv, /*envp=*/nullptr);
278 posix_spawn_file_actions_destroy(&actions);
279 posix_spawnattr_destroy(&attr);
290 log_perror(cmdline.c_str());
296 bool HTTPInput::parse_response(const string &request)
298 vector<string> lines = split_lines(response);
300 if (!suppress_logging) {
301 log(WARNING, "[%s] Empty HTTP response from input.", url.c_str());
306 vector<string> first_line_tokens = split_tokens(lines[0]);
307 if (first_line_tokens.size() < 2) {
308 if (!suppress_logging) {
309 log(WARNING, "[%s] Malformed response line '%s' from input.",
310 url.c_str(), lines[0].c_str());
315 int response = atoi(first_line_tokens[1].c_str());
316 if (response != 200) {
317 if (!suppress_logging) {
318 log(WARNING, "[%s] Non-200 response '%s' from input.",
319 url.c_str(), lines[0].c_str());
324 HTTPHeaderMultimap parameters = extract_headers(lines, url);
326 // Remove “Content-encoding: metacube”.
327 const auto encoding_it = parameters.find("Content-Encoding");
328 if (encoding_it != parameters.end() && encoding_it->second == "metacube") {
329 parameters.erase(encoding_it);
332 // Change “Server: foo” to “Server: metacube/0.1 (reflecting: foo)”
333 // XXX: Use a Via: instead?
334 if (parameters.count("Server") == 0) {
335 parameters.insert(make_pair("Server", SERVER_IDENTIFICATION));
337 for (auto &key_and_value : parameters) {
338 if (key_and_value.first != "Server") {
341 key_and_value.second = SERVER_IDENTIFICATION " (reflecting: " + key_and_value.second + ")";
345 // Erase “Connection: close”; we'll set it on the sending side if needed.
346 parameters.erase("Connection");
348 // Construct the new HTTP header.
349 http_header = "HTTP/1.0 200 OK\r\n";
350 for (const auto &key_and_value : parameters) {
351 http_header.append(key_and_value.first + ": " + key_and_value.second + "\r\n");
354 for (int stream_index : stream_indices) {
355 servers->set_header(stream_index, http_header, stream_header);
361 void HTTPInput::do_work()
363 timespec last_activity;
365 // TODO: Make the timeout persist across restarts.
366 if (state == SENDING_REQUEST || state == RECEIVING_HEADER || state == RECEIVING_DATA) {
367 int err = clock_gettime(CLOCK_MONOTONIC_COARSE, &last_activity);
371 while (!should_stop()) {
372 if (state == SENDING_REQUEST || state == RECEIVING_HEADER || state == RECEIVING_DATA) {
373 // Give the socket 30 seconds since last activity before we time out.
374 static const int timeout_secs = 30;
377 int err = clock_gettime(CLOCK_MONOTONIC_COARSE, &now);
380 timespec elapsed = clock_diff(last_activity, now);
381 if (elapsed.tv_sec >= timeout_secs) {
383 if (!suppress_logging) {
384 log(ERROR, "[%s] Timeout after %d seconds, closing.", url.c_str(), elapsed.tv_sec);
386 state = CLOSING_SOCKET;
390 // Basically calculate (30 - (now - last_activity)) = (30 + (last_activity - now)).
391 // Add a second of slack to account for differences between clocks.
392 timespec timeout = clock_diff(now, last_activity);
393 timeout.tv_sec += timeout_secs + 1;
394 assert(timeout.tv_sec > 0 || (timeout.tv_sec >= 0 && timeout.tv_nsec > 0));
396 bool activity = wait_for_activity(sock, (state == SENDING_REQUEST) ? POLLOUT : POLLIN, &timeout);
398 err = clock_gettime(CLOCK_MONOTONIC_COARSE, &last_activity);
401 // OK. Most likely, should_stop was set, or we have timed out.
407 case NOT_CONNECTED: {
408 // Reap any exited children.
411 err = waitpid(-1, &wstatus, WNOHANG);
413 if (errno == EINTR) {
416 if (errno == ECHILD) {
419 log_perror("waitpid");
426 request_bytes_sent = 0;
428 pending_data.clear();
429 has_metacube_header = false;
430 for (int stream_index : stream_indices) {
431 // Don't zero out the header; it might still be of use to HLS clients.
432 servers->set_unavailable(stream_index);
437 string user; // Thrown away.
438 if (!parse_url(url, &protocol, &user, &host, &port, &path)) {
439 if (!suppress_logging) {
440 log(WARNING, "[%s] Failed to parse URL '%s'", url.c_str(), url.c_str());
445 // Remove the brackets around IPv6 address literals.
446 // TODO: See if we can join this with the code in parse_ip_address(),
447 // or maybe even more it into parse_url().
448 if (!host.empty() && host[0] == '[' && host[host.size() - 1] == ']') {
449 host = host.substr(1, host.size() - 2);
453 if (suppress_logging) {
454 // See if there's more than one minute since last time we made a connection
455 // with logging enabled. If so, turn it on again.
457 int err = clock_gettime(CLOCK_MONOTONIC_COARSE, &now);
460 double elapsed = now.tv_sec - last_verbose_connection.tv_sec +
461 1e-9 * (now.tv_nsec - last_verbose_connection.tv_nsec);
462 if (elapsed > 60.0) {
463 suppress_logging = false;
466 if (!suppress_logging) {
467 int err = clock_gettime(CLOCK_MONOTONIC_COARSE, &last_verbose_connection);
470 ++num_connection_attempts;
471 if (protocol == "pipe") {
472 sock = open_child_process(path.c_str());
475 // Construct a minimal HTTP header.
476 http_header = "HTTP/1.0 200 OK\r\n";
477 for (int stream_index : stream_indices) {
478 servers->set_header(stream_index, http_header, stream_header);
480 state = RECEIVING_DATA;
483 sock = lookup_and_connect(host, port);
485 // Yay, successful connect.
486 state = SENDING_REQUEST;
487 request = "GET " + path + " HTTP/1.0\r\nHost: " + host_header(host, port) + "\r\nUser-Agent: cubemap\r\n\r\n";
488 request_bytes_sent = 0;
492 lock_guard<mutex> lock(stats_mutex);
493 stats.connect_time = time(nullptr);
494 clock_gettime(CLOCK_MONOTONIC_COARSE, &last_activity);
498 case SENDING_REQUEST: {
499 size_t to_send = request.size() - request_bytes_sent;
503 ret = write(sock, request.data() + request_bytes_sent, to_send);
504 } while (ret == -1 && errno == EINTR);
508 state = CLOSING_SOCKET;
513 request_bytes_sent += ret;
515 if (request_bytes_sent == request.size()) {
516 state = RECEIVING_HEADER;
520 case RECEIVING_HEADER: {
525 ret = read(sock, buf, sizeof(buf));
526 } while (ret == -1 && errno == EINTR);
530 state = CLOSING_SOCKET;
535 // This really shouldn't happen...
536 if (!suppress_logging) {
537 log(ERROR, "[%s] Socket unexpectedly closed while reading header",
540 state = CLOSING_SOCKET;
544 RequestParseStatus status = wait_for_double_newline(&response, buf, ret);
546 if (status == RP_OUT_OF_SPACE) {
547 if (!suppress_logging) {
548 log(WARNING, "[%s] Server sent overlong HTTP response!", url.c_str());
550 state = CLOSING_SOCKET;
552 } else if (status == RP_NOT_FINISHED_YET) {
556 // OK, so we're fine, but there might be some of the actual data after the response.
557 // We'll need to deal with that separately.
559 if (status == RP_EXTRA_DATA) {
560 char *ptr = static_cast<char *>(
561 memmem(response.data(), response.size(), "\r\n\r\n", 4));
562 assert(ptr != nullptr);
563 extra_data = string(ptr + 4, &response[0] + response.size());
564 response.resize(ptr - response.data());
567 if (!parse_response(response)) {
568 state = CLOSING_SOCKET;
572 if (!extra_data.empty()) {
573 process_data(&extra_data[0], extra_data.size());
576 if (!suppress_logging) {
577 if (encoding == Input::INPUT_ENCODING_RAW) {
578 log(INFO, "[%s] Connected to '%s', receiving raw data.",
579 url.c_str(), url.c_str());
581 assert(encoding == Input::INPUT_ENCODING_METACUBE);
582 log(INFO, "[%s] Connected to '%s', receiving data.",
583 url.c_str(), url.c_str());
586 state = RECEIVING_DATA;
589 case RECEIVING_DATA: {
594 ret = read(sock, buf, sizeof(buf));
595 } while (ret == -1 && errno == EINTR);
599 state = CLOSING_SOCKET;
604 // This really shouldn't happen...
605 if (!suppress_logging) {
606 log(ERROR, "[%s] Socket unexpectedly closed while reading data",
609 state = CLOSING_SOCKET;
613 num_connection_attempts = 0; // Reset, since we have a successful read.
614 if (suppress_logging) {
615 // This was suppressed earlier, so print it out now.
616 if (encoding == Input::INPUT_ENCODING_RAW) {
617 log(INFO, "[%s] Connected to '%s', receiving raw data.",
618 url.c_str(), url.c_str());
620 assert(encoding == Input::INPUT_ENCODING_METACUBE);
621 log(INFO, "[%s] Connected to '%s', receiving data.",
622 url.c_str(), url.c_str());
624 suppress_logging = false;
627 process_data(buf, ret);
630 case CLOSING_SOCKET: {
632 state = NOT_CONNECTED;
639 // If we are still in NOT_CONNECTED, either something went wrong,
640 // or the connection just got closed.
641 // The earlier steps have already given the error message, if any.
642 if (state == NOT_CONNECTED && !should_stop()) {
643 if (!suppress_logging) {
644 log(INFO, "[%s] Waiting 0.2 seconds and restarting...", url.c_str());
647 if (num_connection_attempts >= 3 && !suppress_logging) {
648 log(INFO, "[%s] %d failed connection attempts, suppressing logging for one minute.",
649 url.c_str(), num_connection_attempts);
650 suppress_logging = true;
653 timeout_ts.tv_sec = 0;
654 timeout_ts.tv_nsec = 200000000;
655 wait_for_wakeup(&timeout_ts);
660 void HTTPInput::process_data(char *ptr, size_t bytes)
663 lock_guard<mutex> lock(stats_mutex);
664 stats.bytes_received += bytes;
667 if (encoding == Input::INPUT_ENCODING_RAW) {
668 for (int stream_index : stream_indices) {
669 servers->add_data(stream_index, ptr, bytes, /*metacube_flags=*/0, /*pts=*/RationalPTS());
674 assert(encoding == Input::INPUT_ENCODING_METACUBE);
675 pending_data.insert(pending_data.end(), ptr, ptr + bytes);
678 // If we don't have enough data (yet) for even the Metacube header, just return.
679 if (pending_data.size() < sizeof(metacube2_block_header)) {
683 // Make sure we have the Metacube sync header at the start.
684 // We may need to skip over junk data (it _should_ not happen, though).
685 if (!has_metacube_header) {
686 char *ptr = static_cast<char *>(
687 memmem(pending_data.data(), pending_data.size(),
688 METACUBE2_SYNC, strlen(METACUBE2_SYNC)));
689 if (ptr == nullptr) {
690 // OK, so we didn't find the sync marker. We know then that
691 // we do not have the _full_ marker in the buffer, but we
692 // could have N-1 bytes. Drop everything before that,
694 drop_pending_data(pending_data.size() - (strlen(METACUBE2_SYNC) - 1));
697 // Yay, we found the header. Drop everything (if anything) before it.
698 drop_pending_data(ptr - pending_data.data());
699 has_metacube_header = true;
701 // Re-check that we have the entire header; we could have dropped data.
702 if (pending_data.size() < sizeof(metacube2_block_header)) {
708 // Now it's safe to read the header.
709 metacube2_block_header hdr;
710 memcpy(&hdr, pending_data.data(), sizeof(hdr));
711 assert(memcmp(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync)) == 0);
712 uint32_t size = ntohl(hdr.size);
713 uint16_t flags = ntohs(hdr.flags);
714 uint16_t expected_csum = metacube2_compute_crc(&hdr);
716 if (expected_csum != ntohs(hdr.csum)) {
717 log(WARNING, "[%s] Metacube checksum failed (expected 0x%x, got 0x%x), "
718 "not reading block claiming to be %d bytes (flags=%x).",
719 url.c_str(), expected_csum, ntohs(hdr.csum),
722 // Drop only the first byte, and let the rest of the code handle resync.
723 pending_data.erase(pending_data.begin(), pending_data.begin() + 1);
724 has_metacube_header = false;
727 if (size > 10485760) {
728 log(WARNING, "[%s] Metacube block of %d bytes (flags=%x); corrupted header??",
729 url.c_str(), size, flags);
732 // See if we have the entire block. If not, wait for more data.
733 if (pending_data.size() < sizeof(metacube2_block_header) + size) {
737 // See if this is a metadata block. If so, we don't want to send it on,
738 // but rather process it ourselves.
739 // TODO: Keep metadata when sending on to other Metacube users.
740 if (flags & METACUBE_FLAGS_METADATA) {
742 lock_guard<mutex> lock(stats_mutex);
743 stats.metadata_bytes_received += size;
745 process_metacube_metadata_block(hdr, pending_data.data() + sizeof(hdr), size);
747 // Send this block on to the servers.
749 lock_guard<mutex> lock(stats_mutex);
750 stats.data_bytes_received += size;
752 char *inner_data = pending_data.data() + sizeof(metacube2_block_header);
753 if (flags & METACUBE_FLAGS_HEADER) {
754 stream_header = string(inner_data, inner_data + size);
755 for (int stream_index : stream_indices) {
756 servers->set_header(stream_index, http_header, stream_header);
759 for (int stream_index : stream_indices) {
760 servers->add_data(stream_index, inner_data, size, flags, next_block_pts);
762 next_block_pts = RationalPTS();
765 // Consume the block. This isn't the most efficient way of dealing with things
766 // should we have many blocks, but these routines don't need to be too efficient
768 pending_data.erase(pending_data.begin(), pending_data.begin() + sizeof(metacube2_block_header) + size);
769 has_metacube_header = false;
773 void HTTPInput::drop_pending_data(size_t num_bytes)
775 if (num_bytes == 0) {
778 log(WARNING, "[%s] Dropping %lld junk bytes; not a Metacube2 stream, or data was dropped from the middle of the stream.",
779 url.c_str(), (long long)num_bytes);
780 assert(pending_data.size() >= num_bytes);
781 pending_data.erase(pending_data.begin(), pending_data.begin() + num_bytes);
784 void HTTPInput::add_destination(int stream_index)
786 stream_indices.push_back(stream_index);
787 servers->set_header(stream_index, http_header, stream_header);
790 InputStats HTTPInput::get_stats() const
792 lock_guard<mutex> lock(stats_mutex);
796 void HTTPInput::process_metacube_metadata_block(const metacube2_block_header &hdr, const char *payload, uint32_t payload_size)
798 if (payload_size < sizeof(uint64_t)) {
799 log(WARNING, "[%s] Undersized Metacube metadata block (%d bytes); corrupted header?",
800 url.c_str(), payload_size);
804 uint64_t type = be64toh(*(const uint64_t *)payload);
805 if (type == METACUBE_METADATA_TYPE_ENCODER_TIMESTAMP) {
807 clock_gettime(CLOCK_REALTIME, &now);
809 const metacube2_timestamp_packet *pkt = (const metacube2_timestamp_packet *)payload;
810 if (payload_size != sizeof(*pkt)) {
811 log(WARNING, "[%s] Metacube timestamp block of wrong size (%d bytes); ignoring.",
812 url.c_str(), payload_size);
816 double elapsed = now.tv_sec - be64toh(pkt->tv_sec) +
817 1e-9 * (now.tv_nsec - long(be64toh(pkt->tv_nsec)));
819 lock_guard<mutex> lock(stats_mutex);
820 stats.latency_sec = elapsed;
822 } else if (type == METACUBE_METADATA_TYPE_NEXT_BLOCK_PTS) {
823 const metacube2_pts_packet *pkt = (const metacube2_pts_packet *)payload;
824 if (payload_size != sizeof(*pkt)) {
825 log(WARNING, "[%s] Metacube pts block of wrong size (%d bytes); ignoring.",
826 url.c_str(), payload_size);
829 next_block_pts.pts = be64toh(pkt->pts);
830 next_block_pts.timebase_num = be64toh(pkt->timebase_num);
831 next_block_pts.timebase_den = be64toh(pkt->timebase_den);
833 // Unknown metadata block, ignore
834 log(INFO, "[%s] Metadata block %llu\n", url.c_str(), type);