X-Git-Url: https://git.sesse.net/?p=cubemap;a=blobdiff_plain;f=stream.cpp;h=6a5ef99f78fa80a8294b6bf9d3bf8ee52d0bb1c0;hp=425477c0bedd424ff1a507220c0fb1fed952665b;hb=980ac162414c9fce62af4fdb9cfc282865b31572;hpb=dc109293b2644d5f1ac85a3f02b3584e82dcf690 diff --git a/stream.cpp b/stream.cpp index 425477c..6a5ef99 100644 --- a/stream.cpp +++ b/stream.cpp @@ -1,27 +1,43 @@ #include #include +#include +#include +#include #include +#include #include #include #include +#include #include +#include #include #include "log.h" -#include "metacube.h" +#include "metacube2.h" #include "state.pb.h" #include "stream.h" #include "util.h" using namespace std; -Stream::Stream(const string &url, size_t backlog_size, Encoding encoding) +Stream::Stream(const string &url, + size_t backlog_size, + uint64_t prebuffering_bytes, + Encoding encoding, + Encoding src_encoding, + unsigned hls_frag_duration, + size_t hls_backlog_margin, + const std::string &allow_origin) : url(url), encoding(encoding), + src_encoding(src_encoding), + allow_origin(allow_origin), data_fd(make_tempfile("")), - backlog_size(backlog_size), - bytes_received(0), - mark_pool(NULL) + backlog_size(backlog_size), + prebuffering_bytes(prebuffering_bytes), + hls_frag_duration(hls_frag_duration), + hls_backlog_margin(hls_backlog_margin) { if (data_fd == -1) { exit(1); @@ -43,23 +59,25 @@ Stream::Stream(const StreamProto &serialized, int data_fd) data_fd(data_fd), backlog_size(serialized.backlog_size()), bytes_received(serialized.bytes_received()), - mark_pool(NULL) + first_fragment_index(serialized.first_fragment_index()), + discontinuity_counter(serialized.discontinuity_counter()) { if (data_fd == -1) { exit(1); } - // Split old-style headers into HTTP and video headers. - if (!serialized.header().empty()) { - string header = serialized.header(); - size_t split = header.find("\r\n\r\n"); - if (split == string::npos) { - http_header = header; - stream_header = ""; - } else { - http_header = header.substr(0, split + 2); // Split off the second \r\n. - stream_header = header.substr(split, string::npos); + for (ssize_t point : serialized.suitable_starting_point()) { + if (point == -1) { + // Can happen when upgrading from before 1.1.3, + // where this was an optional field with -1 signifying + // "no such point". + continue; } + suitable_starting_points.push_back(point); + } + + for (const FragmentStartProto &fragment : serialized.fragment()) { + fragments.push_back(FragmentStart { size_t(fragment.byte_position()), fragment.pts(), fragment.begins_header() }); } } @@ -71,6 +89,18 @@ StreamProto Stream::serialize() serialized.add_data_fds(data_fd); serialized.set_backlog_size(backlog_size); serialized.set_bytes_received(bytes_received); + for (size_t point : suitable_starting_points) { + serialized.add_suitable_starting_point(point); + } + for (const FragmentStart &fragment : fragments) { + FragmentStartProto *proto = serialized.add_fragment(); + proto->set_byte_position(fragment.byte_position); + proto->set_pts(fragment.pts); + proto->set_begins_header(fragment.begins_header); + } + serialized.set_first_fragment_index(first_fragment_index); + serialized.set_discontinuity_counter(discontinuity_counter); + serialized.set_url(url); data_fd = -1; return serialized; @@ -111,13 +141,54 @@ void Stream::set_backlog_size(size_t new_size) // Now cheat a bit by rewinding, and adding all the old data back. bytes_received -= existing_data.size(); - iovec iov; - iov.iov_base = const_cast(existing_data.data()); - iov.iov_len = existing_data.size(); + DataElement data_element; + data_element.data.iov_base = const_cast(existing_data.data()); + data_element.data.iov_len = existing_data.size(); + data_element.metacube_flags = 0; // Ignored by add_data_raw(). - vector iovs; - iovs.push_back(iov); - add_data_raw(iovs); + vector data_elements; + data_elements.push_back(data_element); + add_data_raw(data_elements); + remove_obsolete_starting_points(); +} + +void Stream::set_header(const std::string &new_http_header, const std::string &new_stream_header) +{ + http_header = new_http_header; + if (new_stream_header == stream_header) { + return; + } + + // We cannot start at any of the older starting points anymore, + // since they'd get the wrong header for the stream (not to mention + // that a changed header probably means the stream restarted, + // which means any client starting on the old one would probably + // stop playing properly at the change point). Next block + // should be a suitable starting point (if not, something is + // pretty strange), so it will fill up again soon enough. + suitable_starting_points.clear(); + + // HLS, on the other hand, can deal with discontinuities and multiple + // headers. At least in theory (client support varies wildly). + if (!fragments.empty()) { + // Commit the old header to the backlog, so that we can serve it + // for all the old fragments for as long as they exist. + if (!stream_header.empty()) { + // End the current fragment and make a new one for the header. + fragments.push_back(Stream::FragmentStart { bytes_received, 0.0, true }); + process_queued_data(); + Stream::DataElement elem; + elem.data.iov_base = (char *)stream_header.data(); + elem.data.iov_len = stream_header.size(); + add_data_raw({ elem }); + remove_obsolete_starting_points(); + + // The discontinuity counter will be increased when + // this header goes out of the backlog. + } + clear_hls_playlist_cache(); + } + stream_header = new_stream_header; } void Stream::put_client_to_sleep(Client *client) @@ -126,19 +197,20 @@ void Stream::put_client_to_sleep(Client *client) } // Return a new set of iovecs that contains only the first bytes of . -vector collect_iovecs(const vector &data, size_t bytes_wanted) +vector collect_iovecs(const vector &data, size_t bytes_wanted) { vector ret; - for (size_t i = 0; i < data.size() && bytes_wanted > 0; ++i) { - if (data[i].iov_len <= bytes_wanted) { + size_t max_iovecs = min(data.size(), IOV_MAX); + for (size_t i = 0; i < max_iovecs && bytes_wanted > 0; ++i) { + if (data[i].data.iov_len <= bytes_wanted) { // Consume the entire iovec. - ret.push_back(data[i]); - bytes_wanted -= data[i].iov_len; + ret.push_back(data[i].data); + bytes_wanted -= data[i].data.iov_len; } else { // Take only parts of this iovec. iovec iov; - iov.iov_base = data[i].iov_base; - iov.iov_len = bytes_wanted; + iov.iov_base = data[i].data.iov_base; + iov.iov_len = bytes_wanted; ret.push_back(iov); bytes_wanted = 0; } @@ -147,20 +219,22 @@ vector collect_iovecs(const vector &data, size_t bytes_wanted) } // Return a new set of iovecs that contains all of except the first bytes. -vector remove_iovecs(const vector &data, size_t bytes_wanted) +vector remove_iovecs(const vector &data, size_t bytes_wanted) { - vector ret; + vector ret; size_t i; for (i = 0; i < data.size() && bytes_wanted > 0; ++i) { - if (data[i].iov_len <= bytes_wanted) { + if (data[i].data.iov_len <= bytes_wanted) { // Consume the entire iovec. - bytes_wanted -= data[i].iov_len; + bytes_wanted -= data[i].data.iov_len; } else { // Take only parts of this iovec. - iovec iov; - iov.iov_base = reinterpret_cast(data[i].iov_base) + bytes_wanted; - iov.iov_len = data[i].iov_len - bytes_wanted; - ret.push_back(iov); + Stream::DataElement data_element; + data_element.data.iov_base = reinterpret_cast(data[i].data.iov_base) + bytes_wanted; + data_element.data.iov_len = data[i].data.iov_len - bytes_wanted; + data_element.metacube_flags = METACUBE_FLAGS_NOT_SUITABLE_FOR_STREAM_START; + data_element.pts = RationalPTS(); + ret.push_back(data_element); bytes_wanted = 0; } } @@ -170,9 +244,9 @@ vector remove_iovecs(const vector &data, size_t bytes_wanted) return ret; } -void Stream::add_data_raw(const vector &orig_data) +void Stream::add_data_raw(const vector &orig_data) { - vector data = orig_data; + vector data = orig_data; while (!data.empty()) { size_t pos = bytes_received % backlog_size; @@ -185,7 +259,7 @@ void Stream::add_data_raw(const vector &orig_data) } while (ret == -1 && errno == EINTR); if (ret == -1) { - log_perror("pwrite"); + log_perror("pwritev"); // Dazed and confused, but trying to continue... return; } @@ -196,48 +270,149 @@ void Stream::add_data_raw(const vector &orig_data) } } -void Stream::add_data_deferred(const char *data, size_t bytes) +void Stream::remove_obsolete_starting_points() { + // We could do a binary search here (std::lower_bound), but it seems + // overkill for removing what's probably only a few points. + while (!suitable_starting_points.empty() && + bytes_received - suitable_starting_points[0] > backlog_size) { + suitable_starting_points.pop_front(); + } + assert(backlog_size >= hls_backlog_margin); + while (!fragments.empty() && + bytes_received - fragments[0].byte_position > (backlog_size - hls_backlog_margin)) { + if (fragments[0].begins_header) { + ++discontinuity_counter; + } else { + ++first_fragment_index; + } + fragments.pop_front(); + clear_hls_playlist_cache(); + } +} + +void Stream::add_data_deferred(const char *data, size_t bytes, uint16_t metacube_flags, const RationalPTS &pts) +{ + // For regular output, we don't want to send the client twice + // (it's already sent out together with the HTTP header). + // However, for Metacube output, we need to send it so that + // the Cubemap instance in the other end has a chance to update it. + // It may come twice in its stream, but Cubemap doesn't care. + if (encoding == Stream::STREAM_ENCODING_RAW && + (metacube_flags & METACUBE_FLAGS_HEADER) != 0) { + return; + } + + lock_guard lock(queued_data_mutex); + + DataElement data_element; + data_element.metacube_flags = metacube_flags; + data_element.pts = pts; + if (encoding == Stream::STREAM_ENCODING_METACUBE) { + // Construct a PTS metadata block. (We'll avoid sending it out + // if we don't have a valid PTS.) + metacube2_pts_packet pts_packet; + pts_packet.type = htobe64(METACUBE_METADATA_TYPE_NEXT_BLOCK_PTS); + pts_packet.pts = htobe64(pts.pts); + pts_packet.timebase_num = htobe64(pts.timebase_num); + pts_packet.timebase_den = htobe64(pts.timebase_den); + + metacube2_block_header pts_hdr; + memcpy(pts_hdr.sync, METACUBE2_SYNC, sizeof(pts_hdr.sync)); + pts_hdr.size = htonl(sizeof(pts_packet)); + pts_hdr.flags = htons(METACUBE_FLAGS_METADATA); + pts_hdr.csum = htons(metacube2_compute_crc(&pts_hdr)); + // Add a Metacube block header before the data. - metacube_block_header hdr; - memcpy(hdr.sync, METACUBE_SYNC, sizeof(hdr.sync)); + metacube2_block_header hdr; + memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync)); hdr.size = htonl(bytes); - hdr.flags = htonl(0); + hdr.flags = htons(metacube_flags); + hdr.csum = htons(metacube2_compute_crc(&hdr)); - iovec iov; - iov.iov_base = new char[bytes + sizeof(hdr)]; - iov.iov_len = bytes + sizeof(hdr); + data_element.data.iov_len = bytes + sizeof(hdr); + if (pts.timebase_num != 0) { + data_element.data.iov_len += sizeof(pts_hdr) + sizeof(pts_packet); + } + data_element.data.iov_base = new char[data_element.data.iov_len]; - memcpy(iov.iov_base, &hdr, sizeof(hdr)); - memcpy(reinterpret_cast(iov.iov_base) + sizeof(hdr), data, bytes); + char *ptr = reinterpret_cast(data_element.data.iov_base); + if (pts.timebase_num != 0) { + memcpy(ptr, &pts_hdr, sizeof(pts_hdr)); + ptr += sizeof(pts_hdr); + memcpy(ptr, &pts_packet, sizeof(pts_packet)); + ptr += sizeof(pts_packet); + } + + memcpy(ptr, &hdr, sizeof(hdr)); + ptr += sizeof(hdr); + memcpy(ptr, data, bytes); - queued_data.push_back(iov); + queued_data.push_back(data_element); } else if (encoding == Stream::STREAM_ENCODING_RAW) { // Just add the data itself. - iovec iov; - iov.iov_base = new char[bytes]; - memcpy(iov.iov_base, data, bytes); - iov.iov_len = bytes; + data_element.data.iov_base = new char[bytes]; + memcpy(data_element.data.iov_base, data, bytes); + data_element.data.iov_len = bytes; - queued_data.push_back(iov); + queued_data.push_back(data_element); } else { - assert(encoding == Stream::STREAM_ENCODING_RAW); + assert(false); } } void Stream::process_queued_data() { - if (queued_data.empty()) { - return; + vector queued_data_copy; + + // Hold the lock for as short as possible, since add_data_raw() can possibly + // write to disk, which might disturb the input thread. + { + lock_guard lock(queued_data_mutex); + if (queued_data.empty()) { + return; + } + + swap(queued_data, queued_data_copy); + } + + // Add suitable starting points for the stream, if the queued data + // contains such starting points. Note that we drop starting points + // if they're less than 10 kB apart, so that we don't get a huge + // amount of them for e.g. each and every MPEG-TS 188-byte cell. + // The 10 kB value is somewhat arbitrary, but at least it should make + // the RAM cost of saving the position ~0.1% (or less) of the actual + // data, and 10 kB is a very fine granularity in most streams. + static const int minimum_start_point_distance = 10240; + size_t byte_position = bytes_received; + bool need_hls_clear = false; + for (const DataElement &elem : queued_data_copy) { + if ((elem.metacube_flags & METACUBE_FLAGS_NOT_SUITABLE_FOR_STREAM_START) == 0) { + size_t num_points = suitable_starting_points.size(); + if (num_points >= 2 && + suitable_starting_points[num_points - 1] - suitable_starting_points[num_points - 2] < minimum_start_point_distance) { + // p[n-1] - p[n-2] < 10 kB, so drop p[n-1]. + suitable_starting_points.pop_back(); + } + suitable_starting_points.push_back(byte_position); + + if (elem.pts.timebase_num != 0) { + need_hls_clear |= add_fragment_boundary(byte_position, elem.pts); + } + } + byte_position += elem.data.iov_len; + } + if (need_hls_clear) { + clear_hls_playlist_cache(); } - add_data_raw(queued_data); - for (size_t i = 0; i < queued_data.size(); ++i) { - char *data = reinterpret_cast(queued_data[i].iov_base); + add_data_raw(queued_data_copy); + remove_obsolete_starting_points(); + for (const DataElement &elem : queued_data_copy) { + char *data = reinterpret_cast(elem.data.iov_base); delete[] data; } - queued_data.clear(); // We have more data, so wake up all clients. if (to_process.empty()) { @@ -247,3 +422,140 @@ void Stream::process_queued_data() sleeping_clients.clear(); } } + +bool Stream::add_fragment_boundary(size_t byte_position, const RationalPTS &pts) +{ + double pts_double = double(pts.pts) * pts.timebase_den / pts.timebase_num; + + if (fragments.size() <= 1 || + fragments[fragments.size() - 1].begins_header || + fragments[fragments.size() - 2].begins_header) { + // Just starting up, so try to establish the first in-progress fragment. + fragments.push_back(FragmentStart{ byte_position, pts_double, false }); + return false; + } + + // Keep extending the in-progress fragment as long as we do not + // exceed the target duration by more than half a second + // (RFC 8216 4.3.3.1) and we get closer to the target by doing so. + // Note that in particular, this means we'll always extend + // as long as we don't exceed the target duration. + double current_duration = pts_double - fragments[fragments.size() - 1].pts; + double candidate_duration = pts_double - fragments[fragments.size() - 2].pts; + if (lrintf(candidate_duration) <= hls_frag_duration && + fabs(candidate_duration - hls_frag_duration) < fabs(current_duration - hls_frag_duration)) { + fragments.back() = FragmentStart{ byte_position, pts_double, false }; + return false; + } else { + // Extending the in-progress fragment would make it too long, + // so finalize it and start a new in-progress fragment. + fragments.push_back(FragmentStart{ byte_position, pts_double, false }); + return true; + } +} + +void Stream::clear_hls_playlist_cache() +{ + hls_playlist_http10.reset(); + hls_playlist_http11_close.reset(); + hls_playlist_http11_persistent.reset(); +} + +shared_ptr Stream::generate_hls_playlist(bool http_11, bool close_after_response) +{ + char buf[256]; + snprintf(buf, sizeof(buf), + "#EXTM3U\r\n" + "#EXT-X-VERSION:7\r\n" + "#EXT-X-TARGETDURATION:%u\r\n" + "#EXT-X-MEDIA-SEQUENCE:%" PRIu64 "\r\n" + "#EXT-X-DISCONTINUITY-SEQUENCE:%" PRIu64 "\r\n", + hls_frag_duration, + first_fragment_index, + discontinuity_counter); + + string playlist = buf; + + if (fragments.size() >= 3) { + bool printed_header_for_this_group = false; + bool printed_first_header = false; + for (size_t i = 0; i < fragments.size() - 2; ++i) { + char buf[256]; + + if (fragments[i].begins_header) { + // End of this group. (We've already printed the header + // as part of the previous group.) + printed_header_for_this_group = false; + continue; + } + if (!printed_header_for_this_group) { + // Look forward until we find the header for this group (if any). + for (size_t j = i + 1; j < fragments.size() - 1; ++j) { + if (fragments[j].begins_header) { + if (printed_first_header) { + playlist += "#EXT-X-DISCONTINUITY\r\n"; + } + snprintf(buf, sizeof(buf), + "#EXT-X-MAP:URI=\"%s?frag=%" PRIu64 "-%" PRIu64 "\"\r\n", + url.c_str(), fragments[j].byte_position, + fragments[j + 1].byte_position); + playlist += buf; + printed_first_header = true; + printed_header_for_this_group = true; + break; + } + } + + if (!printed_header_for_this_group && !stream_header.empty()) { + if (printed_first_header) { + playlist += "#EXT-X-DISCONTINUITY\r\n"; + } + snprintf(buf, sizeof(buf), "#EXT-X-MAP:URI=\"%s?frag=header\"\r\n", url.c_str()); + playlist += buf; + } + + // Even if we didn't find anything, we don't want to search again for each fragment. + printed_first_header = true; + printed_header_for_this_group = true; + } + + if (fragments[i + 1].begins_header) { + // Since we only have start pts for each block and not duration, + // we have no idea how long this fragment is; the encoder restarted + // before it got to output the next pts. However, it's likely + // to be very short, so instead of trying to guess, we just skip it. + continue; + } + + snprintf(buf, sizeof(buf), "#EXTINF:%f,\r\n%s?frag=%" PRIu64 "-%" PRIu64 "\r\n", + fragments[i + 1].pts - fragments[i].pts, + url.c_str(), + fragments[i].byte_position, + fragments[i + 1].byte_position); + playlist += buf; + } + } + + string response; + if (http_11) { + response = "HTTP/1.1 200 OK\r\n"; + if (close_after_response) { + response.append("Connection: close\r\n"); + } + } else { + assert(close_after_response); + response = "HTTP/1.0 200 OK\r\n"; + } + snprintf(buf, sizeof(buf), "Content-Length: %zu\r\n", playlist.size()); + response.append(buf); + response.append("Content-Type: application/x-mpegURL\r\n"); + if (!allow_origin.empty()) { + response.append("Access-Control-Allow-Origin: "); + response.append(allow_origin); + response.append("\r\n"); + } + response.append("\r\n"); + response.append(move(playlist)); + + return shared_ptr(new string(move(response))); +}