[cubemap] / server.cpp

#include <assert.h>
#include <errno.h>
#include <inttypes.h>
#include <netinet/in.h>
#include <netinet/tcp.h>
#include <pthread.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/epoll.h>
#include <sys/sendfile.h>
#include <sys/socket.h>
#include <sys/types.h>
#include <unistd.h>
#include <algorithm>
#include <map>
#include <string>
#include <utility>
#include <vector>

#include "ktls.h"
#include "tlse.h"

#include "acceptor.h"
#include "accesslog.h"
#include "log.h"
#include "metacube2.h"
#include "parse.h"
#include "server.h"
#include "state.pb.h"
#include "stream.h"
#include "util.h"

#ifndef SO_MAX_PACING_RATE
#define SO_MAX_PACING_RATE 47
#endif

using namespace std;

extern AccessLogThread *access_log;

namespace {

inline bool is_equal(timespec a, timespec b)
{
        return a.tv_sec == b.tv_sec &&
               a.tv_nsec == b.tv_nsec;
}

inline bool is_earlier(timespec a, timespec b)
{
        if (a.tv_sec != b.tv_sec)
                return a.tv_sec < b.tv_sec;
        return a.tv_nsec < b.tv_nsec;
}

}  // namespace

Server::Server()
{
        epoll_fd = epoll_create(1024);  // Size argument is ignored.
        if (epoll_fd == -1) {
                log_perror("epoll_fd");
                exit(1);
        }
}

Server::~Server()
{
        safe_close(epoll_fd);
}

vector<ClientStats> Server::get_client_stats() const
{
        vector<ClientStats> ret;

        lock_guard<mutex> lock(mu);
        for (const auto &fd_and_client : clients) {
                ret.push_back(fd_and_client.second.get_stats());
        }
        return ret;
}

void Server::do_work()
{
        while (!should_stop()) {
                // Wait until there's activity on at least one of the fds,
                // or 20 ms (about one frame at 50 fps) has elapsed.
                //
                // We could in theory wait forever and rely on wakeup()
                // from add_client_deferred() and add_data_deferred(),
                // but wakeup is a pretty expensive operation, and the
                // two threads might end up fighting over a lock, so it's
                // seemingly (much) more efficient to just have a timeout here.
                int nfds = epoll_pwait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS, &sigset_without_usr1_block);
                if (nfds == -1 && errno != EINTR) {
                        log_perror("epoll_wait");
                        exit(1);
                }

                lock_guard<mutex> lock(mu);  // We release the mutex between iterations.
        
                process_queued_data();

                // Process each client where we have socket activity.
                for (int i = 0; i < nfds; ++i) {
                        Client *client = reinterpret_cast<Client *>(events[i].data.ptr);

                        if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
                                close_client(client);
                                continue;
                        }

                        process_client(client);
                }

                // Process each client where its stream has new data,
                // even if there was no socket activity.
                for (unique_ptr<Stream> &stream : streams) {
                        vector<Client *> to_process;
                        swap(stream->to_process, to_process);
                        for (Client *client : to_process) {
                                process_client(client);
                        }
                }

                // Finally, go through each client to see if it's timed out
                // in the READING_REQUEST state. (Seemingly there are clients
                // that can hold sockets up for days at a time without sending
                // anything at all.)
                timespec timeout_time;
                if (clock_gettime(CLOCK_MONOTONIC_COARSE, &timeout_time) == -1) {
                        log_perror("clock_gettime(CLOCK_MONOTONIC_COARSE)");
                        continue;
                }
                timeout_time.tv_sec -= REQUEST_READ_TIMEOUT_SEC;
                while (!clients_ordered_by_connect_time.empty()) {
                        const pair<timespec, int> &connect_time_and_fd = clients_ordered_by_connect_time.front();

                        // See if we have reached the end of clients to process.
                        if (is_earlier(timeout_time, connect_time_and_fd.first)) {
                                break;
                        }

                        // If this client doesn't exist anymore, just ignore it
                        // (it was deleted earlier).
                        auto client_it = clients.find(connect_time_and_fd.second);
                        if (client_it == clients.end()) {
                                clients_ordered_by_connect_time.pop();
                                continue;
                        }
                        Client *client = &client_it->second;
                        if (!is_equal(client->connect_time, connect_time_and_fd.first)) {
                                // Another client has taken this fd in the meantime.
                                clients_ordered_by_connect_time.pop();
                                continue;
                        }

                        if (client->state != Client::READING_REQUEST) {
                                // Only READING_REQUEST can time out.
                                clients_ordered_by_connect_time.pop();
                                continue;
                        }

                        // OK, it timed out.
                        close_client(client);
                        clients_ordered_by_connect_time.pop();
                }
        }
}

CubemapStateProto Server::serialize(unordered_map<const string *, size_t> *short_response_pool)
{
        // We don't serialize anything queued, so empty the queues.
        process_queued_data();

        // Set all clients in a consistent state before serializing
        // (ie., they have no remaining lost data). Otherwise, increasing
        // the backlog could take clients into a newly valid area of the backlog,
        // sending a stream of zeros instead of skipping the data as it should.
        //
        // TODO: Do this when clients are added back from serialized state instead;
        // it would probably be less wasteful.
        for (auto &fd_and_client : clients) {
                skip_lost_data(&fd_and_client.second);
        }

        CubemapStateProto serialized;
        for (const auto &fd_and_client : clients) {
                serialized.add_clients()->MergeFrom(fd_and_client.second.serialize(short_response_pool));
        }
        for (unique_ptr<Stream> &stream : streams) {
                serialized.add_streams()->MergeFrom(stream->serialize());
        }
        return serialized;
}

void Server::add_client_deferred(int sock, Acceptor *acceptor)
{
        lock_guard<mutex> lock(queued_clients_mutex);
        queued_add_clients.push_back(std::make_pair(sock, acceptor));
}

void Server::add_client(int sock, Acceptor *acceptor)
{
        const bool is_tls = acceptor->is_tls();
        auto inserted = clients.insert(make_pair(sock, Client(sock)));
        assert(inserted.second == true);  // Should not already exist.
        Client *client_ptr = &inserted.first->second;

        // Connection timestamps must be nondecreasing. I can't find any guarantee
        // that even the monotonic clock can't go backwards by a small amount
        // (think switching between CPUs with non-synchronized TSCs), so if
        // this actually should happen, we hack around it by fudging
        // connect_time.
        if (!clients_ordered_by_connect_time.empty() &&
            is_earlier(client_ptr->connect_time, clients_ordered_by_connect_time.back().first)) {
                client_ptr->connect_time = clients_ordered_by_connect_time.back().first;
        }
        clients_ordered_by_connect_time.push(make_pair(client_ptr->connect_time, sock));

        // Start listening on data from this socket.
        epoll_event ev;
        if (is_tls) {
                // Even in the initial state (READING_REQUEST), TLS needs to
                // send data for the handshake, and thus might end up needing
                // to know about EPOLLOUT.
                ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
        } else {
                // EPOLLOUT will be added once we go out of READING_REQUEST.
                ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
        }
        ev.data.ptr = client_ptr;
        if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
                log_perror("epoll_ctl(EPOLL_CTL_ADD)");
                exit(1);
        }

        if (is_tls) {
                assert(tls_server_contexts.count(acceptor));
                client_ptr->tls_context = tls_accept(tls_server_contexts[acceptor]);
                if (client_ptr->tls_context == nullptr) {
                        log(ERROR, "tls_accept() failed");
                        close_client(client_ptr);
                        return;
                }
                tls_make_exportable(client_ptr->tls_context, 1);
        }

        process_client(client_ptr);
}

void Server::add_client_from_serialized(const ClientProto &client, const vector<shared_ptr<const string>> &short_responses)
{
        lock_guard<mutex> lock(mu);
        Stream *stream;
        int stream_index = lookup_stream_by_url(client.url());
        if (stream_index == -1) {
                assert(client.state() != Client::SENDING_DATA);
                stream = nullptr;
        } else {
                stream = streams[stream_index].get();
        }
        auto inserted = clients.insert(make_pair(client.sock(), Client(client, short_responses, stream)));
        assert(inserted.second == true);  // Should not already exist.
        Client *client_ptr = &inserted.first->second;

        // Connection timestamps must be nondecreasing.
        assert(clients_ordered_by_connect_time.empty() ||
               !is_earlier(client_ptr->connect_time, clients_ordered_by_connect_time.back().first));
        clients_ordered_by_connect_time.push(make_pair(client_ptr->connect_time, client.sock()));

        // Start listening on data from this socket.
        epoll_event ev;
        if (client.state() == Client::READING_REQUEST) {
                // See the corresponding comment in Server::add_client().
                if (client.has_tls_context()) {
                        ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
                } else {
                        ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
                }
        } else {
                // If we don't have more data for this client, we'll be putting it into
                // the sleeping array again soon.
                ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
        }
        ev.data.ptr = client_ptr;
        if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
                log_perror("epoll_ctl(EPOLL_CTL_ADD)");
                exit(1);
        }

        if (client_ptr->state == Client::WAITING_FOR_KEYFRAME ||
            client_ptr->state == Client::PREBUFFERING ||
            (client_ptr->state == Client::SENDING_DATA &&
             client_ptr->stream_pos == client_ptr->stream->bytes_received)) {
                client_ptr->stream->put_client_to_sleep(client_ptr);
        } else {
                process_client(client_ptr);
        }
}

int Server::lookup_stream_by_url(const string &url) const
{
        const auto stream_url_it = stream_url_map.find(url);
        if (stream_url_it == stream_url_map.end()) {
                return -1;
        }
        return stream_url_it->second;
}

int Server::add_stream(const string &url,
                       const string &hls_url,
                       size_t backlog_size,
                       size_t prebuffering_bytes,
                       Stream::Encoding encoding,
                       Stream::Encoding src_encoding,
                       unsigned hls_frag_duration,
                       size_t hls_backlog_margin,
                       const string &allow_origin)
{
        lock_guard<mutex> lock(mu);
        stream_url_map.insert(make_pair(url, streams.size()));
        if (!hls_url.empty()) {
                stream_hls_url_map.insert(make_pair(hls_url, streams.size()));
        }
        streams.emplace_back(new Stream(url, backlog_size, prebuffering_bytes, encoding, src_encoding, hls_frag_duration, hls_backlog_margin, allow_origin));
        return streams.size() - 1;
}

int Server::add_stream_from_serialized(const StreamProto &stream, int data_fd)
{
        lock_guard<mutex> lock(mu);
        stream_url_map.insert(make_pair(stream.url(), streams.size()));
        // stream_hls_url_map will be updated in register_hls_url(), since it is not part
        // of the serialized state (it will always be picked out from the configuration).
        streams.emplace_back(new Stream(stream, data_fd));
        return streams.size() - 1;
}
        
void Server::set_backlog_size(int stream_index, size_t new_size)
{
        lock_guard<mutex> lock(mu);
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        streams[stream_index]->set_backlog_size(new_size);
}

void Server::set_prebuffering_bytes(int stream_index, size_t new_amount)
{
        lock_guard<mutex> lock(mu);
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        streams[stream_index]->prebuffering_bytes = new_amount;
}
        
void Server::set_encoding(int stream_index, Stream::Encoding encoding)
{
        lock_guard<mutex> lock(mu);
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        streams[stream_index]->encoding = encoding;
}

void Server::set_src_encoding(int stream_index, Stream::Encoding encoding)
{
        lock_guard<mutex> lock(mu);
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        streams[stream_index]->src_encoding = encoding;
}

void Server::set_hls_frag_duration(int stream_index, unsigned hls_frag_duration)
{
        lock_guard<mutex> lock(mu);
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        streams[stream_index]->hls_frag_duration = hls_frag_duration;
}

void Server::set_hls_backlog_margin(int stream_index, size_t hls_backlog_margin)
{
        lock_guard<mutex> lock(mu);
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        assert(hls_backlog_margin >= 0);
        assert(hls_backlog_margin < streams[stream_index]->backlog_size);
        streams[stream_index]->hls_backlog_margin = hls_backlog_margin;
}

void Server::set_allow_origin(int stream_index, const string &allow_origin)
{
        lock_guard<mutex> lock(mu);
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        streams[stream_index]->allow_origin = allow_origin;
}

void Server::register_hls_url(int stream_index, const string &hls_url)
{
        lock_guard<mutex> lock(mu);
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        assert(!hls_url.empty());
        stream_hls_url_map.insert(make_pair(hls_url, stream_index));
}
        
void Server::set_header(int stream_index, const string &http_header, const string &stream_header)
{
        lock_guard<mutex> lock(mu);
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        Stream *stream = streams[stream_index].get();
        stream->http_header = http_header;

        if (stream_header != stream->stream_header) {
                // 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.
                stream->suitable_starting_points.clear();

                if (!stream->fragments.empty()) {
                        stream->fragments.clear();
                        ++stream->discontinuity_counter;
                        stream->clear_hls_playlist_cache();
                }
        }
        stream->stream_header = stream_header;
}
        
void Server::set_pacing_rate(int stream_index, uint32_t pacing_rate)
{
        lock_guard<mutex> lock(mu);
        assert(clients.empty());
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        streams[stream_index]->pacing_rate = pacing_rate;
}

void Server::add_gen204(const std::string &url, const std::string &allow_origin)
{
        lock_guard<mutex> lock(mu);
        assert(clients.empty());
        ping_url_map[url] = allow_origin;
}

void Server::create_tls_context_for_acceptor(const Acceptor *acceptor)
{
        assert(acceptor->is_tls());

        bool is_server = true;
        TLSContext *server_context = tls_create_context(is_server, TLS_V12);

        const string &cert = acceptor->get_certificate_chain();
        int num_cert = tls_load_certificates(server_context, reinterpret_cast<const unsigned char *>(cert.data()), cert.size());
        assert(num_cert > 0);  // Should have been checked by config earlier.

        const string &key = acceptor->get_private_key();
        int num_key = tls_load_private_key(server_context, reinterpret_cast<const unsigned char *>(key.data()), key.size());
        assert(num_key > 0);  // Should have been checked by config earlier.

        tls_server_contexts.insert(make_pair(acceptor, server_context));
}

void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, uint16_t metacube_flags, const RationalPTS &pts)
{
        assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
        streams[stream_index]->add_data_deferred(data, bytes, metacube_flags, pts);
}

// See the .h file for postconditions after this function.      
void Server::process_client(Client *client)
{
        switch (client->state) {
        case Client::READING_REQUEST: {
                if (client->tls_context != nullptr) {
                        if (send_pending_tls_data(client)) {
                                // send_pending_tls_data() hit postconditions #1 or #4.
                                return;
                        }
                }

read_request_again:
                // Try to read more of the request.
                char buf[1024];
                int ret;
                if (client->tls_context == nullptr) {
                        ret = read_nontls_data(client, buf, sizeof(buf));
                        if (ret == -1) {
                                // read_nontls_data() hit postconditions #1 or #2.
                                return;
                        }
                } else {
                        ret = read_tls_data(client, buf, sizeof(buf));
                        if (ret == -1) {
                                // read_tls_data() hit postconditions #1, #2 or #4.
                                return;
                        }
                }

                RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
        
                switch (status) {
                case RP_OUT_OF_SPACE:
                        log(WARNING, "[%s] Client sent overlong request!", client->remote_addr.c_str());
                        close_client(client);
                        return;
                case RP_NOT_FINISHED_YET:
                        // OK, we don't have the entire header yet. Fine; we'll get it later.
                        // See if there's more data for us.
                        goto read_request_again;
                case RP_EXTRA_DATA:
                        log(WARNING, "[%s] Junk data after request!", client->remote_addr.c_str());
                        close_client(client);
                        return;
                case RP_FINISHED:
                        break;
                }

                assert(status == RP_FINISHED);

                if (client->tls_context && !client->in_ktls_mode && tls_established(client->tls_context)) {
                        // We're ready to enter kTLS mode, unless we still have some
                        // handshake data to send (which then must be sent as non-kTLS).
                        if (send_pending_tls_data(client)) {
                                // send_pending_tls_data() hit postconditions #1 or #4.
                                return;
                        }
                        ret = tls_make_ktls(client->tls_context, client->sock);
                        if (ret < 0) {
                                log_tls_error("tls_make_ktls", ret);
                                close_client(client);
                                return;
                        }
                        client->in_ktls_mode = true;
                }

                int error_code = parse_request(client);
                if (error_code == 200) {
                        if (client->serving_hls_playlist) {
                                construct_hls_playlist(client);
                        } else {
                                construct_stream_header(client);
                        }
                } else if (error_code == 204) {
                        construct_204(client);
                } else {
                        construct_error(client, error_code);
                }

                // We've changed states, so fall through.
                assert(client->state == Client::SENDING_SHORT_RESPONSE ||
                       client->state == Client::SENDING_HEADER);
        }
        case Client::SENDING_SHORT_RESPONSE:
        case Client::SENDING_HEADER: {
sending_header_or_short_response_again:
                int ret;
                do {
                        ret = write(client->sock,
                                    client->header_or_short_response->data() + client->header_or_short_response_bytes_sent,
                                    client->header_or_short_response->size() - client->header_or_short_response_bytes_sent);
                } while (ret == -1 && errno == EINTR);

                if (ret == -1 && errno == EAGAIN) {
                        // We're out of socket space, so now we're at the “low edge” of epoll's
                        // edge triggering. epoll will tell us when there is more room, so for now,
                        // just return.
                        // This is postcondition #4.
                        return;
                }

                if (ret == -1) {
                        // Error! Postcondition #1.
                        log_perror("write");
                        close_client(client);
                        return;
                }
                
                client->header_or_short_response_bytes_sent += ret;
                assert(client->header_or_short_response_bytes_sent <= client->header_or_short_response->size());

                if (client->header_or_short_response_bytes_sent < client->header_or_short_response->size()) {
                        // We haven't sent all yet. Fine; go another round.
                        goto sending_header_or_short_response_again;
                }

                // We're done sending the header or error! Clear it to release some memory.
                client->header_or_short_response = nullptr;
                client->header_or_short_response_holder.clear();
                client->header_or_short_response_ref.reset();

                if (client->state == Client::SENDING_SHORT_RESPONSE) {
                        if (more_requests(client)) {
                                // We're done sending the error, but should keep on reading new requests.
                                goto read_request_again;
                        } else {
                                // We're done sending the error, so now close.
                                // This is postcondition #1.
                                close_client(client);
                        }
                        return;
                }

                Stream *stream = client->stream;
                if (client->stream_pos == Client::STREAM_POS_AT_START) {
                        // Start sending from the beginning of the backlog.
                        client->stream_pos = min<size_t>(
                            stream->bytes_received - stream->backlog_size,
                            0);
                        client->state = Client::SENDING_DATA;
                        goto sending_data;
                } else if (client->stream_pos_end != Client::STREAM_POS_NO_END) {
                        // We're sending a fragment, and should have all of it,
                        // so start sending right away.
                        assert(client->stream_pos >= 0);
                        client->state = Client::SENDING_DATA;
                        goto sending_data;
                } else if (stream->prebuffering_bytes == 0) {
                        // Start sending from the first keyframe we get. In other
                        // words, we won't send any of the backlog, but we'll start
                        // sending immediately as we get the next keyframe block.
                        // Note that this is functionally identical to the next if branch,
                        // except that we save a binary search.
                        assert(client->stream_pos == Client::STREAM_POS_AT_END);
                        assert(client->stream_pos_end == Client::STREAM_POS_NO_END);
                        client->stream_pos = stream->bytes_received;
                        client->state = Client::WAITING_FOR_KEYFRAME;
                } else {
                        // We're not going to send anything to the client before we have
                        // N bytes. However, this wait might be boring; we can just as well
                        // use it to send older data if we have it. We use lower_bound()
                        // so that we are conservative and never add extra latency over just
                        // waiting (assuming CBR or nearly so); otherwise, we could want e.g.
                        // 100 kB prebuffer but end up sending a 10 MB GOP.
                        assert(client->stream_pos == Client::STREAM_POS_AT_END);
                        assert(client->stream_pos_end == Client::STREAM_POS_NO_END);
                        deque<size_t>::const_iterator starting_point_it =
                                lower_bound(stream->suitable_starting_points.begin(),
                                            stream->suitable_starting_points.end(),
                                            stream->bytes_received - stream->prebuffering_bytes);
                        if (starting_point_it == stream->suitable_starting_points.end()) {
                                // None found. Just put us at the end, and then wait for the
                                // first keyframe to appear.
                                client->stream_pos = stream->bytes_received;
                                client->state = Client::WAITING_FOR_KEYFRAME;
                        } else {
                                client->stream_pos = *starting_point_it;
                                client->state = Client::PREBUFFERING;
                                goto prebuffering;
                        }
                }
                // Fall through.
        }
        case Client::WAITING_FOR_KEYFRAME: {
                Stream *stream = client->stream;
                if (stream->suitable_starting_points.empty() ||
                    client->stream_pos > stream->suitable_starting_points.back()) {
                        // We haven't received a keyframe since this stream started waiting,
                        // so keep on waiting for one.
                        // This is postcondition #3.
                        stream->put_client_to_sleep(client);
                        return;
                }
                client->stream_pos = stream->suitable_starting_points.back();
                client->state = Client::PREBUFFERING;
                // Fall through.
        }
        case Client::PREBUFFERING: {
prebuffering:
                Stream *stream = client->stream;
                size_t bytes_to_send = stream->bytes_received - client->stream_pos;
                assert(bytes_to_send <= stream->backlog_size);
                if (bytes_to_send < stream->prebuffering_bytes) {
                        // We don't have enough bytes buffered to start this client yet.
                        // This is postcondition #3.
                        stream->put_client_to_sleep(client);
                        return;
                }
                client->state = Client::SENDING_DATA;
                // Fall through.
        }
        case Client::SENDING_DATA: {
sending_data:
                skip_lost_data(client);
                Stream *stream = client->stream;

sending_data_again:
                size_t bytes_to_send;
                if (client->stream_pos_end == Client::STREAM_POS_NO_END) {
                         bytes_to_send = stream->bytes_received - client->stream_pos;
                } else {
                         bytes_to_send = client->stream_pos_end - client->stream_pos;
                }
                assert(bytes_to_send <= stream->backlog_size);
                if (bytes_to_send == 0) {
                        if (client->stream_pos == client->stream_pos_end) {  // We have a definite end, and we're at it.
                                if (more_requests(client)) {
                                        // We're done sending the fragment, but should keep on reading new requests.
                                        goto read_request_again;
                                } else {
                                        // We're done sending the fragment, so now close.
                                        // This is postcondition #1.
                                        close_client(client);
                                }
                        }
                        return;
                }

                // See if we need to split across the circular buffer.
                bool more_data = false;
                if ((client->stream_pos % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
                        bytes_to_send = stream->backlog_size - (client->stream_pos % stream->backlog_size);
                        more_data = true;
                }

                ssize_t ret;
                do {
                        off_t offset = client->stream_pos % stream->backlog_size;
                        ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
                } while (ret == -1 && errno == EINTR);

                if (ret == -1 && errno == EAGAIN) {
                        // We're out of socket space, so return; epoll will wake us up
                        // when there is more room.
                        // This is postcondition #4.
                        return;
                }
                if (ret == -1) {
                        // Error, close; postcondition #1.
                        log_perror("sendfile");
                        close_client(client);
                        return;
                }
                client->stream_pos += ret;
                client->bytes_sent += ret;

                assert(client->stream_pos_end == Client::STREAM_POS_NO_END || client->stream_pos <= client->stream_pos_end);
                if (client->stream_pos == client->stream_pos_end) {
                        goto sending_data_again;  // Will see that bytes_to_send == 0 and end.
                } else if (client->stream_pos == stream->bytes_received) {
                        // We don't have any more data for this client, so put it to sleep.
                        // This is postcondition #3.
                        stream->put_client_to_sleep(client);
                } else if (more_data && size_t(ret) == bytes_to_send) {
                        goto sending_data_again;
                }
                // We'll also get here for postcondition #4 (similar to the EAGAIN path above).
                break;
        }
        default:
                assert(false);
        }
}

bool Server::send_pending_tls_data(Client *client)
{
        // See if there's data from the TLS library to write.
        if (client->tls_data_to_send == nullptr) {
                client->tls_data_to_send = tls_get_write_buffer(client->tls_context, &client->tls_data_left_to_send);
                if (client->tls_data_to_send == nullptr) {
                        // Really no data to send.
                        return false;
                }
        }

send_data_again:
        int ret;
        do {
                ret = write(client->sock, client->tls_data_to_send, client->tls_data_left_to_send);
        } while (ret == -1 && errno == EINTR);
        assert(ret < 0 || size_t(ret) <= client->tls_data_left_to_send);

        if (ret == -1 && errno == EAGAIN) {
                // We're out of socket space, so now we're at the “low edge” of epoll's
                // edge triggering. epoll will tell us when there is more room, so for now,
                // just return.
                // This is postcondition #4.
                return true;
        }
        if (ret == -1) {
                // Error! Postcondition #1.
                log_perror("write");
                close_client(client);
                return true;
        }
        if (ret > 0 && size_t(ret) == client->tls_data_left_to_send) {
                // All data has been sent, so we don't need to go to sleep.
                tls_buffer_clear(client->tls_context);
                client->tls_data_to_send = nullptr;
                return false;
        }

        // More data to send, so try again.
        client->tls_data_to_send += ret;
        client->tls_data_left_to_send -= ret;
        goto send_data_again;
}

int Server::read_nontls_data(Client *client, char *buf, size_t max_size)
{
        int ret;
        do {
                ret = read(client->sock, buf, max_size);
        } while (ret == -1 && errno == EINTR);

        if (ret == -1 && errno == EAGAIN) {
                // No more data right now. Nothing to do.
                // This is postcondition #2.
                return -1;
        }
        if (ret == -1) {
                log_perror("read");
                close_client(client);
                return -1;
        }
        if (ret == 0) {
                // OK, the socket is closed.
                close_client(client);
                return -1;
        }

        return ret;
}

int Server::read_tls_data(Client *client, char *buf, size_t max_size)
{
read_again:
        int ret;
        do {
                ret = read(client->sock, buf, max_size);
        } while (ret == -1 && errno == EINTR);

        if (ret == -1 && errno == EAGAIN) {
                // No more data right now. Nothing to do.
                // This is postcondition #2.
                return -1;
        }
        if (ret == -1) {
                log_perror("read");
                close_client(client);
                return -1;
        }
        if (ret == 0) {
                // OK, the socket is closed.
                close_client(client);
                return -1;
        }

        // Give it to the TLS library.
        int err = tls_consume_stream(client->tls_context, reinterpret_cast<const unsigned char *>(buf), ret, nullptr);
        if (err < 0) {
                log_tls_error("tls_consume_stream", err);
                close_client(client);
                return -1;
        }
        if (err == 0) {
                // Not consumed any data. See if we can read more.
                goto read_again;
        }

        // Read any decrypted data available for us. (We can reuse buf, since it's free now.)
        ret = tls_read(client->tls_context, reinterpret_cast<unsigned char *>(buf), max_size);
        if (ret == 0) {
                // No decrypted data for us yet, but there might be some more handshaking
                // to send. Do that if needed, then look for more data.
                if (send_pending_tls_data(client)) {
                        // send_pending_tls_data() hit postconditions #1 or #4.
                        return -1;
                }
                goto read_again;
        }
        if (ret < 0) {
                log_tls_error("tls_read", ret);
                close_client(client);
                return -1;
        }

        assert(ret > 0);
        return ret;
}

// See if there's some data we've lost. Ideally, we should drop to a block boundary,
// but resync will be the mux's problem.
void Server::skip_lost_data(Client *client)
{
        Stream *stream = client->stream;
        if (stream == nullptr) {
                return;
        }
        size_t bytes_to_send = stream->bytes_received - client->stream_pos;
        if (bytes_to_send > stream->backlog_size) {
                size_t bytes_lost = bytes_to_send - stream->backlog_size;
                client->bytes_lost += bytes_lost;
                ++client->num_loss_events;
                if (!client->close_after_response) {
                        assert(client->stream_pos_end != Client::STREAM_POS_NO_END);

                        // We've already sent a Content-length, so we can't just skip data.
                        // Close the connection immediately and hope the other side
                        // is able to figure out that there was an error and it needs to skip.
                        client->close_after_response = true;
                        client->stream_pos = client->stream_pos_end;
                } else {
                        client->stream_pos = stream->bytes_received - stream->backlog_size;
                }
        }
}

int Server::parse_request(Client *client)
{
        vector<string> lines = split_lines(client->request);
        client->request.clear();
        if (lines.empty()) {
                return 400;  // Bad request (empty).
        }

        // Parse the headers, for logging purposes.
        // TODO: Case-insensitivity.
        unordered_multimap<string, string> headers = extract_headers(lines, client->remote_addr);
        const auto referer_it = headers.find("Referer");
        if (referer_it != headers.end()) {
                client->referer = referer_it->second;
        }
        const auto user_agent_it = headers.find("User-Agent");
        if (user_agent_it != headers.end()) {
                client->user_agent = user_agent_it->second;
        }

        vector<string> request_tokens = split_tokens(lines[0]);
        if (request_tokens.size() < 3) {
                return 400;  // Bad request (empty).
        }
        if (request_tokens[0] != "GET") {
                return 400;  // Should maybe be 405 instead?
        }

        string url = request_tokens[1];
        client->url = url;
        if (url.size() > 8 && url.find("?backlog") == url.size() - 8) {
                client->stream_pos = Client::STREAM_POS_AT_START;
                url = url.substr(0, url.size() - 8);
        } else {
                size_t pos = url.find("?frag=");
                if (pos != string::npos) {
                        // Parse an endpoint of the type /stream.mp4?frag=1234-5678.
                        const char *ptr = url.c_str() + pos + 6;

                        // "?frag=header" is special.
                        if (strcmp(ptr, "header") == 0) {
                                client->stream_pos = Client::STREAM_POS_HEADER_ONLY;
                                client->stream_pos_end = -1;
                        } else {
                                char *endptr;
                                long long frag_start = strtol(ptr, &endptr, 10);
                                if (ptr == endptr || frag_start < 0 || frag_start == LLONG_MAX) {
                                        return 400;  // Bad request.
                                }
                                if (*endptr != '-') {
                                        return 400;  // Bad request.
                                }
                                ptr = endptr + 1;

                                long long frag_end = strtol(ptr, &endptr, 10);
                                if (ptr == endptr || frag_end < frag_start || frag_end == LLONG_MAX) {
                                        return 400;  // Bad request.
                                }

                                if (*endptr != '\0') {
                                        return 400;  // Bad request.
                                }

                                client->stream_pos = frag_start;
                                client->stream_pos_end = frag_end;
                        }
                        url = url.substr(0, pos);
                } else {
                        client->stream_pos = -1;
                        client->stream_pos_end = -1;
                }
        }

        // Figure out if we're supposed to close the socket after we've delivered the response.
        string protocol = request_tokens[2];
        if (protocol.find("HTTP/") != 0) {
                return 400;  // Bad request.
        }
        client->close_after_response = false;
        client->http_11 = true;
        if (protocol == "HTTP/1.0") {
                // No persistent connections.
                client->close_after_response = true;
                client->http_11 = false;
        } else {
                const auto connection_it = headers.find("Connection");
                if (connection_it != headers.end() && connection_it->second == "close") {
                        client->close_after_response = true;
                }
        }

        const auto stream_url_map_it = stream_url_map.find(url);
        if (stream_url_map_it != stream_url_map.end()) {
                // Serve a regular stream..
                client->stream = streams[stream_url_map_it->second].get();
                client->serving_hls_playlist = false;
        } else {
                const auto stream_hls_url_map_it = stream_hls_url_map.find(url);
                if (stream_hls_url_map_it != stream_hls_url_map.end()) {
                        // Serve HLS playlist.
                        client->stream = streams[stream_hls_url_map_it->second].get();
                        client->serving_hls_playlist = true;
                } else {
                        const auto ping_url_map_it = ping_url_map.find(url);
                        if (ping_url_map_it == ping_url_map.end()) {
                                return 404;  // Not found.
                        } else {
                                // Serve a ping (204 no error).
                                return 204;
                        }
                }
        }

        Stream *stream = client->stream;
        if (stream->http_header.empty()) {
                return 503;  // Service unavailable.
        }

        if (client->serving_hls_playlist) {
                if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
                        // This doesn't make any sense, and is hard to implement, too.
                        return 404;
                } else {
                        return 200;
                }
        }

        if (client->stream_pos_end == Client::STREAM_POS_NO_END) {
                // This stream won't end, so we don't have a content-length,
                // and can just as well tell the client it's Connection: close
                // (otherwise, we'd have to implement chunking TE for no good reason).
                client->close_after_response = true;
        } else {
                if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
                        // This doesn't make any sense, and is hard to implement, too.
                        return 416;  // Range not satisfiable.
                }

                // Check that we have the requested fragment in our backlog.
                size_t buffer_end = stream->bytes_received;
                size_t buffer_start = (buffer_end <= stream->backlog_size) ? 0 : buffer_end - stream->backlog_size;

                if (client->stream_pos_end > buffer_end ||
                    client->stream_pos < buffer_start) {
                        return 416;  // Range not satisfiable.
                }
        }

        client->stream = stream;
        if (setsockopt(client->sock, SOL_SOCKET, SO_MAX_PACING_RATE, &client->stream->pacing_rate, sizeof(client->stream->pacing_rate)) == -1) {
                if (client->stream->pacing_rate != ~0U) {
                        log_perror("setsockopt(SO_MAX_PACING_RATE)");
                }
        }
        client->request.clear();

        return 200;  // OK!
}

void Server::construct_stream_header(Client *client)
{
        Stream *stream = client->stream;
        string response = stream->http_header;
        if (client->stream_pos == Client::STREAM_POS_HEADER_ONLY) {
                char buf[64];
                snprintf(buf, sizeof(buf), "Content-length: %zu\r\n", stream->stream_header.size());
                response.append(buf);
        } else if (client->stream_pos_end != Client::STREAM_POS_NO_END) {
                char buf[64];
                snprintf(buf, sizeof(buf), "Content-length: %" PRIu64 "\r\n", client->stream_pos_end - client->stream_pos);
                response.append(buf);
        }
        if (client->http_11) {
                assert(response.find("HTTP/1.0") == 0);
                response[7] = '1';  // Change to HTTP/1.1.
                if (client->close_after_response) {
                        response.append("Connection: close\r\n");
                }
        } else {
                assert(client->close_after_response);
        }
        if (!stream->allow_origin.empty()) {
                response.append("Access-Control-Allow-Origin: ");
                response.append(stream->allow_origin);
                response.append("\r\n");
        }
        if (stream->encoding == Stream::STREAM_ENCODING_RAW) {
                response.append("\r\n");
        } else if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
                response.append("Content-encoding: metacube\r\n\r\n");
                if (!stream->stream_header.empty()) {
                        metacube2_block_header hdr;
                        memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
                        hdr.size = htonl(stream->stream_header.size());
                        hdr.flags = htons(METACUBE_FLAGS_HEADER);
                        hdr.csum = htons(metacube2_compute_crc(&hdr));
                        response.append(string(reinterpret_cast<char *>(&hdr), sizeof(hdr)));
                }
        } else {
                assert(false);
        }
        if (client->stream_pos == Client::STREAM_POS_HEADER_ONLY) {
                client->state = Client::SENDING_SHORT_RESPONSE;
                response.append(stream->stream_header);
        } else {
                client->state = Client::SENDING_HEADER;
                if (client->stream_pos_end == Client::STREAM_POS_NO_END) {  // Fragments don't contain stream headers.
                        response.append(stream->stream_header);
                }
        }

        client->header_or_short_response_holder = move(response);
        client->header_or_short_response = &client->header_or_short_response_holder;

        // Switch states.
        change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
}
        
void Server::construct_error(Client *client, int error_code)
{
        char error[256];
        if (client->http_11 && client->close_after_response) {
                snprintf(error, sizeof(error),
                        "HTTP/1.1 %d Error\r\nContent-type: text/plain\r\nConnection: close\r\n\r\nSomething went wrong. Sorry.\r\n",
                        error_code);
        } else {
                snprintf(error, sizeof(error),
                        "HTTP/1.%d %d Error\r\nContent-type: text/plain\r\nContent-length: 30\r\n\r\nSomething went wrong. Sorry.\r\n",
                        client->http_11, error_code);
        }
        client->header_or_short_response_holder = error;
        client->header_or_short_response = &client->header_or_short_response_holder;

        // Switch states.
        client->state = Client::SENDING_SHORT_RESPONSE;
        change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
}

void Server::construct_hls_playlist(Client *client)
{
        Stream *stream = client->stream;
        shared_ptr<const string> *cache;
        if (client->http_11) {
                if (client->close_after_response) {
                        cache = &stream->hls_playlist_http11_close;
                } else {
                        cache = &stream->hls_playlist_http11_persistent;
                }
        } else {
                assert(client->close_after_response);
                cache = &stream->hls_playlist_http10;
        }

        if (*cache == nullptr) {
                *cache = stream->generate_hls_playlist(client->http_11, client->close_after_response);
        }
        client->header_or_short_response_ref = *cache;
        client->header_or_short_response = cache->get();

        // Switch states.
        client->state = Client::SENDING_SHORT_RESPONSE;
        change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
}

void Server::construct_204(Client *client)
{
        const auto ping_url_map_it = ping_url_map.find(client->url);
        assert(ping_url_map_it != ping_url_map.end());

        string response;
        if (client->http_11) {
                response = "HTTP/1.1 204 No Content\r\n";
                if (client->close_after_response) {
                        response.append("Connection: close\r\n");
                }
        } else {
                response = "HTTP/1.0 204 No Content\r\n";
                assert(client->close_after_response);
        }
        if (!ping_url_map_it->second.empty()) {
                response.append("Access-Control-Allow-Origin: ");
                response.append(ping_url_map_it->second);
                response.append("\r\n");
        }
        response.append("\r\n");

        client->header_or_short_response_holder = move(response);
        client->header_or_short_response = &client->header_or_short_response_holder;

        // Switch states.
        client->state = Client::SENDING_SHORT_RESPONSE;
        change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
}

namespace {

template<class T>
void delete_from(vector<T> *v, T elem)
{
        typename vector<T>::iterator new_end = remove(v->begin(), v->end(), elem);
        v->erase(new_end, v->end());
}

void send_ktls_close(int sock)
{
        uint8_t record_type = 21;  // Alert.
        uint8_t body[] = {
                1,   // Warning level (but still fatal!).
                0,   // close_notify.
        };

        int cmsg_len = sizeof(record_type);
        char buf[CMSG_SPACE(cmsg_len)];

        msghdr msg = {0};
        msg.msg_control = buf;
        msg.msg_controllen = sizeof(buf);
        cmsghdr *cmsg = CMSG_FIRSTHDR(&msg);
        cmsg->cmsg_level = SOL_TLS;
        cmsg->cmsg_type = TLS_SET_RECORD_TYPE;
        cmsg->cmsg_len = CMSG_LEN(cmsg_len);
        *CMSG_DATA(cmsg) = record_type;
        msg.msg_controllen = cmsg->cmsg_len;

        iovec msg_iov;
        msg_iov.iov_base = body;
        msg_iov.iov_len = sizeof(body);
        msg.msg_iov = &msg_iov;
        msg.msg_iovlen = 1;

        int err;
        do {
                err = sendmsg(sock, &msg, 0);
        } while (err == -1 && errno == EINTR);  // Ignore all other errors.
}

}  // namespace
        
void Server::close_client(Client *client)
{
        if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, nullptr) == -1) {
                log_perror("epoll_ctl(EPOLL_CTL_DEL)");
                exit(1);
        }

        // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
        if (client->stream != nullptr) {
                delete_from(&client->stream->sleeping_clients, client);
                delete_from(&client->stream->to_process, client);
        }

        if (client->tls_context) {
                if (client->in_ktls_mode) {
                        // Keep GnuTLS happy.
                        send_ktls_close(client->sock);
                }
                tls_destroy_context(client->tls_context);
        }

        // Log to access_log.
        access_log->write(client->get_stats());

        // Bye-bye!
        safe_close(client->sock);

        clients.erase(client->sock);
}

void Server::change_epoll_events(Client *client, uint32_t events)
{
        epoll_event ev;
        ev.events = events;
        ev.data.ptr = client;

        if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
                log_perror("epoll_ctl(EPOLL_CTL_MOD)");
                exit(1);
        }
}

bool Server::more_requests(Client *client)
{
        if (client->close_after_response) {
                return false;
        }

        // Log to access_log.
        access_log->write(client->get_stats());

        // Flush pending data; does not cancel out TCP_CORK (since that still takes priority),
        // but does a one-off flush.
        int one = 1;
        if (setsockopt(client->sock, SOL_TCP, TCP_NODELAY, &one, sizeof(one)) == -1) {
                log_perror("setsockopt(TCP_NODELAY)");
                // Can still continue.
        }

        // Switch states and reset the parsers. We don't reset statistics.
        client->state = Client::READING_REQUEST;
        client->url.clear();
        client->stream = NULL;
        client->header_or_short_response = nullptr;
        client->header_or_short_response_holder.clear();
        client->header_or_short_response_ref.reset();
        client->header_or_short_response_bytes_sent = 0;

        change_epoll_events(client, EPOLLIN | EPOLLET | EPOLLRDHUP);  // No TLS handshake, so no EPOLLOUT needed.

        return true;
}

void Server::process_queued_data()
{
        {
                lock_guard<mutex> lock(queued_clients_mutex);

                for (const pair<int, Acceptor *> &id_and_acceptor : queued_add_clients) {
                        add_client(id_and_acceptor.first, id_and_acceptor.second);
                }
                queued_add_clients.clear();
        }

        for (unique_ptr<Stream> &stream : streams) {
                stream->process_queued_data();
        }
}