[cubemap] / main.cpp

#include <stdio.h>
#include <string.h>
#include <stdint.h>
#include <assert.h>
#include <arpa/inet.h>
#include <sys/socket.h>
#include <pthread.h>
#include <sys/types.h>
#include <sys/ioctl.h>
#include <sys/poll.h>
#include <sys/time.h>
#include <signal.h>
#include <errno.h>
#include <ctype.h>
#include <fcntl.h>
#include <vector>
#include <string>
#include <map>
#include <set>

#include "metacube.h"
#include "parse.h"
#include "server.h"
#include "serverpool.h"
#include "input.h"
#include "state.pb.h"

using namespace std;

ServerPool *servers = NULL;
volatile bool hupped = false;

void hup(int ignored)
{
        hupped = true;
}

int create_server_socket(int port)
{
        int server_sock = socket(PF_INET6, SOCK_STREAM, IPPROTO_TCP);
        if (server_sock == -1) {
                perror("socket");
                exit(1);
        }

        int one = 1;
        if (setsockopt(server_sock, SOL_SOCKET, SO_REUSEADDR, &one, sizeof(one)) == -1) {
                perror("setsockopt(SO_REUSEADDR)");
                exit(1);
        }

        // We want dual-stack sockets. (Sorry, OpenBSD and Windows XP...)
        int zero = 0;
        if (setsockopt(server_sock, IPPROTO_IPV6, IPV6_V6ONLY, &zero, sizeof(zero)) == -1) {
                perror("setsockopt(IPV6_V6ONLY)");
                exit(1);
        }

        // Set as non-blocking, so the acceptor thread can notice that we want to shut it down.
        if (ioctl(server_sock, FIONBIO, &one) == -1) {
                perror("ioctl(FIONBIO)");
                exit(1);
        }

        sockaddr_in6 addr;
        memset(&addr, 0, sizeof(addr));
        addr.sin6_family = AF_INET6;
        addr.sin6_port = htons(port);

        if (bind(server_sock, reinterpret_cast<sockaddr *>(&addr), sizeof(addr)) == -1) {
                perror("bind");
                exit(1);
        }

        if (listen(server_sock, 128) == -1) {
                perror("listen");
                exit(1);
        }

        return server_sock;
}

void *acceptor_thread_run(void *arg)
{
        int server_sock = int(intptr_t(arg));
        while (!hupped) {
                // Since we are non-blocking, we need to wait for the right state first.
                // Wait up to 50 ms, then check hupped.
                pollfd pfd;
                pfd.fd = server_sock;
                pfd.events = POLLIN;

                int nfds = poll(&pfd, 1, 50);
                if (nfds == 0 || (nfds == -1 && errno == EINTR)) {
                        continue;
                }
                if (nfds == -1) {
                        perror("poll");
                        usleep(100000);
                        continue;
                }

                sockaddr_in6 addr;
                socklen_t addrlen = sizeof(addr);

                // Get a new socket.
                int sock = accept(server_sock, reinterpret_cast<sockaddr *>(&addr), &addrlen);
                if (sock == -1 && errno == EINTR) {
                        continue;
                }
                if (sock == -1) {
                        perror("accept");
                        usleep(100000);
                        continue;
                }

                // Set the socket as nonblocking.
                int one = 1;
                if (ioctl(sock, FIONBIO, &one) == -1) {
                        perror("FIONBIO");
                        exit(1);
                }

                // Pick a server, round-robin, and hand over the socket to it.
                servers->add_client(sock);
        }
        return NULL;
}

struct StatsThreadParameters {
        string stats_file;
        int stats_interval;
};
                
void *stats_thread_run(void *arg)
{
        const StatsThreadParameters *parms = reinterpret_cast<StatsThreadParameters *>(arg);
        while (!hupped) {
                int fd;
                FILE *fp;
                time_t now;
                vector<ClientStats> client_stats;

                // Open a new, temporary file.
                char *filename = strdup((parms->stats_file + ".new.XXXXXX").c_str());
                fd = mkostemp(filename, O_WRONLY);
                if (fd == -1) {
                        perror(filename);
                        free(filename);
                        goto sleep;
                }

                fp = fdopen(fd, "w");
                if (fp == NULL) {
                        perror("fdopen");
                        close(fd);
                        unlink(filename);
                        free(filename);
                        goto sleep;
                }

                now = time(NULL);
                client_stats = servers->get_client_stats();
                for (size_t i = 0; i < client_stats.size(); ++i) {
                        fprintf(fp, "%s %s %d %llu\n",
                                client_stats[i].remote_addr.c_str(),
                                client_stats[i].stream_id.c_str(),
                                int(now - client_stats[i].connect_time),
                                (long long unsigned)(client_stats[i].bytes_sent));
                }
                if (fclose(fp) == EOF) {
                        perror("fclose");
                        unlink(filename);
                        free(filename);
                        goto sleep;
                }
                
                if (rename(filename, parms->stats_file.c_str()) == -1) {
                        perror("rename");
                        unlink(filename);
                }

sleep:
                int left_to_sleep = parms->stats_interval;
                do {
                        left_to_sleep = sleep(left_to_sleep);
                } while (left_to_sleep > 0 && !hupped);
        }
        return NULL;
}

// Serialize the given state to a file descriptor, and return the (still open)
// descriptor.
int make_tempfile(const CubemapStateProto &state)
{
        char tmpl[] = "/tmp/cubemapstate.XXXXXX";
        int state_fd = mkstemp(tmpl);
        if (state_fd == -1) {
                perror("mkstemp");
                exit(1);
        }

        string serialized;
        state.SerializeToString(&serialized);

        const char *ptr = serialized.data();
        size_t to_write = serialized.size();
        while (to_write > 0) {
                ssize_t ret = write(state_fd, ptr, to_write);
                if (ret == -1) {
                        perror("write");
                        exit(1);
                }

                ptr += ret;
                to_write -= ret;
        }

        return state_fd;
}

// Read the state back from the file descriptor made by make_tempfile,
// and close it.
CubemapStateProto read_tempfile(int state_fd)
{
        if (lseek(state_fd, 0, SEEK_SET) == -1) {
                perror("lseek");
                exit(1);
        }

        string serialized;
        char buf[4096];
        for ( ;; ) {
                ssize_t ret = read(state_fd, buf, sizeof(buf));
                if (ret == -1) {
                        perror("read");
                        exit(1);
                }
                if (ret == 0) {
                        // EOF.
                        break;
                }

                serialized.append(string(buf, buf + ret));
        }

        close(state_fd);  // Implicitly deletes the file.

        CubemapStateProto state;
        if (!state.ParseFromString(serialized)) {
                fprintf(stderr, "PANIC: Failed deserialization of state.\n");
                exit(1);
        }

        return state;
}

int main(int argc, char **argv)
{
        fprintf(stderr, "\nCubemap starting.\n");

        struct timeval serialize_start;
        bool is_reexec = false;

        string config_filename = (argc == 1) ? "cubemap.config" : argv[1];
        vector<ConfigLine> config = parse_config(config_filename);

        int port = fetch_config_int(config, "port", 1, 65535, PARAMATER_MANDATORY);
        int num_servers = fetch_config_int(config, "num_servers", 1, 20000, PARAMATER_MANDATORY);  // Insanely high max limit.

        servers = new ServerPool(num_servers);

        int server_sock = -1, old_port = -1;
        set<string> deserialized_stream_ids;
        map<string, Input *> deserialized_inputs;
        if (argc == 4 && strcmp(argv[2], "-state") == 0) {
                is_reexec = true;

                fprintf(stderr, "Deserializing state from previous process... ");
                int state_fd = atoi(argv[3]);
                CubemapStateProto loaded_state = read_tempfile(state_fd);

                serialize_start.tv_sec = loaded_state.serialize_start_sec();
                serialize_start.tv_usec = loaded_state.serialize_start_usec();

                // Deserialize the streams.
                for (int i = 0; i < loaded_state.streams_size(); ++i) {
                        servers->add_stream_from_serialized(loaded_state.streams(i));
                        deserialized_stream_ids.insert(loaded_state.streams(i).stream_id());
                }

                // Put back the existing clients. It doesn't matter which server we
                // allocate them to, so just do round-robin.
                for (int i = 0; i < loaded_state.clients_size(); ++i) {
                        servers->add_client_from_serialized(loaded_state.clients(i));
                }

                // Deserialize the inputs. Note that we don't actually add them to any state yet.
                for (int i = 0; i < loaded_state.inputs_size(); ++i) {
                        deserialized_inputs.insert(make_pair(
                                loaded_state.inputs(i).stream_id(),
                                new Input(loaded_state.inputs(i))));
                } 

                // Deserialize the server socket.
                server_sock = loaded_state.server_sock();
                old_port = loaded_state.port();

                fprintf(stderr, "done.\n");
        }

        // Find all streams in the configuration file, and create them.
        set<string> expecting_stream_ids = deserialized_stream_ids;
        for (unsigned i = 0; i < config.size(); ++i) {
                if (config[i].keyword != "stream") {
                        continue;
                }
                if (config[i].arguments.size() != 1) {
                        fprintf(stderr, "ERROR: 'stream' takes exactly one argument\n");
                        exit(1);
                }
                string stream_id = config[i].arguments[0];
                if (deserialized_stream_ids.count(stream_id) == 0) {
                        servers->add_stream(stream_id);
                }
                expecting_stream_ids.erase(stream_id);
        }

        // Warn about any servers we've lost.
        // TODO: Make an option (delete=yes?) to actually shut down streams.
        for (set<string>::const_iterator stream_it = expecting_stream_ids.begin();
             stream_it != expecting_stream_ids.end();
             ++stream_it) {
                string stream_id = *stream_it;
                fprintf(stderr, "WARNING: stream '%s' disappeared from the configuration file.\n",
                        stream_id.c_str());
                fprintf(stderr, "         It will not be deleted, but clients will not get any new inputs.\n");
                if (deserialized_inputs.count(stream_id) != 0) {
                        delete deserialized_inputs[stream_id];
                        deserialized_inputs.erase(stream_id);
                }
        }

        // Open a new server socket if we do not already have one, or if we changed ports.
        if (server_sock != -1 && port != old_port) {
                fprintf(stderr, "NOTE: Port changed from %d to %d; opening new socket.\n", old_port, port);
                close(server_sock);
                server_sock = -1;
        }
        if (server_sock == -1) {
                server_sock = create_server_socket(port);
        }

        // See if the user wants stats.
        string stats_file = fetch_config_string(config, "stats_file", PARAMETER_OPTIONAL);
        int stats_interval = fetch_config_int(config, "stats_interval", 1, INT_MAX, PARAMETER_OPTIONAL, -1);
        if (stats_interval != -1 && stats_file.empty()) {
                fprintf(stderr, "WARNING: 'stats_interval' given, but no 'stats_file'. No statistics will be written.\n");
        }

        servers->run();

        pthread_t acceptor_thread;
        pthread_create(&acceptor_thread, NULL, acceptor_thread_run, reinterpret_cast<void *>(server_sock));

        // Find all streams in the configuration file, and create inputs for them.
        vector<Input *> inputs;
        for (unsigned i = 0; i < config.size(); ++i) {
                if (config[i].keyword != "stream") {
                        continue;
                }
                assert(config[i].arguments.size() == 1);
                string stream_id = config[i].arguments[0];

                if (config[i].parameters.count("src") == 0) {
                        fprintf(stderr, "WARNING: stream '%s' has no src= attribute, clients will not get any data.\n",
                                stream_id.c_str());
                        continue;
                }

                string src = config[i].parameters["src"];
                Input *input = NULL;
                if (deserialized_inputs.count(stream_id) != 0) {
                        input = deserialized_inputs[stream_id];
                        if (input->get_url() != src) {
                                fprintf(stderr, "INFO: Stream '%s' has changed URL from '%s' to '%s', restarting input.\n",
                                        stream_id.c_str(), input->get_url().c_str(), src.c_str());
                                delete input;
                                input = NULL;
                        }
                        deserialized_inputs.erase(stream_id);
                }
                if (input == NULL) {
                        input = new Input(stream_id, src);
                }
                input->run();
                inputs.push_back(input);
        }

        // All deserialized inputs should now have been taken care of, one way or the other.
        assert(deserialized_inputs.empty());

        // Start writing statistics.
        pthread_t stats_thread;
        StatsThreadParameters stats_parameters;  // Must live for as long as the stats thread does.
        if (!stats_file.empty()) {
                stats_parameters.stats_file = stats_file;
                stats_parameters.stats_interval = stats_interval;
                pthread_create(&stats_thread, NULL, stats_thread_run, &stats_parameters);
        }

        signal(SIGHUP, hup);
        
        struct timeval server_start;
        gettimeofday(&server_start, NULL);
        if (is_reexec) {
                // Measure time from we started deserializing (below) to now, when basically everything
                // is up and running. This is, in other words, a conservative estimate of how long our
                // “glitch” period was, not counting of course reconnects if the configuration changed.
                double glitch_time = server_start.tv_sec - serialize_start.tv_sec +
                        1e-6 * (server_start.tv_usec - serialize_start.tv_usec);
                fprintf(stderr, "Re-exec happened in approx. %.0f ms.\n", glitch_time * 1000.0);
        }

        while (!hupped) {
                usleep(100000);
        }

        // OK, we've been HUPed. Time to shut down everything, serialize, and re-exec.
        gettimeofday(&serialize_start, NULL);

        if (!stats_file.empty()) {
                pthread_kill(stats_thread, SIGHUP);
                if (pthread_join(stats_thread, NULL) == -1) {
                        perror("pthread_join");
                        exit(1);
                }
        }
        pthread_kill(acceptor_thread, SIGHUP);
        if (pthread_join(acceptor_thread, NULL) == -1) {
                perror("pthread_join");
                exit(1);
        }

        CubemapStateProto state;
        state.set_serialize_start_sec(serialize_start.tv_sec);
        state.set_serialize_start_usec(serialize_start.tv_usec);
        state.set_server_sock(server_sock);
        state.set_port(port);

        for (size_t i = 0; i < inputs.size(); ++i) {
                inputs[i]->stop();
                state.add_inputs()->MergeFrom(inputs[i]->serialize());
        }

        for (int i = 0; i < num_servers; ++i) { 
                servers->get_server(i)->stop();

                CubemapStateProto local_state = servers->get_server(i)->serialize();

                // The stream state should be identical between the servers, so we only store it once.
                if (i == 0) {
                        state.mutable_streams()->MergeFrom(local_state.streams());
                }
                for (int j = 0; j < local_state.clients_size(); ++j) {
                        state.add_clients()->MergeFrom(local_state.clients(j));
                }
        }
        delete servers;

        fprintf(stderr, "Serializing state and re-execing...\n");
        int state_fd = make_tempfile(state);

        char buf[16];
        sprintf(buf, "%d", state_fd);

        for ( ;; ) {
                execlp(argv[0], argv[0], config_filename.c_str(), "-state", buf, NULL);
                perror("execlp");
                fprintf(stderr, "PANIC: re-exec of %s failed. Waiting 0.2 seconds and trying again...\n", argv[0]);
                usleep(200000);
        }
}