-#include <stdio.h>
-#include <string.h>
-#include <stdint.h>
#include <assert.h>
-#include <arpa/inet.h>
-#include <curl/curl.h>
-#include <sys/socket.h>
+#include <errno.h>
#include <pthread.h>
-#include <sys/types.h>
-#include <sys/ioctl.h>
+#include <stdio.h>
+#include <stdlib.h>
#include <sys/epoll.h>
-#include <errno.h>
-#include <vector>
-#include <string>
-#include <map>
+#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 "metacube.h"
-#include "server.h"
+#include "log.h"
+#include "markpool.h"
#include "mutexlock.h"
#include "parse.h"
+#include "server.h"
#include "state.pb.h"
+#include "stream.h"
using namespace std;
-Client::Client(int sock)
- : sock(sock),
- state(Client::READING_REQUEST),
- header_or_error_bytes_sent(0),
- bytes_sent(0)
-{
- request.reserve(1024);
-}
-
-Client::Client(const ClientProto &serialized)
- : sock(serialized.sock()),
- state(State(serialized.state())),
- request(serialized.request()),
- stream_id(serialized.stream_id()),
- header_or_error(serialized.header_or_error()),
- header_or_error_bytes_sent(serialized.header_or_error_bytes_sent()),
- bytes_sent(serialized.bytes_sent())
-{
-}
-
-ClientProto Client::serialize() const
-{
- ClientProto serialized;
- serialized.set_sock(sock);
- serialized.set_state(state);
- serialized.set_request(request);
- serialized.set_stream_id(stream_id);
- serialized.set_header_or_error(header_or_error);
- serialized.set_header_or_error_bytes_sent(serialized.header_or_error_bytes_sent());
- serialized.set_bytes_sent(bytes_sent);
- return serialized;
-}
-
-Stream::Stream(const string &stream_id)
- : stream_id(stream_id),
- data(new char[BACKLOG_SIZE]),
- data_size(0)
-{
- memset(data, 0, BACKLOG_SIZE);
-}
-
-Stream::~Stream()
-{
- delete[] data;
-}
-
-Stream::Stream(const StreamProto &serialized)
- : stream_id(serialized.stream_id()),
- header(serialized.header()),
- data(new char[BACKLOG_SIZE]),
- data_size(serialized.data_size())
-{
- assert(serialized.data().size() == BACKLOG_SIZE);
- memcpy(data, serialized.data().data(), BACKLOG_SIZE);
-}
-
-StreamProto Stream::serialize() const
-{
- StreamProto serialized;
- serialized.set_header(header);
- serialized.set_data(string(data, data + BACKLOG_SIZE));
- serialized.set_data_size(data_size);
- serialized.set_stream_id(stream_id);
- return serialized;
-}
-
Server::Server()
{
pthread_mutex_init(&mutex, NULL);
+ pthread_mutex_init(&queued_data_mutex, NULL);
epoll_fd = epoll_create(1024); // Size argument is ignored.
if (epoll_fd == -1) {
Server::~Server()
{
- close(epoll_fd);
-}
+ int ret;
+ do {
+ ret = close(epoll_fd);
+ } while (ret == -1 && errno == EINTR);
-void Server::run()
-{
- should_stop = false;
-
- // Joinable is already the default, but it's good to be certain.
- pthread_attr_t attr;
- pthread_attr_init(&attr);
- pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
- pthread_create(&worker_thread, &attr, Server::do_work_thunk, this);
-}
-
-void Server::stop()
-{
- {
- MutexLock lock(&mutex);
- should_stop = true;
- }
-
- if (pthread_join(worker_thread, NULL) == -1) {
- perror("pthread_join");
- exit(1);
+ if (ret == -1) {
+ perror("close(epoll_fd)");
}
}
-void *Server::do_work_thunk(void *arg)
+vector<ClientStats> Server::get_client_stats() const
{
- Server *server = static_cast<Server *>(arg);
- server->do_work();
- return NULL;
+ vector<ClientStats> ret;
+
+ MutexLock lock(&mutex);
+ for (map<int, Client>::const_iterator client_it = clients.begin();
+ client_it != clients.end();
+ ++client_it) {
+ ret.push_back(client_it->second.get_stats());
+ }
+ return ret;
}
void Server::do_work()
for ( ;; ) {
int nfds = epoll_wait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS);
if (nfds == -1 && errno == EINTR) {
+ if (should_stop) {
+ return;
+ }
continue;
}
if (nfds == -1) {
MutexLock lock(&mutex); // We release the mutex between iterations.
- if (should_stop) {
- return;
- }
-
+ process_queued_data();
+
for (int i = 0; i < nfds; ++i) {
- int fd = events[i].data.fd;
- assert(clients.count(fd) != 0);
- Client *client = &clients[fd];
+ Client *client = reinterpret_cast<Client *>(events[i].data.u64);
if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
close_client(client);
process_client(client);
}
+
+ for (map<string, Stream *>::iterator stream_it = streams.begin();
+ stream_it != streams.end();
+ ++stream_it) {
+ vector<Client *> to_process;
+ swap(stream_it->second->to_process, to_process);
+ for (size_t i = 0; i < to_process.size(); ++i) {
+ process_client(to_process[i]);
+ }
+ }
+
+ if (should_stop) {
+ return;
+ }
}
}
-CubemapStateProto Server::serialize() const
+CubemapStateProto Server::serialize()
{
+ // We don't serialize anything queued, so empty the queues.
+ process_queued_data();
+
CubemapStateProto serialized;
for (map<int, Client>::const_iterator client_it = clients.begin();
client_it != clients.end();
return serialized;
}
+void Server::add_client_deferred(int sock)
+{
+ MutexLock lock(&queued_data_mutex);
+ queued_add_clients.push_back(sock);
+}
+
void Server::add_client(int sock)
{
- MutexLock lock(&mutex);
clients.insert(make_pair(sock, Client(sock)));
// Start listening on data from this socket.
epoll_event ev;
- ev.events = EPOLLIN | EPOLLRDHUP;
- ev.data.u64 = 0; // Keep Valgrind happy.
- ev.data.fd = sock;
+ ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
+ ev.data.u64 = reinterpret_cast<uint64_t>(&clients[sock]);
if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
perror("epoll_ctl(EPOLL_CTL_ADD)");
exit(1);
}
+
+ process_client(&clients[sock]);
}
void Server::add_client_from_serialized(const ClientProto &client)
{
MutexLock lock(&mutex);
- clients.insert(make_pair(client.sock(), Client(client)));
+ Stream *stream;
+ map<string, Stream *>::iterator stream_it = streams.find(client.stream_id());
+ if (stream_it == streams.end()) {
+ stream = NULL;
+ } else {
+ stream = stream_it->second;
+ }
+ clients.insert(make_pair(client.sock(), Client(client, stream)));
+ Client *client_ptr = &clients[client.sock()];
// Start listening on data from this socket.
epoll_event ev;
if (client.state() == Client::READING_REQUEST) {
- ev.events = EPOLLIN | EPOLLRDHUP;
+ 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 | EPOLLRDHUP;
+ ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
}
ev.data.u64 = 0; // Keep Valgrind happy.
- ev.data.fd = client.sock();
+ ev.data.u64 = reinterpret_cast<uint64_t>(client_ptr);
if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
perror("epoll_ctl(EPOLL_CTL_ADD)");
exit(1);
}
+
+ if (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);
+ }
}
-void Server::add_stream(const string &stream_id)
+void Server::add_stream(const string &stream_id, size_t backlog_size)
{
MutexLock lock(&mutex);
- streams.insert(make_pair(stream_id, new Stream(stream_id)));
+ streams.insert(make_pair(stream_id, new Stream(stream_id, backlog_size)));
}
void Server::add_stream_from_serialized(const StreamProto &stream)
streams.insert(make_pair(stream.stream_id(), new Stream(stream)));
}
+void Server::set_backlog_size(const std::string &stream_id, size_t new_size)
+{
+ MutexLock lock(&mutex);
+ assert(streams.count(stream_id) != 0);
+ streams[stream_id]->set_backlog_size(new_size);
+}
+
void Server::set_header(const string &stream_id, const string &header)
{
MutexLock lock(&mutex);
++client_it) {
Client *client = &client_it->second;
if (client->state == Client::SENDING_DATA &&
- client->bytes_sent == 0) {
+ client->stream_pos == 0) {
construct_header(client);
}
}
}
-void Server::add_data(const string &stream_id, const char *data, size_t bytes)
+void Server::set_mark_pool(const std::string &stream_id, MarkPool *mark_pool)
{
- if (bytes == 0) {
- return;
- }
-
MutexLock lock(&mutex);
- Stream *stream = find_stream(stream_id);
- size_t pos = stream->data_size % BACKLOG_SIZE;
- stream->data_size += bytes;
-
- if (pos + bytes > BACKLOG_SIZE) {
- size_t to_copy = BACKLOG_SIZE - pos;
- memcpy(stream->data + pos, data, to_copy);
- data += to_copy;
- bytes -= to_copy;
- pos = 0;
- }
+ assert(clients.empty());
+ find_stream(stream_id)->mark_pool = mark_pool;
+}
- memcpy(stream->data + pos, data, bytes);
- wake_up_all_clients();
+void Server::add_data_deferred(const string &stream_id, const char *data, size_t bytes)
+{
+ MutexLock lock(&queued_data_mutex);
+ queued_data[stream_id].append(string(data, data + bytes));
}
-
+
+// See the .h file for postconditions after this function.
void Server::process_client(Client *client)
{
switch (client->state) {
case Client::READING_REQUEST: {
+read_request_again:
// Try to read more of the request.
char buf[1024];
- int ret = read(client->sock, buf, sizeof(buf));
+ int ret;
+ do {
+ ret = read(client->sock, buf, sizeof(buf));
+ } while (ret == -1 && errno == EINTR);
+
+ if (ret == -1 && errno == EAGAIN) {
+ // No more data right now. Nothing to do.
+ // This is postcondition #2.
+ return;
+ }
if (ret == -1) {
perror("read");
close_client(client);
return;
}
if (ret == 0) {
- // No data? This really means that we were triggered for something else than
- // POLLIN (which suggests a logic error in epoll).
- fprintf(stderr, "WARNING: fd %d returned unexpectedly 0 bytes!\n", client->sock);
+ // OK, the socket is closed.
close_client(client);
return;
}
- // Guard against overlong requests gobbling up all of our space.
- if (client->request.size() + ret > MAX_CLIENT_REQUEST) {
- fprintf(stderr, "WARNING: fd %d sent overlong request!\n", client->sock);
+ RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
+
+ switch (status) {
+ case RP_OUT_OF_SPACE:
+ log(WARNING, "fd %d sent overlong request!", client->sock);
close_client(client);
return;
- }
-
- // See if we have \r\n\r\n anywhere in the request. We start three bytes
- // before what we just appended, in case we just got the final character.
- size_t existing_req_bytes = client->request.size();
- client->request.append(string(buf, buf + ret));
-
- size_t start_at = (existing_req_bytes >= 3 ? existing_req_bytes - 3 : 0);
- const char *ptr = reinterpret_cast<char *>(
- memmem(client->request.data() + start_at, client->request.size() - start_at,
- "\r\n\r\n", 4));
- if (ptr == NULL) {
+ case RP_NOT_FINISHED_YET:
// OK, we don't have the entire header yet. Fine; we'll get it later.
- return;
- }
-
- if (ptr != client->request.data() + client->request.size() - 4) {
- fprintf(stderr, "WARNING: fd %d had junk data after request!\n", client->sock);
+ // See if there's more data for us.
+ goto read_request_again;
+ case RP_EXTRA_DATA:
+ log(WARNING, "fd %d had junk data after request!", client->sock);
close_client(client);
return;
+ case RP_FINISHED:
+ break;
}
+ assert(status == RP_FINISHED);
+
int error_code = parse_request(client);
if (error_code == 200) {
construct_header(client);
} else {
construct_error(client, error_code);
}
- break;
+
+ // We've changed states, so fall through.
+ assert(client->state == Client::SENDING_ERROR ||
+ client->state == Client::SENDING_HEADER);
}
case Client::SENDING_ERROR:
case Client::SENDING_HEADER: {
- int ret = write(client->sock,
- client->header_or_error.data() + client->header_or_error_bytes_sent,
- client->header_or_error.size() - client->header_or_error_bytes_sent);
+sending_header_or_error_again:
+ int ret;
+ do {
+ ret = write(client->sock,
+ client->header_or_error.data() + client->header_or_error_bytes_sent,
+ client->header_or_error.size() - client->header_or_error_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.
perror("write");
close_client(client);
return;
assert(client->header_or_error_bytes_sent <= client->header_or_error.size());
if (client->header_or_error_bytes_sent < client->header_or_error.size()) {
- // We haven't sent all yet. Fine; we'll do that later.
- return;
+ // We haven't sent all yet. Fine; go another round.
+ goto sending_header_or_error_again;
}
// We're done sending the header or error! Clear it to release some memory.
client->header_or_error.clear();
if (client->state == Client::SENDING_ERROR) {
+ // We're done sending the error, so now close.
+ // This is postcondition #1.
close_client(client);
- } else {
- // Start sending from the end. In other words, we won't send any of the backlog,
- // but we'll start sending immediately as we get data.
- client->state = Client::SENDING_DATA;
- client->bytes_sent = find_stream(client->stream_id)->data_size;
+ return;
}
- break;
+
+ // Start sending from the end. In other words, we won't send any of the backlog,
+ // but we'll start sending immediately as we get data.
+ // This is postcondition #3.
+ client->state = Client::SENDING_DATA;
+ client->stream_pos = client->stream->bytes_received;
+ client->stream->put_client_to_sleep(client);
+ return;
}
case Client::SENDING_DATA: {
+sending_data_again:
// 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.
- const Stream &stream = *find_stream(client->stream_id);
- size_t bytes_to_send = stream.data_size - client->bytes_sent;
- if (bytes_to_send > BACKLOG_SIZE) {
- fprintf(stderr, "WARNING: fd %d lost %lld bytes, maybe too slow connection\n",
+ Stream *stream = client->stream;
+ size_t bytes_to_send = stream->bytes_received - client->stream_pos;
+ if (bytes_to_send == 0) {
+ return;
+ }
+ if (bytes_to_send > stream->backlog_size) {
+ log(WARNING, "fd %d lost %lld bytes, maybe too slow connection",
client->sock,
- (long long int)(bytes_to_send - BACKLOG_SIZE));
- client->bytes_sent = find_stream(client->stream_id)->data_size - BACKLOG_SIZE;
- bytes_to_send = BACKLOG_SIZE;
+ (long long int)(bytes_to_send - stream->backlog_size));
+ client->stream_pos = stream->bytes_received - stream->backlog_size;
+ client->bytes_lost += bytes_to_send - stream->backlog_size;
+ ++client->num_loss_events;
+ bytes_to_send = stream->backlog_size;
}
// See if we need to split across the circular buffer.
- ssize_t ret;
- if ((client->bytes_sent % BACKLOG_SIZE) + bytes_to_send > BACKLOG_SIZE) {
- size_t bytes_first_part = BACKLOG_SIZE - (client->bytes_sent % BACKLOG_SIZE);
-
- iovec iov[2];
- iov[0].iov_base = const_cast<char *>(stream.data + (client->bytes_sent % BACKLOG_SIZE));
- iov[0].iov_len = bytes_first_part;
-
- iov[1].iov_base = const_cast<char *>(stream.data);
- iov[1].iov_len = bytes_to_send - bytes_first_part;
+ 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;
+ }
- ret = writev(client->sock, iov, 2);
- } else {
- ret = write(client->sock,
- stream.data + (client->bytes_sent % BACKLOG_SIZE),
- bytes_to_send);
+ ssize_t ret;
+ do {
+ loff_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) {
- perror("write/writev");
+ // Error, close; postcondition #1.
+ perror("sendfile");
close_client(client);
return;
}
+ client->stream_pos += ret;
client->bytes_sent += ret;
- if (client->bytes_sent == stream.data_size) {
+ if (client->stream_pos == stream->bytes_received) {
// We don't have any more data for this client, so put it to sleep.
- put_client_to_sleep(client);
+ // This is postcondition #3.
+ stream->put_client_to_sleep(client);
+ } else if (more_data && size_t(ret) == bytes_to_send) {
+ goto sending_data_again;
}
break;
}
}
client->stream_id = request_tokens[1];
+ client->stream = find_stream(client->stream_id);
+ if (client->stream->mark_pool != NULL) {
+ client->fwmark = client->stream->mark_pool->get_mark();
+ } else {
+ client->fwmark = 0; // No mark.
+ }
+ if (setsockopt(client->sock, SOL_SOCKET, SO_MARK, &client->fwmark, sizeof(client->fwmark)) == -1) {
+ if (client->fwmark != 0) {
+ perror("setsockopt(SO_MARK)");
+ }
+ }
client->request.clear();
return 200; // OK!
void Server::construct_header(Client *client)
{
- client->header_or_error = "HTTP/1.0 200 OK\r\nContent-type: video/x-flv\r\nCache-Control: no-cache\r\n\r\n" +
- find_stream(client->stream_id)->header;
+ client->header_or_error = find_stream(client->stream_id)->header;
// Switch states.
client->state = Client::SENDING_HEADER;
epoll_event ev;
- ev.events = EPOLLOUT | EPOLLRDHUP;
- ev.data.u64 = 0; // Keep Valgrind happy.
- ev.data.fd = client->sock;
+ ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
+ ev.data.u64 = reinterpret_cast<uint64_t>(client);
if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
perror("epoll_ctl(EPOLL_CTL_MOD)");
client->state = Client::SENDING_ERROR;
epoll_event ev;
- ev.events = EPOLLOUT | EPOLLRDHUP;
- ev.data.u64 = 0; // Keep Valgrind happy.
- ev.data.fd = client->sock;
+ ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
+ ev.data.u64 = reinterpret_cast<uint64_t>(client);
if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
perror("epoll_ctl(EPOLL_CTL_MOD)");
exit(1);
}
}
+
+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 Server::close_client(Client *client)
{
}
// This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
- vector<int>::iterator new_end =
- remove(sleeping_clients.begin(), sleeping_clients.end(), client->sock);
- sleeping_clients.erase(new_end, sleeping_clients.end());
-
- // Bye-bye!
- close(client->sock);
- clients.erase(client->sock);
-}
-
-void Server::put_client_to_sleep(Client *client)
-{
- epoll_event ev;
- ev.events = EPOLLRDHUP;
- ev.data.u64 = 0; // Keep Valgrind happy.
- ev.data.fd = client->sock;
-
- if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
- perror("epoll_ctl(EPOLL_CTL_MOD)");
- exit(1);
+ if (client->stream != NULL) {
+ delete_from(&client->stream->sleeping_clients, client);
+ delete_from(&client->stream->to_process, client);
+ if (client->stream->mark_pool != NULL) {
+ int fwmark = client->fwmark;
+ client->stream->mark_pool->release_mark(fwmark);
+ }
}
- sleeping_clients.push_back(client->sock);
-}
+ // Bye-bye!
+ int ret;
+ do {
+ ret = close(client->sock);
+ } while (ret == -1 && errno == EINTR);
-void Server::wake_up_all_clients()
-{
- for (unsigned i = 0; i < sleeping_clients.size(); ++i) {
- epoll_event ev;
- ev.events = EPOLLOUT | EPOLLRDHUP;
- ev.data.u64 = 0; // Keep Valgrind happy.
- ev.data.fd = sleeping_clients[i];
- if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, sleeping_clients[i], &ev) == -1) {
- perror("epoll_ctl(EPOLL_CTL_MOD)");
- exit(1);
- }
+ if (ret == -1) {
+ perror("close");
}
- sleeping_clients.clear();
+
+ clients.erase(client->sock);
}
Stream *Server::find_stream(const string &stream_id)
assert(it != streams.end());
return it->second;
}
+
+void Server::process_queued_data()
+{
+ MutexLock lock(&queued_data_mutex);
+
+ for (size_t i = 0; i < queued_add_clients.size(); ++i) {
+ add_client(queued_add_clients[i]);
+ }
+ queued_add_clients.clear();
+
+ for (map<string, string>::iterator queued_it = queued_data.begin();
+ queued_it != queued_data.end();
+ ++queued_it) {
+ Stream *stream = find_stream(queued_it->first);
+ stream->add_data(queued_it->second.data(), queued_it->second.size());
+ stream->wake_up_all_clients();
+ }
+ queued_data.clear();
+}