-#include <stdio.h>
-#include <string.h>
-#include <stdint.h>
#include <assert.h>
-#include <arpa/inet.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 <time.h>
-#include <signal.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),
- connect_time(time(NULL)),
- state(Client::READING_REQUEST),
- stream(NULL),
- header_or_error_bytes_sent(0),
- bytes_sent(0)
-{
- request.reserve(1024);
-
- // Find the remote address, and convert it to ASCII.
- sockaddr_in6 addr;
- socklen_t addr_len = sizeof(addr);
-
- if (getpeername(sock, reinterpret_cast<sockaddr *>(&addr), &addr_len) == -1) {
- perror("getpeername");
- remote_addr = "";
- } else {
- char buf[INET6_ADDRSTRLEN];
- if (inet_ntop(addr.sin6_family, &addr.sin6_addr, buf, sizeof(buf)) == NULL) {
- perror("inet_ntop");
- remote_addr = "";
- } else {
- remote_addr = buf;
- }
- }
-}
-
-Client::Client(const ClientProto &serialized, Stream *stream)
- : sock(serialized.sock()),
- remote_addr(serialized.remote_addr()),
- connect_time(serialized.connect_time()),
- state(State(serialized.state())),
- request(serialized.request()),
- stream_id(serialized.stream_id()),
- stream(stream),
- 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_remote_addr(remote_addr);
- serialized.set_connect_time(connect_time);
- 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;
-}
-
-ClientStats Client::get_stats() const
-{
- ClientStats stats;
- stats.stream_id = stream_id;
- stats.remote_addr = remote_addr;
- stats.connect_time = connect_time;
- stats.bytes_sent = bytes_sent;
- return stats;
-}
-
-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;
-}
-
-void Stream::put_client_to_sleep(Client *client)
-{
- sleeping_clients.push_back(client);
-}
-
-void Stream::wake_up_all_clients()
-{
- if (to_process.empty()) {
- swap(sleeping_clients, to_process);
- } else {
- to_process.insert(to_process.end(), sleeping_clients.begin(), sleeping_clients.end());
- sleeping_clients.clear();
- }
-}
-
Server::Server()
{
pthread_mutex_init(&mutex, NULL);
}
}
-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;
- }
-
- pthread_kill(worker_thread, SIGHUP);
- if (pthread_join(worker_thread, NULL) == -1) {
- perror("pthread_join");
- exit(1);
- }
-}
-
vector<ClientStats> Server::get_client_stats() const
{
vector<ClientStats> ret;
return ret;
}
-void *Server::do_work_thunk(void *arg)
-{
- Server *server = static_cast<Server *>(arg);
- server->do_work();
- return NULL;
-}
-
void Server::do_work()
{
for ( ;; ) {
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);
// Start listening on data from this socket.
epoll_event ev;
ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
- ev.data.u64 = 0; // Keep Valgrind happy.
- ev.data.fd = sock;
+ 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);
void Server::add_client_from_serialized(const ClientProto &client)
{
MutexLock lock(&mutex);
- Stream *stream = find_stream(client.stream_id());
+ 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()];
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->bytes_sent == client_ptr->stream->data_size) {
+ 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::set_mark_pool(const std::string &stream_id, MarkPool *mark_pool)
+{
+ MutexLock lock(&mutex);
+ assert(clients.empty());
+ find_stream(stream_id)->mark_pool = mark_pool;
+}
void Server::add_data_deferred(const string &stream_id, const char *data, size_t bytes)
{
queued_data[stream_id].append(string(data, data + bytes));
}
-void Server::add_data(const string &stream_id, const char *data, size_t bytes)
-{
- 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;
- }
-
- memcpy(stream->data + pos, data, bytes);
- stream->wake_up_all_clients();
-}
-
// See the .h file for postconditions after this function.
void Server::process_client(Client *client)
{
switch (status) {
case RP_OUT_OF_SPACE:
- fprintf(stderr, "WARNING: fd %d sent overlong request!\n", client->sock);
+ log(WARNING, "fd %d sent overlong request!", client->sock);
close_client(client);
return;
case RP_NOT_FINISHED_YET:
// See if there's more data for us.
goto read_request_again;
case RP_EXTRA_DATA:
- fprintf(stderr, "WARNING: fd %d had junk data after request!\n", client->sock);
+ log(WARNING, "fd %d had junk data after request!", client->sock);
close_client(client);
return;
case RP_FINISHED:
// but we'll start sending immediately as we get data.
// This is postcondition #3.
client->state = Client::SENDING_DATA;
- client->bytes_sent = client->stream->data_size;
+ 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.
Stream *stream = client->stream;
- size_t bytes_to_send = stream->data_size - client->bytes_sent;
+ size_t bytes_to_send = stream->bytes_received - client->stream_pos;
if (bytes_to_send == 0) {
return;
}
- if (bytes_to_send > BACKLOG_SIZE) {
- fprintf(stderr, "WARNING: fd %d lost %lld bytes, maybe too slow connection\n",
+ 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 = stream->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;
+ 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;
+ }
- iov[1].iov_base = const_cast<char *>(stream->data);
- iov[1].iov_len = bytes_to_send - bytes_first_part;
+ 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);
- do {
- ret = writev(client->sock, iov, 2);
- } while (ret == -1 && errno == EINTR);
- } else {
- do {
- ret = write(client->sock,
- stream->data + (client->bytes_sent % BACKLOG_SIZE),
- 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.
}
if (ret == -1) {
// Error, close; postcondition #1.
- perror("write/writev");
+ 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.
// This is postcondition #3.
stream->put_client_to_sleep(client);
- } else {
- // XXX: Do we need to go another round here to explicitly
- // get the EAGAIN?
+ } 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!
epoll_event ev;
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);
if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
perror("epoll_ctl(EPOLL_CTL_MOD)");
epoll_event ev;
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);
if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
perror("epoll_ctl(EPOLL_CTL_MOD)");
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);
+ }
}
-
+
// Bye-bye!
int ret;
do {
for (map<string, string>::iterator queued_it = queued_data.begin();
queued_it != queued_data.end();
++queued_it) {
- add_data(queued_it->first, queued_it->second.data(), queued_it->second.size());
+ 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();
}