X-Git-Url: https://git.sesse.net/?p=cubemap;a=blobdiff_plain;f=server.cpp;h=99a3925e13f0bca65531f82902012408e0b5556c;hp=d41e3c70730b896e44a9362bf150ae534ee9f892;hb=fb0ec9c7b3fb51c5da24af0183d35f3815e6ca0e;hpb=d9de2a370c08fd6738984e759a1ab4f91015af0c

diff --git a/server.cpp b/server.cpp
index d41e3c7..99a3925 100644
--- a/server.cpp
+++ b/server.cpp
@@ -1,172 +1,115 @@
+#include <assert.h>
+#include <errno.h>
+#include <netinet/in.h>
+#include <pthread.h>
+#include <stdint.h>
 #include <stdio.h>
+#include <stdlib.h>
 #include <string.h>
-#include <stdint.h>
-#include <assert.h>
-#include <arpa/inet.h>
-#include <curl/curl.h>
+#include <sys/epoll.h>
+#include <sys/sendfile.h>
 #include <sys/socket.h>
-#include <pthread.h>
 #include <sys/types.h>
-#include <sys/ioctl.h>
-#include <sys/epoll.h>
-#include <errno.h>
-#include <vector>
-#include <string>
-#include <map>
+#include <unistd.h>
 #include <algorithm>
+#include <map>
+#include <string>
+#include <utility>
+#include <vector>
 
-#include "metacube.h"
-#include "server.h"
+#include "accesslog.h"
+#include "log.h"
+#include "metacube2.h"
 #include "mutexlock.h"
 #include "parse.h"
+#include "server.h"
 #include "state.pb.h"
+#include "stream.h"
+#include "util.h"
 
-using namespace std;
+#ifndef SO_MAX_PACING_RATE
+#define SO_MAX_PACING_RATE 47
+#endif
 
-Client::Client(int sock)
-	: sock(sock),
-	  state(Client::READING_REQUEST),
-	  stream(NULL),
-	  header_or_error_bytes_sent(0),
-	  bytes_sent(0)
-{
-	request.reserve(1024);
-}
-	
-Client::Client(const ClientProto &serialized, Stream *stream)
-	: sock(serialized.sock()),
-	  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())
-{
-}
+using namespace std;
 
-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;
-}
+extern AccessLogThread *access_log;
 
-Stream::Stream(const string &stream_id)
-	: stream_id(stream_id),
-	  data(new char[BACKLOG_SIZE]),
-	  data_size(0)
-{
-	memset(data, 0, BACKLOG_SIZE);
-}
+namespace {
 
-Stream::~Stream()
+inline bool is_equal(timespec a, timespec b)
 {
-	delete[] data;
+	return a.tv_sec == b.tv_sec &&
+	       a.tv_nsec == b.tv_nsec;
 }
 
-Stream::Stream(const StreamProto &serialized)
-	: stream_id(serialized.stream_id()),
-	  header(serialized.header()),
-	  data(new char[BACKLOG_SIZE]),
-	  data_size(serialized.data_size())
+inline bool is_earlier(timespec a, timespec b)
 {
-	assert(serialized.data().size() == BACKLOG_SIZE);
-	memcpy(data, serialized.data().data(), BACKLOG_SIZE);
+	if (a.tv_sec != b.tv_sec)
+		return a.tv_sec < b.tv_sec;
+	return a.tv_nsec < b.tv_nsec;
 }
 
-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;
-}
+}  // namespace
 
 Server::Server()
 {
 	pthread_mutex_init(&mutex, NULL);
+	pthread_mutex_init(&queued_clients_mutex, NULL);
 
 	epoll_fd = epoll_create(1024);  // Size argument is ignored.
 	if (epoll_fd == -1) {
-		perror("epoll_fd");
+		log_perror("epoll_fd");
 		exit(1);
 	}
 }
 
 Server::~Server()
 {
-	int ret;
-	do {
-		ret = close(epoll_fd);
-	} while (ret == -1 && errno == EINTR);
-
-	if (ret == -1) {
-		perror("close(epoll_fd)");
+	for (size_t i = 0; i < streams.size(); ++i) {	
+		delete streams[i];
 	}
-}
 
-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);
+	safe_close(epoll_fd);
 }
-	
-void Server::stop()
+
+vector<ClientStats> Server::get_client_stats() const
 {
-	{
-		MutexLock lock(&mutex);
-		should_stop = true;
-	}
+	vector<ClientStats> ret;
 
-	if (pthread_join(worker_thread, NULL) == -1) {
-		perror("pthread_join");
-		exit(1);
+	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());
 	}
-}
-
-void *Server::do_work_thunk(void *arg)
-{
-	Server *server = static_cast<Server *>(arg);
-	server->do_work();
-	return NULL;
+	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) {
-			continue;
-		}
-		if (nfds == -1) {
-			perror("epoll_wait");
+	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);
 		}
 
 		MutexLock lock(&mutex);  // We release the mutex between iterations.
 	
-		if (should_stop) {
-			return;
-		}
+		process_queued_data();
 
+		// Process each client where we have socket activity.
 		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);
@@ -176,53 +119,147 @@ void Server::do_work()
 			process_client(client);
 		}
 
-		for (unsigned i = 0; i < to_process.size(); ++i) {
-			process_client(to_process[i]);
+		// Process each client where its stream has new data,
+		// even if there was no socket activity.
+		for (size_t i = 0; i < streams.size(); ++i) {	
+			vector<Client *> to_process;
+			swap(streams[i]->to_process, to_process);
+			for (size_t i = 0; i < to_process.size(); ++i) {
+				process_client(to_process[i]);
+			}
+		}
+
+		// 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()) {
+			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).
+			std::map<int, Client>::iterator 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();
 		}
-		to_process.clear();
 	}
 }
 
-CubemapStateProto Server::serialize() const
+CubemapStateProto Server::serialize()
 {
+	// 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 (map<int, Client>::iterator client_it = clients.begin();
+	     client_it != clients.end();
+	     ++client_it) {
+		skip_lost_data(&client_it->second);
+	}
+
 	CubemapStateProto serialized;
 	for (map<int, Client>::const_iterator client_it = clients.begin();
 	     client_it != clients.end();
 	     ++client_it) {
 		serialized.add_clients()->MergeFrom(client_it->second.serialize());
 	}
-	for (map<string, Stream *>::const_iterator stream_it = streams.begin();
-	     stream_it != streams.end();
-	     ++stream_it) {
-		serialized.add_streams()->MergeFrom(stream_it->second->serialize());
+	for (size_t i = 0; i < streams.size(); ++i) {	
+		serialized.add_streams()->MergeFrom(streams[i]->serialize());
 	}
 	return serialized;
 }
 
+void Server::add_client_deferred(int sock)
+{
+	MutexLock lock(&queued_clients_mutex);
+	queued_add_clients.push_back(sock);
+}
+
 void Server::add_client(int sock)
 {
-	MutexLock lock(&mutex);
-	clients.insert(make_pair(sock, Client(sock)));
+	pair<map<int, Client>::iterator, bool> ret =
+		clients.insert(make_pair(sock, Client(sock)));
+	assert(ret.second == true);  // Should not already exist.
+	Client *client_ptr = &ret.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;
 	ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
-	ev.data.u64 = 0;  // Keep Valgrind happy.
-	ev.data.fd = sock;
+	ev.data.u64 = reinterpret_cast<uint64_t>(client_ptr);
 	if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
-		perror("epoll_ctl(EPOLL_CTL_ADD)");
+		log_perror("epoll_ctl(EPOLL_CTL_ADD)");
 		exit(1);
 	}
 
-	process_client(&clients[sock]);
+	process_client(client_ptr);
 }
 
 void Server::add_client_from_serialized(const ClientProto &client)
 {
 	MutexLock lock(&mutex);
-	Stream *stream = find_stream(client.stream_id());
-	clients.insert(make_pair(client.sock(), Client(client, stream)));
-	Client *client_ptr = &clients[client.sock()];
+	Stream *stream;
+	int stream_index = lookup_stream_by_url(client.url());
+	if (stream_index == -1) {
+		assert(client.state() != Client::SENDING_DATA);
+		stream = NULL;
+	} else {
+		stream = streams[stream_index];
+	}
+	pair<map<int, Client>::iterator, bool> ret =
+		clients.insert(make_pair(client.sock(), Client(client, stream)));
+	assert(ret.second == true);  // Should not already exist.
+	Client *client_ptr = &ret.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;
@@ -233,72 +270,88 @@ void Server::add_client_from_serialized(const ClientProto &client)
 		// the sleeping array again soon.
 		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)");
+		log_perror("epoll_ctl(EPOLL_CTL_ADD)");
 		exit(1);
 	}
 
-	if (client_ptr->state == Client::SENDING_DATA && 
-	    client_ptr->bytes_sent == client_ptr->stream->data_size) {
-		put_client_to_sleep(client_ptr);
+	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);
 	}
 }
 
-void Server::add_stream(const string &stream_id)
+int Server::lookup_stream_by_url(const std::string &url) const
+{
+	map<string, int>::const_iterator url_it = url_map.find(url);
+	if (url_it == url_map.end()) {
+		return -1;
+	}
+	return url_it->second;
+}
+
+int Server::add_stream(const string &url, size_t backlog_size, size_t prebuffering_bytes, Stream::Encoding encoding)
 {
 	MutexLock lock(&mutex);
-	streams.insert(make_pair(stream_id, new Stream(stream_id)));
+	url_map.insert(make_pair(url, streams.size()));
+	streams.push_back(new Stream(url, backlog_size, prebuffering_bytes, encoding));
+	return streams.size() - 1;
 }
 
-void Server::add_stream_from_serialized(const StreamProto &stream)
+int Server::add_stream_from_serialized(const StreamProto &stream, int data_fd)
 {
 	MutexLock lock(&mutex);
-	streams.insert(make_pair(stream.stream_id(), new Stream(stream)));
+	url_map.insert(make_pair(stream.url(), streams.size()));
+	streams.push_back(new Stream(stream, data_fd));
+	return streams.size() - 1;
 }
 	
-void Server::set_header(const string &stream_id, const string &header)
+void Server::set_backlog_size(int stream_index, size_t new_size)
 {
 	MutexLock lock(&mutex);
-	find_stream(stream_id)->header = header;
+	assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
+	streams[stream_index]->set_backlog_size(new_size);
+}
 
-	// If there are clients we haven't sent anything to yet, we should give
-	// them the header, so push back into the SENDING_HEADER state.
-	for (map<int, Client>::iterator client_it = clients.begin();
-	     client_it != clients.end();
-	     ++client_it) {
-		Client *client = &client_it->second;
-		if (client->state == Client::SENDING_DATA &&
-		    client->bytes_sent == 0) {
-			construct_header(client);
-		}
-	}
+void Server::set_prebuffering_bytes(int stream_index, size_t new_amount)
+{
+	MutexLock lock(&mutex);
+	assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
+	streams[stream_index]->prebuffering_bytes = new_amount;
 }
 	
-void Server::add_data(const string &stream_id, const char *data, size_t bytes)
+void Server::set_encoding(int stream_index, Stream::Encoding encoding)
 {
-	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(stream_index >= 0 && stream_index < ssize_t(streams.size()));
+	streams[stream_index]->encoding = encoding;
+}
+	
+void Server::set_header(int stream_index, const string &http_header, const string &stream_header)
+{
+	MutexLock lock(&mutex);
+	assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
+	streams[stream_index]->http_header = http_header;
+	streams[stream_index]->stream_header = stream_header;
+}
+	
+void Server::set_pacing_rate(int stream_index, uint32_t pacing_rate)
+{
+	MutexLock lock(&mutex);
+	assert(clients.empty());
+	assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
+	streams[stream_index]->pacing_rate = pacing_rate;
+}
 
-	memcpy(stream->data + pos, data, bytes);
-	wake_up_all_clients();
+void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, StreamStartSuitability suitable_for_stream_start)
+{
+	assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
+	streams[stream_index]->add_data_deferred(data, bytes, suitable_for_stream_start);
 }
 
 // See the .h file for postconditions after this function.	
@@ -320,7 +373,7 @@ read_request_again:
 			return;
 		}
 		if (ret == -1) {
-			perror("read");
+			log_perror("read");
 			close_client(client);
 			return;
 		}
@@ -330,34 +383,27 @@ read_request_again:
 			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, "[%s] Client sent overlong request!", client->remote_addr.c_str());
 			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.
 			// See if there's more data for us.
 			goto read_request_again;
-		}
-
-		if (ptr != client->request.data() + client->request.size() - 4) {
-			fprintf(stderr, "WARNING: fd %d had junk data after request!\n", client->sock);
+		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);
+
 		int error_code = parse_request(client);
 		if (error_code == 200) {
 			construct_header(client);
@@ -389,7 +435,7 @@ sending_header_or_error_again:
 
 		if (ret == -1) {
 			// Error! Postcondition #1.
-			perror("write");
+			log_perror("write");
 			close_client(client);
 			return;
 		}
@@ -412,52 +458,74 @@ sending_header_or_error_again:
 			return;
 		}
 
-		// 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.
+		// 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.
 		// This is postcondition #3.
-		client->state = Client::SENDING_DATA;
-		client->bytes_sent = client->stream->data_size;
-		put_client_to_sleep(client);
+		if (client->stream_pos == size_t(-2)) {
+			client->stream_pos = std::min<size_t>(
+			    client->stream->bytes_received - client->stream->backlog_size,
+			    0);
+			client->state = Client::SENDING_DATA;
+		} else {
+			// client->stream_pos should be -1, but it might not be,
+			// if we have clients from an older version.
+			client->stream_pos = client->stream->bytes_received;
+			client->state = Client::WAITING_FOR_KEYFRAME;
+		}
+		client->stream->put_client_to_sleep(client);
 		return;
 	}
+	case Client::WAITING_FOR_KEYFRAME: {
+		Stream *stream = client->stream;
+		if (ssize_t(client->stream_pos) > stream->last_suitable_starting_point) {
+			// 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->last_suitable_starting_point;
+		client->state = Client::PREBUFFERING;
+		// Fall through.
+	}
+	case Client::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: {
-		// 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 = client->stream;
-		size_t bytes_to_send = stream->data_size - client->bytes_sent;
+		skip_lost_data(client);
+		Stream *stream = client->stream;
+
+sending_data_again:
+		size_t bytes_to_send = stream->bytes_received - client->stream_pos;
+		assert(bytes_to_send <= stream->backlog_size);
 		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",
-			        client->sock,
-			        (long long int)(bytes_to_send - BACKLOG_SIZE));
-			client->bytes_sent = stream->data_size - BACKLOG_SIZE;
-			bytes_to_send = 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 {
+			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);
 
-			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.
@@ -466,19 +534,19 @@ sending_header_or_error_again:
 		}
 		if (ret == -1) {
 			// Error, close; postcondition #1.
-			perror("write/writev");
+			log_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.
-			put_client_to_sleep(client);
-		} else {
-			// XXX: Do we need to go another round here to explicitly
-			// get the EAGAIN?
+			stream->put_client_to_sleep(client);
+		} else if (more_data && size_t(ret) == bytes_to_send) {
+			goto sending_data_again;
 		}
 		break;
 	}
@@ -487,6 +555,23 @@ sending_header_or_error_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.
+void Server::skip_lost_data(Client *client)
+{
+	Stream *stream = client->stream;
+	if (stream == NULL) {
+		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->stream_pos = stream->bytes_received - stream->backlog_size;
+		client->bytes_lost += bytes_lost;
+		++client->num_loss_events;
+	}
+}
+
 int Server::parse_request(Client *client)
 {
 	vector<string> lines = split_lines(client->request);
@@ -494,6 +579,18 @@ int Server::parse_request(Client *client)
 		return 400;  // Bad request (empty).
 	}
 
+	// Parse the headers, for logging purposes.
+	// TODO: Case-insensitivity.
+	multimap<string, string> headers = extract_headers(lines, client->remote_addr);
+	multimap<string, string>::const_iterator referer_it = headers.find("Referer");
+	if (referer_it != headers.end()) {
+		client->referer = referer_it->second;
+	}
+	multimap<string, string>::const_iterator 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() < 2) {
 		return 400;  // Bad request (empty).
@@ -501,12 +598,33 @@ int Server::parse_request(Client *client)
 	if (request_tokens[0] != "GET") {
 		return 400;  // Should maybe be 405 instead?
 	}
-	if (streams.count(request_tokens[1]) == 0) {
+
+	string url = request_tokens[1];
+	if (url.find("?backlog") == url.size() - 8) {
+		client->stream_pos = -2;
+		url = url.substr(0, url.size() - 8);
+	} else {
+		client->stream_pos = -1;
+	}
+
+	map<string, int>::const_iterator url_map_it = url_map.find(url);
+	if (url_map_it == url_map.end()) {
 		return 404;  // Not found.
 	}
 
-	client->stream_id = request_tokens[1];
-	client->stream = find_stream(client->stream_id);
+	Stream *stream = streams[url_map_it->second];
+	if (stream->http_header.empty()) {
+		return 503;  // Service unavailable.
+	}
+
+	client->url = request_tokens[1];
+
+	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!
@@ -514,19 +632,38 @@ int Server::parse_request(Client *client)
 
 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;
+	Stream *stream = client->stream;
+	if (stream->encoding == Stream::STREAM_ENCODING_RAW) {
+		client->header_or_error = stream->http_header +
+			"\r\n" +
+			stream->stream_header;
+	} else if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
+		client->header_or_error = stream->http_header +
+			"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));
+			client->header_or_error.append(
+				string(reinterpret_cast<char *>(&hdr), sizeof(hdr)));
+		}
+		client->header_or_error.append(stream->stream_header);
+	} else {
+		assert(false);
+	}
 
 	// Switch states.
 	client->state = Client::SENDING_HEADER;
 
 	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)");
+		log_perror("epoll_ctl(EPOLL_CTL_MOD)");
 		exit(1);
 	}
 }
@@ -543,58 +680,55 @@ void Server::construct_error(Client *client, int error_code)
 
 	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)");
+		log_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)
 {
 	if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, NULL) == -1) {
-		perror("epoll_ctl(EPOLL_CTL_DEL)");
+		log_perror("epoll_ctl(EPOLL_CTL_DEL)");
 		exit(1);
 	}
 
 	// This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
-	vector<Client *>::iterator new_end =
-		remove(sleeping_clients.begin(), sleeping_clients.end(), client);
-	sleeping_clients.erase(new_end, sleeping_clients.end());
-	
-	// Bye-bye!
-	int ret;
-	do {
-		ret = close(client->sock);
-	} while (ret == -1 && errno == EINTR);
-
-	if (ret == -1) {
-		perror("close");
+	if (client->stream != NULL) {
+		delete_from(&client->stream->sleeping_clients, client);
+		delete_from(&client->stream->to_process, client);
 	}
 
+	// Log to access_log.
+	access_log->write(client->get_stats());
+
+	// Bye-bye!
+	safe_close(client->sock);
+
 	clients.erase(client->sock);
 }
 	
-void Server::put_client_to_sleep(Client *client)
+void Server::process_queued_data()
 {
-	sleeping_clients.push_back(client);
-}
+	{
+		MutexLock lock(&queued_clients_mutex);
 
-void Server::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();
+		for (size_t i = 0; i < queued_add_clients.size(); ++i) {
+			add_client(queued_add_clients[i]);
+		}
+		queued_add_clients.clear();
+	}
+
+	for (size_t i = 0; i < streams.size(); ++i) {	
+		streams[i]->process_queued_data();
 	}
-}
-	
-Stream *Server::find_stream(const string &stream_id)
-{
-	map<string, Stream *>::iterator it = streams.find(stream_id);
-	assert(it != streams.end());
-	return it->second;
 }