X-Git-Url: https://git.sesse.net/?p=cubemap;a=blobdiff_plain;f=server.cpp;h=079b629dcaa9b601b31033c3a4c55626711f7662;hp=5bd0d555c18194ca112a04b21cb3d9344272722e;hb=ce7802e5d5291c414883efbad93a9d165c03f40c;hpb=9abb89bcf7940e2ada9d708f86a218a56334f68d

diff --git a/server.cpp b/server.cpp
index 5bd0d55..079b629 100644
--- a/server.cpp
+++ b/server.cpp
@@ -19,8 +19,7 @@
 
 #include "accesslog.h"
 #include "log.h"
-#include "markpool.h"
-#include "metacube.h"
+#include "metacube2.h"
 #include "mutexlock.h"
 #include "parse.h"
 #include "server.h"
@@ -28,14 +27,35 @@
 #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()
 {
 	pthread_mutex_init(&mutex, NULL);
-	pthread_mutex_init(&queued_data_mutex, NULL);
+	pthread_mutex_init(&queued_clients_mutex, NULL);
 
 	epoll_fd = epoll_create(1024);  // Size argument is ignored.
 	if (epoll_fd == -1) {
@@ -87,6 +107,7 @@ void Server::do_work()
 	
 		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.u64);
 
@@ -98,6 +119,8 @@ void Server::do_work()
 			process_client(client);
 		}
 
+		// 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);
@@ -105,6 +128,49 @@ void Server::do_work()
 				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();
+		}
 	}
 }
 
@@ -113,6 +179,19 @@ 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();
@@ -127,7 +206,7 @@ CubemapStateProto Server::serialize()
 
 void Server::add_client_deferred(int sock)
 {
-	MutexLock lock(&queued_data_mutex);
+	MutexLock lock(&queued_clients_mutex);
 	queued_add_clients.push_back(sock);
 }
 
@@ -138,6 +217,17 @@ void Server::add_client(int 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;
@@ -166,6 +256,11 @@ void Server::add_client_from_serialized(const ClientProto &client)
 	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;
 	if (client.state() == Client::READING_REQUEST) {
@@ -181,8 +276,10 @@ void Server::add_client_from_serialized(const ClientProto &client)
 		exit(1);
 	}
 
-	if (client_ptr->state == Client::SENDING_DATA && 
-	    client_ptr->stream_pos == client_ptr->stream->bytes_received) {
+	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);
@@ -198,11 +295,11 @@ int Server::lookup_stream_by_url(const std::string &url) const
 	return url_it->second;
 }
 
-int Server::add_stream(const string &url, size_t backlog_size, Stream::Encoding encoding)
+int Server::add_stream(const string &url, size_t backlog_size, size_t prebuffering_bytes, Stream::Encoding encoding)
 {
 	MutexLock lock(&mutex);
 	url_map.insert(make_pair(url, streams.size()));
-	streams.push_back(new Stream(url, backlog_size, encoding));
+	streams.push_back(new Stream(url, backlog_size, prebuffering_bytes, encoding));
 	return streams.size() - 1;
 }
 
@@ -234,33 +331,20 @@ void Server::set_header(int stream_index, const string &http_header, const strin
 	assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 	streams[stream_index]->http_header = http_header;
 	streams[stream_index]->stream_header = stream_header;
-
-	// 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->stream_pos == 0) {
-			construct_header(client);
-		}
-	}
 }
 	
-void Server::set_mark_pool(int stream_index, MarkPool *mark_pool)
+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]->mark_pool = mark_pool;
+	streams[stream_index]->pacing_rate = pacing_rate;
 }
 
-void Server::add_data_deferred(int stream_index, const char *data, size_t bytes)
+void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, StreamStartSuitability suitable_for_stream_start)
 {
-	MutexLock lock(&queued_data_mutex);
 	assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
-	streams[stream_index]->add_data_deferred(data, bytes);
+	streams[stream_index]->add_data_deferred(data, bytes, suitable_for_stream_start);
 }
 
 // See the .h file for postconditions after this function.	
@@ -367,36 +451,59 @@ 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->stream_pos = client->stream->bytes_received;
+		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.
+			stream->put_client_to_sleep(client);
+			return;
+		}
+		client->state = Client::SENDING_DATA;
+		// Fall through.
+	}
 	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.
+		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 > 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;
-			bytes_to_send = stream->backlog_size;
-
-			double loss_fraction = double(client->bytes_lost) / double(client->bytes_lost + client->bytes_sent);
-			log(WARNING, "[%s] Client lost %lld bytes (total loss: %.2f%%), maybe too slow connection",
-			        client->remote_addr.c_str(),
-			        (long long int)(bytes_lost),
-				100.0 * loss_fraction);
-		}
 
 		// See if we need to split across the circular buffer.
 		bool more_data = false;
@@ -407,7 +514,7 @@ sending_data_again:
 
 		ssize_t ret;
 		do {
-			loff_t offset = client->stream_pos % stream->backlog_size;
+			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);
 
@@ -440,6 +547,23 @@ 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.
+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);
@@ -455,21 +579,29 @@ int Server::parse_request(Client *client)
 		return 400;  // Should maybe be 405 instead?
 	}
 
-	map<string, int>::const_iterator url_map_it = url_map.find(request_tokens[1]);
+	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->url = request_tokens[1];
-	client->stream = streams[url_map_it->second];
-	if (client->stream->mark_pool != NULL) {
-		client->fwmark = client->stream->mark_pool->get_mark();
-	} else {
-		client->fwmark = 0;  // No mark.
+	Stream *stream = streams[url_map_it->second];
+	if (stream->http_header.empty()) {
+		return 503;  // Service unavailable.
 	}
-	if (setsockopt(client->sock, SOL_SOCKET, SO_MARK, &client->fwmark, sizeof(client->fwmark)) == -1) {                          
-		if (client->fwmark != 0) {
-			log_perror("setsockopt(SO_MARK)");
+
+	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();
@@ -489,10 +621,11 @@ void Server::construct_header(Client *client)
 			"Content-encoding: metacube\r\n" +
 			"\r\n";
 		if (!stream->stream_header.empty()) {
-			metacube_block_header hdr;
-			memcpy(hdr.sync, METACUBE_SYNC, sizeof(hdr.sync));
+			metacube2_block_header hdr;
+			memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
 			hdr.size = htonl(stream->stream_header.size());
-			hdr.flags = htonl(METACUBE_FLAGS_HEADER);
+			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)));
 		}
@@ -552,10 +685,6 @@ void Server::close_client(Client *client)
 	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);
-		}
 	}
 
 	// Log to access_log.
@@ -569,12 +698,14 @@ void Server::close_client(Client *client)
 	
 void Server::process_queued_data()
 {
-	MutexLock lock(&queued_data_mutex);
+	{
+		MutexLock lock(&queued_clients_mutex);
 
-	for (size_t i = 0; i < queued_add_clients.size(); ++i) {
-		add_client(queued_add_clients[i]);
+		for (size_t i = 0; i < queued_add_clients.size(); ++i) {
+			add_client(queued_add_clients[i]);
+		}
+		queued_add_clients.clear();
 	}
-	queued_add_clients.clear();	
 
 	for (size_t i = 0; i < streams.size(); ++i) {	
 		streams[i]->process_queued_data();