#include "accesslog.h"
#include "log.h"
-#include "markpool.h"
-#include "metacube.h"
+#include "metacube2.h"
#include "mutexlock.h"
#include "parse.h"
#include "server.h"
#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) {
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);
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);
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()) {
+ const 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).
+ 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();
+ }
}
}
void Server::add_client_deferred(int sock)
{
- MutexLock lock(&queued_data_mutex);
+ MutexLock lock(&queued_clients_mutex);
queued_add_clients.push_back(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;
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) {
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);
}
}
-int Server::lookup_stream_by_url(const std::string &url) const
+int Server::lookup_stream_by_url(const string &url) const
{
map<string, int>::const_iterator url_it = url_map.find(url);
if (url_it == url_map.end()) {
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;
}
assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
streams[stream_index]->set_backlog_size(new_size);
}
+
+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::set_encoding(int stream_index, Stream::Encoding encoding)
{
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;
- // 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);
- }
+ if (stream_header != streams[stream_index]->stream_header) {
+ // We cannot start at any of the older starting points anymore,
+ // since they'd get the wrong header for the stream (not to mention
+ // that a changed header probably means the stream restarted,
+ // which means any client starting on the old one would probably
+ // stop playing properly at the change point). Next block
+ // should be a suitable starting point (if not, something is
+ // pretty strange), so it will fill up again soon enough.
+ streams[stream_index]->suitable_starting_points.clear();
}
+ streams[stream_index]->stream_header = stream_header;
}
-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.
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.
- // This is postcondition #3.
+ Stream *stream = client->stream;
+ if (client->stream_pos == size_t(-2)) {
+ // Start sending from the beginning of the backlog.
+ client->stream_pos = min<size_t>(
+ stream->bytes_received - stream->backlog_size,
+ 0);
+ client->state = Client::SENDING_DATA;
+ goto sending_data;
+ } else if (stream->prebuffering_bytes == 0) {
+ // 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.
+ // Note that this is functionally identical to the next if branch,
+ // except that we save a binary search.
+ client->stream_pos = stream->bytes_received;
+ client->state = Client::WAITING_FOR_KEYFRAME;
+ } else {
+ // We're not going to send anything to the client before we have
+ // N bytes. However, this wait might be boring; we can just as well
+ // use it to send older data if we have it. We use lower_bound()
+ // so that we are conservative and never add extra latency over just
+ // waiting (assuming CBR or nearly so); otherwise, we could want e.g.
+ // 100 kB prebuffer but end up sending a 10 MB GOP.
+ deque<size_t>::const_iterator starting_point_it =
+ lower_bound(stream->suitable_starting_points.begin(),
+ stream->suitable_starting_points.end(),
+ stream->bytes_received - stream->prebuffering_bytes);
+ if (starting_point_it == stream->suitable_starting_points.end()) {
+ // None found. Just put us at the end, and then wait for the
+ // first keyframe to appear.
+ client->stream_pos = stream->bytes_received;
+ client->state = Client::WAITING_FOR_KEYFRAME;
+ } else {
+ client->stream_pos = *starting_point_it;
+ client->state = Client::PREBUFFERING;
+ goto prebuffering;
+ }
+ }
+ // Fall through.
+ }
+ case Client::WAITING_FOR_KEYFRAME: {
+ Stream *stream = client->stream;
+ if (stream->suitable_starting_points.empty() ||
+ client->stream_pos > stream->suitable_starting_points.back()) {
+ // 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->suitable_starting_points.back();
+ client->state = Client::PREBUFFERING;
+ // Fall through.
+ }
+ case Client::PREBUFFERING: {
+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;
- client->stream_pos = client->stream->bytes_received;
- client->stream->put_client_to_sleep(client);
- return;
+ // Fall through.
}
case Client::SENDING_DATA: {
+sending_data:
skip_lost_data(client);
Stream *stream = client->stream;
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);
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;
-
- 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);
}
}
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).
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();
"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)));
}
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.
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();