Make backlog_size changeable across HUPs.
[cubemap] / server.cpp
1 #include <assert.h>
2 #include <errno.h>
3 #include <pthread.h>
4 #include <stdio.h>
5 #include <stdlib.h>
6 #include <sys/epoll.h>
7 #include <sys/sendfile.h>
8 #include <sys/socket.h>
9 #include <sys/types.h>
10 #include <time.h>
11 #include <unistd.h>
12 #include <algorithm>
13 #include <map>
14 #include <string>
15 #include <utility>
16 #include <vector>
17
18 #include "markpool.h"
19 #include "mutexlock.h"
20 #include "parse.h"
21 #include "server.h"
22 #include "state.pb.h"
23 #include "stream.h"
24
25 using namespace std;
26
27 Server::Server()
28 {
29         pthread_mutex_init(&mutex, NULL);
30         pthread_mutex_init(&queued_data_mutex, NULL);
31
32         epoll_fd = epoll_create(1024);  // Size argument is ignored.
33         if (epoll_fd == -1) {
34                 perror("epoll_fd");
35                 exit(1);
36         }
37 }
38
39 Server::~Server()
40 {
41         int ret;
42         do {
43                 ret = close(epoll_fd);
44         } while (ret == -1 && errno == EINTR);
45
46         if (ret == -1) {
47                 perror("close(epoll_fd)");
48         }
49 }
50
51 vector<ClientStats> Server::get_client_stats() const
52 {
53         vector<ClientStats> ret;
54
55         MutexLock lock(&mutex);
56         for (map<int, Client>::const_iterator client_it = clients.begin();
57              client_it != clients.end();
58              ++client_it) {
59                 ret.push_back(client_it->second.get_stats());
60         }
61         return ret;
62 }
63
64 void Server::do_work()
65 {
66         for ( ;; ) {
67                 int nfds = epoll_wait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS);
68                 if (nfds == -1 && errno == EINTR) {
69                         if (should_stop) {
70                                 return;
71                         }
72                         continue;
73                 }
74                 if (nfds == -1) {
75                         perror("epoll_wait");
76                         exit(1);
77                 }
78
79                 MutexLock lock(&mutex);  // We release the mutex between iterations.
80         
81                 process_queued_data();
82
83                 for (int i = 0; i < nfds; ++i) {
84                         int fd = events[i].data.fd;
85                         assert(clients.count(fd) != 0);
86                         Client *client = &clients[fd];
87
88                         if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
89                                 close_client(client);
90                                 continue;
91                         }
92
93                         process_client(client);
94                 }
95
96                 for (map<string, Stream *>::iterator stream_it = streams.begin();
97                      stream_it != streams.end();
98                      ++stream_it) {
99                         vector<Client *> to_process;
100                         swap(stream_it->second->to_process, to_process);
101                         for (size_t i = 0; i < to_process.size(); ++i) {
102                                 process_client(to_process[i]);
103                         }
104                 }
105
106                 if (should_stop) {
107                         return;
108                 }
109         }
110 }
111
112 CubemapStateProto Server::serialize()
113 {
114         // We don't serialize anything queued, so empty the queues.
115         process_queued_data();
116
117         CubemapStateProto serialized;
118         for (map<int, Client>::const_iterator client_it = clients.begin();
119              client_it != clients.end();
120              ++client_it) {
121                 serialized.add_clients()->MergeFrom(client_it->second.serialize());
122         }
123         for (map<string, Stream *>::const_iterator stream_it = streams.begin();
124              stream_it != streams.end();
125              ++stream_it) {
126                 serialized.add_streams()->MergeFrom(stream_it->second->serialize());
127         }
128         return serialized;
129 }
130
131 void Server::add_client_deferred(int sock)
132 {
133         MutexLock lock(&queued_data_mutex);
134         queued_add_clients.push_back(sock);
135 }
136
137 void Server::add_client(int sock)
138 {
139         clients.insert(make_pair(sock, Client(sock)));
140
141         // Start listening on data from this socket.
142         epoll_event ev;
143         ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
144         ev.data.u64 = 0;  // Keep Valgrind happy.
145         ev.data.fd = sock;
146         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
147                 perror("epoll_ctl(EPOLL_CTL_ADD)");
148                 exit(1);
149         }
150
151         process_client(&clients[sock]);
152 }
153
154 void Server::add_client_from_serialized(const ClientProto &client)
155 {
156         MutexLock lock(&mutex);
157         Stream *stream = find_stream(client.stream_id());
158         clients.insert(make_pair(client.sock(), Client(client, stream)));
159         Client *client_ptr = &clients[client.sock()];
160
161         // Start listening on data from this socket.
162         epoll_event ev;
163         if (client.state() == Client::READING_REQUEST) {
164                 ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
165         } else {
166                 // If we don't have more data for this client, we'll be putting it into
167                 // the sleeping array again soon.
168                 ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
169         }
170         ev.data.u64 = 0;  // Keep Valgrind happy.
171         ev.data.fd = client.sock();
172         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
173                 perror("epoll_ctl(EPOLL_CTL_ADD)");
174                 exit(1);
175         }
176
177         if (client_ptr->state == Client::SENDING_DATA && 
178             client_ptr->bytes_sent == client_ptr->stream->bytes_received) {
179                 client_ptr->stream->put_client_to_sleep(client_ptr);
180         } else {
181                 process_client(client_ptr);
182         }
183 }
184
185 void Server::add_stream(const string &stream_id, size_t backlog_size)
186 {
187         MutexLock lock(&mutex);
188         streams.insert(make_pair(stream_id, new Stream(stream_id, backlog_size)));
189 }
190
191 void Server::add_stream_from_serialized(const StreamProto &stream)
192 {
193         MutexLock lock(&mutex);
194         streams.insert(make_pair(stream.stream_id(), new Stream(stream)));
195 }
196         
197 void Server::set_backlog_size(const std::string &stream_id, size_t new_size)
198 {
199         MutexLock lock(&mutex);
200         assert(streams.count(stream_id) != 0);
201         streams[stream_id]->set_backlog_size(new_size);
202 }
203         
204 void Server::set_header(const string &stream_id, const string &header)
205 {
206         MutexLock lock(&mutex);
207         find_stream(stream_id)->header = header;
208
209         // If there are clients we haven't sent anything to yet, we should give
210         // them the header, so push back into the SENDING_HEADER state.
211         for (map<int, Client>::iterator client_it = clients.begin();
212              client_it != clients.end();
213              ++client_it) {
214                 Client *client = &client_it->second;
215                 if (client->state == Client::SENDING_DATA &&
216                     client->bytes_sent == 0) {
217                         construct_header(client);
218                 }
219         }
220 }
221         
222 void Server::set_mark_pool(const std::string &stream_id, MarkPool *mark_pool)
223 {
224         MutexLock lock(&mutex);
225         assert(clients.empty());
226         find_stream(stream_id)->mark_pool = mark_pool;
227 }
228
229 void Server::add_data_deferred(const string &stream_id, const char *data, size_t bytes)
230 {
231         MutexLock lock(&queued_data_mutex);
232         queued_data[stream_id].append(string(data, data + bytes));
233 }
234
235 // See the .h file for postconditions after this function.      
236 void Server::process_client(Client *client)
237 {
238         switch (client->state) {
239         case Client::READING_REQUEST: {
240 read_request_again:
241                 // Try to read more of the request.
242                 char buf[1024];
243                 int ret;
244                 do {
245                         ret = read(client->sock, buf, sizeof(buf));
246                 } while (ret == -1 && errno == EINTR);
247
248                 if (ret == -1 && errno == EAGAIN) {
249                         // No more data right now. Nothing to do.
250                         // This is postcondition #2.
251                         return;
252                 }
253                 if (ret == -1) {
254                         perror("read");
255                         close_client(client);
256                         return;
257                 }
258                 if (ret == 0) {
259                         // OK, the socket is closed.
260                         close_client(client);
261                         return;
262                 }
263
264                 RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
265         
266                 switch (status) {
267                 case RP_OUT_OF_SPACE:
268                         fprintf(stderr, "WARNING: fd %d sent overlong request!\n", client->sock);
269                         close_client(client);
270                         return;
271                 case RP_NOT_FINISHED_YET:
272                         // OK, we don't have the entire header yet. Fine; we'll get it later.
273                         // See if there's more data for us.
274                         goto read_request_again;
275                 case RP_EXTRA_DATA:
276                         fprintf(stderr, "WARNING: fd %d had junk data after request!\n", client->sock);
277                         close_client(client);
278                         return;
279                 case RP_FINISHED:
280                         break;
281                 }
282
283                 assert(status == RP_FINISHED);
284
285                 int error_code = parse_request(client);
286                 if (error_code == 200) {
287                         construct_header(client);
288                 } else {
289                         construct_error(client, error_code);
290                 }
291
292                 // We've changed states, so fall through.
293                 assert(client->state == Client::SENDING_ERROR ||
294                        client->state == Client::SENDING_HEADER);
295         }
296         case Client::SENDING_ERROR:
297         case Client::SENDING_HEADER: {
298 sending_header_or_error_again:
299                 int ret;
300                 do {
301                         ret = write(client->sock,
302                                     client->header_or_error.data() + client->header_or_error_bytes_sent,
303                                     client->header_or_error.size() - client->header_or_error_bytes_sent);
304                 } while (ret == -1 && errno == EINTR);
305
306                 if (ret == -1 && errno == EAGAIN) {
307                         // We're out of socket space, so now we're at the “low edge” of epoll's
308                         // edge triggering. epoll will tell us when there is more room, so for now,
309                         // just return.
310                         // This is postcondition #4.
311                         return;
312                 }
313
314                 if (ret == -1) {
315                         // Error! Postcondition #1.
316                         perror("write");
317                         close_client(client);
318                         return;
319                 }
320                 
321                 client->header_or_error_bytes_sent += ret;
322                 assert(client->header_or_error_bytes_sent <= client->header_or_error.size());
323
324                 if (client->header_or_error_bytes_sent < client->header_or_error.size()) {
325                         // We haven't sent all yet. Fine; go another round.
326                         goto sending_header_or_error_again;
327                 }
328
329                 // We're done sending the header or error! Clear it to release some memory.
330                 client->header_or_error.clear();
331
332                 if (client->state == Client::SENDING_ERROR) {
333                         // We're done sending the error, so now close.  
334                         // This is postcondition #1.
335                         close_client(client);
336                         return;
337                 }
338
339                 // Start sending from the end. In other words, we won't send any of the backlog,
340                 // but we'll start sending immediately as we get data.
341                 // This is postcondition #3.
342                 client->state = Client::SENDING_DATA;
343                 client->bytes_sent = client->stream->bytes_received;
344                 client->stream->put_client_to_sleep(client);
345                 return;
346         }
347         case Client::SENDING_DATA: {
348 sending_data_again:
349                 // See if there's some data we've lost. Ideally, we should drop to a block boundary,
350                 // but resync will be the mux's problem.
351                 Stream *stream = client->stream;
352                 size_t bytes_to_send = stream->bytes_received - client->bytes_sent;
353                 if (bytes_to_send == 0) {
354                         return;
355                 }
356                 if (bytes_to_send > stream->backlog_size) {
357                         fprintf(stderr, "WARNING: fd %d lost %lld bytes, maybe too slow connection\n",
358                                 client->sock,
359                                 (long long int)(bytes_to_send - stream->backlog_size));
360                         client->bytes_sent = stream->bytes_received - stream->backlog_size;
361                         bytes_to_send = stream->backlog_size;
362                 }
363
364                 // See if we need to split across the circular buffer.
365                 bool more_data = false;
366                 if ((client->bytes_sent % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
367                         bytes_to_send = stream->backlog_size - (client->bytes_sent % stream->backlog_size);
368                         more_data = true;
369                 }
370
371                 ssize_t ret;
372                 do {
373                         loff_t offset = client->bytes_sent % stream->backlog_size;
374                         ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
375                 } while (ret == -1 && errno == EINTR);
376
377                 if (ret == -1 && errno == EAGAIN) {
378                         // We're out of socket space, so return; epoll will wake us up
379                         // when there is more room.
380                         // This is postcondition #4.
381                         return;
382                 }
383                 if (ret == -1) {
384                         // Error, close; postcondition #1.
385                         perror("sendfile");
386                         close_client(client);
387                         return;
388                 }
389                 client->bytes_sent += ret;
390
391                 if (client->bytes_sent == stream->bytes_received) {
392                         // We don't have any more data for this client, so put it to sleep.
393                         // This is postcondition #3.
394                         stream->put_client_to_sleep(client);
395                 } else if (more_data) {
396                         goto sending_data_again;
397                 }
398                 break;
399         }
400         default:
401                 assert(false);
402         }
403 }
404
405 int Server::parse_request(Client *client)
406 {
407         vector<string> lines = split_lines(client->request);
408         if (lines.empty()) {
409                 return 400;  // Bad request (empty).
410         }
411
412         vector<string> request_tokens = split_tokens(lines[0]);
413         if (request_tokens.size() < 2) {
414                 return 400;  // Bad request (empty).
415         }
416         if (request_tokens[0] != "GET") {
417                 return 400;  // Should maybe be 405 instead?
418         }
419         if (streams.count(request_tokens[1]) == 0) {
420                 return 404;  // Not found.
421         }
422
423         client->stream_id = request_tokens[1];
424         client->stream = find_stream(client->stream_id);
425         if (client->stream->mark_pool != NULL) {
426                 client->fwmark = client->stream->mark_pool->get_mark();
427         } else {
428                 client->fwmark = 0;  // No mark.
429         }
430         if (setsockopt(client->sock, SOL_SOCKET, SO_MARK, &client->fwmark, sizeof(client->fwmark)) == -1) {                          
431                 if (client->fwmark != 0) {
432                         perror("setsockopt(SO_MARK)");
433                 }
434         }
435         client->request.clear();
436
437         return 200;  // OK!
438 }
439
440 void Server::construct_header(Client *client)
441 {
442         client->header_or_error = find_stream(client->stream_id)->header;
443
444         // Switch states.
445         client->state = Client::SENDING_HEADER;
446
447         epoll_event ev;
448         ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
449         ev.data.u64 = 0;  // Keep Valgrind happy.
450         ev.data.fd = client->sock;
451
452         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
453                 perror("epoll_ctl(EPOLL_CTL_MOD)");
454                 exit(1);
455         }
456 }
457         
458 void Server::construct_error(Client *client, int error_code)
459 {
460         char error[256];
461         snprintf(error, 256, "HTTP/1.0 %d Error\r\nContent-type: text/plain\r\n\r\nSomething went wrong. Sorry.\r\n",
462                 error_code);
463         client->header_or_error = error;
464
465         // Switch states.
466         client->state = Client::SENDING_ERROR;
467
468         epoll_event ev;
469         ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
470         ev.data.u64 = 0;  // Keep Valgrind happy.
471         ev.data.fd = client->sock;
472
473         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
474                 perror("epoll_ctl(EPOLL_CTL_MOD)");
475                 exit(1);
476         }
477 }
478
479 template<class T>
480 void delete_from(vector<T> *v, T elem)
481 {
482         typename vector<T>::iterator new_end = remove(v->begin(), v->end(), elem);
483         v->erase(new_end, v->end());
484 }
485         
486 void Server::close_client(Client *client)
487 {
488         if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, NULL) == -1) {
489                 perror("epoll_ctl(EPOLL_CTL_DEL)");
490                 exit(1);
491         }
492
493         // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
494         if (client->stream != NULL) {
495                 delete_from(&client->stream->sleeping_clients, client);
496                 delete_from(&client->stream->to_process, client);
497                 if (client->stream->mark_pool != NULL) {
498                         int fwmark = client->fwmark;
499                         client->stream->mark_pool->release_mark(fwmark);
500                 }
501         }
502
503         // Bye-bye!
504         int ret;
505         do {
506                 ret = close(client->sock);
507         } while (ret == -1 && errno == EINTR);
508
509         if (ret == -1) {
510                 perror("close");
511         }
512
513         clients.erase(client->sock);
514 }
515         
516 Stream *Server::find_stream(const string &stream_id)
517 {
518         map<string, Stream *>::iterator it = streams.find(stream_id);
519         assert(it != streams.end());
520         return it->second;
521 }
522
523 void Server::process_queued_data()
524 {
525         MutexLock lock(&queued_data_mutex);
526
527         for (size_t i = 0; i < queued_add_clients.size(); ++i) {
528                 add_client(queued_add_clients[i]);
529         }
530         queued_add_clients.clear();     
531         
532         for (map<string, string>::iterator queued_it = queued_data.begin();
533              queued_it != queued_data.end();
534              ++queued_it) {
535                 Stream *stream = find_stream(queued_it->first);
536                 stream->add_data(queued_it->second.data(), queued_it->second.size());
537                 stream->wake_up_all_clients();
538         }
539         queued_data.clear();
540 }