Some fixes to issues that could cause client hangs.
[cubemap] / server.cpp
index 3e639cf7455a470d45072992cb7efd20eeaf30f9..54e6b5e804d9959632b1ef70df6a53fee7f52fac 100644 (file)
@@ -26,17 +26,19 @@ using namespace std;
 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)
+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())
@@ -211,16 +213,17 @@ void Server::add_client(int sock)
 void Server::add_client_from_serialized(const ClientProto &client)
 {
        MutexLock lock(&mutex);
-       clients.insert(make_pair(client.sock(), Client(client)));
+       Stream *stream = find_stream(client.stream_id());
+       clients.insert(make_pair(client.sock(), Client(client, stream)));
 
        // Start listening on data from this socket.
        epoll_event ev;
        if (client.state() == Client::READING_REQUEST) {
-               ev.events = EPOLLIN | EPOLLRDHUP;
+               ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
        } else {
                // If we don't have more data for this client, we'll be putting it into
                // the sleeping array again soon.
-               ev.events = EPOLLOUT | EPOLLRDHUP;
+               ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
        }
        ev.data.u64 = 0;  // Keep Valgrind happy.
        ev.data.fd = client.sock();
@@ -228,6 +231,8 @@ void Server::add_client_from_serialized(const ClientProto &client)
                perror("epoll_ctl(EPOLL_CTL_ADD)");
                exit(1);
        }
+
+       process_client(&clients[client.sock()]);
 }
 
 void Server::add_stream(const string &stream_id)
@@ -282,26 +287,32 @@ void Server::add_data(const string &stream_id, const char *data, size_t bytes)
        memcpy(stream->data + pos, data, bytes);
        wake_up_all_clients();
 }
-       
+
+// See the .h file for postconditions after this function.     
 void Server::process_client(Client *client)
 {
        switch (client->state) {
        case Client::READING_REQUEST: {
+read_request_again:
                // Try to read more of the request.
                char buf[1024];
                int ret;
                do {
                        ret = read(client->sock, buf, sizeof(buf));
                } while (ret == -1 && errno == EINTR);
+
+               if (ret == -1 && errno == EAGAIN) {
+                       // No more data right now. Nothing to do.
+                       // This is postcondition #2.
+                       return;
+               }
                if (ret == -1) {
                        perror("read");
                        close_client(client);
                        return;
                }
                if (ret == 0) {
-                       // No data? This really means that we were triggered for something else than
-                       // POLLIN (which suggests a logic error in epoll).
-                       fprintf(stderr, "WARNING: fd %d returned unexpectedly 0 bytes!\n", client->sock);
+                       // OK, the socket is closed.
                        close_client(client);
                        return;
                }
@@ -324,7 +335,8 @@ void Server::process_client(Client *client)
                               "\r\n\r\n", 4));
                if (ptr == NULL) {
                        // OK, we don't have the entire header yet. Fine; we'll get it later.
-                       return;
+                       // See if there's more data for us.
+                       goto read_request_again;
                }
 
                if (ptr != client->request.data() + client->request.size() - 4) {
@@ -339,17 +351,31 @@ void Server::process_client(Client *client)
                } else {
                        construct_error(client, error_code);
                }
-               break;
+
+               // We've changed states, so fall through.
+               assert(client->state == Client::SENDING_ERROR ||
+                      client->state == Client::SENDING_HEADER);
        }
        case Client::SENDING_ERROR:
        case Client::SENDING_HEADER: {
+sending_header_or_error_again:
                int ret;
                do {
                        ret = write(client->sock,
                                    client->header_or_error.data() + client->header_or_error_bytes_sent,
                                    client->header_or_error.size() - client->header_or_error_bytes_sent);
                } while (ret == -1 && errno == EINTR);
+
+               if (ret == -1 && errno == EAGAIN) {
+                       // We're out of socket space, so now we're at the “low edge” of epoll's
+                       // edge triggering. epoll will tell us when there is more room, so for now,
+                       // just return.
+                       // This is postcondition #4.
+                       return;
+               }
+
                if (ret == -1) {
+                       // Error! Postcondition #1.
                        perror("write");
                        close_client(client);
                        return;
@@ -359,33 +385,38 @@ void Server::process_client(Client *client)
                assert(client->header_or_error_bytes_sent <= client->header_or_error.size());
 
                if (client->header_or_error_bytes_sent < client->header_or_error.size()) {
-                       // We haven't sent all yet. Fine; we'll do that later.
-                       return;
+                       // We haven't sent all yet. Fine; go another round.
+                       goto sending_header_or_error_again;
                }
 
                // We're done sending the header or error! Clear it to release some memory.
                client->header_or_error.clear();
 
                if (client->state == Client::SENDING_ERROR) {
+                       // We're done sending the error, so now close.  
+                       // This is postcondition #1.
                        close_client(client);
-               } else {
-                       // 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.
-                       client->state = Client::SENDING_DATA;
-                       client->bytes_sent = find_stream(client->stream_id)->data_size;
+                       return;
                }
-               break;
+
+               // 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.
+               client->state = Client::SENDING_DATA;
+               client->bytes_sent = client->stream->data_size;
+               sleeping_clients.push_back(client);
+               return;
        }
        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 = *find_stream(client->stream_id);
-               size_t bytes_to_send = stream.data_size - client->bytes_sent;
+               const Stream *stream = client->stream;
+               size_t bytes_to_send = stream->data_size - client->bytes_sent;
                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 = find_stream(client->stream_id)->data_size - BACKLOG_SIZE;
+                       client->bytes_sent = stream->data_size - BACKLOG_SIZE;
                        bytes_to_send = BACKLOG_SIZE;
                }
 
@@ -395,10 +426,10 @@ void Server::process_client(Client *client)
                        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_base = const_cast<char *>(stream->data + (client->bytes_sent % BACKLOG_SIZE));
                        iov[0].iov_len = bytes_first_part;
 
-                       iov[1].iov_base = const_cast<char *>(stream.data);
+                       iov[1].iov_base = const_cast<char *>(stream->data);
                        iov[1].iov_len = bytes_to_send - bytes_first_part;
 
                        do {
@@ -407,20 +438,31 @@ void Server::process_client(Client *client)
                } else {
                        do {
                                ret = write(client->sock,
-                                           stream.data + (client->bytes_sent % BACKLOG_SIZE),
+                                           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.
+                       // This is postcondition #4.
+                       return;
+               }
                if (ret == -1) {
+                       // Error, close; postcondition #1.
                        perror("write/writev");
                        close_client(client);
                        return;
                }
                client->bytes_sent += ret;
 
-               if (client->bytes_sent == stream.data_size) {
+               if (client->bytes_sent == stream->data_size) {
                        // 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?
                }
                break;
        }
@@ -448,6 +490,7 @@ int Server::parse_request(Client *client)
        }
 
        client->stream_id = request_tokens[1];
+       client->stream = find_stream(client->stream_id);
        client->request.clear();
 
        return 200;  // OK!
@@ -462,7 +505,7 @@ void Server::construct_header(Client *client)
        client->state = Client::SENDING_HEADER;
 
        epoll_event ev;
-       ev.events = EPOLLOUT | EPOLLRDHUP;
+       ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
        ev.data.u64 = 0;  // Keep Valgrind happy.
        ev.data.fd = client->sock;
 
@@ -483,7 +526,7 @@ void Server::construct_error(Client *client, int error_code)
        client->state = Client::SENDING_ERROR;
 
        epoll_event ev;
-       ev.events = EPOLLOUT | EPOLLRDHUP;
+       ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
        ev.data.u64 = 0;  // Keep Valgrind happy.
        ev.data.fd = client->sock;
 
@@ -501,8 +544,8 @@ void Server::close_client(Client *client)
        }
 
        // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
-       vector<int>::iterator new_end =
-               remove(sleeping_clients.begin(), sleeping_clients.end(), client->sock);
+       vector<Client *>::iterator new_end =
+               remove(sleeping_clients.begin(), sleeping_clients.end(), client);
        sleeping_clients.erase(new_end, sleeping_clients.end());
        
        // Bye-bye!
@@ -520,32 +563,16 @@ void Server::close_client(Client *client)
        
 void Server::put_client_to_sleep(Client *client)
 {
-       epoll_event ev;
-       ev.events = EPOLLRDHUP;
-       ev.data.u64 = 0;  // Keep Valgrind happy.
-       ev.data.fd = client->sock;
-
-       if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
-               perror("epoll_ctl(EPOLL_CTL_MOD)");
-               exit(1);
-       }
-
-       sleeping_clients.push_back(client->sock);
+       sleeping_clients.push_back(client);
 }
 
 void Server::wake_up_all_clients()
 {
-       for (unsigned i = 0; i < sleeping_clients.size(); ++i) {
-               epoll_event ev;
-               ev.events = EPOLLOUT | EPOLLRDHUP;
-               ev.data.u64 = 0;  // Keep Valgrind happy.
-               ev.data.fd = sleeping_clients[i];
-               if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, sleeping_clients[i], &ev) == -1) {
-                       perror("epoll_ctl(EPOLL_CTL_MOD)");
-                       exit(1);
-               }
+       vector<Client *> to_process;
+       swap(sleeping_clients, to_process);
+       for (unsigned i = 0; i < to_process.size(); ++i) {
+               process_client(to_process[i]);
        }
-       sleeping_clients.clear();
 }
        
 Stream *Server::find_stream(const string &stream_id)