Fix: don't expose empty streams
[lttng-tools.git] / src / bin / lttng-relayd / main.c
index 93b08fcaa4a32cddfb4deef66682e8a1b8ae6135..adb044f1d3ebd4b6b6d58a035c6e4d9eece706b9 100644 (file)
@@ -2,6 +2,7 @@
  * Copyright (C) 2012 - Julien Desfossez <jdesfossez@efficios.com>
  *                      David Goulet <dgoulet@efficios.com>
  *               2013 - Jérémie Galarneau <jeremie.galarneau@efficios.com>
+ *               2015 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License, version 2 only,
@@ -18,6 +19,7 @@
  */
 
 #define _GNU_SOURCE
+#define _LGPL_SOURCE
 #include <getopt.h>
 #include <grp.h>
 #include <limits.h>
@@ -45,6 +47,7 @@
 #include <common/compat/poll.h>
 #include <common/compat/socket.h>
 #include <common/compat/endian.h>
+#include <common/compat/getenv.h>
 #include <common/defaults.h>
 #include <common/daemonize.h>
 #include <common/futex.h>
@@ -54,6 +57,7 @@
 #include <common/uri.h>
 #include <common/utils.h>
 #include <common/config/config.h>
+#include <urcu/rculist.h>
 
 #include "cmd.h"
 #include "ctf-trace.h"
@@ -112,6 +116,11 @@ static pthread_t dispatcher_thread;
 static pthread_t worker_thread;
 static pthread_t health_thread;
 
+/*
+ * last_relay_stream_id_lock protects last_relay_stream_id increment
+ * atomicity on 32-bit architectures.
+ */
+static pthread_mutex_t last_relay_stream_id_lock = PTHREAD_MUTEX_INITIALIZER;
 static uint64_t last_relay_stream_id;
 
 /*
@@ -126,18 +135,14 @@ static struct relay_conn_queue relay_conn_queue;
 static char *data_buffer;
 static unsigned int data_buffer_size;
 
-/* We need those values for the file/dir creation. */
-static uid_t relayd_uid;
-static gid_t relayd_gid;
-
 /* Global relay stream hash table. */
 struct lttng_ht *relay_streams_ht;
 
 /* Global relay viewer stream hash table. */
 struct lttng_ht *viewer_streams_ht;
 
-/* Global hash table that stores relay index object. */
-struct lttng_ht *indexes_ht;
+/* Global relay sessions hash table. */
+struct lttng_ht *sessions_ht;
 
 /* Relayd health monitoring */
 struct health_app *health_relayd;
@@ -161,8 +166,7 @@ static const char *config_ignore_options[] = { "help", "config" };
 /*
  * usage function on stderr
  */
-static
-void usage(void)
+static void usage(void)
 {
        fprintf(stderr, "Usage: %s OPTIONS\n\nOptions:\n", progname);
        fprintf(stderr, "  -h, --help                Display this usage.\n");
@@ -183,8 +187,7 @@ void usage(void)
  *
  * Return 0 on success else a negative value.
  */
-static
-int set_option(int opt, const char *arg, const char *optname)
+static int set_option(int opt, const char *arg, const char *optname)
 {
        int ret;
 
@@ -196,33 +199,48 @@ int set_option(int opt, const char *arg, const char *optname)
                }
                break;
        case 'C':
-               ret = uri_parse(arg, &control_uri);
-               if (ret < 0) {
-                       ERR("Invalid control URI specified");
-                       goto end;
-               }
-               if (control_uri->port == 0) {
-                       control_uri->port = DEFAULT_NETWORK_CONTROL_PORT;
+               if (lttng_is_setuid_setgid()) {
+                       WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
+                               "-C, --control-port");
+               } else {
+                       ret = uri_parse(arg, &control_uri);
+                       if (ret < 0) {
+                               ERR("Invalid control URI specified");
+                               goto end;
+                       }
+                       if (control_uri->port == 0) {
+                               control_uri->port = DEFAULT_NETWORK_CONTROL_PORT;
+                       }
                }
                break;
        case 'D':
-               ret = uri_parse(arg, &data_uri);
-               if (ret < 0) {
-                       ERR("Invalid data URI specified");
-                       goto end;
-               }
-               if (data_uri->port == 0) {
-                       data_uri->port = DEFAULT_NETWORK_DATA_PORT;
+               if (lttng_is_setuid_setgid()) {
+                       WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
+                               "-D, -data-port");
+               } else {
+                       ret = uri_parse(arg, &data_uri);
+                       if (ret < 0) {
+                               ERR("Invalid data URI specified");
+                               goto end;
+                       }
+                       if (data_uri->port == 0) {
+                               data_uri->port = DEFAULT_NETWORK_DATA_PORT;
+                       }
                }
                break;
        case 'L':
-               ret = uri_parse(arg, &live_uri);
-               if (ret < 0) {
-                       ERR("Invalid live URI specified");
-                       goto end;
-               }
-               if (live_uri->port == 0) {
-                       live_uri->port = DEFAULT_NETWORK_VIEWER_PORT;
+               if (lttng_is_setuid_setgid()) {
+                       WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
+                               "-L, -live-port");
+               } else {
+                       ret = uri_parse(arg, &live_uri);
+                       if (ret < 0) {
+                               ERR("Invalid live URI specified");
+                               goto end;
+                       }
+                       if (live_uri->port == 0) {
+                               live_uri->port = DEFAULT_NETWORK_VIEWER_PORT;
+                       }
                }
                break;
        case 'd':
@@ -232,18 +250,33 @@ int set_option(int opt, const char *arg, const char *optname)
                opt_background = 1;
                break;
        case 'g':
-               tracing_group_name = strdup(arg);
-               tracing_group_name_override = 1;
+               if (lttng_is_setuid_setgid()) {
+                       WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
+                               "-g, --group");
+               } else {
+                       tracing_group_name = strdup(arg);
+                       if (tracing_group_name == NULL) {
+                               ret = -errno;
+                               PERROR("strdup");
+                               goto end;
+                       }
+                       tracing_group_name_override = 1;
+               }
                break;
        case 'h':
                usage();
                exit(EXIT_FAILURE);
        case 'o':
-               ret = asprintf(&opt_output_path, "%s", arg);
-               if (ret < 0) {
-                       ret = -errno;
-                       PERROR("asprintf opt_output_path");
-                       goto end;
+               if (lttng_is_setuid_setgid()) {
+                       WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
+                               "-o, --output");
+               } else {
+                       ret = asprintf(&opt_output_path, "%s", arg);
+                       if (ret < 0) {
+                               ret = -errno;
+                               PERROR("asprintf opt_output_path");
+                               goto end;
+                       }
                }
                break;
        case 'v':
@@ -251,7 +284,10 @@ int set_option(int opt, const char *arg, const char *optname)
                if (arg) {
                        lttng_opt_verbose = config_parse_value(arg);
                } else {
-                       lttng_opt_verbose += 1;
+                       /* Only 3 level of verbosity (-vvv). */
+                       if (lttng_opt_verbose < 3) {
+                               lttng_opt_verbose += 1;
+                       }
                }
                break;
        default:
@@ -273,8 +309,7 @@ end:
  * See config_entry_handler_cb comment in common/config/config.h for the
  * return value conventions.
  */
-static
-int config_entry_handler(const struct config_entry *entry, void *unused)
+static int config_entry_handler(const struct config_entry *entry, void *unused)
 {
        int ret = 0, i;
 
@@ -297,9 +332,9 @@ int config_entry_handler(const struct config_entry *entry, void *unused)
                }
 
                /*
-                * If the option takes no argument on the command line, we have to
-                * check if the value is "true". We support non-zero numeric values,
-                * true, on and yes.
+                * If the option takes no argument on the command line,
+                * we have to check if the value is "true". We support
+                * non-zero numeric values, true, on and yes.
                 */
                if (!long_options[i].has_arg) {
                        ret = config_parse_value(entry->value);
@@ -324,10 +359,9 @@ end:
        return ret;
 }
 
-static
-int set_options(int argc, char **argv)
+static int set_options(int argc, char **argv)
 {
-       int c, ret = 0, option_index = 0;
+       int c, ret = 0, option_index = 0, retval = 0;
        int orig_optopt = optopt, orig_optind = optind;
        char *default_address, *optstring;
        const char *config_path = NULL;
@@ -335,7 +369,7 @@ int set_options(int argc, char **argv)
        optstring = utils_generate_optstring(long_options,
                        sizeof(long_options) / sizeof(struct option));
        if (!optstring) {
-               ret = -ENOMEM;
+               retval = -ENOMEM;
                goto exit;
        }
 
@@ -344,15 +378,20 @@ int set_options(int argc, char **argv)
        while ((c = getopt_long(argc, argv, optstring, long_options,
                                        &option_index)) != -1) {
                if (c == '?') {
-                       ret = -EINVAL;
+                       retval = -EINVAL;
                        goto exit;
                } else if (c != 'f') {
                        continue;
                }
 
-               config_path = utils_expand_path(optarg);
-               if (!config_path) {
-                       ERR("Failed to resolve path: %s", optarg);
+               if (lttng_is_setuid_setgid()) {
+                       WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
+                               "-f, --config");
+               } else {
+                       config_path = utils_expand_path(optarg);
+                       if (!config_path) {
+                               ERR("Failed to resolve path: %s", optarg);
+                       }
                }
        }
 
@@ -361,8 +400,8 @@ int set_options(int argc, char **argv)
        if (ret) {
                if (ret > 0) {
                        ERR("Invalid configuration option at line %i", ret);
-                       ret = -1;
                }
+               retval = -1;
                goto exit;
        }
 
@@ -377,6 +416,7 @@ int set_options(int argc, char **argv)
 
                ret = set_option(c, optarg, long_options[option_index].name);
                if (ret < 0) {
+                       retval = -1;
                        goto exit;
                }
        }
@@ -388,6 +428,7 @@ int set_options(int argc, char **argv)
                        DEFAULT_NETWORK_CONTROL_PORT);
                if (ret < 0) {
                        PERROR("asprintf default data address");
+                       retval = -1;
                        goto exit;
                }
 
@@ -395,6 +436,7 @@ int set_options(int argc, char **argv)
                free(default_address);
                if (ret < 0) {
                        ERR("Invalid control URI specified");
+                       retval = -1;
                        goto exit;
                }
        }
@@ -404,6 +446,7 @@ int set_options(int argc, char **argv)
                        DEFAULT_NETWORK_DATA_PORT);
                if (ret < 0) {
                        PERROR("asprintf default data address");
+                       retval = -1;
                        goto exit;
                }
 
@@ -411,6 +454,7 @@ int set_options(int argc, char **argv)
                free(default_address);
                if (ret < 0) {
                        ERR("Invalid data URI specified");
+                       retval = -1;
                        goto exit;
                }
        }
@@ -420,6 +464,7 @@ int set_options(int argc, char **argv)
                        DEFAULT_NETWORK_VIEWER_PORT);
                if (ret < 0) {
                        PERROR("asprintf default viewer control address");
+                       retval = -1;
                        goto exit;
                }
 
@@ -427,23 +472,43 @@ int set_options(int argc, char **argv)
                free(default_address);
                if (ret < 0) {
                        ERR("Invalid viewer control URI specified");
+                       retval = -1;
                        goto exit;
                }
        }
 
 exit:
        free(optstring);
-       return ret;
+       return retval;
+}
+
+static void print_global_objects(void)
+{
+       rcu_register_thread();
+
+       print_viewer_streams();
+       print_relay_streams();
+       print_sessions();
+
+       rcu_unregister_thread();
 }
 
 /*
  * Cleanup the daemon
  */
-static
-void cleanup(void)
+static void relayd_cleanup(void)
 {
+       print_global_objects();
+
        DBG("Cleaning up");
 
+       if (viewer_streams_ht)
+               lttng_ht_destroy(viewer_streams_ht);
+       if (relay_streams_ht)
+               lttng_ht_destroy(relay_streams_ht);
+       if (sessions_ht)
+               lttng_ht_destroy(sessions_ht);
+
        /* free the dynamically allocated opt_output_path */
        free(opt_output_path);
 
@@ -462,49 +527,61 @@ void cleanup(void)
 /*
  * Write to writable pipe used to notify a thread.
  */
-static
-int notify_thread_pipe(int wpipe)
+static int notify_thread_pipe(int wpipe)
 {
        ssize_t ret;
 
        ret = lttng_write(wpipe, "!", 1);
        if (ret < 1) {
                PERROR("write poll pipe");
+               goto end;
        }
-
+       ret = 0;
+end:
        return ret;
 }
 
-static void notify_health_quit_pipe(int *pipe)
+static int notify_health_quit_pipe(int *pipe)
 {
        ssize_t ret;
 
        ret = lttng_write(pipe[1], "4", 1);
        if (ret < 1) {
                PERROR("write relay health quit");
+               goto end;
        }
+       ret = 0;
+end:
+       return ret;
 }
 
 /*
- * Stop all threads by closing the thread quit pipe.
+ * Stop all relayd and relayd-live threads.
  */
-static
-void stop_threads(void)
+int lttng_relay_stop_threads(void)
 {
-       int ret;
+       int retval = 0;
 
        /* Stopping all threads */
        DBG("Terminating all threads");
-       ret = notify_thread_pipe(thread_quit_pipe[1]);
-       if (ret < 0) {
+       if (notify_thread_pipe(thread_quit_pipe[1])) {
                ERR("write error on thread quit pipe");
+               retval = -1;
        }
 
-       notify_health_quit_pipe(health_quit_pipe);
+       if (notify_health_quit_pipe(health_quit_pipe)) {
+               ERR("write error on health quit pipe");
+       }
 
        /* Dispatch thread */
        CMM_STORE_SHARED(dispatch_thread_exit, 1);
        futex_nto1_wake(&relay_conn_queue.futex);
+
+       if (relayd_live_stop()) {
+               ERR("Error stopping live threads");
+               retval = -1;
+       }
+       return retval;
 }
 
 /*
@@ -513,8 +590,7 @@ void stop_threads(void)
  * Simply stop all worker threads, leaving main() return gracefully after
  * joining all threads and calling cleanup().
  */
-static
-void sighandler(int sig)
+static void sighandler(int sig)
 {
        switch (sig) {
        case SIGPIPE:
@@ -522,11 +598,15 @@ void sighandler(int sig)
                return;
        case SIGINT:
                DBG("SIGINT caught");
-               stop_threads();
+               if (lttng_relay_stop_threads()) {
+                       ERR("Error stopping threads");
+               }
                break;
        case SIGTERM:
                DBG("SIGTERM caught");
-               stop_threads();
+               if (lttng_relay_stop_threads()) {
+                       ERR("Error stopping threads");
+               }
                break;
        case SIGUSR1:
                CMM_STORE_SHARED(recv_child_signal, 1);
@@ -540,8 +620,7 @@ void sighandler(int sig)
  * Setup signal handler for :
  *             SIGINT, SIGTERM, SIGPIPE
  */
-static
-int set_signal_handler(void)
+static int set_signal_handler(void)
 {
        int ret = 0;
        struct sigaction sa;
@@ -595,8 +674,7 @@ void lttng_relay_notify_ready(void)
  *
  * Return -1 on error or 0 if all pipes are created.
  */
-static
-int init_thread_quit_pipe(void)
+static int init_thread_quit_pipe(void)
 {
        int ret;
 
@@ -608,8 +686,7 @@ int init_thread_quit_pipe(void)
 /*
  * Create a poll set with O_CLOEXEC and add the thread quit pipe to the set.
  */
-static
-int create_thread_poll_set(struct lttng_poll_event *events, int size)
+static int create_thread_poll_set(struct lttng_poll_event *events, int size)
 {
        int ret;
 
@@ -640,8 +717,7 @@ error:
  *
  * Return 1 if it was triggered else 0;
  */
-static
-int check_thread_quit_pipe(int fd, uint32_t events)
+static int check_thread_quit_pipe(int fd, uint32_t events)
 {
        if (fd == thread_quit_pipe[0] && (events & LPOLLIN)) {
                return 1;
@@ -653,8 +729,7 @@ int check_thread_quit_pipe(int fd, uint32_t events)
 /*
  * Create and init socket from uri.
  */
-static
-struct lttcomm_sock *relay_init_sock(struct lttng_uri *uri)
+static struct lttcomm_sock *relay_socket_create(struct lttng_uri *uri)
 {
        int ret;
        struct lttcomm_sock *sock = NULL;
@@ -691,64 +766,10 @@ error:
        return NULL;
 }
 
-/*
- * Return nonzero if stream needs to be closed.
- */
-static
-int close_stream_check(struct relay_stream *stream)
-{
-       if (stream->close_flag && stream->prev_seq == stream->last_net_seq_num) {
-               /*
-                * We are about to close the stream so set the data pending flag to 1
-                * which will make the end data pending command skip the stream which
-                * is now closed and ready. Note that after proceeding to a file close,
-                * the written file is ready for reading.
-                */
-               stream->data_pending_check_done = 1;
-               return 1;
-       }
-       return 0;
-}
-
-static void try_close_stream(struct relay_session *session,
-               struct relay_stream *stream)
-{
-       int ret;
-       struct ctf_trace *ctf_trace;
-
-       assert(session);
-       assert(stream);
-
-       if (!close_stream_check(stream)) {
-               /* Can't close it, not ready for that. */
-               goto end;
-       }
-
-       ctf_trace = ctf_trace_find_by_path(session->ctf_traces_ht,
-                       stream->path_name);
-       assert(ctf_trace);
-
-       pthread_mutex_lock(&session->viewer_ready_lock);
-       ctf_trace->invalid_flag = 1;
-       pthread_mutex_unlock(&session->viewer_ready_lock);
-
-       ret = stream_close(session, stream);
-       if (ret || session->snapshot) {
-               /* Already close thus the ctf trace is being or has been destroyed. */
-               goto end;
-       }
-
-       ctf_trace_try_destroy(session, ctf_trace);
-
-end:
-       return;
-}
-
 /*
  * This thread manages the listening for new connections on the network
  */
-static
-void *relay_thread_listener(void *data)
+static void *relay_thread_listener(void *data)
 {
        int i, ret, pollfd, err = -1;
        uint32_t revents, nb_fd;
@@ -761,18 +782,19 @@ void *relay_thread_listener(void *data)
 
        health_code_update();
 
-       control_sock = relay_init_sock(control_uri);
+       control_sock = relay_socket_create(control_uri);
        if (!control_sock) {
                goto error_sock_control;
        }
 
-       data_sock = relay_init_sock(data_uri);
+       data_sock = relay_socket_create(data_uri);
        if (!data_sock) {
                goto error_sock_relay;
        }
 
        /*
-        * Pass 3 as size here for the thread quit pipe, control and data socket.
+        * Pass 3 as size here for the thread quit pipe, control and
+        * data socket.
         */
        ret = create_thread_poll_set(&events, 3);
        if (ret < 0) {
@@ -826,6 +848,14 @@ restart:
                        revents = LTTNG_POLL_GETEV(&events, i);
                        pollfd = LTTNG_POLL_GETFD(&events, i);
 
+                       if (!revents) {
+                               /*
+                                * No activity for this FD (poll
+                                * implementation).
+                                */
+                               continue;
+                       }
+
                        /* Thread quit pipe has been closed. Killing thread. */
                        ret = check_thread_quit_pipe(pollfd, revents);
                        if (ret) {
@@ -838,33 +868,30 @@ restart:
                                goto error;
                        } else if (revents & LPOLLIN) {
                                /*
-                                * Get allocated in this thread, enqueued to a global queue,
-                                * dequeued and freed in the worker thread.
+                                * A new connection is requested, therefore a
+                                * sessiond/consumerd connection is allocated in
+                                * this thread, enqueued to a global queue and
+                                * dequeued (and freed) in the worker thread.
                                 */
                                int val = 1;
                                struct relay_connection *new_conn;
                                struct lttcomm_sock *newsock;
-
-                               new_conn = connection_create();
-                               if (!new_conn) {
-                                       goto error;
-                               }
+                               enum connection_type type;
 
                                if (pollfd == data_sock->fd) {
-                                       new_conn->type = RELAY_DATA;
+                                       type = RELAY_DATA;
                                        newsock = data_sock->ops->accept(data_sock);
                                        DBG("Relay data connection accepted, socket %d",
                                                        newsock->fd);
                                } else {
                                        assert(pollfd == control_sock->fd);
-                                       new_conn->type = RELAY_CONTROL;
+                                       type = RELAY_CONTROL;
                                        newsock = control_sock->ops->accept(control_sock);
                                        DBG("Relay control connection accepted, socket %d",
                                                        newsock->fd);
                                }
                                if (!newsock) {
                                        PERROR("accepting sock");
-                                       connection_free(new_conn);
                                        goto error;
                                }
 
@@ -873,17 +900,22 @@ restart:
                                if (ret < 0) {
                                        PERROR("setsockopt inet");
                                        lttcomm_destroy_sock(newsock);
-                                       connection_free(new_conn);
                                        goto error;
                                }
-                               new_conn->sock = newsock;
+                               new_conn = connection_create(newsock, type);
+                               if (!new_conn) {
+                                       lttcomm_destroy_sock(newsock);
+                                       goto error;
+                               }
 
                                /* Enqueue request for the dispatcher thread. */
-                               cds_wfq_enqueue(&relay_conn_queue.queue, &new_conn->qnode);
+                               cds_wfcq_enqueue(&relay_conn_queue.head, &relay_conn_queue.tail,
+                                                &new_conn->qnode);
 
                                /*
-                                * Wake the dispatch queue futex. Implicit memory barrier with
-                                * the exchange in cds_wfq_enqueue.
+                                * Wake the dispatch queue futex.
+                                * Implicit memory barrier with the
+                                * exchange in cds_wfcq_enqueue.
                                 */
                                futex_nto1_wake(&relay_conn_queue.futex);
                        }
@@ -918,19 +950,18 @@ error_sock_control:
        }
        health_unregister(health_relayd);
        DBG("Relay listener thread cleanup complete");
-       stop_threads();
+       lttng_relay_stop_threads();
        return NULL;
 }
 
 /*
  * This thread manages the dispatching of the requests to worker threads
  */
-static
-void *relay_thread_dispatcher(void *data)
+static void *relay_thread_dispatcher(void *data)
 {
        int err = -1;
        ssize_t ret;
-       struct cds_wfq_node *node;
+       struct cds_wfcq_node *node;
        struct relay_connection *new_conn = NULL;
 
        DBG("[thread] Relay dispatcher started");
@@ -953,7 +984,8 @@ void *relay_thread_dispatcher(void *data)
                        health_code_update();
 
                        /* Dequeue commands */
-                       node = cds_wfq_dequeue_blocking(&relay_conn_queue.queue);
+                       node = cds_wfcq_dequeue_blocking(&relay_conn_queue.head,
+                                                        &relay_conn_queue.tail);
                        if (node == NULL) {
                                DBG("Woken up but nothing in the relay command queue");
                                /* Continue thread execution */
@@ -964,14 +996,15 @@ void *relay_thread_dispatcher(void *data)
                        DBG("Dispatching request waiting on sock %d", new_conn->sock->fd);
 
                        /*
-                        * Inform worker thread of the new request. This call is blocking
-                        * so we can be assured that the data will be read at some point in
-                        * time or wait to the end of the world :)
+                        * Inform worker thread of the new request. This
+                        * call is blocking so we can be assured that
+                        * the data will be read at some point in time
+                        * or wait to the end of the world :)
                         */
                        ret = lttng_write(relay_conn_pipe[1], &new_conn, sizeof(new_conn));
                        if (ret < 0) {
                                PERROR("write connection pipe");
-                               connection_destroy(new_conn);
+                               connection_put(new_conn);
                                goto error;
                        }
                } while (node != NULL);
@@ -993,76 +1026,31 @@ error_testpoint:
        }
        health_unregister(health_relayd);
        DBG("Dispatch thread dying");
-       stop_threads();
+       lttng_relay_stop_threads();
        return NULL;
 }
 
-static void try_close_streams(struct relay_session *session)
-{
-       struct ctf_trace *ctf_trace;
-       struct lttng_ht_iter iter;
-
-       assert(session);
-
-       pthread_mutex_lock(&session->viewer_ready_lock);
-       rcu_read_lock();
-       cds_lfht_for_each_entry(session->ctf_traces_ht->ht, &iter.iter, ctf_trace,
-                       node.node) {
-               struct relay_stream *stream;
-
-               /* Close streams. */
-               cds_list_for_each_entry(stream, &ctf_trace->stream_list, trace_list) {
-                       stream_close(session, stream);
-               }
-
-               ctf_trace->invalid_flag = 1;
-               ctf_trace_try_destroy(session, ctf_trace);
-       }
-       rcu_read_unlock();
-       pthread_mutex_unlock(&session->viewer_ready_lock);
-}
-
-/*
- * Try to destroy a session within a connection.
- */
-static void destroy_session(struct relay_session *session,
-               struct lttng_ht *sessions_ht)
-{
-       assert(session);
-       assert(sessions_ht);
-
-       /* Indicate that this session can be destroyed from now on. */
-       session->close_flag = 1;
-
-       try_close_streams(session);
-
-       /*
-        * This will try to delete and destroy the session if no viewer is attached
-        * to it meaning the refcount is down to zero.
-        */
-       session_try_destroy(sessions_ht, session);
-}
-
 /*
- * Copy index data from the control port to a given index object.
+ * Set index data from the control port to a given index object.
  */
-static void copy_index_control_data(struct relay_index *index,
+static int set_index_control_data(struct relay_index *index,
                struct lttcomm_relayd_index *data)
 {
-       assert(index);
-       assert(data);
+       struct ctf_packet_index index_data;
 
        /*
-        * The index on disk is encoded in big endian, so we don't need to convert
-        * the data received on the network. The data_offset value is NEVER
-        * modified here and is updated by the data thread.
+        * The index on disk is encoded in big endian, so we don't need
+        * to convert the data received on the network. The data_offset
+        * value is NEVER modified here and is updated by the data
+        * thread.
         */
-       index->index_data.packet_size = data->packet_size;
-       index->index_data.content_size = data->content_size;
-       index->index_data.timestamp_begin = data->timestamp_begin;
-       index->index_data.timestamp_end = data->timestamp_end;
-       index->index_data.events_discarded = data->events_discarded;
-       index->index_data.stream_id = data->stream_id;
+       index_data.packet_size = data->packet_size;
+       index_data.content_size = data->content_size;
+       index_data.timestamp_begin = data->timestamp_begin;
+       index_data.timestamp_end = data->timestamp_end;
+       index_data.events_discarded = data->events_discarded;
+       index_data.stream_id = data->stream_id;
+       return relay_index_set_data(index, &index_data);
 }
 
 /*
@@ -1070,31 +1058,22 @@ static void copy_index_control_data(struct relay_index *index,
  *
  * On success, send back the session id or else return a negative value.
  */
-static
-int relay_create_session(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_create_session(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret = 0, send_ret;
        struct relay_session *session;
        struct lttcomm_relayd_status_session reply;
+       char session_name[NAME_MAX];
+       char hostname[HOST_NAME_MAX];
+       uint32_t live_timer = 0;
+       bool snapshot = false;
 
-       assert(recv_hdr);
-       assert(conn);
+       memset(session_name, 0, NAME_MAX);
+       memset(hostname, 0, HOST_NAME_MAX);
 
        memset(&reply, 0, sizeof(reply));
 
-       session = session_create();
-       if (!session) {
-               ret = -1;
-               goto error;
-       }
-       session->minor = conn->minor;
-       session->major = conn->major;
-       conn->session_id = session->id;
-       conn->session = session;
-
-       reply.session_id = htobe64(session->id);
-
        switch (conn->minor) {
        case 1:
        case 2:
@@ -1102,13 +1081,26 @@ int relay_create_session(struct lttcomm_relayd_hdr *recv_hdr,
                break;
        case 4: /* LTTng sessiond 2.4 */
        default:
-               ret = cmd_create_session_2_4(conn, session);
+               ret = cmd_create_session_2_4(conn, session_name,
+                       hostname, &live_timer, &snapshot);
+       }
+       if (ret < 0) {
+               goto send_reply;
        }
 
-       lttng_ht_add_unique_u64(conn->sessions_ht, &session->session_n);
+       session = session_create(session_name, hostname, live_timer,
+                       snapshot, conn->major, conn->minor);
+       if (!session) {
+               ret = -1;
+               goto send_reply;
+       }
+       assert(!conn->session);
+       conn->session = session;
        DBG("Created session %" PRIu64, session->id);
 
-error:
+       reply.session_id = htobe64(session->id);
+
+send_reply:
        if (ret < 0) {
                reply.ret_code = htobe32(LTTNG_ERR_FATAL);
        } else {
@@ -1128,47 +1120,47 @@ error:
  * When we have received all the streams and the metadata for a channel,
  * we make them visible to the viewer threads.
  */
-static
-void set_viewer_ready_flag(struct relay_connection *conn)
+static void publish_connection_local_streams(struct relay_connection *conn)
 {
-       struct relay_stream *stream, *tmp_stream;
+       struct relay_stream *stream;
+       struct relay_session *session = conn->session;
 
-       pthread_mutex_lock(&conn->session->viewer_ready_lock);
-       cds_list_for_each_entry_safe(stream, tmp_stream, &conn->recv_head,
-                       recv_list) {
-               stream->viewer_ready = 1;
-               cds_list_del(&stream->recv_list);
+       /*
+        * We publish all streams belonging to a session atomically wrt
+        * session lock.
+        */
+       pthread_mutex_lock(&session->lock);
+       rcu_read_lock();
+       cds_list_for_each_entry_rcu(stream, &session->recv_list,
+                       recv_node) {
+               stream_publish(stream);
        }
-       pthread_mutex_unlock(&conn->session->viewer_ready_lock);
-       return;
-}
-
-/*
- * Add a recv handle node to the connection recv list with the given stream
- * handle. A new node is allocated thus must be freed when the node is deleted
- * from the list.
- */
-static void queue_stream(struct relay_stream *stream,
-               struct relay_connection *conn)
-{
-       assert(conn);
-       assert(stream);
+       rcu_read_unlock();
 
-       cds_list_add(&stream->recv_list, &conn->recv_head);
+       /*
+        * Inform the viewer that there are new streams in the session.
+        */
+       if (session->viewer_attached) {
+               uatomic_set(&session->new_streams, 1);
+       }
+       pthread_mutex_unlock(&session->lock);
 }
 
 /*
  * relay_add_stream: allocate a new stream for a session
  */
-static
-int relay_add_stream(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_add_stream(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
-       int ret, send_ret;
+       int ret;
+       ssize_t send_ret;
        struct relay_session *session = conn->session;
        struct relay_stream *stream = NULL;
        struct lttcomm_relayd_status_stream reply;
-       struct ctf_trace *trace;
+       struct ctf_trace *trace = NULL;
+       uint64_t stream_handle = -1ULL;
+       char *path_name = NULL, *channel_name = NULL;
+       uint64_t tracefile_size = 0, tracefile_count = 0;
 
        if (!session || conn->version_check_done == 0) {
                ERR("Trying to add a stream before version check");
@@ -1176,102 +1168,48 @@ int relay_add_stream(struct lttcomm_relayd_hdr *recv_hdr,
                goto end_no_session;
        }
 
-       stream = zmalloc(sizeof(struct relay_stream));
-       if (stream == NULL) {
-               PERROR("relay stream zmalloc");
-               ret = -1;
-               goto end_no_session;
-       }
-
-       switch (conn->minor) {
-       case 1: /* LTTng sessiond 2.1 */
-               ret = cmd_recv_stream_2_1(conn, stream);
+       switch (session->minor) {
+       case 1: /* LTTng sessiond 2.1. Allocates path_name and channel_name. */
+               ret = cmd_recv_stream_2_1(conn, &path_name,
+                       &channel_name);
                break;
-       case 2: /* LTTng sessiond 2.2 */
+       case 2: /* LTTng sessiond 2.2. Allocates path_name and channel_name. */
        default:
-               ret = cmd_recv_stream_2_2(conn, stream);
+               ret = cmd_recv_stream_2_2(conn, &path_name,
+                       &channel_name, &tracefile_size, &tracefile_count);
                break;
        }
        if (ret < 0) {
-               goto err_free_stream;
+               goto send_reply;
        }
 
-       rcu_read_lock();
-       stream->stream_handle = ++last_relay_stream_id;
-       stream->prev_seq = -1ULL;
-       stream->session_id = session->id;
-       stream->index_fd = -1;
-       stream->read_index_fd = -1;
-       lttng_ht_node_init_u64(&stream->node, stream->stream_handle);
-       pthread_mutex_init(&stream->lock, NULL);
-
-       ret = utils_mkdir_recursive(stream->path_name, S_IRWXU | S_IRWXG);
-       if (ret < 0) {
-               ERR("relay creating output directory");
-               goto end;
-       }
-
-       /*
-        * No need to use run_as API here because whatever we receives, the relayd
-        * uses its own credentials for the stream files.
-        */
-       ret = utils_create_stream_file(stream->path_name, stream->channel_name,
-                       stream->tracefile_size, 0, relayd_uid, relayd_gid, NULL);
-       if (ret < 0) {
-               ERR("Create output file");
-               goto end;
-       }
-       stream->fd = ret;
-       if (stream->tracefile_size) {
-               DBG("Tracefile %s/%s_0 created", stream->path_name, stream->channel_name);
-       } else {
-               DBG("Tracefile %s/%s created", stream->path_name, stream->channel_name);
-       }
-
-       trace = ctf_trace_find_by_path(session->ctf_traces_ht, stream->path_name);
+       trace = ctf_trace_get_by_path_or_create(session, path_name);
        if (!trace) {
-               trace = ctf_trace_create(stream->path_name);
-               if (!trace) {
-                       ret = -1;
-                       goto end;
-               }
-               ctf_trace_add(session->ctf_traces_ht, trace);
+               goto send_reply;
        }
-       ctf_trace_get_ref(trace);
+       /* This stream here has one reference on the trace. */
 
-       if (!strncmp(stream->channel_name, DEFAULT_METADATA_NAME, NAME_MAX)) {
-               stream->metadata_flag = 1;
-               /* Assign quick reference to the metadata stream in the trace. */
-               trace->metadata_stream = stream;
-       }
+       pthread_mutex_lock(&last_relay_stream_id_lock);
+       stream_handle = ++last_relay_stream_id;
+       pthread_mutex_unlock(&last_relay_stream_id_lock);
 
-       /*
-        * Add the stream in the recv list of the connection. Once the end stream
-        * message is received, this list is emptied and streams are set with the
-        * viewer ready flag.
-        */
-       queue_stream(stream, conn);
+       /* We pass ownership of path_name and channel_name. */
+       stream = stream_create(trace, stream_handle, path_name,
+                       channel_name, tracefile_size, tracefile_count);
+       path_name = NULL;
+       channel_name = NULL;
 
        /*
-        * Both in the ctf_trace object and the global stream ht since the data
-        * side of the relayd does not have the concept of session.
+        * Streams are the owners of their trace. Reference to trace is
+        * kept within stream_create().
         */
-       lttng_ht_add_unique_u64(relay_streams_ht, &stream->node);
-       cds_list_add_tail(&stream->trace_list, &trace->stream_list);
-
-       session->stream_count++;
-
-       DBG("Relay new stream added %s with ID %" PRIu64, stream->channel_name,
-                       stream->stream_handle);
+       ctf_trace_put(trace);
 
-end:
+send_reply:
        memset(&reply, 0, sizeof(reply));
-       reply.handle = htobe64(stream->stream_handle);
-       /* send the session id to the client or a negative return code on error */
-       if (ret < 0) {
+       reply.handle = htobe64(stream_handle);
+       if (!stream) {
                reply.ret_code = htobe32(LTTNG_ERR_UNK);
-               /* stream was not properly added to the ht, so free it */
-               free(stream);
        } else {
                reply.ret_code = htobe32(LTTNG_OK);
        }
@@ -1280,25 +1218,19 @@ end:
                        sizeof(struct lttcomm_relayd_status_stream), 0);
        if (send_ret < 0) {
                ERR("Relay sending stream id");
-               ret = send_ret;
+               ret = (int) send_ret;
        }
-       rcu_read_unlock();
 
 end_no_session:
-       return ret;
-
-err_free_stream:
-       free(stream->path_name);
-       free(stream->channel_name);
-       free(stream);
+       free(path_name);
+       free(channel_name);
        return ret;
 }
 
 /*
  * relay_close_stream: close a specific stream
  */
-static
-int relay_close_stream(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_close_stream(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret, send_ret;
@@ -1328,25 +1260,36 @@ int relay_close_stream(struct lttcomm_relayd_hdr *recv_hdr,
                goto end_no_session;
        }
 
-       rcu_read_lock();
-       stream = stream_find_by_id(relay_streams_ht,
-                       be64toh(stream_info.stream_id));
+       stream = stream_get_by_id(be64toh(stream_info.stream_id));
        if (!stream) {
                ret = -1;
-               goto end_unlock;
+               goto end;
        }
-
+       pthread_mutex_lock(&stream->lock);
+       stream->closed = true;
        stream->last_net_seq_num = be64toh(stream_info.last_net_seq_num);
-       stream->close_flag = 1;
-       session->stream_count--;
-       assert(session->stream_count >= 0);
-
-       /* Check if we can close it or else the data will do it. */
-       try_close_stream(session, stream);
+       if (stream->is_metadata) {
+               struct relay_viewer_stream *vstream;
 
-end_unlock:
-       rcu_read_unlock();
+               vstream = viewer_stream_get_by_id(stream->stream_handle);
+               if (vstream) {
+                       if (vstream->metadata_sent == stream->metadata_received) {
+                               /*
+                                * Since all the metadata has been sent to the
+                                * viewer and that we have a request to close
+                                * its stream, we can safely teardown the
+                                * corresponding metadata viewer stream.
+                                */
+                               viewer_stream_put(vstream);
+                       }
+                       /* Put local reference. */
+                       viewer_stream_put(vstream);
+               }
+       }
+       pthread_mutex_unlock(&stream->lock);
+       stream_put(stream);
 
+end:
        memset(&reply, 0, sizeof(reply));
        if (ret < 0) {
                reply.ret_code = htobe32(LTTNG_ERR_UNK);
@@ -1367,8 +1310,7 @@ end_no_session:
 /*
  * relay_unknown_command: send -1 if received unknown command
  */
-static
-void relay_unknown_command(struct relay_connection *conn)
+static void relay_unknown_command(struct relay_connection *conn)
 {
        struct lttcomm_relayd_generic_reply reply;
        int ret;
@@ -1386,8 +1328,7 @@ void relay_unknown_command(struct relay_connection *conn)
  * relay_start: send an acknowledgment to the client to tell if we are
  * ready to receive data. We are ready if a session is established.
  */
-static
-int relay_start(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_start(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret = htobe32(LTTNG_OK);
@@ -1441,10 +1382,9 @@ end:
 }
 
 /*
- * relay_recv_metadata: receive the metada for the session.
+ * relay_recv_metadata: receive the metadata for the session.
  */
-static
-int relay_recv_metadata(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_recv_metadata(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret = htobe32(LTTNG_OK);
@@ -1453,7 +1393,6 @@ int relay_recv_metadata(struct lttcomm_relayd_hdr *recv_hdr,
        struct lttcomm_relayd_metadata_payload *metadata_struct;
        struct relay_stream *metadata_stream;
        uint64_t data_size, payload_size;
-       struct ctf_trace *ctf_trace;
 
        if (!session) {
                ERR("Metadata sent before version check");
@@ -1498,38 +1437,37 @@ int relay_recv_metadata(struct lttcomm_relayd_hdr *recv_hdr,
        }
        metadata_struct = (struct lttcomm_relayd_metadata_payload *) data_buffer;
 
-       rcu_read_lock();
-       metadata_stream = stream_find_by_id(relay_streams_ht,
-                       be64toh(metadata_struct->stream_id));
+       metadata_stream = stream_get_by_id(be64toh(metadata_struct->stream_id));
        if (!metadata_stream) {
                ret = -1;
-               goto end_unlock;
+               goto end;
        }
 
-       size_ret = lttng_write(metadata_stream->fd, metadata_struct->payload,
+       pthread_mutex_lock(&metadata_stream->lock);
+
+       size_ret = lttng_write(metadata_stream->stream_fd->fd, metadata_struct->payload,
                        payload_size);
        if (size_ret < payload_size) {
                ERR("Relay error writing metadata on file");
                ret = -1;
-               goto end_unlock;
+               goto end_put;
        }
 
-       ret = write_padding_to_file(metadata_stream->fd,
+       ret = write_padding_to_file(metadata_stream->stream_fd->fd,
                        be32toh(metadata_struct->padding_size));
        if (ret < 0) {
-               goto end_unlock;
+               goto end_put;
        }
 
-       ctf_trace = ctf_trace_find_by_path(session->ctf_traces_ht,
-                       metadata_stream->path_name);
-       assert(ctf_trace);
-       ctf_trace->metadata_received +=
+       metadata_stream->metadata_received +=
                payload_size + be32toh(metadata_struct->padding_size);
+       DBG2("Relay metadata written. Updated metadata_received %" PRIu64,
+               metadata_stream->metadata_received);
 
-       DBG2("Relay metadata written");
+end_put:
+       pthread_mutex_unlock(&metadata_stream->lock);
+       stream_put(metadata_stream);
 
-end_unlock:
-       rcu_read_unlock();
 end:
        return ret;
 }
@@ -1537,15 +1475,12 @@ end:
 /*
  * relay_send_version: send relayd version number
  */
-static
-int relay_send_version(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_send_version(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret;
        struct lttcomm_relayd_version reply, msg;
 
-       assert(conn);
-
        conn->version_check_done = 1;
 
        /* Get version from the other side. */
@@ -1569,7 +1504,7 @@ int relay_send_version(struct lttcomm_relayd_hdr *recv_hdr,
        if (reply.major != be32toh(msg.major)) {
                DBG("Incompatible major versions (%u vs %u), deleting session",
                                reply.major, be32toh(msg.major));
-               destroy_session(conn->session, conn->sessions_ht);
+               connection_put(conn);
                ret = 0;
                goto end;
        }
@@ -1600,8 +1535,7 @@ end:
 /*
  * Check for data pending for a given stream id from the session daemon.
  */
-static
-int relay_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        struct relay_session *session = conn->session;
@@ -1635,13 +1569,14 @@ int relay_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
        stream_id = be64toh(msg.stream_id);
        last_net_seq_num = be64toh(msg.last_net_seq_num);
 
-       rcu_read_lock();
-       stream = stream_find_by_id(relay_streams_ht, stream_id);
+       stream = stream_get_by_id(stream_id);
        if (stream == NULL) {
                ret = -1;
-               goto end_unlock;
+               goto end;
        }
 
+       pthread_mutex_lock(&stream->lock);
+
        DBG("Data pending for stream id %" PRIu64 " prev_seq %" PRIu64
                        " and last_seq %" PRIu64, stream_id, stream->prev_seq,
                        last_net_seq_num);
@@ -1655,11 +1590,11 @@ int relay_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
                ret = 1;
        }
 
-       /* Pending check is now done. */
-       stream->data_pending_check_done = 1;
+       stream->data_pending_check_done = true;
+       pthread_mutex_unlock(&stream->lock);
 
-end_unlock:
-       rcu_read_unlock();
+       stream_put(stream);
+end:
 
        memset(&reply, 0, sizeof(reply));
        reply.ret_code = htobe32(ret);
@@ -1675,18 +1610,17 @@ end_no_session:
 /*
  * Wait for the control socket to reach a quiescent state.
  *
- * Note that for now, when receiving this command from the session daemon, this
- * means that every subsequent commands or data received on the control socket
- * has been handled. So, this is why we simply return OK here.
+ * Note that for now, when receiving this command from the session
+ * daemon, this means that every subsequent commands or data received on
+ * the control socket has been handled. So, this is why we simply return
+ * OK here.
  */
-static
-int relay_quiescent_control(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_quiescent_control(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret;
        uint64_t stream_id;
        struct relay_stream *stream;
-       struct lttng_ht_iter iter;
        struct lttcomm_relayd_quiescent_control msg;
        struct lttcomm_relayd_generic_reply reply;
 
@@ -1712,19 +1646,16 @@ int relay_quiescent_control(struct lttcomm_relayd_hdr *recv_hdr,
        }
 
        stream_id = be64toh(msg.stream_id);
-
-       rcu_read_lock();
-       cds_lfht_for_each_entry(relay_streams_ht->ht, &iter.iter, stream,
-                       node.node) {
-               if (stream->stream_handle == stream_id) {
-                       stream->data_pending_check_done = 1;
-                       DBG("Relay quiescent control pending flag set to %" PRIu64,
-                                       stream_id);
-                       break;
-               }
-       }
-       rcu_read_unlock();
-
+       stream = stream_get_by_id(stream_id);
+       if (!stream) {
+               goto reply;
+       }
+       pthread_mutex_lock(&stream->lock);
+       stream->data_pending_check_done = true;
+       pthread_mutex_unlock(&stream->lock);
+       DBG("Relay quiescent control pending flag set to %" PRIu64, stream_id);
+       stream_put(stream);
+reply:
        memset(&reply, 0, sizeof(reply));
        reply.ret_code = htobe32(LTTNG_OK);
        ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
@@ -1737,14 +1668,13 @@ end_no_session:
 }
 
 /*
- * Initialize a data pending command. This means that a client is about to ask
- * for data pending for each stream he/she holds. Simply iterate over all
- * streams of a session and set the data_pending_check_done flag.
+ * Initialize a data pending command. This means that a consumer is about
+ * to ask for data pending for each stream it holds. Simply iterate over
+ * all streams of a session and set the data_pending_check_done flag.
  *
  * This command returns to the client a LTTNG_OK code.
  */
-static
-int relay_begin_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_begin_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret;
@@ -1781,18 +1711,25 @@ int relay_begin_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
        session_id = be64toh(msg.session_id);
 
        /*
-        * Iterate over all streams to set the begin data pending flag. For now, the
-        * streams are indexed by stream handle so we have to iterate over all
-        * streams to find the one associated with the right session_id.
+        * Iterate over all streams to set the begin data pending flag.
+        * For now, the streams are indexed by stream handle so we have
+        * to iterate over all streams to find the one associated with
+        * the right session_id.
         */
        rcu_read_lock();
        cds_lfht_for_each_entry(relay_streams_ht->ht, &iter.iter, stream,
                        node.node) {
-               if (stream->session_id == session_id) {
-                       stream->data_pending_check_done = 0;
+               if (!stream_get(stream)) {
+                       continue;
+               }
+               if (stream->trace->session->id == session_id) {
+                       pthread_mutex_lock(&stream->lock);
+                       stream->data_pending_check_done = false;
+                       pthread_mutex_unlock(&stream->lock);
                        DBG("Set begin data pending flag to stream %" PRIu64,
                                        stream->stream_handle);
                }
+               stream_put(stream);
        }
        rcu_read_unlock();
 
@@ -1810,16 +1747,15 @@ end_no_session:
 }
 
 /*
- * End data pending command. This will check, for a given session id, if each
- * stream associated with it has its data_pending_check_done flag set. If not,
- * this means that the client lost track of the stream but the data is still
- * being streamed on our side. In this case, we inform the client that data is
- * inflight.
+ * End data pending command. This will check, for a given session id, if
+ * each stream associated with it has its data_pending_check_done flag
+ * set. If not, this means that the client lost track of the stream but
+ * the data is still being streamed on our side. In this case, we inform
+ * the client that data is in flight.
  *
  * Return to the client if there is data in flight or not with a ret_code.
  */
-static
-int relay_end_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_end_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret;
@@ -1830,9 +1766,6 @@ int relay_end_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
        uint64_t session_id;
        uint32_t is_data_inflight = 0;
 
-       assert(recv_hdr);
-       assert(conn);
-
        DBG("End data pending command");
 
        if (!conn->session || conn->version_check_done == 0) {
@@ -1856,17 +1789,33 @@ int relay_end_data_pending(struct lttcomm_relayd_hdr *recv_hdr,
 
        session_id = be64toh(msg.session_id);
 
-       /* Iterate over all streams to see if the begin data pending flag is set. */
+       /*
+        * Iterate over all streams to see if the begin data pending
+        * flag is set.
+        */
        rcu_read_lock();
        cds_lfht_for_each_entry(relay_streams_ht->ht, &iter.iter, stream,
                        node.node) {
-               if (stream->session_id == session_id &&
-                               !stream->data_pending_check_done && !stream->terminated_flag) {
-                       is_data_inflight = 1;
-                       DBG("Data is still in flight for stream %" PRIu64,
-                                       stream->stream_handle);
-                       break;
+               if (!stream_get(stream)) {
+                       continue;
+               }
+               if (stream->trace->session->id != session_id) {
+                       stream_put(stream);
+                       continue;
                }
+               pthread_mutex_lock(&stream->lock);
+               if (!stream->data_pending_check_done) {
+                       if (!stream->closed || !(((int64_t) (stream->prev_seq - stream->last_net_seq_num)) >= 0)) {
+                               is_data_inflight = 1;
+                               DBG("Data is still in flight for stream %" PRIu64,
+                                               stream->stream_handle);
+                               pthread_mutex_unlock(&stream->lock);
+                               stream_put(stream);
+                               break;
+                       }
+               }
+               pthread_mutex_unlock(&stream->lock);
+               stream_put(stream);
        }
        rcu_read_unlock();
 
@@ -1888,14 +1837,13 @@ end_no_session:
  *
  * Return 0 on success else a negative value.
  */
-static
-int relay_recv_index(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_recv_index(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
-       int ret, send_ret, index_created = 0;
+       int ret, send_ret;
        struct relay_session *session = conn->session;
        struct lttcomm_relayd_index index_info;
-       struct relay_index *index, *wr_index = NULL;
+       struct relay_index *index;
        struct lttcomm_relayd_generic_reply reply;
        struct relay_stream *stream;
        uint64_t net_seq_num;
@@ -1925,69 +1873,66 @@ int relay_recv_index(struct lttcomm_relayd_hdr *recv_hdr,
 
        net_seq_num = be64toh(index_info.net_seq_num);
 
-       rcu_read_lock();
-       stream = stream_find_by_id(relay_streams_ht,
-                       be64toh(index_info.relay_stream_id));
+       stream = stream_get_by_id(be64toh(index_info.relay_stream_id));
        if (!stream) {
+               ERR("stream_get_by_id not found");
                ret = -1;
-               goto end_rcu_unlock;
+               goto end;
        }
+       pthread_mutex_lock(&stream->lock);
 
        /* Live beacon handling */
        if (index_info.packet_size == 0) {
-               DBG("Received live beacon for stream %" PRIu64, stream->stream_handle);
+               DBG("Received live beacon for stream %" PRIu64,
+                               stream->stream_handle);
 
                /*
-                * Only flag a stream inactive when it has already received data.
+                * Only flag a stream inactive when it has already
+                * received data and no indexes are in flight.
                 */
-               if (stream->total_index_received > 0) {
-                       stream->beacon_ts_end = be64toh(index_info.timestamp_end);
+               if (stream->total_index_received > 0
+                               && stream->indexes_in_flight == 0) {
+                       stream->beacon_ts_end =
+                               be64toh(index_info.timestamp_end);
                }
                ret = 0;
-               goto end_rcu_unlock;
+               goto end_stream_put;
        } else {
                stream->beacon_ts_end = -1ULL;
        }
 
-       index = relay_index_find(stream->stream_handle, net_seq_num);
+       if (stream->ctf_stream_id == -1ULL) {
+               stream->ctf_stream_id = be64toh(index_info.stream_id);
+       }
+       index = relay_index_get_by_id_or_create(stream, net_seq_num);
        if (!index) {
-               /* A successful creation will add the object to the HT. */
-               index = relay_index_create(stream->stream_handle, net_seq_num);
-               if (!index) {
-                       goto end_rcu_unlock;
-               }
-               index_created = 1;
+               ret = -1;
+               ERR("relay_index_get_by_id_or_create index NULL");
+               goto end_stream_put;
        }
-
-       copy_index_control_data(index, &index_info);
-
-       if (index_created) {
-               /*
-                * Try to add the relay index object to the hash table. If an object
-                * already exist, destroy back the index created, set the data in this
-                * object and write it on disk.
-                */
-               relay_index_add(index, &wr_index);
-               if (wr_index) {
-                       copy_index_control_data(wr_index, &index_info);
-                       free(index);
-               }
-       } else {
-               /* The index already exists so write it on disk. */
-               wr_index = index;
+       if (set_index_control_data(index, &index_info)) {
+               ERR("set_index_control_data error");
+               relay_index_put(index);
+               ret = -1;
+               goto end_stream_put;
        }
-
-       /* Do we have a writable ready index to write on disk. */
-       if (wr_index) {
-               ret = relay_index_write(wr_index->fd, wr_index);
-               if (ret < 0) {
-                       goto end_rcu_unlock;
-               }
+       ret = relay_index_try_flush(index);
+       if (ret == 0) {
                stream->total_index_received++;
+       } else if (ret > 0) {
+               /* no flush. */
+               ret = 0;
+       } else {
+               ERR("relay_index_try_flush error %d", ret);
+               relay_index_put(index);
+               ret = -1;
        }
 
-end_rcu_unlock:
-       rcu_read_unlock();
+end_stream_put:
+       pthread_mutex_unlock(&stream->lock);
+       stream_put(stream);
+
+end:
 
        memset(&reply, 0, sizeof(reply));
        if (ret < 0) {
@@ -2010,8 +1955,7 @@ end_no_session:
  *
  * Return 0 on success else a negative value.
  */
-static
-int relay_streams_sent(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_streams_sent(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret, send_ret;
@@ -2028,17 +1972,10 @@ int relay_streams_sent(struct lttcomm_relayd_hdr *recv_hdr,
        }
 
        /*
-        * Flag every pending stream in the connection recv list that they are
-        * ready to be used by the viewer.
+        * Publish every pending stream in the connection recv list which are
+        * now ready to be used by the viewer.
         */
-       set_viewer_ready_flag(conn);
-
-       /*
-        * Inform the viewer that there are new streams in the session.
-        */
-       if (conn->session->viewer_refcount) {
-               uatomic_set(&conn->session->new_streams, 1);
-       }
+       publish_connection_local_streams(conn);
 
        memset(&reply, 0, sizeof(reply));
        reply.ret_code = htobe32(LTTNG_OK);
@@ -2058,8 +1995,7 @@ end_no_session:
 /*
  * Process the commands received on the control socket
  */
-static
-int relay_process_control(struct lttcomm_relayd_hdr *recv_hdr,
+static int relay_process_control(struct lttcomm_relayd_hdr *recv_hdr,
                struct relay_connection *conn)
 {
        int ret = 0;
@@ -2116,90 +2052,91 @@ end:
 /*
  * Handle index for a data stream.
  *
- * RCU read side lock MUST be acquired.
+ * Called with the stream lock held.
  *
  * Return 0 on success else a negative value.
  */
 static int handle_index_data(struct relay_stream *stream, uint64_t net_seq_num,
                int rotate_index)
 {
-       int ret = 0, index_created = 0;
-       uint64_t stream_id, data_offset;
-       struct relay_index *index, *wr_index = NULL;
-
-       assert(stream);
+       int ret = 0;
+       uint64_t data_offset;
+       struct relay_index *index;
 
-       stream_id = stream->stream_handle;
        /* Get data offset because we are about to update the index. */
        data_offset = htobe64(stream->tracefile_size_current);
 
+       DBG("handle_index_data: stream %" PRIu64 " data offset %" PRIu64,
+               stream->stream_handle, stream->tracefile_size_current);
+
        /*
-        * Lookup for an existing index for that stream id/sequence number. If on
-        * exists, the control thread already received the data for it thus we need
-        * to write it on disk.
+        * Lookup for an existing index for that stream id/sequence
+        * number. If it exists, the control thread has already received the
+        * data for it, thus we need to write it to disk.
         */
-       index = relay_index_find(stream_id, net_seq_num);
+       index = relay_index_get_by_id_or_create(stream, net_seq_num);
        if (!index) {
-               /* A successful creation will add the object to the HT. */
-               index = relay_index_create(stream_id, net_seq_num);
-               if (!index) {
-                       ret = -1;
-                       goto error;
-               }
-               index_created = 1;
+               ret = -1;
+               goto end;
        }
 
-       if (rotate_index || stream->index_fd < 0) {
-               index->to_close_fd = stream->index_fd;
-               ret = index_create_file(stream->path_name, stream->channel_name,
-                               relayd_uid, relayd_gid, stream->tracefile_size,
-                               stream->tracefile_count_current);
-               if (ret < 0) {
-                       /* This will close the stream's index fd if one. */
-                       relay_index_free_safe(index);
-                       goto error;
+       if (rotate_index || !stream->index_fd) {
+               int fd;
+
+               /* Put ref on previous index_fd. */
+               if (stream->index_fd) {
+                       stream_fd_put(stream->index_fd);
+                       stream->index_fd = NULL;
                }
-               stream->index_fd = ret;
-       }
-       index->fd = stream->index_fd;
-       index->index_data.offset = data_offset;
 
-       if (index_created) {
-               /*
-                * Try to add the relay index object to the hash table. If an object
-                * already exist, destroy back the index created and set the data.
-                */
-               relay_index_add(index, &wr_index);
-               if (wr_index) {
-                       /* Copy back data from the created index. */
-                       wr_index->fd = index->fd;
-                       wr_index->to_close_fd = index->to_close_fd;
-                       wr_index->index_data.offset = data_offset;
-                       free(index);
+               fd = index_create_file(stream->path_name, stream->channel_name,
+                               -1, -1, stream->tracefile_size,
+                               stream->current_tracefile_id);
+               if (fd < 0) {
+                       ret = -1;
+                       /* Put self-ref for this index due to error. */
+                       relay_index_put(index);
+                       goto end;
+               }
+               stream->index_fd = stream_fd_create(fd);
+               if (!stream->index_fd) {
+                       ret = -1;
+                       if (close(fd)) {
+                               PERROR("Error closing FD %d", fd);
+                       }
+                       /* Put self-ref for this index due to error. */
+                       relay_index_put(index);
+                       /* Will put the local ref. */
+                       goto end;
                }
-       } else {
-               /* The index already exists so write it on disk. */
-               wr_index = index;
        }
 
-       /* Do we have a writable ready index to write on disk. */
-       if (wr_index) {
-               ret = relay_index_write(wr_index->fd, wr_index);
-               if (ret < 0) {
-                       goto error;
-               }
-               stream->total_index_received++;
+       if (relay_index_set_fd(index, stream->index_fd, data_offset)) {
+               ret = -1;
+               /* Put self-ref for this index due to error. */
+               relay_index_put(index);
+               goto end;
        }
 
-error:
+       ret = relay_index_try_flush(index);
+       if (ret == 0) {
+               stream->total_index_received++;
+       } else if (ret > 0) {
+               /* No flush. */
+               ret = 0;
+       } else {
+               /* Put self-ref for this index due to error. */
+               relay_index_put(index);
+               ret = -1;
+       }
+end:
        return ret;
 }
 
 /*
  * relay_process_data: Process the data received on the data socket
  */
-static
-int relay_process_data(struct relay_connection *conn)
+static int relay_process_data(struct relay_connection *conn)
 {
        int ret = 0, rotate_index = 0;
        ssize_t size_ret;
@@ -2209,8 +2146,7 @@ int relay_process_data(struct relay_connection *conn)
        uint64_t net_seq_num;
        uint32_t data_size;
        struct relay_session *session;
-
-       assert(conn);
+       bool new_stream = false;
 
        ret = conn->sock->ops->recvmsg(conn->sock, &data_hdr,
                        sizeof(struct lttcomm_relayd_data_hdr), 0);
@@ -2226,17 +2162,12 @@ int relay_process_data(struct relay_connection *conn)
        }
 
        stream_id = be64toh(data_hdr.stream_id);
-
-       rcu_read_lock();
-       stream = stream_find_by_id(relay_streams_ht, stream_id);
+       stream = stream_get_by_id(stream_id);
        if (!stream) {
                ret = -1;
-               goto end_rcu_unlock;
+               goto end;
        }
-
-       session = session_find_by_id(conn->sessions_ht, stream->session_id);
-       assert(session);
-
+       session = stream->trace->session;
        data_size = be32toh(data_hdr.data_size);
        if (data_buffer_size < data_size) {
                char *tmp_data_ptr;
@@ -2246,7 +2177,7 @@ int relay_process_data(struct relay_connection *conn)
                        ERR("Allocating data buffer");
                        free(data_buffer);
                        ret = -1;
-                       goto end_rcu_unlock;
+                       goto end_stream_put;
                }
                data_buffer = tmp_data_ptr;
                data_buffer_size = data_size;
@@ -2264,115 +2195,104 @@ int relay_process_data(struct relay_connection *conn)
                        DBG("Socket %d did an orderly shutdown", conn->sock->fd);
                }
                ret = -1;
-               goto end_rcu_unlock;
+               goto end_stream_put;
        }
 
+       pthread_mutex_lock(&stream->lock);
+
        /* Check if a rotation is needed. */
        if (stream->tracefile_size > 0 &&
                        (stream->tracefile_size_current + data_size) >
                        stream->tracefile_size) {
-               struct relay_viewer_stream *vstream;
                uint64_t new_id;
 
-               new_id = (stream->tracefile_count_current + 1) %
+               new_id = (stream->current_tracefile_id + 1) %
                        stream->tracefile_count;
                /*
-                * When we wrap-around back to 0, we start overwriting old
-                * trace data.
+                * Move viewer oldest available data position forward if
+                * we are overwriting a tracefile.
                 */
-               if (!stream->tracefile_overwrite && new_id == 0) {
-                       stream->tracefile_overwrite = 1;
-               }
-               pthread_mutex_lock(&stream->viewer_stream_rotation_lock);
-               if (stream->tracefile_overwrite) {
+               if (new_id == stream->oldest_tracefile_id) {
                        stream->oldest_tracefile_id =
                                (stream->oldest_tracefile_id + 1) %
                                stream->tracefile_count;
                }
-               vstream = viewer_stream_find_by_id(stream->stream_handle);
-               if (vstream) {
-                       /*
-                        * The viewer is reading a file about to be
-                        * overwritten. Close the FDs it is
-                        * currently using and let it handle the fault.
-                        */
-                       if (vstream->tracefile_count_current == new_id) {
-                               pthread_mutex_lock(&vstream->overwrite_lock);
-                               vstream->abort_flag = 1;
-                               pthread_mutex_unlock(&vstream->overwrite_lock);
-                               DBG("Streaming side setting abort_flag on stream %s_%lu\n",
-                                               stream->channel_name, new_id);
-                       } else if (vstream->tracefile_count_current ==
-                                       stream->tracefile_count_current) {
-                               /*
-                                * The reader and writer were in the
-                                * same trace file, inform the viewer
-                                * that no new index will ever be added
-                                * to this file.
-                                */
-                               vstream->close_write_flag = 1;
-                       }
-               }
-               ret = utils_rotate_stream_file(stream->path_name, stream->channel_name,
-                               stream->tracefile_size, stream->tracefile_count,
-                               relayd_uid, relayd_gid, stream->fd,
-                               &(stream->tracefile_count_current), &stream->fd);
-               stream->total_index_received = 0;
-               pthread_mutex_unlock(&stream->viewer_stream_rotation_lock);
+               ret = utils_rotate_stream_file(stream->path_name,
+                               stream->channel_name, stream->tracefile_size,
+                               stream->tracefile_count, -1,
+                               -1, stream->stream_fd->fd,
+                               &stream->current_tracefile_id,
+                               &stream->stream_fd->fd);
                if (ret < 0) {
                        ERR("Rotating stream output file");
-                       goto end_rcu_unlock;
+                       goto end_stream_unlock;
+               }
+               stream->current_tracefile_seq++;
+               if (stream->current_tracefile_seq
+                       - stream->oldest_tracefile_seq >=
+                               stream->tracefile_count) {
+                       stream->oldest_tracefile_seq++;
                }
-               /* Reset current size because we just perform a stream rotation. */
+               /*
+                * Reset current size because we just performed a stream
+                * rotation.
+                */
                stream->tracefile_size_current = 0;
                rotate_index = 1;
        }
 
        /*
-        * Index are handled in protocol version 2.4 and above. Also, snapshot and
-        * index are NOT supported.
+        * Index are handled in protocol version 2.4 and above. Also,
+        * snapshot and index are NOT supported.
         */
        if (session->minor >= 4 && !session->snapshot) {
                ret = handle_index_data(stream, net_seq_num, rotate_index);
                if (ret < 0) {
-                       goto end_rcu_unlock;
+                       goto end_stream_unlock;
                }
        }
 
        /* Write data to stream output fd. */
-       size_ret = lttng_write(stream->fd, data_buffer, data_size);
+       size_ret = lttng_write(stream->stream_fd->fd, data_buffer, data_size);
        if (size_ret < data_size) {
                ERR("Relay error writing data to file");
                ret = -1;
-               goto end_rcu_unlock;
+               goto end_stream_unlock;
        }
 
-       DBG2("Relay wrote %d bytes to tracefile for stream id %" PRIu64,
-                       ret, stream->stream_handle);
+       DBG2("Relay wrote %zd bytes to tracefile for stream id %" PRIu64,
+                       size_ret, stream->stream_handle);
 
-       ret = write_padding_to_file(stream->fd, be32toh(data_hdr.padding_size));
+       ret = write_padding_to_file(stream->stream_fd->fd,
+                       be32toh(data_hdr.padding_size));
        if (ret < 0) {
-               goto end_rcu_unlock;
+               goto end_stream_unlock;
+       }
+       stream->tracefile_size_current +=
+                       data_size + be32toh(data_hdr.padding_size);
+       if (stream->prev_seq == -1ULL) {
+               new_stream = true;
        }
-       stream->tracefile_size_current += data_size + be32toh(data_hdr.padding_size);
 
        stream->prev_seq = net_seq_num;
 
-       try_close_stream(session, stream);
-
-end_rcu_unlock:
-       rcu_read_unlock();
+end_stream_unlock:
+       pthread_mutex_unlock(&stream->lock);
+       if (new_stream) {
+               pthread_mutex_lock(&session->lock);
+               uatomic_set(&session->new_streams, 1);
+               pthread_mutex_unlock(&session->lock);
+       }
+end_stream_put:
+       stream_put(stream);
 end:
        return ret;
 }
 
-static
-void cleanup_connection_pollfd(struct lttng_poll_event *events, int pollfd)
+static void cleanup_connection_pollfd(struct lttng_poll_event *events, int pollfd)
 {
        int ret;
 
-       assert(events);
-
        (void) lttng_poll_del(events, pollfd);
 
        ret = close(pollfd);
@@ -2381,37 +2301,44 @@ void cleanup_connection_pollfd(struct lttng_poll_event *events, int pollfd)
        }
 }
 
-static void destroy_connection(struct lttng_ht *relay_connections_ht,
-               struct relay_connection *conn)
+static void relay_thread_close_connection(struct lttng_poll_event *events,
+               int pollfd, struct relay_connection *conn)
 {
-       assert(relay_connections_ht);
-       assert(conn);
-
-       connection_delete(relay_connections_ht, conn);
+       const char *type_str;
 
-       /* For the control socket, we try to destroy the session. */
-       if (conn->type == RELAY_CONTROL && conn->session) {
-               destroy_session(conn->session, conn->sessions_ht);
+       switch (conn->type) {
+       case RELAY_DATA:
+               type_str = "Data";
+               break;
+       case RELAY_CONTROL:
+               type_str = "Control";
+               break;
+       case RELAY_VIEWER_COMMAND:
+               type_str = "Viewer Command";
+               break;
+       case RELAY_VIEWER_NOTIFICATION:
+               type_str = "Viewer Notification";
+               break;
+       default:
+               type_str = "Unknown";
        }
-
-       connection_destroy(conn);
+       cleanup_connection_pollfd(events, pollfd);
+       connection_put(conn);
+       DBG("%s connection closed with %d", type_str, pollfd);
 }
 
 /*
  * This thread does the actual work
  */
-static
-void *relay_thread_worker(void *data)
+static void *relay_thread_worker(void *data)
 {
        int ret, err = -1, last_seen_data_fd = -1;
        uint32_t nb_fd;
-       struct relay_connection *conn;
        struct lttng_poll_event events;
        struct lttng_ht *relay_connections_ht;
        struct lttng_ht_iter iter;
        struct lttcomm_relayd_hdr recv_hdr;
-       struct relay_local_data *relay_ctx = (struct relay_local_data *) data;
-       struct lttng_ht *sessions_ht = relay_ctx->sessions_ht;
+       struct relay_connection *destroy_conn = NULL;
 
        DBG("[thread] Relay worker started");
 
@@ -2431,12 +2358,6 @@ void *relay_thread_worker(void *data)
                goto relay_connections_ht_error;
        }
 
-       /* Tables of received indexes indexed by index handle and net_seq_num. */
-       indexes_ht = lttng_ht_new(0, LTTNG_HT_TYPE_TWO_U64);
-       if (!indexes_ht) {
-               goto indexes_ht_error;
-       }
-
        ret = create_thread_poll_set(&events, 2);
        if (ret < 0) {
                goto error_poll_create;
@@ -2471,9 +2392,9 @@ restart:
                nb_fd = ret;
 
                /*
-                * Process control. The control connection is prioritised so we don't
-                * starve it with high throughout put tracing data on the data
-                * connection.
+                * Process control. The control connection is
+                * prioritized so we don't starve it with high
+                * throughput tracing data on the data connection.
                 */
                for (i = 0; i < nb_fd; i++) {
                        /* Fetch once the poll data */
@@ -2482,6 +2403,14 @@ restart:
 
                        health_code_update();
 
+                       if (!revents) {
+                               /*
+                                * No activity for this FD (poll
+                                * implementation).
+                                */
+                               continue;
+                       }
+
                        /* Thread quit pipe has been closed. Killing thread. */
                        ret = check_thread_quit_pipe(pollfd, revents);
                        if (ret) {
@@ -2495,48 +2424,43 @@ restart:
                                        ERR("Relay connection pipe error");
                                        goto error;
                                } else if (revents & LPOLLIN) {
+                                       struct relay_connection *conn;
+
                                        ret = lttng_read(relay_conn_pipe[0], &conn, sizeof(conn));
                                        if (ret < 0) {
                                                goto error;
                                        }
-                                       conn->sessions_ht = sessions_ht;
-                                       connection_init(conn);
                                        lttng_poll_add(&events, conn->sock->fd,
                                                        LPOLLIN | LPOLLRDHUP);
-                                       rcu_read_lock();
-                                       lttng_ht_add_unique_ulong(relay_connections_ht,
-                                                       &conn->sock_n);
-                                       rcu_read_unlock();
+                                       connection_ht_add(relay_connections_ht, conn);
                                        DBG("Connection socket %d added", conn->sock->fd);
                                }
                        } else {
-                               rcu_read_lock();
-                               conn = connection_find_by_sock(relay_connections_ht, pollfd);
+                               struct relay_connection *ctrl_conn;
+
+                               ctrl_conn = connection_get_by_sock(relay_connections_ht, pollfd);
                                /* If not found, there is a synchronization issue. */
-                               assert(conn);
+                               assert(ctrl_conn);
 
                                if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
-                                       cleanup_connection_pollfd(&events, pollfd);
-                                       destroy_connection(relay_connections_ht, conn);
+                                       relay_thread_close_connection(&events, pollfd, ctrl_conn);
                                        if (last_seen_data_fd == pollfd) {
                                                last_seen_data_fd = last_notdel_data_fd;
                                        }
                                } else if (revents & LPOLLIN) {
-                                       if (conn->type == RELAY_CONTROL) {
-                                               ret = conn->sock->ops->recvmsg(conn->sock, &recv_hdr,
+                                       if (ctrl_conn->type == RELAY_CONTROL) {
+                                               ret = ctrl_conn->sock->ops->recvmsg(ctrl_conn->sock, &recv_hdr,
                                                                sizeof(recv_hdr), 0);
                                                if (ret <= 0) {
                                                        /* Connection closed */
-                                                       cleanup_connection_pollfd(&events, pollfd);
-                                                       destroy_connection(relay_connections_ht, conn);
-                                                       DBG("Control connection closed with %d", pollfd);
+                                                       relay_thread_close_connection(&events, pollfd,
+                                                               ctrl_conn);
                                                } else {
-                                                       ret = relay_process_control(&recv_hdr, conn);
+                                                       ret = relay_process_control(&recv_hdr, ctrl_conn);
                                                        if (ret < 0) {
                                                                /* Clear the session on error. */
-                                                               cleanup_connection_pollfd(&events, pollfd);
-                                                               destroy_connection(relay_connections_ht, conn);
-                                                               DBG("Connection closed with %d", pollfd);
+                                                               relay_thread_close_connection(&events, pollfd,
+                                                                       ctrl_conn);
                                                        }
                                                        seen_control = 1;
                                                }
@@ -2551,7 +2475,7 @@ restart:
                                } else {
                                        ERR("Unknown poll events %u for sock %d", revents, pollfd);
                                }
-                               rcu_read_unlock();
+                               connection_put(ctrl_conn);
                        }
                }
 
@@ -2581,48 +2505,50 @@ restart:
                        /* Fetch the poll data. */
                        uint32_t revents = LTTNG_POLL_GETEV(&events, i);
                        int pollfd = LTTNG_POLL_GETFD(&events, i);
+                       struct relay_connection *data_conn;
 
                        health_code_update();
 
+                       if (!revents) {
+                               /* No activity for this FD (poll implementation). */
+                               continue;
+                       }
+
                        /* Skip the command pipe. It's handled in the first loop. */
                        if (pollfd == relay_conn_pipe[0]) {
                                continue;
                        }
 
-                       if (revents) {
-                               rcu_read_lock();
-                               conn = connection_find_by_sock(relay_connections_ht, pollfd);
-                               if (!conn) {
-                                       /* Skip it. Might be removed before. */
-                                       rcu_read_unlock();
-                                       continue;
-                               }
+                       data_conn = connection_get_by_sock(relay_connections_ht, pollfd);
+                       if (!data_conn) {
+                               /* Skip it. Might be removed before. */
+                               continue;
+                       }
 
-                               if (revents & LPOLLIN) {
-                                       if (conn->type != RELAY_DATA) {
-                                               continue;
-                                       }
+                       if (revents & LPOLLIN) {
+                               if (data_conn->type != RELAY_DATA) {
+                                       goto put_connection;
+                               }
 
-                                       ret = relay_process_data(conn);
-                                       /* Connection closed */
-                                       if (ret < 0) {
-                                               cleanup_connection_pollfd(&events, pollfd);
-                                               destroy_connection(relay_connections_ht, conn);
-                                               DBG("Data connection closed with %d", pollfd);
-                                               /*
-                                                * Every goto restart call sets the last seen fd where
-                                                * here we don't really care since we gracefully
-                                                * continue the loop after the connection is deleted.
-                                                */
-                                       } else {
-                                               /* Keep last seen port. */
-                                               last_seen_data_fd = pollfd;
-                                               rcu_read_unlock();
-                                               goto restart;
-                                       }
+                               ret = relay_process_data(data_conn);
+                               /* Connection closed */
+                               if (ret < 0) {
+                                       relay_thread_close_connection(&events, pollfd,
+                                               data_conn);
+                                       /*
+                                        * Every goto restart call sets the last seen fd where
+                                        * here we don't really care since we gracefully
+                                        * continue the loop after the connection is deleted.
+                                        */
+                               } else {
+                                       /* Keep last seen port. */
+                                       last_seen_data_fd = pollfd;
+                                       connection_put(data_conn);
+                                       goto restart;
                                }
-                               rcu_read_unlock();
                        }
+               put_connection:
+                       connection_put(data_conn);
                }
                last_seen_data_fd = -1;
        }
@@ -2632,19 +2558,23 @@ restart:
 
 exit:
 error:
-       lttng_poll_clean(&events);
-
        /* Cleanup reamaining connection object. */
        rcu_read_lock();
-       cds_lfht_for_each_entry(relay_connections_ht->ht, &iter.iter, conn,
+       cds_lfht_for_each_entry(relay_connections_ht->ht, &iter.iter,
+                       destroy_conn,
                        sock_n.node) {
                health_code_update();
-               destroy_connection(relay_connections_ht, conn);
+               /*
+                * No need to grab another ref, because we own
+                * destroy_conn.
+                */
+               relay_thread_close_connection(&events, destroy_conn->sock->fd,
+                               destroy_conn);
        }
        rcu_read_unlock();
+
+       lttng_poll_clean(&events);
 error_poll_create:
-       lttng_ht_destroy(indexes_ht);
-indexes_ht_error:
        lttng_ht_destroy(relay_connections_ht);
 relay_connections_ht_error:
        /* Close relay conn pipes */
@@ -2661,7 +2591,7 @@ error_testpoint:
        }
        health_unregister(health_relayd);
        rcu_unregister_thread();
-       stop_threads();
+       lttng_relay_stop_threads();
        return NULL;
 }
 
@@ -2683,31 +2613,35 @@ static int create_relay_conn_pipe(void)
  */
 int main(int argc, char **argv)
 {
-       int ret = 0;
+       int ret = 0, retval = 0;
        void *status;
-       struct relay_local_data *relay_ctx;
 
        /* Parse arguments */
        progname = argv[0];
-       if ((ret = set_options(argc, argv)) < 0) {
-               goto exit;
+       if (set_options(argc, argv)) {
+               retval = -1;
+               goto exit_options;
        }
 
-       if ((ret = set_signal_handler()) < 0) {
-               goto exit;
+       if (set_signal_handler()) {
+               retval = -1;
+               goto exit_options;
        }
 
        /* Try to create directory if -o, --output is specified. */
        if (opt_output_path) {
                if (*opt_output_path != '/') {
                        ERR("Please specify an absolute path for -o, --output PATH");
-                       goto exit;
+                       retval = -1;
+                       goto exit_options;
                }
 
-               ret = utils_mkdir_recursive(opt_output_path, S_IRWXU | S_IRWXG);
+               ret = utils_mkdir_recursive(opt_output_path, S_IRWXU | S_IRWXG,
+                               -1, -1);
                if (ret < 0) {
                        ERR("Unable to create %s", opt_output_path);
-                       goto exit;
+                       retval = -1;
+                       goto exit_options;
                }
        }
 
@@ -2718,7 +2652,8 @@ int main(int argc, char **argv)
                ret = lttng_daemonize(&child_ppid, &recv_child_signal,
                        !opt_background);
                if (ret < 0) {
-                       goto exit;
+                       retval = -1;
+                       goto exit_options;
                }
 
                /*
@@ -2731,170 +2666,181 @@ int main(int argc, char **argv)
                }
        }
 
-       /* Create thread quit pipe */
-       if ((ret = init_thread_quit_pipe()) < 0) {
-               goto error;
+
+       /* Initialize thread health monitoring */
+       health_relayd = health_app_create(NR_HEALTH_RELAYD_TYPES);
+       if (!health_relayd) {
+               PERROR("health_app_create error");
+               retval = -1;
+               goto exit_health_app_create;
        }
 
-       /* We need those values for the file/dir creation. */
-       relayd_uid = getuid();
-       relayd_gid = getgid();
+       /* Create thread quit pipe */
+       if (init_thread_quit_pipe()) {
+               retval = -1;
+               goto exit_init_data;
+       }
 
        /* Check if daemon is UID = 0 */
-       if (relayd_uid == 0) {
+       if (!getuid()) {
                if (control_uri->port < 1024 || data_uri->port < 1024 || live_uri->port < 1024) {
                        ERR("Need to be root to use ports < 1024");
-                       ret = -1;
-                       goto exit;
+                       retval = -1;
+                       goto exit_init_data;
                }
        }
 
        /* Setup the thread apps communication pipe. */
-       if ((ret = create_relay_conn_pipe()) < 0) {
-               goto exit;
+       if (create_relay_conn_pipe()) {
+               retval = -1;
+               goto exit_init_data;
        }
 
        /* Init relay command queue. */
-       cds_wfq_init(&relay_conn_queue.queue);
+       cds_wfcq_init(&relay_conn_queue.head, &relay_conn_queue.tail);
 
        /* Set up max poll set size */
-       lttng_poll_set_max_size();
+       if (lttng_poll_set_max_size()) {
+               retval = -1;
+               goto exit_init_data;
+       }
 
        /* Initialize communication library */
        lttcomm_init();
        lttcomm_inet_init();
 
-       relay_ctx = zmalloc(sizeof(struct relay_local_data));
-       if (!relay_ctx) {
-               PERROR("relay_ctx");
-               goto exit;
-       }
-
        /* tables of sessions indexed by session ID */
-       relay_ctx->sessions_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
-       if (!relay_ctx->sessions_ht) {
-               goto exit_relay_ctx_sessions;
+       sessions_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
+       if (!sessions_ht) {
+               retval = -1;
+               goto exit_init_data;
        }
 
        /* tables of streams indexed by stream ID */
        relay_streams_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
        if (!relay_streams_ht) {
-               goto exit_relay_ctx_streams;
+               retval = -1;
+               goto exit_init_data;
        }
 
        /* tables of streams indexed by stream ID */
        viewer_streams_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
        if (!viewer_streams_ht) {
-               goto exit_relay_ctx_viewer_streams;
-       }
-
-       /* Initialize thread health monitoring */
-       health_relayd = health_app_create(NR_HEALTH_RELAYD_TYPES);
-       if (!health_relayd) {
-               PERROR("health_app_create error");
-               goto exit_health_app_create;
+               retval = -1;
+               goto exit_init_data;
        }
 
        ret = utils_create_pipe(health_quit_pipe);
-       if (ret < 0) {
-               goto error_health_pipe;
+       if (ret) {
+               retval = -1;
+               goto exit_health_quit_pipe;
        }
 
        /* Create thread to manage the client socket */
        ret = pthread_create(&health_thread, NULL,
                        thread_manage_health, (void *) NULL);
-       if (ret != 0) {
+       if (ret) {
+               errno = ret;
                PERROR("pthread_create health");
-               goto health_error;
+               retval = -1;
+               goto exit_health_thread;
        }
 
        /* Setup the dispatcher thread */
        ret = pthread_create(&dispatcher_thread, NULL,
                        relay_thread_dispatcher, (void *) NULL);
-       if (ret != 0) {
+       if (ret) {
+               errno = ret;
                PERROR("pthread_create dispatcher");
-               goto exit_dispatcher;
+               retval = -1;
+               goto exit_dispatcher_thread;
        }
 
        /* Setup the worker thread */
        ret = pthread_create(&worker_thread, NULL,
-                       relay_thread_worker, (void *) relay_ctx);
-       if (ret != 0) {
+                       relay_thread_worker, NULL);
+       if (ret) {
+               errno = ret;
                PERROR("pthread_create worker");
-               goto exit_worker;
+               retval = -1;
+               goto exit_worker_thread;
        }
 
        /* Setup the listener thread */
        ret = pthread_create(&listener_thread, NULL,
                        relay_thread_listener, (void *) NULL);
-       if (ret != 0) {
+       if (ret) {
+               errno = ret;
                PERROR("pthread_create listener");
-               goto exit_listener;
+               retval = -1;
+               goto exit_listener_thread;
        }
 
-       ret = live_start_threads(live_uri, relay_ctx);
-       if (ret != 0) {
+       ret = relayd_live_create(live_uri);
+       if (ret) {
                ERR("Starting live viewer threads");
+               retval = -1;
                goto exit_live;
        }
 
+       /*
+        * This is where we start awaiting program completion (e.g. through
+        * signal that asks threads to teardown).
+        */
+
+       ret = relayd_live_join();
+       if (ret) {
+               retval = -1;
+       }
 exit_live:
+
        ret = pthread_join(listener_thread, &status);
-       if (ret != 0) {
-               PERROR("pthread_join");
-               goto error;     /* join error, exit without cleanup */
+       if (ret) {
+               errno = ret;
+               PERROR("pthread_join listener_thread");
+               retval = -1;
        }
 
-exit_listener:
+exit_listener_thread:
        ret = pthread_join(worker_thread, &status);
-       if (ret != 0) {
-               PERROR("pthread_join");
-               goto error;     /* join error, exit without cleanup */
+       if (ret) {
+               errno = ret;
+               PERROR("pthread_join worker_thread");
+               retval = -1;
        }
 
-exit_worker:
+exit_worker_thread:
        ret = pthread_join(dispatcher_thread, &status);
-       if (ret != 0) {
-               PERROR("pthread_join");
-               goto error;     /* join error, exit without cleanup */
+       if (ret) {
+               errno = ret;
+               PERROR("pthread_join dispatcher_thread");
+               retval = -1;
        }
+exit_dispatcher_thread:
 
-exit_dispatcher:
        ret = pthread_join(health_thread, &status);
-       if (ret != 0) {
-               PERROR("pthread_join health thread");
-               goto error;     /* join error, exit without cleanup */
+       if (ret) {
+               errno = ret;
+               PERROR("pthread_join health_thread");
+               retval = -1;
        }
+exit_health_thread:
 
-       /*
-        * Stop live threads only after joining other threads.
-        */
-       live_stop_threads();
-
-health_error:
        utils_close_pipe(health_quit_pipe);
+exit_health_quit_pipe:
 
-error_health_pipe:
+exit_init_data:
        health_app_destroy(health_relayd);
-
 exit_health_app_create:
-       lttng_ht_destroy(viewer_streams_ht);
-
-exit_relay_ctx_viewer_streams:
-       lttng_ht_destroy(relay_streams_ht);
+exit_options:
+       relayd_cleanup();
 
-exit_relay_ctx_streams:
-       lttng_ht_destroy(relay_ctx->sessions_ht);
+       /* Ensure all prior call_rcu are done. */
+       rcu_barrier();
 
-exit_relay_ctx_sessions:
-       free(relay_ctx);
-
-exit:
-       cleanup();
-       if (!ret) {
+       if (!retval) {
                exit(EXIT_SUCCESS);
+       } else {
+               exit(EXIT_FAILURE);
        }
-
-error:
-       exit(EXIT_FAILURE);
 }
This page took 0.083957 seconds and 4 git commands to generate.