Fix: notification thread: RCU-safe reclaim of hash table nodes
[lttng-tools.git] / src / bin / lttng-sessiond / notification-thread-events.c
index 751e665bdc624936a8401fd307ae75ff93656978..4799c718641bfb0be35bf9433405abf71ab82ba4 100644 (file)
@@ -58,11 +58,15 @@ struct lttng_channel_trigger_list {
        struct channel_key channel_key;
        struct cds_list_head list;
        struct cds_lfht_node channel_triggers_ht_node;
+       /* call_rcu delayed reclaim. */
+       struct rcu_head rcu_node;
 };
 
 struct lttng_trigger_ht_element {
        struct lttng_trigger *trigger;
        struct cds_lfht_node node;
+       /* call_rcu delayed reclaim. */
+       struct rcu_head rcu_node;
 };
 
 struct lttng_condition_list_element {
@@ -79,6 +83,8 @@ struct notification_client_list {
        struct lttng_trigger *trigger;
        struct cds_list_head list;
        struct cds_lfht_node notification_trigger_ht_node;
+       /* call_rcu delayed reclaim. */
+       struct rcu_head rcu_node;
 };
 
 struct notification_client {
@@ -88,7 +94,7 @@ struct notification_client {
        uid_t uid;
        gid_t gid;
        /*
-        * Indicates if the credentials and versions of the client has been
+        * Indicates if the credentials and versions of the client have been
         * checked.
         */
        bool validated;
@@ -145,6 +151,8 @@ struct notification_client {
                        struct lttng_dynamic_buffer buffer;
                } outbound;
        } communication;
+       /* call_rcu delayed reclaim. */
+       struct rcu_head rcu_node;
 };
 
 struct channel_state_sample {
@@ -152,8 +160,24 @@ struct channel_state_sample {
        struct cds_lfht_node channel_state_ht_node;
        uint64_t highest_usage;
        uint64_t lowest_usage;
+       /* call_rcu delayed reclaim. */
+       struct rcu_head rcu_node;
 };
 
+static unsigned long hash_channel_key(struct channel_key *key);
+static int evaluate_condition(struct lttng_condition *condition,
+               struct lttng_evaluation **evaluation,
+               struct notification_thread_state *state,
+               struct channel_state_sample *previous_sample,
+               struct channel_state_sample *latest_sample,
+               uint64_t buffer_capacity);
+static
+int send_evaluation_to_clients(struct lttng_trigger *trigger,
+               struct lttng_evaluation *evaluation,
+               struct notification_client_list *client_list,
+               struct notification_thread_state *state,
+               uid_t channel_uid, gid_t channel_gid);
+
 static
 int match_client(struct cds_lfht_node *node, const void *key)
 {
@@ -281,7 +305,7 @@ unsigned long lttng_condition_buffer_usage_hash(
 
                val = condition->threshold_ratio.value * (double) UINT32_MAX;
                hash ^= hash_key_u64(&val, lttng_ht_seed);
-       } else if (condition->threshold_ratio.set) {
+       } else if (condition->threshold_bytes.set) {
                uint64_t val;
 
                val = condition->threshold_bytes.value;
@@ -308,6 +332,22 @@ unsigned long lttng_condition_hash(struct lttng_condition *condition)
        }
 }
 
+static
+unsigned long hash_channel_key(struct channel_key *key)
+{
+       unsigned long key_hash = hash_key_u64(&key->key, lttng_ht_seed);
+       unsigned long domain_hash = hash_key_ulong(
+               (void *) (unsigned long) key->domain, lttng_ht_seed);
+
+       return key_hash ^ domain_hash;
+}
+
+static
+void free_channel_info_rcu(struct rcu_head *node)
+{
+       free(caa_container_of(node, struct channel_info, rcu_node));
+}
+
 static
 void channel_info_destroy(struct channel_info *channel_info)
 {
@@ -321,7 +361,7 @@ void channel_info_destroy(struct channel_info *channel_info)
        if (channel_info->channel_name) {
                free(channel_info->channel_name);
        }
-       free(channel_info);
+       call_rcu(&channel_info->rcu_node, free_channel_info_rcu);
 }
 
 static
@@ -357,6 +397,126 @@ error:
        return NULL;
 }
 
+/* This function must be called with the RCU read lock held. */
+static
+int evaluate_condition_for_client(struct lttng_trigger *trigger,
+               struct lttng_condition *condition,
+               struct notification_client *client,
+               struct notification_thread_state *state)
+{
+       int ret;
+       struct cds_lfht_iter iter;
+       struct cds_lfht_node *node;
+       struct channel_info *channel_info = NULL;
+       struct channel_key *channel_key = NULL;
+       struct channel_state_sample *last_sample = NULL;
+       struct lttng_channel_trigger_list *channel_trigger_list = NULL;
+       struct lttng_evaluation *evaluation = NULL;
+       struct notification_client_list client_list = { 0 };
+       struct notification_client_list_element client_list_element = { 0 };
+
+       assert(trigger);
+       assert(condition);
+       assert(client);
+       assert(state);
+
+       /* Find the channel associated with the trigger. */
+       cds_lfht_for_each_entry(state->channel_triggers_ht, &iter,
+                       channel_trigger_list , channel_triggers_ht_node) {
+               struct lttng_trigger_list_element *element;
+
+               cds_list_for_each_entry(element, &channel_trigger_list->list, node) {
+                       struct lttng_condition *current_condition =
+                               lttng_trigger_get_condition(
+                                               element->trigger);
+
+                       assert(current_condition);
+                       if (!lttng_condition_is_equal(condition,
+                                               current_condition)) {
+                               continue;
+                       }
+
+                       /* Found the trigger, save the channel key. */
+                       channel_key = &channel_trigger_list->channel_key;
+                       break;
+               }
+               if (channel_key) {
+                       /* The channel key was found stop iteration. */
+                       break;
+               }
+       }
+
+       if (!channel_key){
+               /* No channel found; normal exit. */
+               DBG("[notification-thread] No channel associated with newly subscribed-to condition");
+               ret = 0;
+               goto end;
+       }
+
+       /* Fetch channel info for the matching channel. */
+       cds_lfht_lookup(state->channels_ht,
+                       hash_channel_key(channel_key),
+                       match_channel_info,
+                       channel_key,
+                       &iter);
+       node = cds_lfht_iter_get_node(&iter);
+       assert(node);
+       channel_info = caa_container_of(node, struct channel_info,
+                       channels_ht_node);
+
+       /* Retrieve the channel's last sample, if it exists. */
+       cds_lfht_lookup(state->channel_state_ht,
+                       hash_channel_key(channel_key),
+                       match_channel_state_sample,
+                       channel_key,
+                       &iter);
+       node = cds_lfht_iter_get_node(&iter);
+       if (node) {
+               last_sample = caa_container_of(node,
+                               struct channel_state_sample,
+                               channel_state_ht_node);
+       } else {
+               /* Nothing to evaluate, no sample was ever taken. Normal exit */
+               DBG("[notification-thread] No channel sample associated with newly subscribed-to condition");
+               ret = 0;
+               goto end;
+       }
+
+       ret = evaluate_condition(condition, &evaluation, state, NULL,
+                       last_sample, channel_info->capacity);
+       if (ret) {
+               WARN("[notification-thread] Fatal error occurred while evaluating a newly subscribed-to condition");
+               goto end;
+       }
+
+       if (!evaluation) {
+               /* Evaluation yielded nothing. Normal exit. */
+               DBG("[notification-thread] Newly subscribed-to condition evaluated to false, nothing to report to client");
+               ret = 0;
+               goto end;
+       }
+
+       /*
+        * Create a temporary client list with the client currently
+        * subscribing.
+        */
+       cds_lfht_node_init(&client_list.notification_trigger_ht_node);
+       CDS_INIT_LIST_HEAD(&client_list.list);
+       client_list.trigger = trigger;
+
+       CDS_INIT_LIST_HEAD(&client_list_element.node);
+       client_list_element.client = client;
+       cds_list_add(&client_list_element.node, &client_list.list);
+
+       /* Send evaluation result to the newly-subscribed client. */
+       DBG("[notification-thread] Newly subscribed-to condition evaluated to true, notifying client");
+       ret = send_evaluation_to_clients(trigger, evaluation, &client_list,
+                       state, channel_info->uid, channel_info->gid);
+
+end:
+       return ret;
+}
+
 static
 int notification_thread_client_subscribe(struct notification_client *client,
                struct lttng_condition *condition,
@@ -404,11 +564,6 @@ int notification_thread_client_subscribe(struct notification_client *client,
        condition_list_element->condition = condition;
        cds_list_add(&condition_list_element->node, &client->condition_list);
 
-       /*
-        * Add the client to the list of clients interested in a given trigger
-        * if a "notification" trigger with a corresponding condition was
-        * added prior.
-        */
        cds_lfht_lookup(state->notification_trigger_clients_ht,
                        lttng_condition_hash(condition),
                        match_client_list_condition,
@@ -422,6 +577,19 @@ int notification_thread_client_subscribe(struct notification_client *client,
 
        client_list = caa_container_of(node, struct notification_client_list,
                        notification_trigger_ht_node);
+       if (evaluate_condition_for_client(client_list->trigger, condition,
+                       client, state)) {
+               WARN("[notification-thread] Evaluation of a condition on client subscription failed, aborting.");
+               ret = -1;
+               free(client_list_element);
+               goto end_unlock;
+       }
+
+       /*
+        * Add the client to the list of clients interested in a given trigger
+        * if a "notification" trigger with a corresponding condition was
+        * added prior.
+        */
        client_list_element->client = client;
        CDS_INIT_LIST_HEAD(&client_list_element->node);
        cds_list_add(&client_list_element->node, &client_list->list);
@@ -520,6 +688,12 @@ end:
        return 0;
 }
 
+static
+void free_notification_client_rcu(struct rcu_head *node)
+{
+       free(caa_container_of(node, struct notification_client, rcu_node));
+}
+
 static
 void notification_client_destroy(struct notification_client *client,
                struct notification_thread_state *state)
@@ -542,7 +716,7 @@ void notification_client_destroy(struct notification_client *client,
        }
        lttng_dynamic_buffer_reset(&client->communication.inbound.buffer);
        lttng_dynamic_buffer_reset(&client->communication.outbound.buffer);
-       free(client);
+       call_rcu(&client->rcu_node, free_notification_client_rcu);
 }
 
 /*
@@ -642,13 +816,6 @@ bool trigger_applies_to_client(struct lttng_trigger *trigger,
        return applies;
 }
 
-static
-unsigned long hash_channel_key(struct channel_key *key)
-{
-       return hash_key_u64(&key->key, lttng_ht_seed) ^ hash_key_ulong(
-               (void *) (unsigned long) key->domain, lttng_ht_seed);
-}
-
 static
 int handle_notification_thread_command_add_channel(
        struct notification_thread_state *state,
@@ -676,6 +843,7 @@ int handle_notification_thread_command_add_channel(
 
        channel_key = &new_channel_info->key;
 
+       rcu_read_lock();
        /* Build a list of all triggers applying to the new channel. */
        cds_lfht_for_each_entry(state->triggers_ht, &iter, trigger_ht_element,
                        node) {
@@ -688,7 +856,7 @@ int handle_notification_thread_command_add_channel(
 
                new_element = zmalloc(sizeof(*new_element));
                if (!new_element) {
-                       goto error;
+                       goto error_unlock;
                }
                CDS_INIT_LIST_HEAD(&new_element->node);
                new_element->trigger = trigger_ht_element->trigger;
@@ -700,14 +868,13 @@ int handle_notification_thread_command_add_channel(
                        trigger_count);
        channel_trigger_list = zmalloc(sizeof(*channel_trigger_list));
        if (!channel_trigger_list) {
-               goto error;
+               goto error_unlock;
        }
        channel_trigger_list->channel_key = *channel_key;
        CDS_INIT_LIST_HEAD(&channel_trigger_list->list);
        cds_lfht_node_init(&channel_trigger_list->channel_triggers_ht_node);
        cds_list_splice(&trigger_list, &channel_trigger_list->list);
 
-       rcu_read_lock();
        /* Add channel to the channel_ht which owns the channel_infos. */
        cds_lfht_add(state->channels_ht,
                        hash_channel_key(channel_key),
@@ -722,12 +889,28 @@ int handle_notification_thread_command_add_channel(
        rcu_read_unlock();
        *cmd_result = LTTNG_OK;
        return 0;
+error_unlock:
+       rcu_read_unlock();
 error:
        /* Empty trigger list */
        channel_info_destroy(new_channel_info);
        return 1;
 }
 
+static
+void free_channel_trigger_list_rcu(struct rcu_head *node)
+{
+       free(caa_container_of(node, struct lttng_channel_trigger_list,
+                       rcu_node));
+}
+
+static
+void free_channel_state_sample_rcu(struct rcu_head *node)
+{
+       free(caa_container_of(node, struct channel_state_sample,
+                       rcu_node));
+}
+
 static
 int handle_notification_thread_command_remove_channel(
        struct notification_thread_state *state,
@@ -770,7 +953,7 @@ int handle_notification_thread_command_remove_channel(
                free(trigger_list_element);
        }
        cds_lfht_del(state->channel_triggers_ht, node);
-       free(trigger_list);
+       call_rcu(&trigger_list->rcu_node, free_channel_trigger_list_rcu);
 
        /* Free sampled channel state. */
        cds_lfht_lookup(state->channel_state_ht,
@@ -789,7 +972,7 @@ int handle_notification_thread_command_remove_channel(
                                channel_state_ht_node);
 
                cds_lfht_del(state->channel_state_ht, node);
-               free(sample);
+               call_rcu(&sample->rcu_node, free_channel_state_sample_rcu);
        }
 
        /* Remove the channel from the channels_ht and free it. */
@@ -974,6 +1157,7 @@ int handle_notification_thread_command_register_trigger(
                        channels_ht_node) {
                struct lttng_trigger_list_element *trigger_list_element;
                struct lttng_channel_trigger_list *trigger_list;
+               struct cds_lfht_iter lookup_iter;
 
                if (!trigger_applies_to_channel(trigger, channel)) {
                        continue;
@@ -983,10 +1167,9 @@ int handle_notification_thread_command_register_trigger(
                                hash_channel_key(&channel->key),
                                match_channel_trigger_list,
                                &channel->key,
-                               &iter);
-               node = cds_lfht_iter_get_node(&iter);
+                               &lookup_iter);
+               node = cds_lfht_iter_get_node(&lookup_iter);
                assert(node);
-               /* Free the list of triggers associated with this channel. */
                trigger_list = caa_container_of(node,
                                struct lttng_channel_trigger_list,
                                channel_triggers_ht_node);
@@ -999,6 +1182,7 @@ int handle_notification_thread_command_register_trigger(
                CDS_INIT_LIST_HEAD(&trigger_list_element->node);
                trigger_list_element->trigger = trigger;
                cds_list_add(&trigger_list_element->node, &trigger_list->list);
+
                /* A trigger can only apply to one channel. */
                break;
        }
@@ -1026,6 +1210,20 @@ error:
        return ret;
 }
 
+static
+void free_notification_client_list_rcu(struct rcu_head *node)
+{
+       free(caa_container_of(node, struct notification_client_list,
+                       rcu_node));
+}
+
+static
+void free_lttng_trigger_ht_element_rcu(struct rcu_head *node)
+{
+       free(caa_container_of(node, struct lttng_trigger_ht_element,
+                       rcu_node));
+}
+
 static
 int handle_notification_thread_command_unregister_trigger(
                struct notification_thread_state *state,
@@ -1077,6 +1275,8 @@ int handle_notification_thread_command_unregister_trigger(
 
                        DBG("[notification-thread] Removed trigger from channel_triggers_ht");
                        cds_list_del(&trigger_element->node);
+                       /* A trigger can only appear once per channel */
+                       break;
                }
        }
 
@@ -1098,7 +1298,7 @@ int handle_notification_thread_command_unregister_trigger(
                free(client_list_element);
        }
        cds_lfht_del(state->notification_trigger_clients_ht, node);
-       free(client_list);
+       call_rcu(&client_list->rcu_node, free_notification_client_list_rcu);
 
        /* Remove trigger from triggers_ht. */
        trigger_ht_element = caa_container_of(triggers_ht_node,
@@ -1110,7 +1310,7 @@ int handle_notification_thread_command_unregister_trigger(
        action = lttng_trigger_get_action(trigger_ht_element->trigger);
        lttng_action_destroy(action);
        lttng_trigger_destroy(trigger_ht_element->trigger);
-       free(trigger_ht_element);
+       call_rcu(&trigger_ht_element->rcu_node, free_lttng_trigger_ht_element_rcu);
 end:
        rcu_read_unlock();
        if (_cmd_reply) {
@@ -1129,7 +1329,7 @@ int handle_notification_thread_command(
        struct notification_thread_command *cmd;
 
        /* Read event_fd to put it back into a quiescent state. */
-       ret = read(handle->cmd_queue.event_fd, &counter, sizeof(counter));
+       ret = read(lttng_pipe_get_readfd(handle->cmd_queue.event_pipe), &counter, sizeof(counter));
        if (ret == -1) {
                goto error;
        }
This page took 0.029569 seconds and 5 git commands to generate.