Fix: Lock FD tracker across fork
[lttng-ust.git] / liblttng-ust / lttng-ust-comm.c
index 61dbb41b4553c9a9eee21829ba269e4d5c93c1cb..47ba36e5c9f1d9f1871c267469049708a7dd1e1e 100644 (file)
@@ -86,6 +86,8 @@ static int initialized;
  *
  * ust_lock nests within the dynamic loader lock (within glibc) because
  * it is taken within the library constructor.
+ *
+ * The ust fd tracker lock nests within the ust_mutex.
  */
 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
 
@@ -228,7 +230,11 @@ static sem_t constructor_wait;
 /*
  * Doing this for both the global and local sessiond.
  */
-static int sem_count = { 2 };
+enum {
+       sem_count_initial_value = 4,
+};
+
+static int sem_count = sem_count_initial_value;
 
 /*
  * Counting nesting within lttng-ust. Used to ensure that calling fork()
@@ -243,7 +249,7 @@ struct sock_info {
        const char *name;
        pthread_t ust_listener; /* listener thread */
        int root_handle;
-       int constructor_sem_posted;
+       int registration_done;
        int allowed;
        int global;
        int thread_active;
@@ -256,6 +262,7 @@ struct sock_info {
        char *wait_shm_mmap;
        /* Keep track of lazy state dump not performed yet. */
        int statedump_pending;
+       int initial_statedump_done;
 };
 
 /* Socket from app (connect) to session daemon (listen) for communication */
@@ -264,6 +271,7 @@ struct sock_info global_apps = {
        .global = 1,
 
        .root_handle = -1,
+       .registration_done = 0,
        .allowed = 0,
        .thread_active = 0,
 
@@ -274,6 +282,7 @@ struct sock_info global_apps = {
        .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 
        .statedump_pending = 0,
+       .initial_statedump_done = 0,
 };
 
 /* TODO: allow global_apps_sock_path override */
@@ -282,6 +291,7 @@ struct sock_info local_apps = {
        .name = "local",
        .global = 0,
        .root_handle = -1,
+       .registration_done = 0,
        .allowed = 0,   /* Check setuid bit first */
        .thread_active = 0,
 
@@ -289,6 +299,7 @@ struct sock_info local_apps = {
        .notify_socket = -1,
 
        .statedump_pending = 0,
+       .initial_statedump_done = 0,
 };
 
 static int wait_poll_fallback;
@@ -621,45 +632,85 @@ int send_reply(int sock, struct ustcomm_ust_reply *lur)
 }
 
 static
-int handle_register_done(struct sock_info *sock_info)
+void decrement_sem_count(unsigned int count)
 {
        int ret;
 
-       if (sock_info->constructor_sem_posted)
-               return 0;
-       sock_info->constructor_sem_posted = 1;
+       assert(uatomic_read(&sem_count) >= count);
+
        if (uatomic_read(&sem_count) <= 0) {
-               return 0;
+               return;
        }
-       ret = uatomic_add_return(&sem_count, -1);
+
+       ret = uatomic_add_return(&sem_count, -count);
        if (ret == 0) {
                ret = sem_post(&constructor_wait);
                assert(!ret);
        }
+}
+
+static
+int handle_register_done(struct sock_info *sock_info)
+{
+       if (sock_info->registration_done)
+               return 0;
+       sock_info->registration_done = 1;
+
+       decrement_sem_count(1);
+       if (!sock_info->statedump_pending) {
+               sock_info->initial_statedump_done = 1;
+               decrement_sem_count(1);
+       }
+
+       return 0;
+}
+
+static
+int handle_register_failed(struct sock_info *sock_info)
+{
+       if (sock_info->registration_done)
+               return 0;
+       sock_info->registration_done = 1;
+       sock_info->initial_statedump_done = 1;
+
+       decrement_sem_count(2);
+
        return 0;
 }
 
 /*
  * Only execute pending statedump after the constructor semaphore has
- * been posted by each listener thread. This means statedump will only
- * be performed after the "registration done" command is received from
- * each session daemon the application is connected to.
+ * been posted by the current listener thread. This means statedump will
+ * only be performed after the "registration done" command is received
+ * from this thread's session daemon.
  *
  * This ensures we don't run into deadlock issues with the dynamic
  * loader mutex, which is held while the constructor is called and
  * waiting on the constructor semaphore. All operations requiring this
  * dynamic loader lock need to be postponed using this mechanism.
+ *
+ * In a scenario with two session daemons connected to the application,
+ * it is possible that the first listener thread which receives the
+ * registration done command issues its statedump while the dynamic
+ * loader lock is still held by the application constructor waiting on
+ * the semaphore. It will however be allowed to proceed when the
+ * second session daemon sends the registration done command to the
+ * second listener thread. This situation therefore does not produce
+ * a deadlock.
  */
 static
 void handle_pending_statedump(struct sock_info *sock_info)
 {
-       int ctor_passed = sock_info->constructor_sem_posted;
-
-       if (ctor_passed && sock_info->statedump_pending) {
+       if (sock_info->registration_done && sock_info->statedump_pending) {
                sock_info->statedump_pending = 0;
                pthread_mutex_lock(&ust_fork_mutex);
                lttng_handle_pending_statedump(sock_info);
                pthread_mutex_unlock(&ust_fork_mutex);
+
+               if (!sock_info->initial_statedump_done) {
+                       sock_info->initial_statedump_done = 1;
+                       decrement_sem_count(1);
+               }
        }
 }
 
@@ -877,12 +928,13 @@ int handle_message(struct sock_info *sock_info,
        {
                /* Receive shm_fd, wakeup_fd */
                ret = ustcomm_recv_stream_from_sessiond(sock,
-                       &lum->u.stream.len,
+                       NULL,
                        &args.stream.shm_fd,
                        &args.stream.wakeup_fd);
                if (ret) {
                        goto error;
                }
+
                if (ops->cmd)
                        ret = ops->cmd(lum->handle, lum->cmd,
                                        (unsigned long) &lum->u,
@@ -1068,7 +1120,8 @@ void cleanup_sock_info(struct sock_info *sock_info, int exiting)
                }
                sock_info->root_handle = -1;
        }
-       sock_info->constructor_sem_posted = 0;
+       sock_info->registration_done = 0;
+       sock_info->initial_statedump_done = 0;
 
        /*
         * wait_shm_mmap, socket and notify socket are used by listener
@@ -1476,7 +1529,7 @@ restart:
                 * If we cannot find the sessiond daemon, don't delay
                 * constructor execution.
                 */
-               ret = handle_register_done(sock_info);
+               ret = handle_register_failed(sock_info);
                assert(!ret);
                ust_unlock();
                goto restart;
@@ -1530,7 +1583,7 @@ restart:
                 * If we cannot register to the sessiond daemon, don't
                 * delay constructor execution.
                 */
-               ret = handle_register_done(sock_info);
+               ret = handle_register_failed(sock_info);
                assert(!ret);
                ust_unlock();
                goto restart;
@@ -1559,7 +1612,7 @@ restart:
                 * If we cannot find the sessiond daemon, don't delay
                 * constructor execution.
                 */
-               ret = handle_register_done(sock_info);
+               ret = handle_register_failed(sock_info);
                assert(!ret);
                ust_unlock();
                goto restart;
@@ -1623,7 +1676,7 @@ restart:
                 * If we cannot register to the sessiond daemon, don't
                 * delay constructor execution.
                 */
-               ret = handle_register_done(sock_info);
+               ret = handle_register_failed(sock_info);
                assert(!ret);
                ust_unlock();
                goto restart;
@@ -1653,7 +1706,7 @@ restart:
                         * If we cannot register to the sessiond daemon, don't
                         * delay constructor execution.
                         */
-                       ret = handle_register_done(sock_info);
+                       ret = handle_register_failed(sock_info);
                        assert(!ret);
                        ust_unlock();
                        goto end;
@@ -1922,7 +1975,7 @@ void lttng_ust_cleanup(int exiting)
        exit_tracepoint();
        if (!exiting) {
                /* Reinitialize values for fork */
-               sem_count = 2;
+               sem_count = sem_count_initial_value;
                lttng_ust_comm_should_quit = 0;
                initialized = 0;
        }
@@ -2014,6 +2067,7 @@ void ust_before_fork(sigset_t *save_sigset)
 
        ust_lock_nocheck();
        urcu_bp_before_fork();
+       lttng_ust_lock_fd_tracker();
 }
 
 static void ust_after_fork_common(sigset_t *restore_sigset)
@@ -2021,6 +2075,7 @@ static void ust_after_fork_common(sigset_t *restore_sigset)
        int ret;
 
        DBG("process %d", getpid());
+       lttng_ust_unlock_fd_tracker();
        ust_unlock();
 
        pthread_mutex_unlock(&ust_fork_mutex);
This page took 0.026367 seconds and 4 git commands to generate.