Fix: add urcu-bp TLS fixup
[lttng-ust.git] / liblttng-ust / lttng-ust-comm.c
index b99bf00edec2df760df075f0ad77597a3e1a54d1..378ca21cfc71f354a282fe4a9a521cf77319ef02 100644 (file)
@@ -34,7 +34,6 @@
 #include <time.h>
 #include <assert.h>
 #include <signal.h>
-#include <dlfcn.h>
 #include <urcu/uatomic.h>
 #include <urcu/futex.h>
 #include <urcu/compiler.h>
@@ -63,11 +62,119 @@ static int initialized;
  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  * Held when handling a command, also held by fork() to deal with
  * removal of threads, and by exit path.
+ *
+ * The UST lock is the centralized mutex across UST tracing control and
+ * probe registration.
+ *
+ * ust_exit_mutex must never nest in ust_mutex.
+ *
+ * ust_fork_mutex must never nest in ust_mutex.
+ *
+ * ust_mutex_nest is a per-thread nesting counter, allowing the perf
+ * counter lazy initialization called by events within the statedump,
+ * which traces while the ust_mutex is held.
+ *
+ * ust_lock nests within the dynamic loader lock (within glibc) because
+ * it is taken within the library constructor.
  */
+static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+/* Allow nesting the ust_mutex within the same thread. */
+static DEFINE_URCU_TLS(int, ust_mutex_nest);
+
+/*
+ * ust_exit_mutex protects thread_active variable wrt thread exit. It
+ * cannot be done by ust_mutex because pthread_cancel(), which takes an
+ * internal libc lock, cannot nest within ust_mutex.
+ *
+ * It never nests within a ust_mutex.
+ */
+static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+/*
+ * ust_fork_mutex protects base address statedump tracing against forks. It
+ * prevents the dynamic loader lock to be taken (by base address statedump
+ * tracing) while a fork is happening, thus preventing deadlock issues with
+ * the dynamic loader lock.
+ */
+static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
 
 /* Should the ust comm thread quit ? */
 static int lttng_ust_comm_should_quit;
 
+/*
+ * Return 0 on success, -1 if should quit.
+ * The lock is taken in both cases.
+ * Signal-safe.
+ */
+int ust_lock(void)
+{
+       sigset_t sig_all_blocked, orig_mask;
+       int ret;
+
+       sigfillset(&sig_all_blocked);
+       ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+       if (!URCU_TLS(ust_mutex_nest)++)
+               pthread_mutex_lock(&ust_mutex);
+       ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+       if (lttng_ust_comm_should_quit) {
+               return -1;
+       } else {
+               return 0;
+       }
+}
+
+/*
+ * ust_lock_nocheck() can be used in constructors/destructors, because
+ * they are already nested within the dynamic loader lock, and therefore
+ * have exclusive access against execution of liblttng-ust destructor.
+ * Signal-safe.
+ */
+void ust_lock_nocheck(void)
+{
+       sigset_t sig_all_blocked, orig_mask;
+       int ret;
+
+       sigfillset(&sig_all_blocked);
+       ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+       if (!URCU_TLS(ust_mutex_nest)++)
+               pthread_mutex_lock(&ust_mutex);
+       ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+}
+
+/*
+ * Signal-safe.
+ */
+void ust_unlock(void)
+{
+       sigset_t sig_all_blocked, orig_mask;
+       int ret;
+
+       sigfillset(&sig_all_blocked);
+       ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+       if (!--URCU_TLS(ust_mutex_nest))
+               pthread_mutex_unlock(&ust_mutex);
+       ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+}
+
 /*
  * Wait for either of these before continuing to the main
  * program:
@@ -108,7 +215,8 @@ struct sock_info {
 
        char wait_shm_path[PATH_MAX];
        char *wait_shm_mmap;
-       struct lttng_session *session_enabled;
+       /* Keep track of lazy state dump not performed yet. */
+       int statedump_pending;
 };
 
 /* Socket from app (connect) to session daemon (listen) for communication */
@@ -126,7 +234,7 @@ struct sock_info global_apps = {
 
        .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 
-       .session_enabled = NULL,
+       .statedump_pending = 0,
 };
 
 /* TODO: allow global_apps_sock_path override */
@@ -141,7 +249,7 @@ struct sock_info local_apps = {
        .socket = -1,
        .notify_socket = -1,
 
-       .session_enabled = NULL,
+       .statedump_pending = 0,
 };
 
 static int wait_poll_fallback;
@@ -221,6 +329,22 @@ void lttng_fixup_nest_count_tls(void)
        asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 }
 
+static
+void lttng_fixup_ust_mutex_nest_tls(void)
+{
+       asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
+}
+
+/*
+ * Fixup urcu bp TLS.
+ */
+static
+void lttng_fixup_urcu_bp_tls(void)
+{
+       rcu_read_lock();
+       rcu_read_unlock();
+}
+
 int lttng_get_notify_socket(void *owner)
 {
        struct sock_info *info = owner;
@@ -387,6 +511,30 @@ int handle_register_done(struct sock_info *sock_info)
        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.
+ *
+ * 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.
+ */
+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) {
+               sock_info->statedump_pending = 0;
+               pthread_mutex_lock(&ust_fork_mutex);
+               lttng_handle_pending_statedump(sock_info);
+               pthread_mutex_unlock(&ust_fork_mutex);
+       }
+}
+
 static
 int handle_message(struct sock_info *sock_info,
                int sock, struct ustcomm_ust_msg *lum)
@@ -397,11 +545,9 @@ int handle_message(struct sock_info *sock_info,
        union ust_args args;
        ssize_t len;
 
-       ust_lock();
-
        memset(&lur, 0, sizeof(lur));
 
-       if (lttng_ust_comm_should_quit) {
+       if (ust_lock()) {
                ret = -LTTNG_UST_ERR_EXITING;
                goto end;
        }
@@ -705,6 +851,14 @@ end:
 
 error:
        ust_unlock();
+
+       /*
+        * Performed delayed statedump operations outside of the UST
+        * lock. We need to take the dynamic loader lock before we take
+        * the UST lock internally within handle_pending_statedump().
+         */
+       handle_pending_statedump(sock_info);
+
        return ret;
 }
 
@@ -747,9 +901,14 @@ void cleanup_sock_info(struct sock_info *sock_info, int exiting)
                sock_info->notify_socket = -1;
        }
        if (sock_info->wait_shm_mmap) {
-               ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
-               if (ret) {
-                       ERR("Error unmapping wait shm");
+               long page_size;
+
+               page_size = sysconf(_SC_PAGE_SIZE);
+               if (page_size > 0) {
+                       ret = munmap(sock_info->wait_shm_mmap, page_size);
+                       if (ret) {
+                               ERR("Error unmapping wait shm");
+                       }
                }
                sock_info->wait_shm_mmap = NULL;
        }
@@ -922,15 +1081,20 @@ error_close:
 static
 char *get_map_shm(struct sock_info *sock_info)
 {
-       size_t mmap_size = sysconf(_SC_PAGE_SIZE);
+       long page_size;
        int wait_shm_fd, ret;
        char *wait_shm_mmap;
 
-       wait_shm_fd = get_wait_shm(sock_info, mmap_size);
+       page_size = sysconf(_SC_PAGE_SIZE);
+       if (page_size < 0) {
+               goto error;
+       }
+
+       wait_shm_fd = get_wait_shm(sock_info, page_size);
        if (wait_shm_fd < 0) {
                goto error;
        }
-       wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
+       wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
                  MAP_SHARED, wait_shm_fd, 0);
        /* close shm fd immediately after taking the mmap reference */
        ret = close(wait_shm_fd);
@@ -952,8 +1116,7 @@ void wait_for_sessiond(struct sock_info *sock_info)
 {
        int ret;
 
-       ust_lock();
-       if (lttng_ust_comm_should_quit) {
+       if (ust_lock()) {
                goto quit;
        }
        if (wait_poll_fallback) {
@@ -1058,9 +1221,7 @@ restart:
                DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
                prev_connect_failed = 1;
 
-               ust_lock();
-
-               if (lttng_ust_comm_should_quit) {
+               if (ust_lock()) {
                        goto quit;
                }
 
@@ -1075,9 +1236,7 @@ restart:
        }
        sock_info->socket = ret;
 
-       ust_lock();
-
-       if (lttng_ust_comm_should_quit) {
+       if (ust_lock()) {
                goto quit;
        }
 
@@ -1118,9 +1277,7 @@ restart:
                DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
                prev_connect_failed = 1;
 
-               ust_lock();
-
-               if (lttng_ust_comm_should_quit) {
+               if (ust_lock()) {
                        goto quit;
                }
 
@@ -1157,9 +1314,7 @@ restart:
                WARN("Unsupported timeout value %ld", timeout);
        }
 
-       ust_lock();
-
-       if (lttng_ust_comm_should_quit) {
+       if (ust_lock()) {
                goto quit;
        }
 
@@ -1190,8 +1345,7 @@ restart:
                switch (len) {
                case 0: /* orderly shutdown */
                        DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
-                       ust_lock();
-                       if (lttng_ust_comm_should_quit) {
+                       if (ust_lock()) {
                                goto quit;
                        }
                        /*
@@ -1213,14 +1367,6 @@ restart:
                        ret = handle_message(sock_info, sock, &lum);
                        if (ret) {
                                ERR("Error handling message for %s socket", sock_info->name);
-                       } else {
-                               struct lttng_session *session;
-
-                               session = sock_info->session_enabled;
-                               if (session) {
-                                       sock_info->session_enabled = NULL;
-                                       lttng_ust_baddr_statedump(session);
-                               }
                        }
                        continue;
                default:
@@ -1238,8 +1384,7 @@ restart:
 
        }
 end:
-       ust_lock();
-       if (lttng_ust_comm_should_quit) {
+       if (ust_lock()) {
                goto quit;
        }
        /* Cleanup socket handles before trying to reconnect */
@@ -1248,11 +1393,22 @@ end:
        goto restart;   /* try to reconnect */
 
 quit:
-       sock_info->thread_active = 0;
        ust_unlock();
+
+       pthread_mutex_lock(&ust_exit_mutex);
+       sock_info->thread_active = 0;
+       pthread_mutex_unlock(&ust_exit_mutex);
        return NULL;
 }
 
+/*
+ * Weak symbol to call when the ust malloc wrapper is not loaded.
+ */
+__attribute__((weak))
+void lttng_ust_malloc_wrapper_init(void)
+{
+}
+
 /*
  * sessiond monitoring thread: monitor presence of global and per-user
  * sessiond by polling the application common named pipe.
@@ -1273,10 +1429,12 @@ void __attribute__((constructor)) lttng_ust_init(void)
         * to be the dynamic linker mutex) and ust_lock, taken within
         * the ust lock.
         */
+       lttng_fixup_urcu_bp_tls();
        lttng_fixup_ringbuffer_tls();
        lttng_fixup_vtid_tls();
        lttng_fixup_nest_count_tls();
        lttng_fixup_procname_tls();
+       lttng_fixup_ust_mutex_nest_tls();
 
        /*
         * We want precise control over the order in which we construct
@@ -1286,12 +1444,18 @@ void __attribute__((constructor)) lttng_ust_init(void)
         */
        init_usterr();
        init_tracepoint();
+       lttng_ust_baddr_statedump_init();
        lttng_ring_buffer_metadata_client_init();
        lttng_ring_buffer_client_overwrite_init();
        lttng_ring_buffer_client_overwrite_rt_init();
        lttng_ring_buffer_client_discard_init();
        lttng_ring_buffer_client_discard_rt_init();
+       lttng_perf_counter_init();
        lttng_context_init();
+       /*
+        * Invoke ust malloc wrapper init before starting other threads.
+        */
+       lttng_ust_malloc_wrapper_init();
 
        timeout_mode = get_constructor_timeout(&constructor_timeout);
 
@@ -1323,24 +1487,24 @@ void __attribute__((constructor)) lttng_ust_init(void)
                ERR("pthread_attr_setdetachstate: %s", strerror(ret));
        }
 
-       ust_lock();
+       pthread_mutex_lock(&ust_exit_mutex);
        ret = pthread_create(&global_apps.ust_listener, &thread_attr,
                        ust_listener_thread, &global_apps);
        if (ret) {
                ERR("pthread_create global: %s", strerror(ret));
        }
        global_apps.thread_active = 1;
-       ust_unlock();
+       pthread_mutex_unlock(&ust_exit_mutex);
 
        if (local_apps.allowed) {
-               ust_lock();
+               pthread_mutex_lock(&ust_exit_mutex);
                ret = pthread_create(&local_apps.ust_listener, &thread_attr,
                                ust_listener_thread, &local_apps);
                if (ret) {
                        ERR("pthread_create local: %s", strerror(ret));
                }
                local_apps.thread_active = 1;
-               ust_unlock();
+               pthread_mutex_unlock(&ust_exit_mutex);
        } else {
                handle_register_done(&local_apps);
        }
@@ -1395,11 +1559,13 @@ void lttng_ust_cleanup(int exiting)
        lttng_ust_abi_exit();
        lttng_ust_events_exit();
        lttng_context_exit();
+       lttng_perf_counter_exit();
        lttng_ring_buffer_client_discard_rt_exit();
        lttng_ring_buffer_client_discard_exit();
        lttng_ring_buffer_client_overwrite_rt_exit();
        lttng_ring_buffer_client_overwrite_exit();
        lttng_ring_buffer_metadata_client_exit();
+       lttng_ust_baddr_statedump_destroy();
        exit_tracepoint();
        if (!exiting) {
                /* Reinitialize values for fork */
@@ -1424,9 +1590,11 @@ void __attribute__((destructor)) lttng_ust_exit(void)
         * mutexes to ensure it is not in a mutex critical section when
         * pthread_cancel is later called.
         */
-       ust_lock();
+       ust_lock_nocheck();
        lttng_ust_comm_should_quit = 1;
+       ust_unlock();
 
+       pthread_mutex_lock(&ust_exit_mutex);
        /* cancel threads */
        if (global_apps.thread_active) {
                ret = pthread_cancel(global_apps.ust_listener);
@@ -1446,7 +1614,7 @@ void __attribute__((destructor)) lttng_ust_exit(void)
                        local_apps.thread_active = 0;
                }
        }
-       ust_unlock();
+       pthread_mutex_unlock(&ust_exit_mutex);
 
        /*
         * Do NOT join threads: use of sys_futex makes it impossible to
@@ -1485,7 +1653,10 @@ void ust_before_fork(sigset_t *save_sigset)
        if (ret == -1) {
                PERROR("sigprocmask");
        }
-       ust_lock();
+
+       pthread_mutex_lock(&ust_fork_mutex);
+
+       ust_lock_nocheck();
        rcu_bp_before_fork();
 }
 
@@ -1495,6 +1666,9 @@ static void ust_after_fork_common(sigset_t *restore_sigset)
 
        DBG("process %d", getpid());
        ust_unlock();
+
+       pthread_mutex_unlock(&ust_fork_mutex);
+
        /* Restore signals */
        ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
        if (ret == -1) {
@@ -1535,9 +1709,8 @@ void ust_after_fork_child(sigset_t *restore_sigset)
        lttng_ust_init();
 }
 
-void lttng_ust_sockinfo_session_enabled(void *owner,
-               struct lttng_session *session_enabled)
+void lttng_ust_sockinfo_session_enabled(void *owner)
 {
        struct sock_info *sock_info = owner;
-       sock_info->session_enabled = session_enabled;
+       sock_info->statedump_pending = 1;
 }
This page took 0.029432 seconds and 4 git commands to generate.