Fix: add urcu-bp TLS fixup
[lttng-ust.git] / liblttng-ust / lttng-ust-comm.c
index 4724fabc0d6a72cf57923cea7e6da44d94563052..8e1a0f35ff74c820f0cc195a55ae38bf27850b1a 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 +62,55 @@ 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.
+ */
+static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+/*
+ * 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;
 
 /* Should the ust comm thread quit ? */
 static int lttng_ust_comm_should_quit;
 
+/*
+ * Return 0 on success, -1 if should quilt.
+ * The lock is taken in both cases.
+ */
+int ust_lock(void)
+{
+       pthread_mutex_lock(&ust_mutex);
+       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.
+ */
+void ust_lock_nocheck(void)
+{
+       pthread_mutex_lock(&ust_mutex);
+}
+
+void ust_unlock(void)
+{
+       pthread_mutex_unlock(&ust_mutex);
+}
+
 /*
  * Wait for either of these before continuing to the main
  * program:
@@ -108,7 +151,8 @@ struct sock_info {
 
        char wait_shm_path[PATH_MAX];
        char *wait_shm_mmap;
-       int 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 +170,7 @@ struct sock_info global_apps = {
 
        .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 
-       .session_enabled = 0,
+       .statedump_pending = 0,
 };
 
 /* TODO: allow global_apps_sock_path override */
@@ -141,7 +185,7 @@ struct sock_info local_apps = {
        .socket = -1,
        .notify_socket = -1,
 
-       .session_enabled = 0,
+       .statedump_pending = 0,
 };
 
 static int wait_poll_fallback;
@@ -221,6 +265,16 @@ void lttng_fixup_nest_count_tls(void)
        asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 }
 
+/*
+ * 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 +441,28 @@ 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;
+               lttng_handle_pending_statedump(sock_info);
+       }
+}
+
 static
 int handle_message(struct sock_info *sock_info,
                int sock, struct ustcomm_ust_msg *lum)
@@ -397,11 +473,9 @@ 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;
        }
@@ -705,18 +779,15 @@ end:
 
 error:
        ust_unlock();
-       return ret;
-}
 
-static
-void handle_pending_statedumps(struct sock_info *sock_info)
-{
-       int ctor_passed = sock_info->constructor_sem_posted;
+       /*
+        * 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);
 
-       if (ctor_passed && sock_info->session_enabled) {
-               sock_info->session_enabled = 0;
-               lttng_handle_pending_statedumps(&lttng_ust_baddr_statedump);
-       }
+       return ret;
 }
 
 static
@@ -758,9 +829,14 @@ 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;
        }
@@ -933,15 +1009,20 @@ 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);
@@ -963,8 +1044,7 @@ 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) {
@@ -1069,9 +1149,7 @@ 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;
                }
 
@@ -1086,9 +1164,7 @@ restart:
        }
        sock_info->socket = ret;
 
-       ust_lock();
-
-       if (lttng_ust_comm_should_quit) {
+       if (ust_lock()) {
                goto quit;
        }
 
@@ -1129,9 +1205,7 @@ 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;
                }
 
@@ -1168,9 +1242,7 @@ restart:
                WARN("Unsupported timeout value %ld", timeout);
        }
 
-       ust_lock();
-
-       if (lttng_ust_comm_should_quit) {
+       if (ust_lock()) {
                goto quit;
        }
 
@@ -1201,8 +1273,7 @@ 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;
                        }
                        /*
@@ -1224,8 +1295,6 @@ restart:
                        ret = handle_message(sock_info, sock, &lum);
                        if (ret) {
                                ERR("Error handling message for %s socket", sock_info->name);
-                       } else {
-                               handle_pending_statedumps(sock_info);
                        }
                        continue;
                default:
@@ -1243,8 +1312,7 @@ restart:
 
        }
 end:
-       ust_lock();
-       if (lttng_ust_comm_should_quit) {
+       if (ust_lock()) {
                goto quit;
        }
        /* Cleanup socket handles before trying to reconnect */
@@ -1253,11 +1321,22 @@ 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.
@@ -1278,6 +1357,7 @@ 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();
@@ -1291,12 +1371,17 @@ 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_context_init();
+       /*
+        * Invoke ust malloc wrapper init before starting other threads.
+        */
+       lttng_ust_malloc_wrapper_init();
 
        timeout_mode = get_constructor_timeout(&constructor_timeout);
 
@@ -1328,24 +1413,24 @@ 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);
        }
@@ -1405,6 +1490,7 @@ void lttng_ust_cleanup(int exiting)
        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 */
@@ -1429,9 +1515,11 @@ 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);
@@ -1451,7 +1539,7 @@ 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
@@ -1490,7 +1578,7 @@ void ust_before_fork(sigset_t *save_sigset)
        if (ret == -1) {
                PERROR("sigprocmask");
        }
-       ust_lock();
+       ust_lock_nocheck();
        rcu_bp_before_fork();
 }
 
@@ -1543,5 +1631,5 @@ void ust_after_fork_child(sigset_t *restore_sigset)
 void lttng_ust_sockinfo_session_enabled(void *owner)
 {
        struct sock_info *sock_info = owner;
-       sock_info->session_enabled = 1;
+       sock_info->statedump_pending = 1;
 }
This page took 0.028764 seconds and 4 git commands to generate.