Bugfix for #745 deadlock with baddr statedump+fork
[lttng-ust.git] / liblttng-ust / lttng-ust-comm.c
index 74d805aefb06c23e42c048498cf2861183d29dfd..a73808e1128f7b63fbb62422e5f7bccc5115a9b4 100644 (file)
@@ -67,9 +67,16 @@ static int initialized;
  * probe registration.
  *
  * ust_exit_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.
  */
 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
@@ -79,16 +86,38 @@ static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  */
 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 quilt.
+ * Return 0 on success, -1 if should quit.
  * The lock is taken in both cases.
+ * Signal-safe.
  */
 int ust_lock(void)
 {
-       pthread_mutex_lock(&ust_mutex);
+       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 {
@@ -100,15 +129,45 @@ int ust_lock(void)
  * 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)
 {
-       pthread_mutex_lock(&ust_mutex);
+       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)
 {
-       pthread_mutex_unlock(&ust_mutex);
+       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));
+       }
 }
 
 /*
@@ -265,6 +324,12 @@ 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)));
+}
+
 int lttng_get_notify_socket(void *owner)
 {
        struct sock_info *info = owner;
@@ -448,8 +513,10 @@ void handle_pending_statedump(struct sock_info *sock_info)
        int ctor_passed = sock_info->constructor_sem_posted;
 
        if (ctor_passed && sock_info->statedump_pending) {
+               pthread_mutex_lock(&ust_fork_mutex);
                sock_info->statedump_pending = 0;
                lttng_handle_pending_statedump(sock_info);
+               pthread_mutex_unlock(&ust_fork_mutex);
        }
 }
 
@@ -819,9 +886,14 @@ 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;
        }
@@ -994,15 +1066,20 @@ 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);
@@ -1341,6 +1418,7 @@ void __attribute__((constructor)) lttng_ust_init(void)
        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
@@ -1356,6 +1434,7 @@ void __attribute__((constructor)) lttng_ust_init(void)
        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.
@@ -1464,6 +1543,7 @@ 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();
@@ -1557,6 +1637,9 @@ void ust_before_fork(sigset_t *save_sigset)
        if (ret == -1) {
                PERROR("sigprocmask");
        }
+
+       pthread_mutex_lock(&ust_fork_mutex);
+
        ust_lock_nocheck();
        rcu_bp_before_fork();
 }
@@ -1567,6 +1650,9 @@ 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) {
This page took 0.024902 seconds and 4 git commands to generate.