Fix: Cleanup local_apps sock_info in lttng_ust_cleanup
[lttng-ust.git] / liblttng-ust / lttng-ust-comm.c
index a1ebcbe6fc83564a5fc4d9c0c0642403938da35f..049667bf298d2122ffb904fc554b96744eb0083f 100644 (file)
@@ -51,6 +51,8 @@
 #include "lttng-tracer-core.h"
 #include "compat.h"
 #include "../libringbuffer/tlsfixup.h"
+#include "lttng-ust-baddr.h"
+#include "getenv.h"
 
 /*
  * Has lttng ust comm constructor been called ?
@@ -61,11 +63,119 @@ 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:
@@ -102,9 +212,12 @@ struct sock_info {
 
        char sock_path[PATH_MAX];
        int socket;
+       int notify_socket;
 
        char wait_shm_path[PATH_MAX];
        char *wait_shm_mmap;
+       /* Keep track of lazy state dump not performed yet. */
+       int statedump_pending;
 };
 
 /* Socket from app (connect) to session daemon (listen) for communication */
@@ -116,10 +229,13 @@ struct sock_info global_apps = {
        .allowed = 1,
        .thread_active = 0,
 
-       .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
+       .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
        .socket = -1,
+       .notify_socket = -1,
 
-       .wait_shm_path = DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH,
+       .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
+
+       .statedump_pending = 0,
 };
 
 /* TODO: allow global_apps_sock_path override */
@@ -132,6 +248,9 @@ struct sock_info local_apps = {
        .thread_active = 0,
 
        .socket = -1,
+       .notify_socket = -1,
+
+       .statedump_pending = 0,
 };
 
 static int wait_poll_fallback;
@@ -169,15 +288,39 @@ static const char *cmd_name_mapping[] = {
 
        /* Event FD commands */
        [ LTTNG_UST_FILTER ] = "Create Filter",
+       [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 };
 
+static const char *str_timeout;
+static int got_timeout_env;
+
 extern void lttng_ring_buffer_client_overwrite_init(void);
+extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 extern void lttng_ring_buffer_client_discard_init(void);
+extern void lttng_ring_buffer_client_discard_rt_init(void);
 extern void lttng_ring_buffer_metadata_client_init(void);
 extern void lttng_ring_buffer_client_overwrite_exit(void);
+extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 extern void lttng_ring_buffer_client_discard_exit(void);
+extern void lttng_ring_buffer_client_discard_rt_exit(void);
 extern void lttng_ring_buffer_metadata_client_exit(void);
 
+/*
+ * Returns the HOME directory path. Caller MUST NOT free(3) the returned
+ * pointer.
+ */
+static
+const char *get_lttng_home_dir(void)
+{
+       const char *val;
+
+       val = (const char *) lttng_secure_getenv("LTTNG_HOME");
+       if (val != NULL) {
+               return val;
+       }
+       return (const char *) lttng_secure_getenv("HOME");
+}
+
 /*
  * Force a read (imply TLS fixup for dlopen) of TLS variables.
  */
@@ -187,15 +330,40 @@ 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;
+
+       return info->notify_socket;
+}
+
 static
 void print_cmd(int cmd, int handle)
 {
        const char *cmd_name = "Unknown";
 
-       if (cmd_name_mapping[cmd]) {
+       if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
+                       && cmd_name_mapping[cmd]) {
                cmd_name = cmd_name_mapping[cmd];
        }
-       DBG("Message Received \"%s\", Handle \"%s\" (%d)", cmd_name,
+       DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
+               cmd_name, cmd,
                lttng_ust_obj_get_name(handle), handle);
 }
 
@@ -213,48 +381,96 @@ int setup_local_apps(void)
                assert(local_apps.allowed == 0);
                return 0;
        }
-       home_dir = (const char *) getenv("HOME");
+       home_dir = get_lttng_home_dir();
        if (!home_dir) {
                WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
                assert(local_apps.allowed == 0);
                return -ENOENT;
        }
        local_apps.allowed = 1;
-       snprintf(local_apps.sock_path, PATH_MAX,
-                DEFAULT_HOME_APPS_UNIX_SOCK, home_dir);
-       snprintf(local_apps.wait_shm_path, PATH_MAX,
-                DEFAULT_HOME_APPS_WAIT_SHM_PATH, uid);
+       snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
+               home_dir,
+               LTTNG_DEFAULT_HOME_RUNDIR,
+               LTTNG_UST_SOCK_FILENAME);
+       snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
+               LTTNG_UST_WAIT_FILENAME,
+               uid);
        return 0;
 }
 
+/*
+ * Get notify_sock timeout, in ms.
+ * -1: wait forever. 0: don't wait. >0: timeout, in ms.
+ */
 static
-int register_app_to_sessiond(int socket)
+long get_timeout(void)
 {
-       ssize_t ret;
-       struct {
-               uint32_t major;
-               uint32_t minor;
-               pid_t pid;
-               pid_t ppid;
-               uid_t uid;
-               gid_t gid;
-               uint32_t bits_per_long;
-               char name[16];  /* process name */
-       } reg_msg;
-
-       reg_msg.major = LTTNG_UST_COMM_VERSION_MAJOR;
-       reg_msg.minor = LTTNG_UST_COMM_VERSION_MINOR;
-       reg_msg.pid = getpid();
-       reg_msg.ppid = getppid();
-       reg_msg.uid = getuid();
-       reg_msg.gid = getgid();
-       reg_msg.bits_per_long = CAA_BITS_PER_LONG;
-       lttng_ust_getprocname(reg_msg.name);
-
-       ret = ustcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
-       if (ret >= 0 && ret != sizeof(reg_msg))
-               return -EIO;
-       return ret;
+       long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
+
+       if (!got_timeout_env) {
+               str_timeout = getenv("LTTNG_UST_REGISTER_TIMEOUT");
+               got_timeout_env = 1;
+       }
+       if (str_timeout)
+               constructor_delay_ms = strtol(str_timeout, NULL, 10);
+       return constructor_delay_ms;
+}
+
+static
+long get_notify_sock_timeout(void)
+{
+       return get_timeout();
+}
+
+/*
+ * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
+ */
+static
+int get_constructor_timeout(struct timespec *constructor_timeout)
+{
+       long constructor_delay_ms;
+       int ret;
+
+       constructor_delay_ms = get_timeout();
+
+       switch (constructor_delay_ms) {
+       case -1:/* fall-through */
+       case 0:
+               return constructor_delay_ms;
+       default:
+               break;
+       }
+
+       /*
+        * If we are unable to find the current time, don't wait.
+        */
+       ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
+       if (ret) {
+               /* Don't wait. */
+               return 0;
+       }
+       constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
+       constructor_timeout->tv_nsec +=
+               (constructor_delay_ms % 1000UL) * 1000000UL;
+       if (constructor_timeout->tv_nsec >= 1000000000UL) {
+               constructor_timeout->tv_sec++;
+               constructor_timeout->tv_nsec -= 1000000000UL;
+       }
+       /* Timeout wait (constructor_delay_ms). */
+       return 1;
+}
+
+static
+int register_to_sessiond(int socket, enum ustctl_socket_type type)
+{
+       return ustcomm_send_reg_msg(socket,
+               type,
+               CAA_BITS_PER_LONG,
+               lttng_alignof(uint8_t) * CHAR_BIT,
+               lttng_alignof(uint16_t) * CHAR_BIT,
+               lttng_alignof(uint32_t) * CHAR_BIT,
+               lttng_alignof(uint64_t) * CHAR_BIT,
+               lttng_alignof(unsigned long) * CHAR_BIT);
 }
 
 static
@@ -298,6 +514,30 @@ 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)
@@ -308,11 +548,9 @@ 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;
        }
@@ -334,7 +572,7 @@ int handle_message(struct sock_info *sock_info,
                if (lum->handle == LTTNG_UST_ROOT_HANDLE)
                        ret = -EPERM;
                else
-                       ret = lttng_ust_objd_unref(lum->handle);
+                       ret = lttng_ust_objd_unref(lum->handle, 1);
                break;
        case LTTNG_UST_FILTER:
        {
@@ -380,6 +618,7 @@ int handle_message(struct sock_info *sock_info,
                                        goto error;
                                }
                                ret = len;
+                               free(bytecode);
                                goto end;
                        } else {
                                DBG("incorrect filter data message size: %zd", len);
@@ -405,12 +644,76 @@ int handle_message(struct sock_info *sock_info,
                }
                break;
        }
+       case LTTNG_UST_EXCLUSION:
+       {
+               /* Receive exclusion names */
+               struct lttng_ust_excluder_node *node;
+               unsigned int count;
+
+               count = lum->u.exclusion.count;
+               if (count == 0) {
+                       /* There are no names to read */
+                       ret = 0;
+                       goto error;
+               }
+               node = zmalloc(sizeof(*node) +
+                               count * LTTNG_UST_SYM_NAME_LEN);
+               if (!node) {
+                       ret = -ENOMEM;
+                       goto error;
+               }
+               node->excluder.count = count;
+               len = ustcomm_recv_unix_sock(sock, node->excluder.names,
+                               count * LTTNG_UST_SYM_NAME_LEN);
+               switch (len) {
+               case 0: /* orderly shutdown */
+                       ret = 0;
+                       free(node);
+                       goto error;
+               default:
+                       if (len == count * LTTNG_UST_SYM_NAME_LEN) {
+                               DBG("Exclusion data received");
+                               break;
+                       } else if (len < 0) {
+                               DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
+                               if (len == -ECONNRESET) {
+                                       ERR("%s remote end closed connection", sock_info->name);
+                                       ret = len;
+                                       free(node);
+                                       goto error;
+                               }
+                               ret = len;
+                               free(node);
+                               goto end;
+                       } else {
+                               DBG("Incorrect exclusion data message size: %zd", len);
+                               ret = -EINVAL;
+                               free(node);
+                               goto end;
+                       }
+               }
+               if (ops->cmd) {
+                       ret = ops->cmd(lum->handle, lum->cmd,
+                                       (unsigned long) node,
+                                       &args, sock_info);
+                       if (ret) {
+                               free(node);
+                       }
+                       /* Don't free exclusion data if everything went fine. */
+               } else {
+                       ret = -ENOSYS;
+                       free(node);
+               }
+               break;
+       }
        case LTTNG_UST_CHANNEL:
        {
                void *chan_data;
+               int wakeup_fd;
 
                len = ustcomm_recv_channel_from_sessiond(sock,
-                               &chan_data, lum->u.channel.len);
+                               &chan_data, lum->u.channel.len,
+                               &wakeup_fd);
                switch (len) {
                case 0: /* orderly shutdown */
                        ret = 0;
@@ -435,6 +738,7 @@ int handle_message(struct sock_info *sock_info,
                        }
                }
                args.channel.chan_data = chan_data;
+               args.channel.wakeup_fd = wakeup_fd;
                if (ops->cmd)
                        ret = ops->cmd(lum->handle, lum->cmd,
                                        (unsigned long) &lum->u,
@@ -550,6 +854,14 @@ 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;
 }
 
@@ -558,31 +870,48 @@ void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 {
        int ret;
 
-       if (sock_info->socket != -1) {
-               ret = ustcomm_close_unix_sock(sock_info->socket);
-               if (ret) {
-                       ERR("Error closing apps socket");
-               }
-               sock_info->socket = -1;
-       }
        if (sock_info->root_handle != -1) {
-               ret = lttng_ust_objd_unref(sock_info->root_handle);
+               ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
                if (ret) {
                        ERR("Error unref root handle");
                }
                sock_info->root_handle = -1;
        }
        sock_info->constructor_sem_posted = 0;
+
        /*
-        * wait_shm_mmap is used by listener threads outside of the
-        * ust lock, so we cannot tear it down ourselves, because we
-        * cannot join on these threads. Leave this task to the OS
+        * wait_shm_mmap, socket and notify socket are used by listener
+        * threads outside of the ust lock, so we cannot tear them down
+        * ourselves, because we cannot join on these threads. Leave
+        * responsibility of cleaning up these resources to the OS
         * process exit.
         */
-       if (!exiting && sock_info->wait_shm_mmap) {
-               ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
+       if (exiting)
+               return;
+
+       if (sock_info->socket != -1) {
+               ret = ustcomm_close_unix_sock(sock_info->socket);
+               if (ret) {
+                       ERR("Error closing ust cmd socket");
+               }
+               sock_info->socket = -1;
+       }
+       if (sock_info->notify_socket != -1) {
+               ret = ustcomm_close_unix_sock(sock_info->notify_socket);
                if (ret) {
-                       ERR("Error unmapping wait shm");
+                       ERR("Error closing ust notify socket");
+               }
+               sock_info->notify_socket = -1;
+       }
+       if (sock_info->wait_shm_mmap) {
+               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;
        }
@@ -608,6 +937,30 @@ int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
         */
        wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
        if (wait_shm_fd >= 0) {
+               int32_t tmp_read;
+               ssize_t len;
+               size_t bytes_read = 0;
+
+               /*
+                * Try to read the fd. If unable to do so, try opening
+                * it in write mode.
+                */
+               do {
+                       len = read(wait_shm_fd,
+                               &((char *) &tmp_read)[bytes_read],
+                               sizeof(tmp_read) - bytes_read);
+                       if (len > 0) {
+                               bytes_read += len;
+                       }
+               } while ((len < 0 && errno == EINTR)
+                       || (len > 0 && bytes_read < sizeof(tmp_read)));
+               if (bytes_read != sizeof(tmp_read)) {
+                       ret = close(wait_shm_fd);
+                       if (ret) {
+                               ERR("close wait_shm_fd");
+                       }
+                       goto open_write;
+               }
                goto end;
        } else if (wait_shm_fd < 0 && errno != ENOENT) {
                /*
@@ -618,9 +971,11 @@ int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
                ERR("Error opening shm %s", sock_info->wait_shm_path);
                goto end;
        }
+
+open_write:
        /*
-        * If the open failed because the file did not exist, try
-        * creating it ourself.
+        * If the open failed because the file did not exist, or because
+        * the file was not truncated yet, try creating it ourself.
         */
        URCU_TLS(lttng_ust_nest_count)++;
        pid = fork();
@@ -729,15 +1084,20 @@ 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);
@@ -757,10 +1117,7 @@ error:
 static
 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) {
@@ -775,23 +1132,32 @@ void wait_for_sessiond(struct sock_info *sock_info)
 
        DBG("Waiting for %s apps sessiond", sock_info->name);
        /* Wait for futex wakeup */
-       if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
-               ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
-                       FUTEX_WAIT, 0, NULL, NULL, 0);
-               if (ret < 0) {
-                       if (errno == EFAULT) {
-                               wait_poll_fallback = 1;
-                               DBG(
+       if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
+               goto end_wait;
+
+       while (futex_async((int32_t *) sock_info->wait_shm_mmap,
+                       FUTEX_WAIT, 0, NULL, NULL, 0)) {
+               switch (errno) {
+               case EWOULDBLOCK:
+                       /* Value already changed. */
+                       goto end_wait;
+               case EINTR:
+                       /* Retry if interrupted by signal. */
+                       break;  /* Get out of switch. */
+               case EFAULT:
+                       wait_poll_fallback = 1;
+                       DBG(
 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
 "Please upgrade your kernel "
 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
 "mainline). LTTng-UST will use polling mode fallback.");
-                               if (ust_debug())
-                                       PERROR("futex");
-                       }
+                       if (ust_debug())
+                               PERROR("futex");
+                       goto end_wait;
                }
        }
+end_wait:
        return;
 
 quit:
@@ -815,6 +1181,7 @@ void *ust_listener_thread(void *arg)
 {
        struct sock_info *sock_info = arg;
        int sock, ret, prev_connect_failed = 0, has_waited = 0;
+       long timeout;
 
        /* Restart trying to connect to the session daemon */
 restart:
@@ -833,25 +1200,41 @@ restart:
                has_waited = 1;
                prev_connect_failed = 0;
        }
-       ust_lock();
-
-       if (lttng_ust_comm_should_quit) {
-               goto quit;
-       }
 
        if (sock_info->socket != -1) {
                ret = ustcomm_close_unix_sock(sock_info->socket);
                if (ret) {
-                       ERR("Error closing %s apps socket", sock_info->name);
+                       ERR("Error closing %s ust cmd socket",
+                               sock_info->name);
                }
                sock_info->socket = -1;
        }
+       if (sock_info->notify_socket != -1) {
+               ret = ustcomm_close_unix_sock(sock_info->notify_socket);
+               if (ret) {
+                       ERR("Error closing %s ust notify socket",
+                               sock_info->name);
+               }
+               sock_info->notify_socket = -1;
+       }
 
-       /* Register */
+       /*
+        * Register. We need to perform both connect and sending
+        * registration message before doing the next connect otherwise
+        * we may reach unix socket connect queue max limits and block
+        * on the 2nd connect while the session daemon is awaiting the
+        * first connect registration message.
+        */
+       /* Connect cmd socket */
        ret = ustcomm_connect_unix_sock(sock_info->sock_path);
        if (ret < 0) {
                DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
                prev_connect_failed = 1;
+
+               if (ust_lock()) {
+                       goto quit;
+               }
+
                /*
                 * If we cannot find the sessiond daemon, don't delay
                 * constructor execution.
@@ -861,8 +1244,11 @@ restart:
                ust_unlock();
                goto restart;
        }
+       sock_info->socket = ret;
 
-       sock_info->socket = sock = ret;
+       if (ust_lock()) {
+               goto quit;
+       }
 
        /*
         * Create only one root handle per listener thread for the whole
@@ -878,9 +1264,75 @@ restart:
                sock_info->root_handle = ret;
        }
 
-       ret = register_app_to_sessiond(sock);
+       ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
+       if (ret < 0) {
+               ERR("Error registering to %s ust cmd socket",
+                       sock_info->name);
+               prev_connect_failed = 1;
+               /*
+                * If we cannot register to the sessiond daemon, don't
+                * delay constructor execution.
+                */
+               ret = handle_register_done(sock_info);
+               assert(!ret);
+               ust_unlock();
+               goto restart;
+       }
+
+       ust_unlock();
+
+       /* Connect notify socket */
+       ret = ustcomm_connect_unix_sock(sock_info->sock_path);
        if (ret < 0) {
-               ERR("Error registering to %s apps socket", sock_info->name);
+               DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
+               prev_connect_failed = 1;
+
+               if (ust_lock()) {
+                       goto quit;
+               }
+
+               /*
+                * If we cannot find the sessiond daemon, don't delay
+                * constructor execution.
+                */
+               ret = handle_register_done(sock_info);
+               assert(!ret);
+               ust_unlock();
+               goto restart;
+       }
+       sock_info->notify_socket = ret;
+
+       timeout = get_notify_sock_timeout();
+       if (timeout >= 0) {
+               /*
+                * Give at least 10ms to sessiond to reply to
+                * notifications.
+                */
+               if (timeout < 10)
+                       timeout = 10;
+               ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
+                               timeout);
+               if (ret < 0) {
+                       WARN("Error setting socket receive timeout");
+               }
+               ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
+                               timeout);
+               if (ret < 0) {
+                       WARN("Error setting socket send timeout");
+               }
+       } else if (timeout < -1) {
+               WARN("Unsupported timeout value %ld", timeout);
+       }
+
+       if (ust_lock()) {
+               goto quit;
+       }
+
+       ret = register_to_sessiond(sock_info->notify_socket,
+                       USTCTL_SOCKET_NOTIFY);
+       if (ret < 0) {
+               ERR("Error registering to %s ust notify socket",
+                       sock_info->name);
                prev_connect_failed = 1;
                /*
                 * If we cannot register to the sessiond daemon, don't
@@ -891,6 +1343,8 @@ restart:
                ust_unlock();
                goto restart;
        }
+       sock = sock_info->socket;
+
        ust_unlock();
 
        for (;;) {
@@ -901,8 +1355,7 @@ 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;
                        }
                        /*
@@ -941,8 +1394,7 @@ restart:
 
        }
 end:
-       ust_lock();
-       if (lttng_ust_comm_should_quit) {
+       if (ust_lock()) {
                goto quit;
        }
        /* Cleanup socket handles before trying to reconnect */
@@ -951,49 +1403,20 @@ 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;
 }
 
 /*
- * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
+ * Weak symbol to call when the ust malloc wrapper is not loaded.
  */
-static
-int get_timeout(struct timespec *constructor_timeout)
+__attribute__((weak))
+void lttng_ust_malloc_wrapper_init(void)
 {
-       long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
-       char *str_delay;
-       int ret;
-
-       str_delay = getenv("LTTNG_UST_REGISTER_TIMEOUT");
-       if (str_delay) {
-               constructor_delay_ms = strtol(str_delay, NULL, 10);
-       }
-
-       switch (constructor_delay_ms) {
-       case -1:/* fall-through */
-       case 0:
-               return constructor_delay_ms;
-       default:
-               break;
-       }
-
-       /*
-        * If we are unable to find the current time, don't wait.
-        */
-       ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
-       if (ret) {
-               return -1;
-       }
-       constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
-       constructor_timeout->tv_nsec +=
-               (constructor_delay_ms % 1000UL) * 1000000UL;
-       if (constructor_timeout->tv_nsec >= 1000000000UL) {
-               constructor_timeout->tv_sec++;
-               constructor_timeout->tv_nsec -= 1000000000UL;
-       }
-       return 1;
 }
 
 /*
@@ -1016,10 +1439,12 @@ 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
@@ -1029,11 +1454,20 @@ 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_timeout(&constructor_timeout);
+       timeout_mode = get_constructor_timeout(&constructor_timeout);
 
        ret = sem_init(&constructor_wait, 0, 0);
        assert(!ret);
@@ -1063,24 +1497,24 @@ 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);
        }
@@ -1122,9 +1556,7 @@ static
 void lttng_ust_cleanup(int exiting)
 {
        cleanup_sock_info(&global_apps, exiting);
-       if (local_apps.allowed) {
-               cleanup_sock_info(&local_apps, exiting);
-       }
+       cleanup_sock_info(&local_apps, exiting);
        /*
         * The teardown in this function all affect data structures
         * accessed under the UST lock by the listener thread. This
@@ -1134,9 +1566,14 @@ 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 */
@@ -1161,9 +1598,11 @@ 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);
@@ -1183,7 +1622,7 @@ 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
@@ -1222,7 +1661,10 @@ 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();
 }
 
@@ -1232,6 +1674,9 @@ 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) {
@@ -1271,3 +1716,9 @@ void ust_after_fork_child(sigset_t *restore_sigset)
        ust_after_fork_common(restore_sigset);
        lttng_ust_init();
 }
+
+void lttng_ust_sockinfo_session_enabled(void *owner)
+{
+       struct sock_info *sock_info = owner;
+       sock_info->statedump_pending = 1;
+}
This page took 0.035155 seconds and 4 git commands to generate.