Centralize locking with ust_lock, fork handling
[lttng-ust.git] / libust / lttng-ust-comm.c
index 037b314765ab12b6974f1ca26e294d1771fde504..b6c5de51e456e5e3d8f501fd75265819eeebf199 100644 (file)
 
 #include <sys/types.h>
 #include <sys/socket.h>
+#include <sys/prctl.h>
 #include <unistd.h>
 #include <errno.h>
-#include <ust/lttng-ust-abi.h>
-#include <lttng-ust-comm.h>
-#include <ust/usterr-signal-safe.h>
 #include <pthread.h>
+#include <semaphore.h>
+#include <time.h>
 #include <assert.h>
+#include <signal.h>
+#include <urcu/uatomic.h>
+
+#include <lttng-ust-comm.h>
+#include <ust/usterr-signal-safe.h>
+#include <ust/lttng-ust-abi.h>
+#include <ust/tracepoint.h>
+#include <ust/tracepoint-internal.h>
+#include <ust/ust.h>
 
 /*
- * communication thread mutex. Held when handling a command, also held
- * by fork() to deal with removal of threads, and by exit path.
+ * Has lttng ust comm constructor been called ?
+ */
+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.
  */
-static pthread_mutex_t lttng_ust_comm_mutex = PTHREAD_MUTEX_INITIALIZER;
 
 /* Should the ust comm thread quit ? */
 static int lttng_ust_comm_should_quit;
 
+/*
+ * Wait for either of these before continuing to the main
+ * program:
+ * - the register_done message from sessiond daemon
+ *   (will let the sessiond daemon enable sessions before main
+ *   starts.)
+ * - sessiond daemon is not reachable.
+ * - timeout (ensuring applications are resilient to session
+ *   daemon problems).
+ */
+static sem_t constructor_wait;
+/*
+ * Doing this for both the global and local sessiond.
+ */
+static int sem_count = { 2 };
+
 /*
  * Info about socket and associated listener thread.
  */
 struct sock_info {
+       const char *name;
        char sock_path[PATH_MAX];
        int socket;
        pthread_t ust_listener; /* listener thread */
        int root_handle;
+       int constructor_sem_posted;
+       int allowed;
 };
 
 /* Socket from app (connect) to session daemon (listen) for communication */
 struct sock_info global_apps = {
+       .name = "global",
        .sock_path = DEFAULT_GLOBAL_APPS_UNIX_SOCK,
        .socket = -1,
        .root_handle = -1,
+       .allowed = 1,
 };
 
 /* TODO: allow global_apps_sock_path override */
 
 struct sock_info local_apps = {
+       .name = "local",
        .socket = -1,
        .root_handle = -1,
+       .allowed = 0,   /* Check setuid bit first */
 };
 
+extern void ltt_ring_buffer_client_overwrite_init(void);
+extern void ltt_ring_buffer_client_discard_init(void);
+extern void ltt_ring_buffer_metadata_client_init(void);
+extern void ltt_ring_buffer_client_overwrite_exit(void);
+extern void ltt_ring_buffer_client_discard_exit(void);
+extern void ltt_ring_buffer_metadata_client_exit(void);
+
 static
-int setup_local_apps_socket(void)
+int setup_local_apps(void)
 {
        const char *home_dir;
 
+       /*
+        * Disallow per-user tracing for setuid binaries.
+        */
+       if (getuid() != geteuid()) {
+               local_apps.allowed = 0;
+               return 0;
+       } else {
+               local_apps.allowed = 1;
+       }
        home_dir = (const char *) getenv("HOME");
        if (!home_dir)
                return -ENOENT;
@@ -79,17 +132,28 @@ static
 int register_app_to_sessiond(int socket)
 {
        ssize_t ret;
+       int prctl_ret;
        struct {
                uint32_t major;
                uint32_t minor;
                pid_t pid;
+               pid_t ppid;
                uid_t uid;
+               gid_t gid;
+               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();
+       prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0);
+       if (prctl_ret) {
+               ERR("Error executing prctl");
+               return -errno;
+       }
 
        ret = lttcomm_send_unix_sock(socket, &reg_msg, sizeof(reg_msg));
        if (ret >= 0 && ret != sizeof(reg_msg))
@@ -120,13 +184,30 @@ int send_reply(int sock, struct lttcomm_ust_reply *lur)
 }
 
 static
-int handle_message(int sock, struct lttcomm_ust_msg *lum)
+int handle_register_done(struct sock_info *sock_info)
+{
+       int ret;
+
+       if (sock_info->constructor_sem_posted)
+               return 0;
+       sock_info->constructor_sem_posted = 1;
+       ret = uatomic_add_return(&sem_count, -1);
+       if (ret == 0) {
+               ret = sem_post(&constructor_wait);
+               assert(!ret);
+       }
+       return 0;
+}
+
+static
+int handle_message(struct sock_info *sock_info,
+               int sock, struct lttcomm_ust_msg *lum)
 {
        int ret = 0;
        const struct objd_ops *ops;
        struct lttcomm_ust_reply lur;
 
-       pthread_mutex_lock(&lttng_ust_comm_mutex);
+       ust_lock();
 
        memset(&lur, 0, sizeof(lur));
 
@@ -142,6 +223,12 @@ int handle_message(int sock, struct lttcomm_ust_msg *lum)
        }
 
        switch (lum->cmd) {
+       case LTTNG_UST_REGISTER_DONE:
+               if (lum->handle == LTTNG_UST_ROOT_HANDLE)
+                       ret = handle_register_done(sock_info);
+               else
+                       ret = -EINVAL;
+               break;
        case LTTNG_UST_RELEASE:
                if (lum->handle == LTTNG_UST_ROOT_HANDLE)
                        ret = -EPERM;
@@ -168,7 +255,7 @@ end:
        }
        ret = send_reply(sock, &lur);
 
-       pthread_mutex_unlock(&lttng_ust_comm_mutex);
+       ust_unlock();
        return ret;
 }
 
@@ -208,17 +295,17 @@ void *ust_listener_thread(void *arg)
 
        /* Restart trying to connect to the session daemon */
 restart:
-       pthread_mutex_lock(&lttng_ust_comm_mutex);
+       ust_lock();
 
        if (lttng_ust_comm_should_quit) {
-               pthread_mutex_unlock(&lttng_ust_comm_mutex);
+               ust_unlock();
                goto quit;
        }
 
        if (sock_info->socket != -1) {
                ret = close(sock_info->socket);
                if (ret) {
-                       ERR("Error closing local apps socket");
+                       ERR("Error closing %s apps socket", sock_info->name);
                }
                sock_info->socket = -1;
        }
@@ -228,8 +315,14 @@ restart:
        /* Register */
        ret = lttcomm_connect_unix_sock(sock_info->sock_path);
        if (ret < 0) {
-               ERR("Error connecting to global apps socket");
-               pthread_mutex_unlock(&lttng_ust_comm_mutex);
+               ERR("Error connecting to %s apps socket", sock_info->name);
+               /*
+                * If we cannot find the sessiond daemon, don't delay
+                * constructor execution.
+                */
+               ret = handle_register_done(sock_info);
+               assert(!ret);
+               ust_unlock();
                sleep(5);
                goto restart;
        }
@@ -244,7 +337,7 @@ restart:
                ret = lttng_abi_create_root_handle();
                if (ret) {
                        ERR("Error creating root handle");
-                       pthread_mutex_unlock(&lttng_ust_comm_mutex);
+                       ust_unlock();
                        goto quit;
                }
                sock_info->root_handle = ret;
@@ -252,12 +345,18 @@ restart:
 
        ret = register_app_to_sessiond(sock);
        if (ret < 0) {
-               ERR("Error registering app to local apps socket");
-               pthread_mutex_unlock(&lttng_ust_comm_mutex);
+               ERR("Error registering to %s apps socket", sock_info->name);
+               /*
+                * If we cannot register to the sessiond daemon, don't
+                * delay constructor execution.
+                */
+               ret = handle_register_done(sock_info);
+               assert(!ret);
+               ust_unlock();
                sleep(5);
                goto restart;
        }
-       pthread_mutex_unlock(&lttng_ust_comm_mutex);
+       ust_unlock();
 
        for (;;) {
                ssize_t len;
@@ -266,23 +365,23 @@ restart:
                len = lttcomm_recv_unix_sock(sock, &lum, sizeof(lum));
                switch (len) {
                case 0: /* orderly shutdown */
-                       DBG("ltt-sessiond has performed an orderly shutdown\n");
+                       DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
                        goto end;
                case sizeof(lum):
                        DBG("message received\n");
-                       ret = handle_message(sock, &lum);
+                       ret = handle_message(sock_info, sock, &lum);
                        if (ret < 0) {
-                               ERR("Error handling message\n");
+                               ERR("Error handling message for %s socket", sock_info->name);
                        }
                        continue;
                case -1:
                        if (errno == ECONNRESET) {
-                               ERR("remote end closed connection\n");
+                               ERR("%s remote end closed connection\n", sock_info->name);
                                goto end;
                        }
                        goto end;
                default:
-                       ERR("incorrect message size: %zd\n", len);
+                       ERR("incorrect message size (%s socket): %zd\n", sock_info->name, len);
                        continue;
                }
 
@@ -293,6 +392,45 @@ quit:
        return NULL;
 }
 
+/*
+ * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
+ */
+static
+int get_timeout(struct timespec *constructor_timeout)
+{
+       long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
+       char *str_delay;
+       int ret;
+
+       str_delay = getenv("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;
+}
 
 /*
  * sessiond monitoring thread: monitor presence of global and per-user
@@ -300,25 +438,91 @@ quit:
  */
 /* TODO */
 
-void __attribute__((constructor)) lttng_ust_comm_init(void)
+void __attribute__((constructor)) lttng_ust_init(void)
 {
+       struct timespec constructor_timeout;
+       int timeout_mode;
        int ret;
 
+       if (uatomic_xchg(&initialized, 1) == 1)
+               return;
+
+       /*
+        * We want precise control over the order in which we construct
+        * our sub-libraries vs starting to receive commands from
+        * sessiond (otherwise leading to errors when trying to create
+        * sessiond before the init functions are completed).
+        */
        init_usterr();
+       init_tracepoint();
+       ltt_ring_buffer_metadata_client_init();
+       ltt_ring_buffer_client_overwrite_init();
+       ltt_ring_buffer_client_discard_init();
+
+       timeout_mode = get_timeout(&constructor_timeout);
+
+       ret = sem_init(&constructor_wait, 0, 0);
+       assert(!ret);
 
-       ret = setup_local_apps_socket();
+       ret = setup_local_apps();
        if (ret) {
-               ERR("Error setting up to local apps socket");
+               ERR("Error setting up to local apps");
        }
-#if 0
-       ret = pthread_create(&global_apps.ust_listener, NULL,
-                       ust_listener_thread, &global_apps);
-#endif //0
        ret = pthread_create(&local_apps.ust_listener, NULL,
                        ust_listener_thread, &local_apps);
+
+       if (local_apps.allowed) {
+               ret = pthread_create(&global_apps.ust_listener, NULL,
+                               ust_listener_thread, &global_apps);
+       } else {
+               handle_register_done(&local_apps);
+       }
+
+       switch (timeout_mode) {
+       case 1: /* timeout wait */
+               do {
+                       ret = sem_timedwait(&constructor_wait,
+                                       &constructor_timeout);
+               } while (ret < 0 && errno == EINTR);
+               if (ret < 0 && errno == ETIMEDOUT) {
+                       ERR("Timed out waiting for ltt-sessiond");
+               } else {
+                       assert(!ret);
+               }
+               break;
+       case -1:/* wait forever */
+               do {
+                       ret = sem_wait(&constructor_wait);
+               } while (ret < 0 && errno == EINTR);
+               assert(!ret);
+               break;
+       case 0: /* no timeout */
+               break;
+       }
 }
 
-void __attribute__((destructor)) lttng_ust_comm_exit(void)
+static
+void lttng_ust_cleanup(int exiting)
+{
+       cleanup_sock_info(&global_apps);
+       if (local_apps.allowed) {
+               cleanup_sock_info(&local_apps);
+       }
+       lttng_ust_abi_exit();
+       ltt_events_exit();
+       ltt_ring_buffer_client_discard_exit();
+       ltt_ring_buffer_client_overwrite_exit();
+       ltt_ring_buffer_metadata_client_exit();
+       exit_tracepoint();
+       if (!exiting) {
+               /* Reinitialize values for fork */
+               sem_count = 2;
+               lttng_ust_comm_should_quit = 0;
+               initialized = 0;
+       }
+}
+
+void __attribute__((destructor)) lttng_ust_exit(void)
 {
        int ret;
 
@@ -333,26 +537,88 @@ void __attribute__((destructor)) lttng_ust_comm_exit(void)
         * mutexes to ensure it is not in a mutex critical section when
         * pthread_cancel is later called.
         */
-       pthread_mutex_lock(&lttng_ust_comm_mutex);
+       ust_lock();
        lttng_ust_comm_should_quit = 1;
-       pthread_mutex_unlock(&lttng_ust_comm_mutex);
+       ust_unlock();
 
-#if 0
        ret = pthread_cancel(global_apps.ust_listener);
        if (ret) {
                ERR("Error cancelling global ust listener thread");
        }
-#endif //0
+       if (local_apps.allowed) {
+               ret = pthread_cancel(local_apps.ust_listener);
+               if (ret) {
+                       ERR("Error cancelling local ust listener thread");
+               }
+       }
+       lttng_ust_cleanup(1);
+}
 
-       cleanup_sock_info(&global_apps);
+/*
+ * We exclude the worker threads across fork and clone (except
+ * CLONE_VM), because these system calls only keep the forking thread
+ * running in the child.  Therefore, we don't want to call fork or clone
+ * in the middle of an tracepoint or ust tracing state modification.
+ * Holding this mutex protects these structures across fork and clone.
+ */
+void ust_before_fork(ust_fork_info_t *fork_info)
+{
+       /*
+        * Disable signals. This is to avoid that the child intervenes
+        * before it is properly setup for tracing. It is safer to
+        * disable all signals, because then we know we are not breaking
+        * anything by restoring the original mask.
+         */
+       sigset_t all_sigs;
+       int ret;
 
-       ret = pthread_cancel(local_apps.ust_listener);
-       if (ret) {
-               ERR("Error cancelling local ust listener thread");
+       /* Disable signals */
+       sigfillset(&all_sigs);
+       ret = sigprocmask(SIG_BLOCK, &all_sigs, &fork_info->orig_sigs);
+       if (ret == -1) {
+               PERROR("sigprocmask");
        }
+       ust_lock();
+       rcu_bp_before_fork();
+}
 
-       cleanup_sock_info(&local_apps);
+static void ust_after_fork_common(ust_fork_info_t *fork_info)
+{
+       int ret;
 
-       lttng_ust_abi_exit();
-       ltt_events_exit();
+       DBG("process %d", getpid());
+       ust_unlock();
+       /* Restore signals */
+       ret = sigprocmask(SIG_SETMASK, &fork_info->orig_sigs, NULL);
+       if (ret == -1) {
+               PERROR("sigprocmask");
+       }
+}
+
+void ust_after_fork_parent(ust_fork_info_t *fork_info)
+{
+       DBG("process %d", getpid());
+       rcu_bp_after_fork_parent();
+       /* Release mutexes and reenable signals */
+       ust_after_fork_common(fork_info);
+}
+
+/*
+ * After fork, in the child, we need to cleanup all the leftover state,
+ * except the worker thread which already magically disappeared thanks
+ * to the weird Linux fork semantics. After tyding up, we call
+ * lttng_ust_init() again to start over as a new PID.
+ *
+ * This is meant for forks() that have tracing in the child between the
+ * fork and following exec call (if there is any).
+ */
+void ust_after_fork_child(ust_fork_info_t *fork_info)
+{
+       DBG("process %d", getpid());
+       /* Release urcu mutexes */
+       rcu_bp_after_fork_child();
+       lttng_ust_cleanup(0);
+       lttng_ust_init();
+       /* Release mutexes and reenable signals */
+       ust_after_fork_common(fork_info);
 }
This page took 0.029856 seconds and 4 git commands to generate.