Fix: handle pthread errors
[lttng-ust.git] / liblttng-ust / lttng-ust-comm.c
index d0641c5c6e510b01d0538d9c36b07c5c430644c6..37f21c80d01ea4881e54b74c5b49e998a8588879 100644 (file)
@@ -46,6 +46,7 @@
 #include "tracepoint-internal.h"
 #include "ltt-tracer-core.h"
 #include "compat.h"
+#include "../libringbuffer/tlsfixup.h"
 
 /*
  * Has lttng ust comm constructor been called ?
@@ -235,6 +236,7 @@ int handle_message(struct sock_info *sock_info,
        struct ustcomm_ust_reply lur;
        int shm_fd, wait_fd;
        union ust_args args;
+       ssize_t len;
 
        ust_lock();
 
@@ -284,28 +286,30 @@ end:
                //lur.ret_code = USTCOMM_SESSION_FAIL;
                lur.ret_code = ret;
        }
-       switch (lum->cmd) {
-       case LTTNG_UST_STREAM:
-               /*
-                * Special-case reply to send stream info.
-                * Use lum.u output.
-                */
-               lur.u.stream.memory_map_size = *args.stream.memory_map_size;
-               shm_fd = *args.stream.shm_fd;
-               wait_fd = *args.stream.wait_fd;
-               break;
-       case LTTNG_UST_METADATA:
-       case LTTNG_UST_CHANNEL:
-               lur.u.channel.memory_map_size = *args.channel.memory_map_size;
-               shm_fd = *args.channel.shm_fd;
-               wait_fd = *args.channel.wait_fd;
-               break;
-       case LTTNG_UST_TRACER_VERSION:
-               lur.u.version = lum->u.version;
-               break;
-       case LTTNG_UST_TRACEPOINT_LIST_GET:
-               memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
-               break;
+       if (ret >= 0) {
+               switch (lum->cmd) {
+               case LTTNG_UST_STREAM:
+                       /*
+                        * Special-case reply to send stream info.
+                        * Use lum.u output.
+                        */
+                       lur.u.stream.memory_map_size = *args.stream.memory_map_size;
+                       shm_fd = *args.stream.shm_fd;
+                       wait_fd = *args.stream.wait_fd;
+                       break;
+               case LTTNG_UST_METADATA:
+               case LTTNG_UST_CHANNEL:
+                       lur.u.channel.memory_map_size = *args.channel.memory_map_size;
+                       shm_fd = *args.channel.shm_fd;
+                       wait_fd = *args.channel.wait_fd;
+                       break;
+               case LTTNG_UST_TRACER_VERSION:
+                       lur.u.version = lum->u.version;
+                       break;
+               case LTTNG_UST_TRACEPOINT_LIST_GET:
+                       memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
+                       break;
+               }
        }
        ret = send_reply(sock, &lur);
        if (ret < 0) {
@@ -317,14 +321,20 @@ end:
             || lum->cmd == LTTNG_UST_CHANNEL
             || lum->cmd == LTTNG_UST_METADATA)
                        && lur.ret_code == USTCOMM_OK) {
+               int sendret = 0;
+
                /* we also need to send the file descriptors. */
                ret = ustcomm_send_fds_unix_sock(sock,
                        &shm_fd, &shm_fd,
                        1, sizeof(int));
                if (ret < 0) {
                        perror("send shm_fd");
-                       goto error;
+                       sendret = ret;
                }
+               /*
+                * The sessiond expects 2 file descriptors, even upon
+                * error.
+                */
                ret = ustcomm_send_fds_unix_sock(sock,
                        &wait_fd, &wait_fd,
                        1, sizeof(int));
@@ -332,6 +342,26 @@ end:
                        perror("send wait_fd");
                        goto error;
                }
+               if (sendret) {
+                       ret = sendret;
+                       goto error;
+               }
+       }
+       /*
+        * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
+        * after the reply.
+        */
+       if (lur.ret_code == USTCOMM_OK) {
+               switch (lum->cmd) {
+               case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
+                       len = ustcomm_send_unix_sock(sock,
+                               &args.field_list.entry,
+                               sizeof(args.field_list.entry));
+                       if (len != sizeof(args.field_list.entry)) {
+                               ret = -1;
+                               goto error;
+                       }
+               }
        }
        /*
         * We still have the memory map reference, and the fds have been
@@ -383,12 +413,12 @@ error:
 }
 
 static
-void cleanup_sock_info(struct sock_info *sock_info)
+void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 {
        int ret;
 
        if (sock_info->socket != -1) {
-               ret = close(sock_info->socket);
+               ret = ustcomm_close_unix_sock(sock_info->socket);
                if (ret) {
                        ERR("Error closing apps socket");
                }
@@ -402,7 +432,13 @@ void cleanup_sock_info(struct sock_info *sock_info)
                sock_info->root_handle = -1;
        }
        sock_info->constructor_sem_posted = 0;
-       if (sock_info->wait_shm_mmap) {
+       /*
+        * 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
+        * process exit.
+        */
+       if (!exiting && sock_info->wait_shm_mmap) {
                ret = munmap(sock_info->wait_shm_mmap, sysconf(_SC_PAGE_SIZE));
                if (ret) {
                        ERR("Error unmapping wait shm");
@@ -662,7 +698,7 @@ restart:
        }
 
        if (sock_info->socket != -1) {
-               ret = close(sock_info->socket);
+               ret = ustcomm_close_unix_sock(sock_info->socket);
                if (ret) {
                        ERR("Error closing %s apps socket", sock_info->name);
                }
@@ -813,12 +849,22 @@ int get_timeout(struct timespec *constructor_timeout)
 void __attribute__((constructor)) lttng_ust_init(void)
 {
        struct timespec constructor_timeout;
+       sigset_t sig_all_blocked, orig_parent_mask;
        int timeout_mode;
        int ret;
 
        if (uatomic_xchg(&initialized, 1) == 1)
                return;
 
+       /*
+        * Fixup interdependency between TLS fixup mutex (which happens
+        * to be the dynamic linker mutex) and ust_lock, taken within
+        * the ust lock.
+        */
+       lttng_fixup_event_tls();
+       lttng_fixup_ringbuffer_tls();
+       lttng_fixup_vtid_tls();
+
        /*
         * We want precise control over the order in which we construct
         * our sub-libraries vs starting to receive commands from
@@ -840,16 +886,39 @@ void __attribute__((constructor)) lttng_ust_init(void)
        if (ret) {
                ERR("Error setting up to local apps");
        }
-       ret = pthread_create(&local_apps.ust_listener, NULL,
-                       ust_listener_thread, &local_apps);
 
+       /* A new thread created by pthread_create inherits the signal mask
+        * from the parent. To avoid any signal being received by the
+        * listener thread, we block all signals temporarily in the parent,
+        * while we create the listener thread.
+        */
+       sigfillset(&sig_all_blocked);
+       ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+
+       ret = pthread_create(&global_apps.ust_listener, NULL,
+                       ust_listener_thread, &global_apps);
+       if (ret) {
+               ERR("pthread_create global: %s", strerror(ret));
+       }
        if (local_apps.allowed) {
-               ret = pthread_create(&global_apps.ust_listener, NULL,
-                               ust_listener_thread, &global_apps);
+               ret = pthread_create(&local_apps.ust_listener, NULL,
+                               ust_listener_thread, &local_apps);
+               if (ret) {
+                       ERR("pthread_create local: %s", strerror(ret));
+               }
        } else {
                handle_register_done(&local_apps);
        }
 
+       /* Restore original signal mask in parent */
+       ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
+       if (ret) {
+               ERR("pthread_sigmask: %s", strerror(ret));
+       }
+
        switch (timeout_mode) {
        case 1: /* timeout wait */
                do {
@@ -876,10 +945,17 @@ void __attribute__((constructor)) lttng_ust_init(void)
 static
 void lttng_ust_cleanup(int exiting)
 {
-       cleanup_sock_info(&global_apps);
+       cleanup_sock_info(&global_apps, exiting);
        if (local_apps.allowed) {
-               cleanup_sock_info(&local_apps);
+               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
+        * lock, along with the lttng_ust_comm_should_quit flag, ensure
+        * that none of these threads are accessing this data at this
+        * point.
+        */
        lttng_ust_abi_exit();
        lttng_ust_events_exit();
        ltt_ring_buffer_client_discard_exit();
@@ -913,16 +989,27 @@ void __attribute__((destructor)) lttng_ust_exit(void)
        lttng_ust_comm_should_quit = 1;
        ust_unlock();
 
+       /* cancel threads */
        ret = pthread_cancel(global_apps.ust_listener);
        if (ret) {
-               ERR("Error cancelling global ust listener thread");
+               ERR("Error cancelling global ust listener thread: %s",
+                       strerror(ret));
        }
        if (local_apps.allowed) {
                ret = pthread_cancel(local_apps.ust_listener);
                if (ret) {
-                       ERR("Error cancelling local ust listener thread");
+                       ERR("Error cancelling local ust listener thread: %s",
+                               strerror(ret));
                }
        }
+       /*
+        * Do NOT join threads: use of sys_futex makes it impossible to
+        * join the threads without using async-cancel, but async-cancel
+        * is delivered by a signal, which could hit the target thread
+        * anywhere in its code path, including while the ust_lock() is
+        * held, causing a deadlock for the other thread. Let the OS
+        * cleanup the threads if there are stalled in a syscall.
+        */
        lttng_ust_cleanup(1);
 }
 
This page took 0.026344 seconds and 4 git commands to generate.