fix: on exit, leave thread/mmap reclaim to OS
[lttng-ust.git] / liblttng-ust / lttng-ust-comm.c
index 71bf716a3ab263e6ee34b5443f64e7ece712be41..86cce18d57fe658588db907ed8016f39e95f975f 100644 (file)
@@ -22,7 +22,6 @@
 #define _LGPL_SOURCE
 #include <sys/types.h>
 #include <sys/socket.h>
-#include <sys/prctl.h>
 #include <sys/mman.h>
 #include <sys/stat.h>
 #include <sys/types.h>
 #include <signal.h>
 #include <urcu/uatomic.h>
 #include <urcu/futex.h>
+#include <urcu/compiler.h>
 
-#include <lttng/ust-comm.h>
 #include <lttng/ust-events.h>
-#include <lttng/usterr-signal-safe.h>
 #include <lttng/ust-abi.h>
-#include <lttng/tracepoint.h>
-#include <lttng/tracepoint-internal.h>
 #include <lttng/ust.h>
+#include <ust-comm.h>
+#include <usterr-signal-safe.h>
+#include "tracepoint-internal.h"
 #include "ltt-tracer-core.h"
+#include "compat.h"
 
 /*
  * Has lttng ust comm constructor been called ?
@@ -159,7 +159,6 @@ static
 int register_app_to_sessiond(int socket)
 {
        ssize_t ret;
-       int prctl_ret;
        struct {
                uint32_t major;
                uint32_t minor;
@@ -167,6 +166,7 @@ int register_app_to_sessiond(int socket)
                pid_t ppid;
                uid_t uid;
                gid_t gid;
+               uint32_t bits_per_long;
                char name[16];  /* process name */
        } reg_msg;
 
@@ -176,11 +176,8 @@ int register_app_to_sessiond(int socket)
        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;
-       }
+       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))
@@ -237,6 +234,7 @@ int handle_message(struct sock_info *sock_info,
        const struct lttng_ust_objd_ops *ops;
        struct ustcomm_ust_reply lur;
        int shm_fd, wait_fd;
+       union ust_args args;
 
        ust_lock();
 
@@ -269,7 +267,8 @@ int handle_message(struct sock_info *sock_info,
        default:
                if (ops->cmd)
                        ret = ops->cmd(lum->handle, lum->cmd,
-                                       (unsigned long) &lum->u);
+                                       (unsigned long) &lum->u,
+                                       &args);
                else
                        ret = -ENOSYS;
                break;
@@ -291,23 +290,21 @@ end:
                 * Special-case reply to send stream info.
                 * Use lum.u output.
                 */
-               lur.u.stream.memory_map_size = lum->u.stream.memory_map_size;
-               shm_fd = lum->u.stream.shm_fd;
-               wait_fd = lum->u.stream.wait_fd;
+               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 = lum->u.channel.memory_map_size;
-               shm_fd = lum->u.channel.shm_fd;
-               wait_fd = lum->u.channel.wait_fd;
+               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_VERSION:
+       case LTTNG_UST_TRACER_VERSION:
                lur.u.version = lum->u.version;
                break;
        case LTTNG_UST_TRACEPOINT_LIST_GET:
-               memcpy(lur.u.tracepoint_list_entry,
-                       lum->u.tracepoint_list_entry,
-                       LTTNG_UST_SYM_NAME_LEN);
+               memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
                break;
        }
        ret = send_reply(sock, &lur);
@@ -336,13 +333,57 @@ end:
                        goto error;
                }
        }
+       /*
+        * We still have the memory map reference, and the fds have been
+        * sent to the sessiond. We can therefore close those fds. Note
+        * that we keep the write side of the wait_fd open, but close
+        * the read side.
+        */
+       if (lur.ret_code == USTCOMM_OK) {
+               switch (lum->cmd) {
+               case LTTNG_UST_STREAM:
+                       if (shm_fd >= 0) {
+                               ret = close(shm_fd);
+                               if (ret) {
+                                       PERROR("Error closing stream shm_fd");
+                               }
+                               *args.stream.shm_fd = -1;
+                       }
+                       if (wait_fd >= 0) {
+                               ret = close(wait_fd);
+                               if (ret) {
+                                       PERROR("Error closing stream wait_fd");
+                               }
+                               *args.stream.wait_fd = -1;
+                       }
+                       break;
+               case LTTNG_UST_METADATA:
+               case LTTNG_UST_CHANNEL:
+                       if (shm_fd >= 0) {
+                               ret = close(shm_fd);
+                               if (ret) {
+                                       PERROR("Error closing channel shm_fd");
+                               }
+                               *args.channel.shm_fd = -1;
+                       }
+                       if (wait_fd >= 0) {
+                               ret = close(wait_fd);
+                               if (ret) {
+                                       PERROR("Error closing channel wait_fd");
+                               }
+                               *args.channel.wait_fd = -1;
+                       }
+                       break;
+               }
+       }
+
 error:
        ust_unlock();
        return ret;
 }
 
 static
-void cleanup_sock_info(struct sock_info *sock_info)
+void cleanup_sock_info(struct sock_info *sock_info, int exiting)
 {
        int ret;
 
@@ -361,7 +402,13 @@ 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");
@@ -413,7 +460,7 @@ int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
                 * shared memory map will have been created.
                 */
                pid = wait(&status);
-               if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
+               if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
                        wait_shm_fd = -1;
                        goto end;
                }
@@ -561,14 +608,15 @@ void wait_for_sessiond(struct sock_info *sock_info)
                if (ret < 0) {
                        if (errno == EFAULT) {
                                wait_poll_fallback = 1;
-                               WARN(
+                               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");
                        }
-                       PERROR("futex");
                }
        }
        return;
@@ -681,6 +729,20 @@ restart:
                switch (len) {
                case 0: /* orderly shutdown */
                        DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info->name);
+                       ust_lock();
+                       /*
+                        * Either sessiond has shutdown or refused us by closing the socket.
+                        * In either case, we don't want to delay construction execution,
+                        * and we need to wait before retry.
+                        */
+                       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 end;
                case sizeof(lum):
                        DBG("message received\n");
@@ -690,6 +752,7 @@ restart:
                        }
                        continue;
                case -1:
+                       DBG("Receive failed from lttng-sessiond with errno %d", errno);
                        if (errno == ECONNRESET) {
                                ERR("%s remote end closed connection\n", sock_info->name);
                                goto end;
@@ -819,10 +882,17 @@ 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();
@@ -856,6 +926,7 @@ 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");
@@ -866,6 +937,14 @@ void __attribute__((destructor)) lttng_ust_exit(void)
                        ERR("Error cancelling local ust listener thread");
                }
        }
+       /*
+        * 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);
 }
 
@@ -876,7 +955,7 @@ void __attribute__((destructor)) lttng_ust_exit(void)
  * 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)
+void ust_before_fork(sigset_t *save_sigset)
 {
        /*
         * Disable signals. This is to avoid that the child intervenes
@@ -889,7 +968,7 @@ void ust_before_fork(ust_fork_info_t *fork_info)
 
        /* Disable signals */
        sigfillset(&all_sigs);
-       ret = sigprocmask(SIG_BLOCK, &all_sigs, &fork_info->orig_sigs);
+       ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
        if (ret == -1) {
                PERROR("sigprocmask");
        }
@@ -897,25 +976,25 @@ void ust_before_fork(ust_fork_info_t *fork_info)
        rcu_bp_before_fork();
 }
 
-static void ust_after_fork_common(ust_fork_info_t *fork_info)
+static void ust_after_fork_common(sigset_t *restore_sigset)
 {
        int ret;
 
        DBG("process %d", getpid());
        ust_unlock();
        /* Restore signals */
-       ret = sigprocmask(SIG_SETMASK, &fork_info->orig_sigs, NULL);
+       ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
        if (ret == -1) {
                PERROR("sigprocmask");
        }
 }
 
-void ust_after_fork_parent(ust_fork_info_t *fork_info)
+void ust_after_fork_parent(sigset_t *restore_sigset)
 {
        DBG("process %d", getpid());
        rcu_bp_after_fork_parent();
        /* Release mutexes and reenable signals */
-       ust_after_fork_common(fork_info);
+       ust_after_fork_common(restore_sigset);
 }
 
 /*
@@ -927,7 +1006,7 @@ void ust_after_fork_parent(ust_fork_info_t *fork_info)
  * 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)
+void ust_after_fork_child(sigset_t *restore_sigset)
 {
        DBG("process %d", getpid());
        /* Release urcu mutexes */
@@ -935,6 +1014,6 @@ void ust_after_fork_child(ust_fork_info_t *fork_info)
        lttng_ust_cleanup(0);
        lttng_context_vtid_reset();
        /* Release mutexes and reenable signals */
-       ust_after_fork_common(fork_info);
+       ust_after_fork_common(restore_sigset);
        lttng_ust_init();
 }
This page took 0.028546 seconds and 4 git commands to generate.