X-Git-Url: http://git.lttng.org/?a=blobdiff_plain;f=liblttng-ust%2Flttng-ust-comm.c;h=44e8bab4c24ca4aa2a77b72727288c9b24c8e67d;hb=e6ea14c54c9b1dd79428e61e0343d8a34d87b5a8;hp=ad02dd9d366faaf043f9b13bc638eef371b99a4a;hpb=4d3c95233beec21079b6f895a7a55ae1421ee5cc;p=lttng-ust.git diff --git a/liblttng-ust/lttng-ust-comm.c b/liblttng-ust/lttng-ust-comm.c index ad02dd9d..44e8bab4 100644 --- a/liblttng-ust/lttng-ust-comm.c +++ b/liblttng-ust/lttng-ust-comm.c @@ -37,14 +37,14 @@ #include #include #include +#include -#include #include -#include #include -#include -#include #include +#include +#include +#include "tracepoint-internal.h" #include "ltt-tracer-core.h" /* @@ -167,6 +167,7 @@ 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; @@ -176,6 +177,7 @@ int register_app_to_sessiond(int socket) reg_msg.ppid = getppid(); reg_msg.uid = getuid(); reg_msg.gid = getgid(); + reg_msg.bits_per_long = CAA_BITS_PER_LONG; prctl_ret = prctl(PR_GET_NAME, (unsigned long) reg_msg.name, 0, 0, 0); if (prctl_ret) { ERR("Error executing prctl"); @@ -237,6 +239,7 @@ 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 +272,8 @@ 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; @@ -285,25 +289,30 @@ 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 = lum->u.stream.memory_map_size; - shm_fd = lum->u.stream.shm_fd; - wait_fd = lum->u.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; - break; - case LTTNG_UST_VERSION: - lur.u.version = lum->u.version; - 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) { @@ -331,13 +340,57 @@ 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; @@ -356,7 +409,13 @@ 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"); @@ -408,7 +467,7 @@ 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; } @@ -556,14 +615,15 @@ 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; @@ -676,6 +736,20 @@ 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"); @@ -685,6 +759,7 @@ 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; @@ -814,10 +889,17 @@ 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(); @@ -851,6 +933,7 @@ 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"); @@ -861,6 +944,14 @@ 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); } @@ -871,7 +962,7 @@ 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 @@ -884,7 +975,7 @@ 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"); } @@ -892,25 +983,25 @@ 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); } /* @@ -922,7 +1013,7 @@ 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 */ @@ -930,6 +1021,6 @@ 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(); }