#include <grp.h>
#include <limits.h>
#include <pthread.h>
-#include <semaphore.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include "fd-limit.h"
#include "filter.h"
#include "health.h"
+#include "testpoint.h"
#define CONSUMERD_FILE "lttng-consumerd"
static pid_t ppid; /* Parent PID for --sig-parent option */
static char *rundir;
-/* Consumer daemon specific control data */
+/*
+ * Consumer daemon specific control data. Every value not initialized here is
+ * set to 0 by the static definition.
+ */
static struct consumer_data kconsumer_data = {
.type = LTTNG_CONSUMER_KERNEL,
.err_unix_sock_path = DEFAULT_KCONSUMERD_ERR_SOCK_PATH,
.cmd_sock = -1,
.pid_mutex = PTHREAD_MUTEX_INITIALIZER,
.lock = PTHREAD_MUTEX_INITIALIZER,
+ .cond = PTHREAD_COND_INITIALIZER,
+ .cond_mutex = PTHREAD_MUTEX_INITIALIZER,
};
static struct consumer_data ustconsumer64_data = {
.type = LTTNG_CONSUMER64_UST,
.cmd_sock = -1,
.pid_mutex = PTHREAD_MUTEX_INITIALIZER,
.lock = PTHREAD_MUTEX_INITIALIZER,
+ .cond = PTHREAD_COND_INITIALIZER,
+ .cond_mutex = PTHREAD_MUTEX_INITIALIZER,
};
static struct consumer_data ustconsumer32_data = {
.type = LTTNG_CONSUMER32_UST,
.cmd_sock = -1,
.pid_mutex = PTHREAD_MUTEX_INITIALIZER,
.lock = PTHREAD_MUTEX_INITIALIZER,
+ .cond = PTHREAD_COND_INITIALIZER,
+ .cond_mutex = PTHREAD_MUTEX_INITIALIZER,
};
/* Shared between threads */
ERR("Unable to clean %s", rundir);
}
free(cmd);
+ free(rundir);
DBG("Cleaning up all sessions");
DBG("Thread manage kernel started");
+ testpoint(thread_manage_kernel);
+
health_code_update(&health_thread_kernel);
+ testpoint(thread_manage_kernel_before_loop);
+
ret = create_thread_poll_set(&events, 2);
if (ret < 0) {
goto error_poll_create;
return NULL;
}
+/*
+ * Signal pthread condition of the consumer data that the thread.
+ */
+static void signal_consumer_condition(struct consumer_data *data, int state)
+{
+ pthread_mutex_lock(&data->cond_mutex);
+
+ /*
+ * The state is set before signaling. It can be any value, it's the waiter
+ * job to correctly interpret this condition variable associated to the
+ * consumer pthread_cond.
+ *
+ * A value of 0 means that the corresponding thread of the consumer data
+ * was not started. 1 indicates that the thread has started and is ready
+ * for action. A negative value means that there was an error during the
+ * thread bootstrap.
+ */
+ data->consumer_thread_is_ready = state;
+ (void) pthread_cond_signal(&data->cond);
+
+ pthread_mutex_unlock(&data->cond_mutex);
+}
+
/*
* This thread manage the consumer error sent back to the session daemon.
*/
DBG("[thread] Manage consumer started");
+ /*
+ * Since the consumer thread can be spawned at any moment in time, we init
+ * the health to a poll status (1, which is a valid health over time).
+ * When the thread starts, we update here the health to a "code" path being
+ * an even value so this thread, when reaching a poll wait, does not
+ * trigger an error with an even value.
+ *
+ * Here is the use case we avoid.
+ *
+ * +1: the first poll update during initialization (main())
+ * +2 * x: multiple code update once in this thread.
+ * +1: poll wait in this thread (being a good health state).
+ * == even number which after the wait period shows as a bad health.
+ *
+ * In a nutshell, the following poll update to the health state brings back
+ * the state to an even value meaning a code path.
+ */
+ health_poll_update(&consumer_data->health);
+
health_code_update(&consumer_data->health);
ret = lttcomm_listen_unix_sock(consumer_data->err_sock);
/* Inifinite blocking call, waiting for transmission */
restart:
health_poll_update(&consumer_data->health);
+
+ testpoint(thread_manage_consumer);
+
ret = lttng_poll_wait(&events, -1);
health_poll_update(&consumer_data->health);
if (ret < 0) {
goto error;
}
+ /*
+ * Set the CLOEXEC flag. Return code is useless because either way, the
+ * show must go on.
+ */
+ (void) utils_set_fd_cloexec(sock);
+
health_code_update(&consumer_data->health);
DBG2("Receiving code from consumer err_sock");
consumer_data->cmd_sock =
lttcomm_connect_unix_sock(consumer_data->cmd_unix_sock_path);
if (consumer_data->cmd_sock < 0) {
- sem_post(&consumer_data->sem);
+ /* On error, signal condition and quit. */
+ signal_consumer_condition(consumer_data, -1);
PERROR("consumer connect");
goto error;
}
- /* Signal condition to tell that the kconsumerd is ready */
- sem_post(&consumer_data->sem);
- DBG("consumer command socket ready");
+ signal_consumer_condition(consumer_data, 1);
+ DBG("Consumer command socket ready");
} else {
ERR("consumer error when waiting for SOCK_READY : %s",
lttcomm_get_readable_code(-code));
DBG("[thread] Manage application started");
+ testpoint(thread_manage_apps);
+
rcu_register_thread();
rcu_thread_online();
goto error;
}
+ testpoint(thread_manage_apps_before_loop);
+
health_code_update(&health_thread_app_manage);
while (1) {
ust_app_unregister(ust_cmd.sock);
} else {
/*
- * We just need here to monitor the close of the UST
- * socket and poll set monitor those by default.
- * Listen on POLLIN (even if we never expect any
- * data) to ensure that hangup wakes us.
+ * We only monitor the error events of the socket. This
+ * thread does not handle any incoming data from UST
+ * (POLLIN).
*/
- ret = lttng_poll_add(&events, ust_cmd.sock, LPOLLIN);
+ ret = lttng_poll_add(&events, ust_cmd.sock,
+ LPOLLERR & LPOLLHUP & LPOLLRDHUP);
if (ret < 0) {
goto error;
}
DBG("[thread] Manage application registration started");
+ testpoint(thread_registration_apps);
+
ret = lttcomm_listen_unix_sock(apps_sock);
if (ret < 0) {
goto error_listen;
goto error;
}
+ /*
+ * Set the CLOEXEC flag. Return code is useless because
+ * either way, the show must go on.
+ */
+ (void) utils_set_fd_cloexec(sock);
+
/* Create UST registration command for enqueuing */
ust_cmd = zmalloc(sizeof(struct ust_command));
if (ust_cmd == NULL) {
*/
static int spawn_consumer_thread(struct consumer_data *consumer_data)
{
- int ret;
+ int ret, clock_ret;
struct timespec timeout;
- timeout.tv_sec = DEFAULT_SEM_WAIT_TIMEOUT;
- timeout.tv_nsec = 0;
+ /* Make sure we set the readiness flag to 0 because we are NOT ready */
+ consumer_data->consumer_thread_is_ready = 0;
- /* Setup semaphore */
- ret = sem_init(&consumer_data->sem, 0, 0);
- if (ret < 0) {
- PERROR("sem_init consumer semaphore");
+ /* Setup pthread condition */
+ ret = pthread_condattr_init(&consumer_data->condattr);
+ if (ret != 0) {
+ errno = ret;
+ PERROR("pthread_condattr_init consumer data");
+ goto error;
+ }
+
+ /*
+ * Set the monotonic clock in order to make sure we DO NOT jump in time
+ * between the clock_gettime() call and the timedwait call. See bug #324
+ * for a more details and how we noticed it.
+ */
+ ret = pthread_condattr_setclock(&consumer_data->condattr, CLOCK_MONOTONIC);
+ if (ret != 0) {
+ errno = ret;
+ PERROR("pthread_condattr_setclock consumer data");
+ goto error;
+ }
+
+ ret = pthread_cond_init(&consumer_data->cond, &consumer_data->condattr);
+ if (ret != 0) {
+ errno = ret;
+ PERROR("pthread_cond_init consumer data");
goto error;
}
- ret = pthread_create(&consumer_data->thread, NULL,
- thread_manage_consumer, consumer_data);
+ ret = pthread_create(&consumer_data->thread, NULL, thread_manage_consumer,
+ consumer_data);
if (ret != 0) {
PERROR("pthread_create consumer");
ret = -1;
goto error;
}
+ /* We are about to wait on a pthread condition */
+ pthread_mutex_lock(&consumer_data->cond_mutex);
+
/* Get time for sem_timedwait absolute timeout */
- ret = clock_gettime(CLOCK_REALTIME, &timeout);
- if (ret < 0) {
- PERROR("clock_gettime spawn consumer");
- /* Infinite wait for the kconsumerd thread to be ready */
- ret = sem_wait(&consumer_data->sem);
- } else {
- /* Normal timeout if the gettime was successful */
- timeout.tv_sec += DEFAULT_SEM_WAIT_TIMEOUT;
- ret = sem_timedwait(&consumer_data->sem, &timeout);
+ clock_ret = clock_gettime(CLOCK_MONOTONIC, &timeout);
+ /*
+ * Set the timeout for the condition timed wait even if the clock gettime
+ * call fails since we might loop on that call and we want to avoid to
+ * increment the timeout too many times.
+ */
+ timeout.tv_sec += DEFAULT_SEM_WAIT_TIMEOUT;
+
+ /*
+ * The following loop COULD be skipped in some conditions so this is why we
+ * set ret to 0 in order to make sure at least one round of the loop is
+ * done.
+ */
+ ret = 0;
+
+ /*
+ * Loop until the condition is reached or when a timeout is reached. Note
+ * that the pthread_cond_timedwait(P) man page specifies that EINTR can NOT
+ * be returned but the pthread_cond(3), from the glibc-doc, says that it is
+ * possible. This loop does not take any chances and works with both of
+ * them.
+ */
+ while (!consumer_data->consumer_thread_is_ready && ret != ETIMEDOUT) {
+ if (clock_ret < 0) {
+ PERROR("clock_gettime spawn consumer");
+ /* Infinite wait for the consumerd thread to be ready */
+ ret = pthread_cond_wait(&consumer_data->cond,
+ &consumer_data->cond_mutex);
+ } else {
+ ret = pthread_cond_timedwait(&consumer_data->cond,
+ &consumer_data->cond_mutex, &timeout);
+ }
}
- if (ret < 0) {
- if (errno == ETIMEDOUT) {
+ /* Release the pthread condition */
+ pthread_mutex_unlock(&consumer_data->cond_mutex);
+
+ if (ret != 0) {
+ errno = ret;
+ if (ret == ETIMEDOUT) {
/*
* Call has timed out so we kill the kconsumerd_thread and return
* an error.
*/
- ERR("The consumer thread was never ready. Killing it");
+ ERR("Condition timed out. The consumer thread was never ready."
+ " Killing it");
ret = pthread_cancel(consumer_data->thread);
if (ret < 0) {
PERROR("pthread_cancel consumer thread");
}
} else {
- PERROR("semaphore wait failed consumer thread");
+ PERROR("pthread_cond_wait failed consumer thread");
}
goto error;
}
pthread_mutex_lock(&consumer_data->pid_mutex);
if (consumer_data->pid == 0) {
- ERR("Kconsumerd did not start");
+ ERR("Consumerd did not start");
pthread_mutex_unlock(&consumer_data->pid_mutex);
goto error;
}
switch (domain) {
case LTTNG_DOMAIN_KERNEL:
DBG3("Copying tracing session consumer output in kernel session");
+ /*
+ * XXX: We should audit the session creation and what this function
+ * does "extra" in order to avoid a destroy since this function is used
+ * in the domain session creation (kernel and ust) only. Same for UST
+ * domain.
+ */
+ if (session->kernel_session->consumer) {
+ consumer_destroy_output(session->kernel_session->consumer);
+ }
session->kernel_session->consumer =
consumer_copy_output(session->consumer);
/* Ease our life a bit for the next part */
break;
case LTTNG_DOMAIN_UST:
DBG3("Copying tracing session consumer output in UST session");
+ if (session->ust_session->consumer) {
+ consumer_destroy_output(session->ust_session->consumer);
+ }
session->ust_session->consumer =
consumer_copy_output(session->consumer);
/* Ease our life a bit for the next part */
case LTTNG_LIST_DOMAINS:
case LTTNG_START_TRACE:
case LTTNG_STOP_TRACE:
+ case LTTNG_DATA_PENDING:
need_domain = 0;
break;
default:
ret = cmd_add_context(cmd_ctx->session, cmd_ctx->lsm->domain.type,
cmd_ctx->lsm->u.context.channel_name,
cmd_ctx->lsm->u.context.event_name,
- &cmd_ctx->lsm->u.context.ctx);
+ &cmd_ctx->lsm->u.context.ctx, kernel_poll_pipe[1]);
break;
}
case LTTNG_DISABLE_CHANNEL:
DBG("No URIs received from client... continuing");
*sock_error = 1;
ret = LTTNG_ERR_SESSION_FAIL;
+ free(uris);
goto error;
}
ret = cmd_set_consumer_uri(cmd_ctx->lsm->domain.type, cmd_ctx->session,
nb_uri, uris);
if (ret != LTTNG_OK) {
+ free(uris);
goto error;
}
}
}
+ free(uris);
+
break;
}
case LTTNG_START_TRACE:
DBG("No URIs received from client... continuing");
*sock_error = 1;
ret = LTTNG_ERR_SESSION_FAIL;
+ free(uris);
goto error;
}
if (nb_uri == 1 && uris[0].dtype != LTTNG_DST_PATH) {
DBG("Creating session with ONE network URI is a bad call");
ret = LTTNG_ERR_SESSION_FAIL;
+ free(uris);
goto error;
}
}
ret = cmd_create_session_uri(cmd_ctx->lsm->session.name, uris, nb_uri,
&cmd_ctx->creds);
+ free(uris);
+
break;
}
case LTTNG_DESTROY_SESSION:
{
struct lttng_filter_bytecode *bytecode;
- if (cmd_ctx->lsm->u.filter.bytecode_len > 65336) {
+ if (cmd_ctx->lsm->u.filter.bytecode_len > LTTNG_FILTER_MAX_LEN) {
ret = LTTNG_ERR_FILTER_INVAL;
goto error;
}
bytecode);
break;
}
+ case LTTNG_DATA_PENDING:
+ {
+ ret = cmd_data_pending(cmd_ctx->session);
+ break;
+ }
default:
ret = LTTNG_ERR_UND;
break;
goto error;
}
+ /*
+ * Set the CLOEXEC flag. Return code is useless because either way, the
+ * show must go on.
+ */
+ (void) utils_set_fd_cloexec(sock);
+
ret = lttcomm_listen_unix_sock(sock);
if (ret < 0) {
goto error;
goto error;
}
+ /*
+ * Set the CLOEXEC flag. Return code is useless because either way, the
+ * show must go on.
+ */
+ (void) utils_set_fd_cloexec(new_sock);
+
DBG("Receiving data from client for health...");
ret = lttcomm_recv_unix_sock(new_sock, (void *)&msg, sizeof(msg));
if (ret <= 0) {
DBG("[thread] Manage client started");
+ testpoint(thread_manage_clients);
+
rcu_register_thread();
health_code_update(&health_thread_cmd);
kill(ppid, SIGUSR1);
}
+ testpoint(thread_manage_clients_before_loop);
+
health_code_update(&health_thread_cmd);
while (1) {
goto error;
}
+ /*
+ * Set the CLOEXEC flag. Return code is useless because either way, the
+ * show must go on.
+ */
+ (void) utils_set_fd_cloexec(sock);
+
/* Set socket option for credentials retrieval */
ret = lttcomm_setsockopt_creds_unix_sock(sock);
if (ret < 0) {
goto end;
}
+ /* Set the cloexec flag */
+ ret = utils_set_fd_cloexec(client_sock);
+ if (ret < 0) {
+ ERR("Unable to set CLOEXEC flag to the client Unix socket (fd: %d). "
+ "Continuing but note that the consumer daemon will have a "
+ "reference to this socket on exec()", client_sock);
+ }
+
/* File permission MUST be 660 */
ret = chmod(client_unix_sock_path, S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP);
if (ret < 0) {
goto end;
}
+ /* Set the cloexec flag */
+ ret = utils_set_fd_cloexec(apps_sock);
+ if (ret < 0) {
+ ERR("Unable to set CLOEXEC flag to the app Unix socket (fd: %d). "
+ "Continuing but note that the consumer daemon will have a "
+ "reference to this socket on exec()", apps_sock);
+ }
+
/* File permission MUST be 666 */
ret = chmod(apps_unix_sock_path,
S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP | S_IROTH | S_IWOTH);
goto end;
}
+ DBG3("Session daemon client socket %d and application socket %d created",
+ client_sock, apps_sock);
+
end:
umask(old_umask);
return ret;
/*
* Init health counters of the consumer thread. We do a quick hack here to
* the state of the consumer health is fine even if the thread is not
- * started. This is simply to ease our life and has no cost what so ever.
+ * started. Once the thread starts, the health state is updated with a poll
+ * value to set a health code path. This is simply to ease our life and has
+ * no cost what so ever.
*/
health_init(&kconsumer_data.health);
health_poll_update(&kconsumer_data.health);
goto error; /* join error, exit without cleanup */
}
+ ret = join_consumer_thread(&ustconsumer32_data);
+ if (ret != 0) {
+ PERROR("join_consumer ust32");
+ goto error; /* join error, exit without cleanup */
+ }
+
+ ret = join_consumer_thread(&ustconsumer64_data);
+ if (ret != 0) {
+ PERROR("join_consumer ust64");
+ goto error; /* join error, exit without cleanup */
+ }
+
exit_client:
+ ret = pthread_join(health_thread, &status);
+ if (ret != 0) {
+ PERROR("pthread_join health thread");
+ goto error; /* join error, exit without cleanup */
+ }
+
exit_health:
exit:
/*