[lttng-dev] [PATCH lttng-tools] Fix: Change sempahore to pthread conditions

Mathieu Desnoyers mathieu.desnoyers at efficios.com
Tue Sep 25 17:23:11 EDT 2012


* David Goulet (dgoulet at efficios.com) wrote:
> Fixes #324
> 
> Signed-off-by: David Goulet <dgoulet at efficios.com>
> ---
>  src/bin/lttng-sessiond/consumer.h |   21 +++++-
>  src/bin/lttng-sessiond/main.c     |  127 ++++++++++++++++++++++++++++---------
>  2 files changed, 116 insertions(+), 32 deletions(-)
> 
> diff --git a/src/bin/lttng-sessiond/consumer.h b/src/bin/lttng-sessiond/consumer.h
> index 1337f32..6639bd8 100644
> --- a/src/bin/lttng-sessiond/consumer.h
> +++ b/src/bin/lttng-sessiond/consumer.h
> @@ -18,8 +18,6 @@
>  #ifndef _CONSUMER_H
>  #define _CONSUMER_H
>  
> -#include <semaphore.h>
> -
>  #include <common/consumer.h>
>  #include <common/hashtable/hashtable.h>
>  #include <lttng/lttng.h>
> @@ -54,7 +52,24 @@ struct consumer_data {
>  	enum lttng_consumer_type type;
>  
>  	pthread_t thread;	/* Worker thread interacting with the consumer */
> -	sem_t sem;
> +
> +	/* Conditions used by the consumer thread to indicate readiness. */
> +	pthread_cond_t cond;
> +	pthread_condattr_t condattr;
> +	pthread_mutex_t cond_mutex;
> +
> +	/*
> +	 * This is a flag condition indicating that the consumer thread is ready
> +	 * and connected to the lttng-consumerd daemon. This flag MUST only be
> +	 * updated by locking the condition mutex above or before spawning a
> +	 * consumer thread.
> +	 *
> +	 * A value of 0 means that the thread is NOT ready. A value of 1 means that
> +	 * the thread consumer did connect successfully to the lttng-consumerd
> +	 * daemon. A negative value indicates that there is been an error and the
> +	 * thread as likely quit.

as -> has

> +	 */
> +	int consumer_thread_is_ready;
>  
>  	/* Mutex to control consumerd pid assignation */
>  	pthread_mutex_t pid_mutex;
> diff --git a/src/bin/lttng-sessiond/main.c b/src/bin/lttng-sessiond/main.c
> index 730ac65..2fb8700 100644
> --- a/src/bin/lttng-sessiond/main.c
> +++ b/src/bin/lttng-sessiond/main.c
> @@ -21,7 +21,6 @@
>  #include <grp.h>
>  #include <limits.h>
>  #include <pthread.h>
> -#include <semaphore.h>
>  #include <signal.h>
>  #include <stdio.h>
>  #include <stdlib.h>
> @@ -81,7 +80,10 @@ static int is_root;			/* Set to 1 if the daemon is running as root */
>  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,
> @@ -90,6 +92,8 @@ static struct consumer_data kconsumer_data = {
>  	.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,
> @@ -99,6 +103,8 @@ static struct consumer_data ustconsumer64_data = {
>  	.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,
> @@ -108,6 +114,8 @@ static struct consumer_data ustconsumer32_data = {
>  	.cmd_sock = -1,
>  	.pid_mutex = PTHREAD_MUTEX_INITIALIZER,
>  	.lock = PTHREAD_MUTEX_INITIALIZER,
> +	.cond = PTHREAD_COND_INITIALIZER,
> +	.cond_mutex = PTHREAD_MUTEX_INITIALIZER,
>  };
>  
>  /* Shared between threads */
> @@ -789,6 +797,20 @@ error_poll_create:
>  }
>  
>  /*
> + * 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);
> +
> +	/* Indicates that the thread is ready for action! */
> +	data->consumer_thread_is_ready = state;

the comment does not match the fact that state can be set to 0 here,
thus meaning "not ready" for action anymore ?

> +	(void) pthread_cond_signal(&data->cond);
> +
> +	pthread_mutex_unlock(&data->cond_mutex);
> +}
> +
> +/*
>   * This thread manage the consumer error sent back to the session daemon.
>   */
>  static void *thread_manage_consumer(void *data)
> @@ -886,13 +908,13 @@ restart:
>  		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));
> @@ -1446,59 +1468,106 @@ error_create_poll:
>   */
>  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_create(&consumer_data->thread, NULL,
> -			thread_manage_consumer, consumer_data);
> +	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);
>  	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
> +	 * incremente the timeout too many times.

incremente -> increment

> +	 */
> +	timeout.tv_sec += DEFAULT_SEM_WAIT_TIMEOUT;
> +
> +	/* Make sure that ret is never set to ETIMEDOUT before reaching the loop */

Just commenting that we reset it would be fine. What matters here is
that the loop could be skipped in some conditions, which is why setting
it to 0 matters.

The rest looks good. So after fixing those cosmetic issues, please add
my Acked-by.

Thanks!

Mathieu

> +	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;
>  	}
> -- 
> 1.7.10.4
> 
> 
> _______________________________________________
> lttng-dev mailing list
> lttng-dev at lists.lttng.org
> http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com



More information about the lttng-dev mailing list