[Intel-gfx] [PATCH i-g-t] lib/dummyload: Handle timeout in a new thread instead of signal handler

Chris Wilson chris at chris-wilson.co.uk
Mon Mar 27 09:09:30 UTC 2017


On Mon, Mar 27, 2017 at 11:45:07AM +0300, Ander Conselvan de Oliveira wrote:
> Currently, the main thread needs to wakeup to run the signal handler
> that ends a spin batch. When testing whether a function call succesfully
> waits for a batch to complete, this behavior is undesired. It actually
> invalidates the test.
> 
> Fix this by spawning a new thread to handle the timeout.

Very neat!

> Cc: Chris Wilson <chris at chris-wilson.co.uk>
> Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira at intel.com>
> ---
>  lib/igt_dummyload.c | 55 +++++++++++++++++------------------------------------
>  lib/igt_dummyload.h |  3 ++-
>  2 files changed, 19 insertions(+), 39 deletions(-)
> 
> diff --git a/lib/igt_dummyload.c b/lib/igt_dummyload.c
> index 019c1fb..38b1930 100644
> --- a/lib/igt_dummyload.c
> +++ b/lib/igt_dummyload.c
> @@ -24,6 +24,7 @@
>  
>  #include "igt.h"
>  #include "igt_dummyload.h"
> +#include <pthread.h>
>  #include <time.h>
>  #include <signal.h>
>  #include <sys/syscall.h>
> @@ -166,6 +167,8 @@ igt_spin_batch_new(int fd, int engine, unsigned int dep_handle)
>  	spin = calloc(1, sizeof(struct igt_spin));
>  	igt_assert(spin);
>  
> +	pthread_mutex_init(&spin->mutex, NULL);
> +
>  	emit_recursive_batch(spin, fd, engine, dep_handle);
>  	igt_assert(gem_bo_busy(fd, spin->handle));
>  
> @@ -174,27 +177,11 @@ igt_spin_batch_new(int fd, int engine, unsigned int dep_handle)
>  	return spin;
>  }
>  
> -static void clear_sig_handler(int sig)
> -{
> -	struct sigaction act;
> -
> -	memset(&act, 0, sizeof(act));
> -	act.sa_handler = SIG_DFL;
> -	igt_assert(sigaction(sig, &act, NULL) == 0);
> -}
> -
> -static void sig_handler(int sig, siginfo_t *info, void *arg)
> +static void notify(union sigval arg)
>  {
> -	struct igt_spin *iter;
> +	igt_spin_t *spin = arg.sival_ptr;
>  
> -	igt_list_for_each(iter, &spin_list, link) {
> -		if (iter->signo == info->si_signo) {
> -			igt_spin_batch_end(iter);
> -			return;
> -		}
> -	}
> -
> -	clear_sig_handler(sig);
> +	igt_spin_batch_end(spin);
>  }
>  
>  /**
> @@ -208,43 +195,33 @@ static void sig_handler(int sig, siginfo_t *info, void *arg)
>   */
>  void igt_spin_batch_set_timeout(igt_spin_t *spin, int64_t ns)
>  {
> -	static int spin_signo = 48; /* Midpoint of SIGRTMIN, SIGRTMAX */
>  	timer_t timer;
>  	struct sigevent sev;
> -	struct sigaction act, oldact;
>  	struct itimerspec its;
>  
> +	pthread_mutex_lock(&spin->mutex);
> +
>  	igt_assert(ns > 0);
>  	if (!spin)
>  		return;

A mutex here is interesting, but only for detecting a race in setting
the timeout twice. Make it assert instead, and we then don't need a
mutex at all.

>  	igt_assert(!spin->timer);
>  
> -	/* SIGRTMAX is used by valgrind, SIGRTMAX - 1 by igt_fork_hang_detector */
> -	if (spin_signo >= SIGRTMAX - 2)
> -		spin_signo = SIGRTMIN;
> -	spin->signo = ++spin_signo;
> -
>  	memset(&sev, 0, sizeof(sev));
> -	sev.sigev_notify = SIGEV_SIGNAL | SIGEV_THREAD_ID;
> -	sev.sigev_notify_thread_id = gettid();
> -	sev.sigev_signo = spin->signo;
> +	sev.sigev_notify = SIGEV_THREAD;
> +	sev.sigev_value.sival_ptr = spin;
> +	sev.sigev_notify_function = notify;
>  	igt_assert(timer_create(CLOCK_MONOTONIC, &sev, &timer) == 0);
>  	igt_assert(timer);
>  
> -	memset(&oldact, 0, sizeof(oldact));
> -	memset(&act, 0, sizeof(act));
> -	act.sa_sigaction = sig_handler;
> -	act.sa_flags = SA_SIGINFO;
> -	igt_assert(sigaction(spin->signo, &act, &oldact) == 0);
> -	igt_assert(oldact.sa_sigaction == NULL);
> -
>  	memset(&its, 0, sizeof(its));
>  	its.it_value.tv_sec = ns / NSEC_PER_SEC;
>  	its.it_value.tv_nsec = ns % NSEC_PER_SEC;
>  	igt_assert(timer_settime(timer, 0, &its, NULL) == 0);
>  
>  	spin->timer = timer;
> +
> +	pthread_mutex_unlock(&spin->mutex);
>  }
>  
>  /**
> @@ -258,11 +235,12 @@ void igt_spin_batch_end(igt_spin_t *spin)
>  	if (!spin)
>  		return;
>  
> +	pthread_mutex_lock(&spin->mutex);
> +
>  	*spin->batch = MI_BATCH_BUFFER_END;
>  	__sync_synchronize();
>  
> -	if (spin->signo)
> -		clear_sig_handler(spin->signo);
> +	pthread_mutex_unlock(&spin->mutex);

This doesn't require a mutex.

>  }
>  
>  /**
> @@ -287,6 +265,7 @@ void igt_spin_batch_free(int fd, igt_spin_t *spin)
>  	gem_munmap(spin->batch, BATCH_SIZE);
>  
>  	gem_close(fd, spin->handle);
> +	pthread_mutex_destroy(&spin->mutex);

We need to cancel the timer (before the unmap).

>  	free(spin);
>  }
>  
> diff --git a/lib/igt_dummyload.h b/lib/igt_dummyload.h
> index 2adfadf..8fdc3fe 100644
> --- a/lib/igt_dummyload.h
> +++ b/lib/igt_dummyload.h
> @@ -31,11 +31,12 @@
>  #include "igt_aux.h"
>  
>  typedef struct igt_spin {
> +	pthread_mutex_t mutex;
>  	unsigned int handle;
>  	timer_t timer;
> -	int signo;
>  	struct igt_list link;
>  	uint32_t *batch;
> +

Oi!

>  } igt_spin_t;

-- 
Chris Wilson, Intel Open Source Technology Centre


More information about the Intel-gfx mailing list