Re: [LTP] [PATCH] Fix soft lockups due to gtod_latency test case

[Subrata M. <su...@li...>]

On Mon, 2008-11-24 at 22:00 +0530, Sripathi Kodi wrote:
> Hi,
> 
> The test case gtod_latency in realtime causes soft lockups on some 
> machines. This is because the test hogs the cpus for a long time, 
> because it's main loop runs at 
> SCHED_FIFO, 99. We have seen soft lockups mainly on LS20 machines 
> (x86_64, Opterons). 
> 
> The following patch introduces a sleep after some iterations of the 
> test. This ensures that the test doesn't hog the cpu completely and 
> hence avoids soft lockups.
> 
> Testing done:
> Compiled with the patch and observed that the soft lockups are gone. 
> Also observed that the latencies are not affected. In fact latencies 
> improve with this patch on most hardware.
> 
> Changelog:
>  * Introduce periodic sleeps in the busy loop of gtod_latency to avoid soft lockups.
> 
> Signed-off-by: Sripathi Kodi <sri...@in...>
> Acked-by: John Stultz <joh...@us...>

Thanks.

Regards--
Subrata

> 
> diff -uprN ltp_cvs_24nov_org/testcases/realtime/func/gtod_latency/gtod_latency.c ltp_cvs_24nov/testcases/realtime/func/gtod_latency/gtod_latency.c
> --- ltp_cvs_24nov_org/testcases/realtime/func/gtod_latency/gtod_latency.c	2008-11-24 21:16:20.000000000 +0530
> +++ ltp_cvs_24nov/testcases/realtime/func/gtod_latency/gtod_latency.c	2008-11-24 21:25:33.000000000 +0530
> @@ -216,7 +216,7 @@ struct timespec stop_data[ITERATIONS];
> 
>  int main(int argc, char *argv[])
>  {
> -	int i, err;
> +	int i, j, k, err;
>  	unsigned long long delta;
>  	unsigned long long max, min;
>  	struct sched_param param;
> @@ -262,9 +262,15 @@ int main(int argc, char *argv[])
>  		latency_trace_enable();
>  		latency_trace_start();
>  	}
> -	for (i = 0; i < ITERATIONS; i++) {
> -		clock_gettime(CLOCK_MONOTONIC,&start_data[i]);
> -		clock_gettime(CLOCK_MONOTONIC,&stop_data[i]);
> +	/* This loop runs for a long time, hence can cause soft lockups.
> +	   Calling sleep periodically avoids this. */
> +	for (i=0; i<1000; i++) {
> +		for (j=0; j < ITERATIONS/1000; j++) {
> +			k = (i * ITERATIONS/1000) + j;
> +			clock_gettime(CLOCK_MONOTONIC,&start_data[k]);
> +			clock_gettime(CLOCK_MONOTONIC,&stop_data[k]);
> +		}
> +		usleep(1000);
>  	}
>  	for (i = 0; i < ITERATIONS; i++) {
>  		delta = timespec_subtract(&start_data[i], &stop_data[i]);