[LTP] [PATCH] syscalls/perf_event_open03: skip test on slower systems

[Cyril Hrubis]

Hi!
> Some systems (specially with combination of -debug kernel
> with KASAN enabled) have trouble completing this test
> in specified timeout.
> 
> Lowering number of iterations would make the test condition
> less accurate as it's based on global counter.
> 
> Instead, calculate the rate of iterations system can do in
> first 5 seconds and used that to decide whether to continue
> to run the test. If the rate is too slow, TCONF after 5
> seconds.
> 
> Signed-off-by: Jan Stancek <jstancek@redhat.com>
> ---
>  .../perf_event_open/perf_event_open03.c       | 43 ++++++++++++++++++-
>  1 file changed, 41 insertions(+), 2 deletions(-)
> 
> .needs_cmds = NULL gets rid of compile warning.

Please no workarounds for compiler bugs like this one.

This has been unfortunatelly broken in gcc for at least four releases
now but I do not think that it makes sense to add random field
initializers to most of the tests in the LTP sources.

> diff --git a/testcases/kernel/syscalls/perf_event_open/perf_event_open03.c b/testcases/kernel/syscalls/perf_event_open/perf_event_open03.c
> index dcb70962771c..c7bf123a04b4 100644
> --- a/testcases/kernel/syscalls/perf_event_open/perf_event_open03.c
> +++ b/testcases/kernel/syscalls/perf_event_open/perf_event_open03.c
> @@ -16,13 +16,16 @@
>  
>  #include "config.h"
>  #include "tst_test.h"
> +#include "tst_timer_test.h"
>  #include "lapi/syscalls.h"
>  
>  #include "perf_event_open.h"
>  
>  #define INTEL_PT_PATH "/sys/bus/event_source/devices/intel_pt/type"
>  
> +const int iterations = 12000000;
>  static int fd = -1;
> +static int timeout;
>  
>  static void setup(void)
>  {
> @@ -39,6 +42,38 @@ static void setup(void)
>  
>  	SAFE_FILE_SCANF(INTEL_PT_PATH, "%d", &ev.type);
>  	fd = perf_event_open(&ev, getpid(), -1, -1, 0);
> +
> +	timeout = tst_timeout_remaining();
> +}
> +
> +/*
> + * Check how fast we can do the iterations after 5 seconds of runtime.
> + * If the rate is too small to complete for current timeout then
> + * stop the test.
> + */
> +static void check_progress(int i)
> +{
> +	static float iter_per_ms;
> +	long long elapsed_ms;
> +
> +	if (iter_per_ms)
> +		return;
> +
> +	if (i % 1000 != 0)
> +		return;
> +
> +	tst_timer_stop();
> +	elapsed_ms = tst_timer_elapsed_ms();
> +	if (elapsed_ms > 5000) {
> +		iter_per_ms = (float) i / elapsed_ms;
> +		tst_res(TINFO, "rate: %f iters/ms", iter_per_ms);
> +		tst_res(TINFO, "needed rate for current test timeout: %f iters/ms",
> +			(float) iterations / (timeout * 1000));
> +
> +		if (iter_per_ms * 1000 * (timeout - 1) < iterations)
> +			tst_brk(TCONF, "System too slow to complete"
> +				" test in specified timeout");

String shouldn't be split like that even if they are over 80 chars, at
least if I remmeber LKML coding style correctly, the reason is that
splitting them like this breaks git grep for the output messages.

> +	}
>  }
>  
>  static void run(void)
> @@ -47,10 +82,13 @@ static void run(void)
>  	int i;
>  
>  	diff = SAFE_READ_MEMINFO("MemAvailable:");
> +	tst_timer_start(CLOCK_MONOTONIC);
>  
>  	/* leak about 100MB of RAM */
> -	for (i = 0; i < 12000000; i++)
> +	for (i = 0; i < iterations; i++) {
>  		ioctl(fd, PERF_EVENT_IOC_SET_FILTER, "filter,0/0@abcd");
> +		check_progress(i);
> +	}

Generally looks good. I guess that I will have to consider turning this
into a generic functionality when I return to the runtime patchset, but
for now this should be good to go.

Reviewed-by: Cyril Hrubis <chrubis@suse.cz>


-- 
Cyril Hrubis
chrubis@suse.cz