[LTP] [PATCH 1/2] perf_event_open02: stop groups with hw counters first

[LTP] [PATCH 1/2] perf_event_open02: stop groups with hw counters first

[Jan Stancek]

Nick reports that prctl(PR_TASK_PERF_EVENTS_DISABLE) doesn't stop tsk0
as last one, which sometimes causes test to fail, because of way test
calculates 'ratio' value and very low tolerance.

Looking at implementation, prctl() walks the list and does appear to
disable events in order they were created (tsk0 first).

Stop groups with hw counters before tsk0.

ratio     Before        After
mean      4.99921       4.99871
stdev     0.00006315    0.0000746

Fixes: #588
Signed-off-by: Jan Stancek <jstancek@redhat.com>
---
 testcases/kernel/syscalls/perf_event_open/perf_event_open02.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
index 95e30519fe42..1cfe29bb3d8a 100644
--- a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
+++ b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
@@ -324,6 +324,12 @@ static void verify(void)
 
 	do_work();
 
+	/* stop groups with hw counters first before tsk0 */
+	for (i = 0; i < n; i++) {
+		ioctl(hwfd[i], PERF_EVENT_IOC_DISABLE);
+		ioctl(tskfd[i], PERF_EVENT_IOC_DISABLE);
+	}
+
 	if (prctl(PR_TASK_PERF_EVENTS_DISABLE) == -1) {
 		tst_brkm(TBROK | TERRNO, cleanup,
 			 "prctl(PR_TASK_PERF_EVENTS_DISABLE) failed");
-- 
1.8.3.1

[LTP] [PATCH 2/2] perf_event_open02: make test more reliable on -rt kernels

[Jan Stancek]

Test is failing on -rt kernels quite reliably as reported here:
  https://lkml.org/lkml/2019/8/14/714

Suspected cause is extra time spend in scheduler, which counts towards
test process hw counters.

Pin test process to single CPU (doesn't matter which one) to lessen
effects from scheduling:

ratio     Before        After
mean      5.02582       4.99844
stdev     0.0090435     0.0000272
          95+% FAIL     100% PASS

Tested with "perf_event_open02: stop groups with hw counters first" patch
on RHEL8 kernel-rt 4.18.0-147.rt24.93.el8.x86_64 (HP dl380g7).

Signed-off-by: Jan Stancek <jstancek@redhat.com>
---
 .../kernel/syscalls/perf_event_open/perf_event_open02.c   | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
index 1cfe29bb3d8a..6fc6f4afa119 100644
--- a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
+++ b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
@@ -52,6 +52,7 @@ Usage is: ./performance_counter02  [-v]
 The -v flag makes it print out the values of each counter.
 */
 
+#define _GNU_SOURCE
 #include <stdio.h>
 #include <stddef.h>
 #include <stdlib.h>
@@ -222,6 +223,20 @@ static void setup(void)
 	int i;
 	struct perf_event_attr tsk_event, hw_event;
 
+#ifdef HAVE_SCHED_GETCPU
+	cpu_set_t mask;
+	int cpu = sched_getcpu();
+
+	if (cpu == -1)
+		tst_brkm(TBROK | TERRNO, NULL, "sched_getcpu() failed");
+
+	CPU_ZERO(&mask);
+	CPU_SET(cpu, &mask);
+
+	if (sched_setaffinity(0, sizeof(cpu_set_t), &mask) == -1)
+		tst_brkm(TBROK | TERRNO, NULL, "sched_setaffinity() failed");
+#endif
+
 	/*
 	 * According to perf_event_open's manpage, the official way of
 	 * knowing if perf_event_open() support is enabled is checking for
-- 
1.8.3.1

[LTP] [PATCH 2/2] perf_event_open02: make test more reliable on -rt kernels

[Cyril Hrubis]

Hi!
> Signed-off-by: Jan Stancek <jstancek@redhat.com>
> ---
>  .../kernel/syscalls/perf_event_open/perf_event_open02.c   | 15 +++++++++++++++
>  1 file changed, 15 insertions(+)
> 
> diff --git a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
> index 1cfe29bb3d8a..6fc6f4afa119 100644
> --- a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
> +++ b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
> @@ -52,6 +52,7 @@ Usage is: ./performance_counter02  [-v]
>  The -v flag makes it print out the values of each counter.
>  */
>  
> +#define _GNU_SOURCE
>  #include <stdio.h>
>  #include <stddef.h>
>  #include <stdlib.h>
> @@ -222,6 +223,20 @@ static void setup(void)
>  	int i;
>  	struct perf_event_attr tsk_event, hw_event;
>  
> +#ifdef HAVE_SCHED_GETCPU
> +	cpu_set_t mask;

Don't we have to allocate the mask dynamically so that we do not fail on
systems with more than CPU_SETSIZE (1024) CPUs?

We already have fallback macros for older libc in include/lapi/cpuset.h
so that we can use CPU_ALLOC() unconditionally.

> +	int cpu = sched_getcpu();
> +
> +	if (cpu == -1)
> +		tst_brkm(TBROK | TERRNO, NULL, "sched_getcpu() failed");
> +
> +	CPU_ZERO(&mask);
> +	CPU_SET(cpu, &mask);
> +
> +	if (sched_setaffinity(0, sizeof(cpu_set_t), &mask) == -1)
> +		tst_brkm(TBROK | TERRNO, NULL, "sched_setaffinity() failed");
> +#endif
> +
>  	/*
>  	 * According to perf_event_open's manpage, the official way of
>  	 * knowing if perf_event_open() support is enabled is checking for
> -- 
> 1.8.3.1
> 
> 
> -- 
> Mailing list info: https://lists.linux.it/listinfo/ltp

-- 
Cyril Hrubis
chrubis@suse.cz

[LTP] [PATCH 2/2] perf_event_open02: make test more reliable on -rt kernels

[Jan Stancek]

----- Original Message -----
> Hi!
> > Signed-off-by: Jan Stancek <jstancek@redhat.com>
> > ---
> >  .../kernel/syscalls/perf_event_open/perf_event_open02.c   | 15
> >  +++++++++++++++
> >  1 file changed, 15 insertions(+)
> > 
> > diff --git a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
> > b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
> > index 1cfe29bb3d8a..6fc6f4afa119 100644
> > --- a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
> > +++ b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
> > @@ -52,6 +52,7 @@ Usage is: ./performance_counter02  [-v]
> >  The -v flag makes it print out the values of each counter.
> >  */
> >  
> > +#define _GNU_SOURCE
> >  #include <stdio.h>
> >  #include <stddef.h>
> >  #include <stdlib.h>
> > @@ -222,6 +223,20 @@ static void setup(void)
> >  	int i;
> >  	struct perf_event_attr tsk_event, hw_event;
> >  
> > +#ifdef HAVE_SCHED_GETCPU
> > +	cpu_set_t mask;
> 
> Don't we have to allocate the mask dynamically so that we do not fail on
> systems with more than CPU_SETSIZE (1024) CPUs?
> 
> We already have fallback macros for older libc in include/lapi/cpuset.h
> so that we can use CPU_ALLOC() unconditionally.

Haven't thought of that. I'll look into adding that for v2.

[LTP] [PATCH v2] perf_event_open02: make test more reliable on -rt kernels

[Jan Stancek]

Test is failing on -rt kernels quite reliably as reported here:
  https://lkml.org/lkml/2019/8/14/714

Suspected cause is extra time spend in scheduler, which counts towards
test process hw counters.

Pin test process to single CPU (doesn't matter which one) to lessen
effects from scheduling:

ratio     Before        After
mean      5.02582       4.99844
stdev     0.0090435     0.0000272
          95+% FAIL     100% PASS

Tested with "perf_event_open02: stop groups with hw counters first" patch
on RHEL8 kernel-rt 4.18.0-147.rt24.93.el8.x86_64 (HP dl380g7).

Signed-off-by: Jan Stancek <jstancek@redhat.com>
---
 .../syscalls/perf_event_open/perf_event_open02.c      | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
index 1cfe29bb3d8a..c2831aa20d40 100644
--- a/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
+++ b/testcases/kernel/syscalls/perf_event_open/perf_event_open02.c
@@ -52,6 +52,7 @@ Usage is: ./performance_counter02  [-v]
 The -v flag makes it print out the values of each counter.
 */
 
+#define _GNU_SOURCE
 #include <stdio.h>
 #include <stddef.h>
 #include <stdlib.h>
@@ -72,6 +73,7 @@ The -v flag makes it print out the values of each counter.
 
 #include "test.h"
 #include "safe_macros.h"
+#include "lapi/cpuset.h"
 #include "lapi/syscalls.h"
 
 char *TCID = "perf_event_open02";
@@ -222,6 +224,23 @@ static void setup(void)
 	int i;
 	struct perf_event_attr tsk_event, hw_event;
 
+#ifdef HAVE_SCHED_GETCPU
+	int cpu = sched_getcpu();
+	size_t mask_size;
+	cpu_set_t *mask;
+
+	if (cpu == -1)
+		tst_brkm(TBROK | TERRNO, NULL, "sched_getcpu() failed");
+
+	mask = CPU_ALLOC(cpu + 1);
+	mask_size = CPU_ALLOC_SIZE(cpu + 1);
+	CPU_ZERO_S(mask_size, mask);
+	CPU_SET(cpu, mask);
+	if (sched_setaffinity(0, mask_size, mask) == -1)
+		tst_brkm(TBROK | TERRNO, NULL, "sched_setaffinity() failed");
+	CPU_FREE(mask);
+#endif
+
 	/*
 	 * According to perf_event_open's manpage, the official way of
 	 * knowing if perf_event_open() support is enabled is checking for
-- 
1.8.3.1

[LTP] [PATCH v2] perf_event_open02: make test more reliable on -rt kernels

[Cyril Hrubis]

Hi!
This version looks good, acked.

-- 
Cyril Hrubis
chrubis@suse.cz

[LTP] [PATCH v2] perf_event_open02: make test more reliable on -rt kernels

[Jan Stancek]

----- Original Message -----
> Hi!
> This version looks good, acked.

Both pushed.

[LTP] [PATCH 1/2] perf_event_open02: stop groups with hw counters first

[Cyril Hrubis]

Hi!
Acked.

-- 
Cyril Hrubis
chrubis@suse.cz