libperf: lack of interface

libperf: lack of interface

[nakamura.shun]

Hello.

I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
Does anyone plan to implement any of these libperf features?

- Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
- Interfaces that can run fcntl (SIGIO) from userland

[1] https://github.com/deater/perf_event_tests/tree/master/tests/rdpmc


Thanks.
Shunsuke

Re: libperf: lack of interface

[Peter Zijlstra]

On Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com wrote:
> Hello.
> 
> I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> Does anyone plan to implement any of these libperf features?
> 
> - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> - Interfaces that can run fcntl (SIGIO) from userland
> 
> [1] https://github.com/deater/perf_event_tests/tree/master/tests/rdpmc

While IOC_RESET has it's uses, it really shouldn't be used in
combination with RDPMC, it destroys the whole 'dont do
syscalls'/performance thing.

The typical RDPMC usage is:

	/* setup counter on self */

	...

	start = RDPMC
	/* your code */
	end = RDPMC
	delta = end - start;

	...

Nowhere do we need a reset...

Re: libperf: lack of interface

[Arnaldo Carvalho de Melo]

Em Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com escreveu:
> Hello.
> 
> I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> Does anyone plan to implement any of these libperf features?
> 
> - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> - Interfaces that can run fcntl (SIGIO) from userland
> 
> [1] https://github.com/deater/perf_event_tests/tree/master/tests/rdpmc

So, while in this specific case you should probably follow PeterZ's
advice, feel free to submit patches moving stuff that is in
tools/perf/util/ to tools/lib/perf/ (libperf) when you have a reasonable
use case, such as Vince's test suite.

We can then discuss if any adjustment is needed, but doing it this
piecemeal way, with justification, should be a good way to enrich
libperf with things that are in quiescent state in tools/perf/util/ and
that have external users.

Thanks,

- Arnaldo

Re: libperf: lack of interface

[Jiri Olsa]

On Wed, Sep 01, 2021 at 10:36:40AM -0300, Arnaldo Carvalho de Melo wrote:
> Em Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com escreveu:
> > Hello.
> > 
> > I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> > Does anyone plan to implement any of these libperf features?
> > 
> > - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> > - Interfaces that can run fcntl (SIGIO) from userland

hi,
we could add perf_evsel__fd like below, would it help your usecase?

if you described your usecases in more details we could
see if we could add/move something to libperf for that

as Arnaldo said below it could be already in tools/perf/util/*.c
somewhere ;-)

jirka


> > 
> > [1] https://github.com/deater/perf_event_tests/tree/master/tests/rdpmc
> 
> So, while in this specific case you should probably follow PeterZ's
> advice, feel free to submit patches moving stuff that is in
> tools/perf/util/ to tools/lib/perf/ (libperf) when you have a reasonable
> use case, such as Vince's test suite.
> 
> We can then discuss if any adjustment is needed, but doing it this
> piecemeal way, with justification, should be a good way to enrich
> libperf with things that are in quiescent state in tools/perf/util/ and
> that have external users.
> 
> Thanks,
> 
> - Arnaldo
> 

---
 tools/lib/perf/evsel.c              | 13 +++++++++++++
 tools/lib/perf/include/perf/evsel.h |  1 +
 2 files changed, 14 insertions(+)

diff --git a/tools/lib/perf/evsel.c b/tools/lib/perf/evsel.c
index d8886720e83d..8cbdaa78749c 100644
--- a/tools/lib/perf/evsel.c
+++ b/tools/lib/perf/evsel.c
@@ -413,3 +413,16 @@ void perf_evsel__free_id(struct perf_evsel *evsel)
 	zfree(&evsel->id);
 	evsel->ids = 0;
 }
+
+int perf_evsel__fd(struct perf_evsel *evsel, int cpu, int thread)
+{
+	int max_thread = xyarray__max_y(evsel->fd);
+	int max_cpu = xyarray__max_x(evsel->fd);
+
+	if (thread >= max_thread)
+		return -EINVAL;
+	if (cpu >= max_cpu)
+		return -EINVAL;
+
+	return FD(evsel, cpu, thread);
+}
diff --git a/tools/lib/perf/include/perf/evsel.h b/tools/lib/perf/include/perf/evsel.h
index 60eae25076d3..5f29935efcaf 100644
--- a/tools/lib/perf/include/perf/evsel.h
+++ b/tools/lib/perf/include/perf/evsel.h
@@ -39,5 +39,6 @@ LIBPERF_API int perf_evsel__disable_cpu(struct perf_evsel *evsel, int cpu);
 LIBPERF_API struct perf_cpu_map *perf_evsel__cpus(struct perf_evsel *evsel);
 LIBPERF_API struct perf_thread_map *perf_evsel__threads(struct perf_evsel *evsel);
 LIBPERF_API struct perf_event_attr *perf_evsel__attr(struct perf_evsel *evsel);
+LIBPERF_API int perf_evsel__fd(struct perf_evsel *evsel, int cpu, int thread);
 
 #endif /* __LIBPERF_EVSEL_H */
-- 
2.31.1

Re: libperf: lack of interface

[nakamura.shun]

Hi, Jirka, Arnaldo, and Peter
Sorry for the late reply.

> On Wed, Sep 01, 2021 at 10:36:40AM -0300, Arnaldo Carvalho de Melo wrote:
> > Em Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com escreveu:
> > > Hello.
> > > 
> > > I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> > > Does anyone plan to implement any of these libperf features?
> > > 
> > > - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> > > - Interfaces that can run fcntl (SIGIO) from userland
> 
> hi,
> we could add perf_evsel__fd like below, would it help your usecase?
> 
> if you described your usecases in more details we could
> see if we could add/move something to libperf for that
> 
> as Arnaldo said below it could be already in tools/perf/util/*.c
> somewhere ;-)

As Peter says, I understood that for rdpmc, no reset is needed.

However, PAPI resets it explicitly, for example, at PAPI_reset.
In other, PAPI also has the ability to call PERF_EVENT_IOC_REFLESH on overflow to call a user-registered handler, using SIGIO.

I think it is desirable to be able to achieve similar functionality.

Best Regards
Shunsuke

Re: libperf: lack of interface

[nakamura.shun]

Hi

> > On Wed, Sep 01, 2021 at 10:36:40AM -0300, Arnaldo Carvalho de Melo wrote:
> > > Em Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com escreveu:
> > > > Hello.
> > > >
> > > > I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> > > > Does anyone plan to implement any of these libperf features?
> > > >
> > > > - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> > > > - Interfaces that can run fcntl (SIGIO) from userland
> >
> > hi,
> > we could add perf_evsel__fd like below, would it help your usecase?
> >
> > if you described your usecases in more details we could
> > see if we could add/move something to libperf for that
> >
> > as Arnaldo said below it could be already in tools/perf/util/*.c
> > somewhere ;-)
> 
> As Peter says, I understood that for rdpmc, no reset is needed.
> 
> However, PAPI resets it explicitly, for example, at PAPI_reset.
> In other, PAPI also has the ability to call PERF_EVENT_IOC_REFLESH on overflow to call a user-registered handler, using SIGIO.
> 
> I think it is desirable to be able to achieve similar functionality.

Does anyone have any comments?

Best Regards
Shunsuke

Re: libperf: lack of interface

[Jiri Olsa]

On Mon, Oct 18, 2021 at 08:57:20AM +0000, nakamura.shun@fujitsu.com wrote:
> Hi
> 
> > > On Wed, Sep 01, 2021 at 10:36:40AM -0300, Arnaldo Carvalho de Melo wrote:
> > > > Em Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com escreveu:
> > > > > Hello.
> > > > >
> > > > > I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> > > > > Does anyone plan to implement any of these libperf features?
> > > > >
> > > > > - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> > > > > - Interfaces that can run fcntl (SIGIO) from userland
> > >
> > > hi,
> > > we could add perf_evsel__fd like below, would it help your usecase?
> > >
> > > if you described your usecases in more details we could
> > > see if we could add/move something to libperf for that
> > >
> > > as Arnaldo said below it could be already in tools/perf/util/*.c
> > > somewhere ;-)
> > 
> > As Peter says, I understood that for rdpmc, no reset is needed.
> > 
> > However, PAPI resets it explicitly, for example, at PAPI_reset.
> > In other, PAPI also has the ability to call PERF_EVENT_IOC_REFLESH on overflow to call a user-registered handler, using SIGIO.
> > 
> > I think it is desirable to be able to achieve similar functionality.
> 
> Does anyone have any comments?

I'll try to add that functionality soon,
I'll cc you on patch

jirka

Re: libperf: lack of interface

[nakamura.shun]

Hi jirka

> On Mon, Oct 18, 2021 at 08:57:20AM +0000, nakamura.shun@fujitsu.com wrote:
> > Hi
> > 
> > > > On Wed, Sep 01, 2021 at 10:36:40AM -0300, Arnaldo Carvalho de Melo wrote:
> > > > > Em Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com escreveu:
> > > > > > Hello.
> > > > > >
> > > > > > I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> > > > > > Does anyone plan to implement any of these libperf features?
> > > > > >
> > > > > > - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> > > > > > - Interfaces that can run fcntl (SIGIO) from userland
> > > >
> > > > hi,
> > > > we could add perf_evsel__fd like below, would it help your usecase?
> > > >
> > > > if you described your usecases in more details we could
> > > > see if we could add/move something to libperf for that
> > > >
> > > > as Arnaldo said below it could be already in tools/perf/util/*.c
> > > > somewhere ;-)
> > > 
> > > As Peter says, I understood that for rdpmc, no reset is needed.
> > > 
> > > However, PAPI resets it explicitly, for example, at PAPI_reset.
> > > In other, PAPI also has the ability to call PERF_EVENT_IOC_REFLESH on overflow to call a user-registered handler, using SIGIO.
> > > 
> > > I think it is desirable to be able to achieve similar functionality.
> > 
> > Does anyone have any comments?
> 
> I'll try to add that functionality soon,
> I'll cc you on patch

Thank you.
I look forward to your patch.

Best Regards
Shunsuke

Re: libperf: lack of interface

[nakamura.shun]

Hi jirka

> > On Mon, Oct 18, 2021 at 08:57:20AM +0000, nakamura.shun@fujitsu.com wrote:
> > > Hi
> > > 
> > > > > On Wed, Sep 01, 2021 at 10:36:40AM -0300, Arnaldo Carvalho de Melo wrote:
> > > > > > Em Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com escreveu:
> > > > > > > Hello.
> > > > > > >
> > > > > > > I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> > > > > > > Does anyone plan to implement any of these libperf features?
> > > > > > >
> > > > > > > - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> > > > > > > - Interfaces that can run fcntl (SIGIO) from userland
> > > > >
> > > > > hi,
> > > > > we could add perf_evsel__fd like below, would it help your usecase?
> > > > >
> > > > > if you described your usecases in more details we could
> > > > > see if we could add/move something to libperf for that
> > > > >
> > > > > as Arnaldo said below it could be already in tools/perf/util/*.c
> > > > > somewhere ;-)
> > > > 
> > > > As Peter says, I understood that for rdpmc, no reset is needed.
> > > > 
> > > > However, PAPI resets it explicitly, for example, at PAPI_reset.
> > > > In other, PAPI also has the ability to call PERF_EVENT_IOC_REFLESH on overflow to call a user-registered handler, using SIGIO.
> > > > 
> > > > I think it is desirable to be able to achieve similar functionality.
> > > 
> > > Does anyone have any comments?
> > 
> > I'll try to add that functionality soon,
> > I'll cc you on patch
> 
> Thank you.
> I look forward to your patch.

Do you have a specific plan? 
I would like to know the rough period until your patch is made.

Best Regards
Shunsuke

Re: libperf: lack of interface

[nakamura.shun]

Hi jirka

> > > On Mon, Oct 18, 2021 at 08:57:20AM +0000, nakamura.shun@fujitsu.com wrote:
> > > > Hi
> > > > 
> > > > > > On Wed, Sep 01, 2021 at 10:36:40AM -0300, Arnaldo Carvalho de Melo wrote:
> > > > > > > Em Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com escreveu:
> > > > > > > > Hello.
> > > > > > > >
> > > > > > > > I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> > > > > > > > Does anyone plan to implement any of these libperf features?
> > > > > > > >
> > > > > > > > - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> > > > > > > > - Interfaces that can run fcntl (SIGIO) from userland
> > > > > >
> > > > > > hi,
> > > > > > we could add perf_evsel__fd like below, would it help your usecase?
> > > > > >
> > > > > > if you described your usecases in more details we could
> > > > > > see if we could add/move something to libperf for that
> > > > > >
> > > > > > as Arnaldo said below it could be already in tools/perf/util/*.c
> > > > > > somewhere ;-)
> > > > > 
> > > > > As Peter says, I understood that for rdpmc, no reset is needed.
> > > > > 
> > > > > However, PAPI resets it explicitly, for example, at PAPI_reset.
> > > > > In other, PAPI also has the ability to call PERF_EVENT_IOC_REFLESH on overflow to call a user-registered handler, using SIGIO.
> > > > > 
> > > > > I think it is desirable to be able to achieve similar functionality.
> > > > 
> > > > Does anyone have any comments?
> > > 
> > > I'll try to add that functionality soon,
> > > I'll cc you on patch
> > 
> > Thank you.
> > I look forward to your patch.
> 
> Do you have a specific plan? 
> I would like to know the rough period until your patch is made.
 
I will also try to implement these interfaces.

Best Regards
Shunsuke

Re: libperf: lack of interface

[Jiri Olsa]

On Tue, Dec 07, 2021 at 05:49:16AM +0000, nakamura.shun@fujitsu.com wrote:
> Hi jirka
> 
> > > > On Mon, Oct 18, 2021 at 08:57:20AM +0000, nakamura.shun@fujitsu.com wrote:
> > > > > Hi
> > > > > 
> > > > > > > On Wed, Sep 01, 2021 at 10:36:40AM -0300, Arnaldo Carvalho de Melo wrote:
> > > > > > > > Em Wed, Sep 01, 2021 at 09:45:10AM +0000, nakamura.shun@fujitsu.com escreveu:
> > > > > > > > > Hello.
> > > > > > > > >
> > > > > > > > > I'm trying to change rdpmc test in perf_event_tests[1] to use libperf, but libperf doesn't have enough interfaces.
> > > > > > > > > Does anyone plan to implement any of these libperf features?
> > > > > > > > >
> > > > > > > > > - Interfaces that can run ioctl (PERF_EVENT_IOC_RESET) from userland
> > > > > > > > > - Interfaces that can run fcntl (SIGIO) from userland
> > > > > > >
> > > > > > > hi,
> > > > > > > we could add perf_evsel__fd like below, would it help your usecase?
> > > > > > >
> > > > > > > if you described your usecases in more details we could
> > > > > > > see if we could add/move something to libperf for that
> > > > > > >
> > > > > > > as Arnaldo said below it could be already in tools/perf/util/*.c
> > > > > > > somewhere ;-)
> > > > > > 
> > > > > > As Peter says, I understood that for rdpmc, no reset is needed.
> > > > > > 
> > > > > > However, PAPI resets it explicitly, for example, at PAPI_reset.
> > > > > > In other, PAPI also has the ability to call PERF_EVENT_IOC_REFLESH on overflow to call a user-registered handler, using SIGIO.
> > > > > > 
> > > > > > I think it is desirable to be able to achieve similar functionality.
> > > > > 
> > > > > Does anyone have any comments?
> > > > 
> > > > I'll try to add that functionality soon,
> > > > I'll cc you on patch
> > > 
> > > Thank you.
> > > I look forward to your patch.
> > 
> > Do you have a specific plan? 
> > I would like to know the rough period until your patch is made.
>  
> I will also try to implement these interfaces.

sorry for late reply, please do

thanks,
jirka