Re: [PATCH V3 20/37] perf script: Add 'synth' event type for synthesized events

Re: [PATCH V3 20/37] perf script: Add 'synth' event type for synthesized events

[Arnaldo Carvalho de Melo]

Em Wed, Jun 21, 2017 at 01:17:19PM +0300, Adrian Hunter escreveu:
> +++ b/tools/perf/util/event.h
> @@ -252,6 +252,9 @@ enum auxtrace_error_type {
>        PERF_AUXTRACE_ERROR_MAX
> };

> +/* Attribute type for custom synthesized events */
> +#define PERF_TYPE_SYNTH		3000000000

Why don't you make it PERF_TYPE_MAX and bump PERF_TYPE_MAX by one? I.e.
this way we have what can be in attr.type in a nice enumeration, can
validate it more easily (attr.type < PERF_TYPE_MAX) and will not need to 
do those conversions to/from OUTPUT_TYPE_/PERF_TYPE_).

Peter: now its not the PERF_RECORD_ namespace that userspaces want a
chunk of, its PERF_TYPE_, which so far has been pretty stable, grabing
_one_ for event synthesizing things like Intel PT (and ARM's coresight,
I think) directly at include/uapi/linux/perf_event.h's perf_type_id enum
looks cleaner, no?

- Arnaldo

Re: [PATCH V3 20/37] perf script: Add 'synth' event type for synthesized events

[Adrian Hunter]

On 06/21/2017 04:51 PM, Arnaldo Carvalho de Melo wrote:
> Em Wed, Jun 21, 2017 at 01:17:19PM +0300, Adrian Hunter escreveu:
>> +++ b/tools/perf/util/event.h
>> @@ -252,6 +252,9 @@ enum auxtrace_error_type {
>>        PERF_AUXTRACE_ERROR_MAX
>> };
> 
>> +/* Attribute type for custom synthesized events */
>> +#define PERF_TYPE_SYNTH		3000000000
> 
> Why don't you make it PERF_TYPE_MAX and bump PERF_TYPE_MAX by one? I.e.
> this way we have what can be in attr.type in a nice enumeration, can
> validate it more easily (attr.type < PERF_TYPE_MAX) and will not need to 
> do those conversions to/from OUTPUT_TYPE_/PERF_TYPE_).

PERF_TYPE_ is dynamically allocated above PERF_TYPE_MAX for PMUs.  Presently
perf_pmu_register() calls idr_alloc() with end=0 which limits the allocation
to INT_MAX.

> Peter: now its not the PERF_RECORD_ namespace that userspaces want a
> chunk of, its PERF_TYPE_, which so far has been pretty stable, grabing
> _one_ for event synthesizing things like Intel PT (and ARM's coresight,
> I think) directly at include/uapi/linux/perf_event.h's perf_type_id enum
> looks cleaner, no?

Re: [PATCH V3 20/37] perf script: Add 'synth' event type for synthesized events

[Arnaldo Carvalho de Melo]

Em Wed, Jun 21, 2017 at 07:41:04PM +0300, Adrian Hunter escreveu:
> On 06/21/2017 04:51 PM, Arnaldo Carvalho de Melo wrote:
> > Em Wed, Jun 21, 2017 at 01:17:19PM +0300, Adrian Hunter escreveu:
> >> +++ b/tools/perf/util/event.h
> >> @@ -252,6 +252,9 @@ enum auxtrace_error_type {
> >>        PERF_AUXTRACE_ERROR_MAX
> >> };
> > 
> >> +/* Attribute type for custom synthesized events */
> >> +#define PERF_TYPE_SYNTH		3000000000
> > 
> > Why don't you make it PERF_TYPE_MAX and bump PERF_TYPE_MAX by one? I.e.
> > this way we have what can be in attr.type in a nice enumeration, can
> > validate it more easily (attr.type < PERF_TYPE_MAX) and will not need to 
> > do those conversions to/from OUTPUT_TYPE_/PERF_TYPE_).
 
> PERF_TYPE_ is dynamically allocated above PERF_TYPE_MAX for PMUs.  Presently
> perf_pmu_register() calls idr_alloc() with end=0 which limits the allocation
> to INT_MAX.

Oh, forgot about that, guess a comment right beside PERF_TYPE_MAX is in
demand :-\

So why not:

/*
 * PERF_TYPE_ is dynamically allocated above PERF_TYPE_MAX for PMUs.  Presently
 * perf_pmu_register() calls idr_alloc() with end=0 which limits the allocation
 * to INT_MAX.
 */
#define PERF_TYPE_SYNTH		(INT_MAX + 1L)

I.e. wouldn't be some arbitrarily huge value, but one right after what
was defined as the area for the dynamicly allocated PERF_TYPE_
"namespace" for PMUs, right?

> > Peter: now its not the PERF_RECORD_ namespace that userspaces want a
> > chunk of, its PERF_TYPE_, which so far has been pretty stable, grabing
> > _one_ for event synthesizing things like Intel PT (and ARM's coresight,
> > I think) directly at include/uapi/linux/perf_event.h's perf_type_id enum
> > looks cleaner, no?

- Arnaldo

Re: [PATCH V3 20/37] perf script: Add 'synth' event type for synthesized events

[Adrian Hunter]

On 06/21/2017 08:29 PM, Arnaldo Carvalho de Melo wrote:
> Em Wed, Jun 21, 2017 at 07:41:04PM +0300, Adrian Hunter escreveu:
>> On 06/21/2017 04:51 PM, Arnaldo Carvalho de Melo wrote:
>>> Em Wed, Jun 21, 2017 at 01:17:19PM +0300, Adrian Hunter escreveu:
>>>> +++ b/tools/perf/util/event.h
>>>> @@ -252,6 +252,9 @@ enum auxtrace_error_type {
>>>>        PERF_AUXTRACE_ERROR_MAX
>>>> };
>>>
>>>> +/* Attribute type for custom synthesized events */
>>>> +#define PERF_TYPE_SYNTH		3000000000
>>>
>>> Why don't you make it PERF_TYPE_MAX and bump PERF_TYPE_MAX by one? I.e.
>>> this way we have what can be in attr.type in a nice enumeration, can
>>> validate it more easily (attr.type < PERF_TYPE_MAX) and will not need to 
>>> do those conversions to/from OUTPUT_TYPE_/PERF_TYPE_).
>  
>> PERF_TYPE_ is dynamically allocated above PERF_TYPE_MAX for PMUs.  Presently
>> perf_pmu_register() calls idr_alloc() with end=0 which limits the allocation
>> to INT_MAX.
> 
> Oh, forgot about that, guess a comment right beside PERF_TYPE_MAX is in
> demand :-\
> 
> So why not:
> 
> /*
>  * PERF_TYPE_ is dynamically allocated above PERF_TYPE_MAX for PMUs.  Presently
>  * perf_pmu_register() calls idr_alloc() with end=0 which limits the allocation
>  * to INT_MAX.
>  */
> #define PERF_TYPE_SYNTH		(INT_MAX + 1L)
> 
> I.e. wouldn't be some arbitrarily huge value, but one right after what
> was defined as the area for the dynamicly allocated PERF_TYPE_
> "namespace" for PMUs, right?

That seems fine.  Pedantically it should be (INT_MAX + 1U) otherwise it will
be negative on a 32-bit system.

> 
>>> Peter: now its not the PERF_RECORD_ namespace that userspaces want a
>>> chunk of, its PERF_TYPE_, which so far has been pretty stable, grabing
>>> _one_ for event synthesizing things like Intel PT (and ARM's coresight,
>>> I think) directly at include/uapi/linux/perf_event.h's perf_type_id enum
>>> looks cleaner, no?
> 
> - Arnaldo
> 

Re: [PATCH V3 20/37] perf script: Add 'synth' event type for synthesized events

[Arnaldo Carvalho de Melo]

Em Wed, Jun 21, 2017 at 11:20:03PM +0300, Adrian Hunter escreveu:
> On 06/21/2017 08:29 PM, Arnaldo Carvalho de Melo wrote:
> > Em Wed, Jun 21, 2017 at 07:41:04PM +0300, Adrian Hunter escreveu:
> >> PERF_TYPE_ is dynamically allocated above PERF_TYPE_MAX for PMUs.  Presently
> >> perf_pmu_register() calls idr_alloc() with end=0 which limits the allocation
> >> to INT_MAX.

> > Oh, forgot about that, guess a comment right beside PERF_TYPE_MAX is in
> > demand :-\

> > So why not:

> > /*
> >  * PERF_TYPE_ is dynamically allocated above PERF_TYPE_MAX for PMUs.  Presently
> >  * perf_pmu_register() calls idr_alloc() with end=0 which limits the allocation
> >  * to INT_MAX.
> >  */
> > #define PERF_TYPE_SYNTH		(INT_MAX + 1L)

> > I.e. wouldn't be some arbitrarily huge value, but one right after what
> > was defined as the area for the dynamicly allocated PERF_TYPE_
> > "namespace" for PMUs, right?
 
> That seems fine.  Pedantically it should be (INT_MAX + 1U) otherwise it will
> be negative on a 32-bit system.

oh, not pedantic at all, thanks for the fix, will update your patch and
continue from there, Ingo already pulled everything up to this point,
btw.

- Arnaldo