Messages in this thread | | 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 >
| |