Re: [PATCH] scripts/checksyscalls.sh: only whine perf_counter_open when supported

[Mike Frysinger <vapier.adi@gmail.com>]

On Fri, Jun 12, 2009 at 09:09, Ingo Molnar wrote:
> * Mike Frysinger <vapier.adi@gmail.com> wrote:
>> On Fri, Jun 12, 2009 at 08:59, Ingo Molnar wrote:
>> > * Mike Frysinger <vapier.adi@gmail.com> wrote:
>> >> On Fri, Jun 12, 2009 at 08:31, Ingo Molnar wrote:
>> >> > * Mike Frysinger <vapier.adi@gmail.com> wrote:
>> >> >> On Fri, Jun 12, 2009 at 08:17, Ingo Molnar wrote:
>> >> >> > * Mike Frysinger <vapier.adi@gmail.com> wrote:
>> >> >> >> On Fri, Jun 12, 2009 at 08:05, Ingo Molnar wrote:
>> >> >> >> > * Mike Frysinger <vapier@gentoo.org> wrote:
>> >> >> >> >> If the port does not support HAVE_PERF_COUNTERS, then they can't
>> >> >> >> >> support the perf_counter_open syscall either.  Rather than forcing
>> >> >> >> >> everyone to add an ignore (or suffer the warning until they get
>> >> >> >> >> around to implementing support), only whine about the syscall when
>> >> >> >> >> applicable.
>> >> >> >> >
>> >> >> >> > No, this patch is wrong - it's really easy to add support: just hook
>> >> >> >> > up the syscall. This should happen for every architecture really, so
>> >> >> >> > the warning is correct and it should not be patched out.
>> >> >> >> >
>> >> >> >> > PMU support is not required to get perfcounters support: if an
>> >> >> >> > architecture hooks up the syscall it will get generic software
>> >> >> >> > counters and the tools will work as well.
>> >> >> >> >
>> >> >> >> > Profiling falls back to a hrtimer-based sampling method - this is a
>> >> >> >> > much better fallback than oprofile's fall-back to the timer tick.
>> >> >> >> > This hrtimer based sampling is dynticks/nohz-correct and can go
>> >> >> >> > beyond HZ if the architecture supports hrtimers.
>> >> >> >>
>> >> >> >> if there is generic support available, why must every arch select
>> >> >> >> HAVE_PERF_COUNTERS in their Kconfig ?
>> >> >> >
>> >> >> > Because we only want to enable it on architectures that have tested
>> >> >> > it. It should only need a syscall addition, but nothing beats having
>> >> >> > tested things, hence we have that additional Kconfig symbol.
>> >> >>
>> >> >> that is a pretty weak reason. [...]
>> >> >
>> >> > It isnt - this is proper isolation - dont offer something to the
>> >> > user to enable that 1) cannot be used due to the lack of a syscall
>> >> > 2) has not been tested by anyone on that architecture, ever.
>> >> >
>> >> > That way say build breakages or runtime failures due to perfcounters
>> >> > only become possible on an architecture if the architecture
>> >> > maintainer has hooked up the syscall and has provided
>> >> > HAVE_PERF_COUNTERS explicitly.
>> >>
>> >> except that the syscall presence is trivial to detect in the code by
>> >> something like:
>> >> #ifndef __NR_perf_counter_open
>> >> # error sorry, your arch has not hooked up perf_counter_open syscall yet
>> >> #endif
>> >>
>> >> as for "no arch testing yet", there are plenty of drivers which lack
>> >> arch depends in the Kconfig specifically so that it can be *easily*
>> >> tested on random systems out there without requiring manual twiddling.
>> >
>> > This is a new kernel subsystem, not just yet another driver.
>>
>> so what ?  if it has generic pieces, it is exactly the same as yet
>> another generic driver.  people should be able to randomly test
>> build it when possible to discover latent issues that your testing
>> limited to one arch did not find.
>>
>> > Which bit of: "we dont want perfcounters to be enabled in the
>> > Kconfig on architectures that have no syscalls and no testing for
>> > it" is hard to understand? It is a valid technical concern.
>>
>> your (1) is valid but i already pointed out a simple fix for that.
>> your (2) is not.
>
> Uhm, your 'fix':
>
>  #ifndef __NR_perf_counter_open
>  # error sorry, your arch has not hooked up perf_counter_open syscall yet
>  #endif
>
> is completely unacceptable. We dont propagate build failures via
> user-enable config options, we never did. There's a lot of people
> doing randconfig builds - if it randomly failed due to your 'fix'
> that would upset a lot of testing for no good reason.

accept that is a valid bug: the arch is missing the syscall and it
should hook it up
-mike