Re: [RFC][PATCH 09/11] perf_counter: revamp syscall input ABI

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2009-03-19 at 12:41 +0100, Peter Zijlstra wrote:
> On Thu, 2009-03-19 at 09:15 +1100, Paul Mackerras wrote:
> > Peter Zijlstra writes:
> > 
> > > Ah, I thought we should keep a pure 64 bit raw value. You never know
> > > what hardware will do.
> > 
> > Oh I see, you use hw_event->raw_event if hw_event->raw is set.  I
> > missed that before.
> > 
> > Still, you're putting that into hwc->config along with other bits like
> > ARCH_PERFMON_EVENTSEL_USR and ARCH_PERFMON_EVENTSEL_OS, so I would
> > think we could spare two bits for the type, leaving 62 bits for the
> > raw event code.  And if that isn't enough, there's the
> > hw_event.extra_config_len field, which allows userspace to pass in
> > arbitrary amounts of extra PMU configuration data.
> 
> Good point, overflow interrupt, and usr/os/hv event filter and enable
> bits are usually in the config word.
> 
> OK, how about the below? I didn't cut it to 2 bits, as that would
> already exhaust the TYPE space -- then again, 60 does feel cramped a
> bit..

Hmm, we could play dirty and do:

union {
  struct {
    __u64 raw_event_id : 63,
          raw          :  1;
  };
  struct {
    __u64 event_id     : 56,
          type            8;
  };
};

Giving us 7 bit type space.