On Mon, Nov 07, 2011 at 03:07:50PM +0100, Peter Zijlstra wrote:
> On Thu, 2011-11-03 at 14:33 +0200, Gleb Natapov wrote:
> > @@ -1580,6 +1580,8 @@ __init int intel_pmu_init(void)
> > x86_pmu.num_counters = eax.split.num_counters;
> > x86_pmu.cntval_bits = eax.split.bit_width;
> > x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1;
> > + x86_pmu.events_mask = ebx;
> > + x86_pmu.events_mask_len = eax.split.mask_length;
> >
> > /*
> > * Quirk: v2 perfmon does not report fixed-purpose events, so
> > @@ -1651,6 +1653,7 @@ __init int intel_pmu_init(void)
> > * architectural event which is often completely bogus:
> > */
> > intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
> > + x86_pmu.events_mask &= ~0x40;
> >
> > pr_cont("erratum AAJ80 worked around, ");
> > }
>
> It might make sense to introduce cpuid10_ebx or so, also I think the
> removal of the branch-miss-retired event is either unwanted or
> incomplete. As seen software already expects that bit to be set, even
> though its known broken.
>
I removed branch-miss-retired here because for perf user it exists. Perf
approximates it by other event but perf user shouldn't know that. A
guest is not always runs with exactly same cpu model number as a host,
so if we will not drop the bit here if guest will see cpu model other
than the one that has this quirk it will not be able to use the event.
BTW why perf does not check event mask to see if architectural event is
available before programming it?
> At the very least add a full ebx iteration to disable unsupported events
> in the intel-v1 case.
--
Gleb.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
