Re: [Patch v2 10/24] perf/x86/intel: Process arch-PEBS records or record fragments

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Feb 25, 2025 at 11:39:27AM +0100, Peter Zijlstra wrote:
> On Tue, Feb 18, 2025 at 03:28:04PM +0000, Dapeng Mi wrote:
> > A significant difference with adaptive PEBS is that arch-PEBS record
> > supports fragments which means an arch-PEBS record could be split into
> > several independent fragments which have its own arch-PEBS header in
> > each fragment.
> > 
> > This patch defines architectural PEBS record layout structures and add
> > helpers to process arch-PEBS records or fragments. Only legacy PEBS
> > groups like basic, GPR, XMM and LBR groups are supported in this patch,
> > the new added YMM/ZMM/OPMASK vector registers capturing would be
> > supported in subsequent patches.
> > 
> > Signed-off-by: Dapeng Mi <dapeng1.mi@linux.intel.com>
> > ---
> >  arch/x86/events/intel/core.c      |   9 ++
> >  arch/x86/events/intel/ds.c        | 219 ++++++++++++++++++++++++++++++
> >  arch/x86/include/asm/msr-index.h  |   6 +
> >  arch/x86/include/asm/perf_event.h | 100 ++++++++++++++
> >  4 files changed, 334 insertions(+)
> > 
> > diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
> > index 37540eb80029..184f69afde08 100644
> > --- a/arch/x86/events/intel/core.c
> > +++ b/arch/x86/events/intel/core.c
> > @@ -3124,6 +3124,15 @@ static int handle_pmi_common(struct pt_regs *regs, u64 status)
> >  			wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
> >  	}
> >  
> > +	/*
> > +	 * Arch PEBS sets bit 54 in the global status register
> > +	 */
> > +	if (__test_and_clear_bit(GLOBAL_STATUS_ARCH_PEBS_THRESHOLD_BIT,
> > +				 (unsigned long *)&status)) {
> 
> Will arch_pebs hardware ever toggle bit 62?

This had me looking at the bit 62 handling, and I noticed the thing from
commit 8077eca079a2 ("perf/x86/pebs: Add workaround for broken OVFL
status on HSW+").

Did that ever get fixed in later chips; notably I'm assuming ARCH PEBS
does not suffer this?

Also, should that workaround have been extended to also include
GLOBAL_STATUS_PERF_METRICS_OVF in that mask, or was that defect fixed
for every chip capable of metrics stuff?

In any case, I think we want a patch clarifying the situation with a
comment.


> > +		handled++;
> > +		x86_pmu.drain_pebs(regs, &data);
> 
> 		static_call(x86_pmu_drain_pebs)(regs, &data);
> 
> > +	}
>