Re: [PATCH 2/5] x86, perf: Add option to disable reading branch flags/cycles

[Peter Zijlstra <peterz@infradead.org>]

On Wed, May 27, 2015 at 09:13:15PM -0700, Andi Kleen wrote:
> From: Andi Kleen <ak@linux.intel.com>
> 
> With LBRv5 reading the extra LBR flags like mispredict, TSX, cycles
> is not free anymore, as it has moved to a separate MSR.
> 
> For callstack mode we don't need any of this information; so we can
> avoid the unnecessary MSR read. Add flags to the perf interface
> where perf record can request not collecting this information.
> I added sample_type flags for CYCLES and FLAGS. It's a bit unusual for
> sample_types to be negative (disable), not positive (enable), but
> since the legacy ABI reported the flags we need some form of explicit
> disabling to avoid breaking the ABI. In theory it would be possible
> to make CYCLES opt-in (as it's not deployed yet), but I also made it
> opt-out to be symmetric to FLAGS.
> 
> After we have the flags the x86 perf code can keep track if any
> users need the flags. If noone needs it the information is not
> collected.

> diff --git a/arch/x86/kernel/cpu/perf_event_intel_lbr.c b/arch/x86/kernel/cpu/perf_event_intel_lbr.c
> index 1fd8b5a..5de2048 100644
> --- a/arch/x86/kernel/cpu/perf_event_intel_lbr.c
> +++ b/arch/x86/kernel/cpu/perf_event_intel_lbr.c
> @@ -340,6 +340,10 @@ void intel_pmu_lbr_enable(struct perf_event *event)
>  	}
>  
>  	cpuc->lbr_users++;
> +	if (!(event->attr.sample_type & PERF_SAMPLE_BRANCH_NO_FLAGS))
> +		cpuc->lbr_flags_users++;
> +	if (!(event->attr.sample_type & PERF_SAMPLE_BRANCH_NO_CYCLES))
> +		cpuc->lbr_cycles_users++;
>  	perf_sched_cb_inc(event->ctx->pmu);
>  }

This patch seems to add an unfortunate amount of branches. And while I
appreciate it'll all be cheaper than the SKL MSR read, it does add
overhead to the older chips.

Also, I think its broken. We can now have two events, with the 'same'
LBR config but with different flags/cycles settings.
__intel_shared_reg_get_constraints() will find a match, but the enable
code above will disable flags/cycles for both of them.

> diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
> index 1b8bd4a..8dd5765 100644
> --- a/include/uapi/linux/perf_event.h
> +++ b/include/uapi/linux/perf_event.h
> @@ -138,6 +138,8 @@ enum perf_event_sample_format {
>  	PERF_SAMPLE_IDENTIFIER			= 1U << 16,
>  	PERF_SAMPLE_TRANSACTION			= 1U << 17,
>  	PERF_SAMPLE_REGS_INTR			= 1U << 18,
> +	PERF_SAMPLE_BRANCH_NO_FLAGS		= 1U << 19,
> +	PERF_SAMPLE_BRANCH_NO_CYCLES		= 1U << 20,
>  
>  	PERF_SAMPLE_MAX = 1U << 19,		/* non-ABI */
>  };

Should this not be part of perf_branch_sample_type instead? That seems
to otherwise specify all the branch stack details.