Re: [PATCH 2/3] perf/x86/pebs: add workaround for broken OVFL status on HSW

[Peter Zijlstra]

Just spotted this again, ping?

On Thu, Mar 10, 2016 at 11:42:36AM +0100, Peter Zijlstra wrote:
> On Wed, Mar 09, 2016 at 09:40:07AM -0800, Stephane Eranian wrote:
> > With your queue.tip perf/core branch, I run into another problem.
> > I am monitoring with 2 PEBS events and I have the NMI watchdog enabled.
> > 
> > I see non-EXACT PEBS records again, despite my change (which is in).
> > I tracked it down to the following issue after the testing of bit 62:
> > 
> > [31137.273061] CPU71 status=0x200000001 orig_status=0x200000001 bit62=0
> > 
> > The IRQ handler is called because the fixed counter for the NMI has overflowed
> > and it sees this in bit 33, but it also sees that one of the PEBS
> > events has also
> > overflowed, yet bit 62 is not set. Therefore both overflows are
> > treated as regular
> > and the drain_pebs() is not called generating a non-EXACT record for the PEBS
> > counter (counter 0). So something is wrong still and this is on Broadwell.
> > 
> > First, I don't understand why the OVF bit for counter 0 is set. It
> > should not according
> > to specs because the counter is in PEBS mode. There must be a race there. So we
> > have to handle it by relying on cpuc->pebs_enabled. I will try that.
> > We likely also
> > need to force OVF bit 62 to 1 so we can ack it in the end (and in case
> > it gets set).
> 
> How about we make the clear of pebs_enabled unconditional?
> 
> ---
>  arch/x86/events/intel/core.c | 20 ++++++++++----------
>  1 file changed, 10 insertions(+), 10 deletions(-)
> 
> diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
> index 68fa55b4d42e..dc9579665425 100644
> --- a/arch/x86/events/intel/core.c
> +++ b/arch/x86/events/intel/core.c
> @@ -1883,6 +1883,16 @@ static int intel_pmu_handle_irq(struct pt_regs *regs)
>  	status &= ~(GLOBAL_STATUS_COND_CHG |
>  		    GLOBAL_STATUS_ASIF |
>  		    GLOBAL_STATUS_LBRS_FROZEN);
> +	/*
> +	 * There are cases where, even though, the PEBS ovfl bit is set
> +	 * in GLOBAL_OVF_STATUS, the PEBS events may also have their
> +	 * overflow bits set for their counters. We must clear them
> +	 * here because they have been processed as exact samples in
> +	 * the drain_pebs() routine. They must not be processed again
> +	 * in the for_each_bit_set() loop for regular samples below.
> +	 */
> +	status &= ~cpuc->pebs_enabled;
> +
>  	if (!status)
>  		goto done;
>  
> @@ -1892,16 +1902,6 @@ static int intel_pmu_handle_irq(struct pt_regs *regs)
>  	if (__test_and_clear_bit(62, (unsigned long *)&status)) {
>  		handled++;
>  		x86_pmu.drain_pebs(regs);
> -		/*
> -		 * There are cases where, even though, the PEBS ovfl bit is set
> -		 * in GLOBAL_OVF_STATUS, the PEBS events may also have their
> -		 * overflow bits set for their counters. We must clear them
> -		 * here because they have been processed as exact samples in
> -		 * the drain_pebs() routine. They must not be processed again
> -		 * in the for_each_bit_set() loop for regular samples below.
> -		 */
> -		status &= ~cpuc->pebs_enabled;
> -		status &= x86_pmu.intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI;
>  	}
>  
>  	/*