Re: [PATCH v3 6/9] perf, x86: handle multiple records in PEBS buffer

[Yan, Zheng]

On 07/26/2014 12:40 AM, Andi Kleen wrote:
>> Suppose two pebs events, one has exclude_kernel set. It overflows,
>> before entering the kernel, the other event generates PEBS records from
>> inside the kernel with both events marked in the overflow field.
>>
>> And only once we leave the kernel can the exclude_kernel event tick
>> again and trigger the assist, finalyl clearing the bit.
>>
>> If you were to report the records to both events, one would get a lot of
>> kernel info he was not entitled to.
> 
> Ok that case can be filtered in software. Shouldn't be too difficult.
> Perhaps just using ip
> 
>       if (event->attr.exclude_kernel && pebs->ip >= __PAGE_OFFSET)
>           skip;
>       if (event->attr.exclude_user && pebs->ip < __PAGE_OFFSET)
>           skip;
> 
> This would also help with the existing skid.
> 
> Any other concerns?
> 
> -Andi
> 

how about following patch
diff --git a/arch/x86/kernel/cpu/perf_event_intel_ds.c 
b/arch/x86/kernel/cpu/perf_event_intel_ds.c
index 33b4c0e..ea76507 100644
--- a/arch/x86/kernel/cpu/perf_event_intel_ds.c
+++ b/arch/x86/kernel/cpu/perf_event_intel_ds.c
@@ -1016,6 +1016,16 @@ static void setup_pebs_sample_data(struct perf_event 
*event,
                data->br_stack = &cpuc->lbr_stack;
 }
 
+static inline bool intel_pmu_pebs_filter(struct perf_event *event,
+                                        struct pebs_record_nhm *record)
+{
+       if (event->attr.exclude_user && !kernel_ip(record->ip))
+               return true;
+       if (event->attr.exclude_kernel && kernel_ip(record->ip))
+               return true;
+       return false;
+}
+
 static void __intel_pmu_pebs_event(struct perf_event *event,
                                   struct pt_regs *iregs,
                                   void *at, void *top, int count)
@@ -1052,6 +1062,8 @@ static void __intel_pmu_pebs_event(struct perf_event 
*event,
                struct pebs_record_nhm *p = at;
                if (!(p->status & (1 << event->hw.idx)))
                        continue;
+               if (intel_pmu_pebs_filter(event, p))
+                       continue;
 
                setup_pebs_sample_data(event, iregs, at, &data, &regs);
                perf_output_sample(&handle, &header, &data, event);
@@ -1139,6 +1151,8 @@ static void intel_pmu_drain_pebs_nhm(struct pt_regs 
*iregs)
                        WARN_ON_ONCE(!event);
                        if (!event->attr.precise_ip)
                                continue;
+                       if (intel_pmu_pebs_filter(event, p))
+                               continue;
                        counts[bit]++;
                }
        }
@@ -1149,7 +1163,8 @@ static void intel_pmu_drain_pebs_nhm(struct pt_regs 
*iregs)
                event = cpuc->events[bit];
                for (at = base; at < top; at += x86_pmu.pebs_record_size) {
                        struct pebs_record_nhm *p = at;
-                       if (p->status & (1 << bit))
+                       if ((p->status & (1 << bit)) &&
+                           !intel_pmu_pebs_filter(event, p))
                                break;
                }
 
---

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/