Re: [PATCH 2/2] perf/x86/amd: Add support for Large Increment per Cycle Events

[Peter Zijlstra]

On Wed, Jan 08, 2020 at 04:26:47PM -0600, Kim Phillips wrote:
> On 12/20/19 6:09 AM, Peter Zijlstra wrote:
> > On Thu, Nov 14, 2019 at 12:37:20PM -0600, Kim Phillips wrote:

> >> @@ -926,10 +944,14 @@ int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
> >>  			break;
> >>  
> >>  		/* not already used */
> >> -		if (test_bit(hwc->idx, used_mask))
> >> +		if (test_bit(hwc->idx, used_mask) || (is_large_inc(hwc) &&
> >> +		    test_bit(hwc->idx + 1, used_mask)))
> >>  			break;
> >>  
> >>  		__set_bit(hwc->idx, used_mask);
> >> +		if (is_large_inc(hwc))
> >> +			__set_bit(hwc->idx + 1, used_mask);
> >> +
> >>  		if (assign)
> >>  			assign[i] = hwc->idx;
> >>  	}
> > 
> > This is just really sad.. fixed that too.
> 
> [*]

> If I undo re-adding my perf_assign_events code, and re-add my "not
> already used" code that you removed - see [*] above - the problem DOES
> go away, and all the counts are all accurate.
> 
> One problem I see with your change in the "not already used" fastpath
> area, is that the new mask variable gets updated with position 'i'
> regardless of any previous Large Increment event assignments.

Urgh, I completely messed that up. Find the below delta (I'll push out a
new version to queue.git as well).

> I.e., a
> successfully scheduled large increment event assignment may have
> already consumed that 'i' slot for its Merge event in a previous
> iteration of the loop.  So if the fastpath scheduler fails to assign
> that following event, the slow path is wrongly entered due to a wrong
> 'i' comparison with 'n', for example.

That should only be part of the story though; the fast-path is purely
optional. False-negatives on the fast path should not affect
functionality, only performance. False-positives on the fast path are a
no-no of course.

---
diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 222f172cbaf5..3bb738f5a472 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -937,7 +937,7 @@ int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
 	 * fastpath, try to reuse previous register
 	 */
 	for (i = 0; i < n; i++) {
-		u64 mask = BIT_ULL(i);
+		u64 mask;
 
 		hwc = &cpuc->event_list[i]->hw;
 		c = cpuc->event_constraint[i];
@@ -950,6 +950,7 @@ int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
 		if (!test_bit(hwc->idx, c->idxmsk))
 			break;
 
+		mask = BIT_ULL(hwc->idx);
 		if (is_counter_pair(hwc))
 			mask |= mask << 1;