perf, x86: Prefer fixed-purpose counters when scheduling

This avoids a scheduling failure for cases like:

  cycles, cycles, instructions, instructions (on Core2)

Which would end up being programmed like:

  PMC0, PMC1, FP-instructions, fail

Because all events will have the same weight.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/n/tip-8tnwb92asqj7xajqqoty4gel@git.kernel.orgSigned-off-by: Ingo Molnar <mingo@elte.hu>

arch/x86/kernel/cpu/perf_event.c

@@ -574,17 +574,26 @@ static bool __perf_sched_find_counter(struct perf_sched *sched)
 
 	c = sched->constraints[sched->state.event];
 
+	/* Prefer fixed purpose counters */
+	if (x86_pmu.num_counters_fixed) {
+		idx = X86_PMC_IDX_FIXED;
+		for_each_set_bit_cont(idx, c->idxmsk, X86_PMC_IDX_MAX) {
+			if (!__test_and_set_bit(idx, sched->state.used))
+				goto done;
+		}
+	}
 	/* Grab the first unused counter starting with idx */
 	idx = sched->state.counter;
-	for_each_set_bit_cont(idx, c->idxmsk, X86_PMC_IDX_MAX) {
+	for_each_set_bit_cont(idx, c->idxmsk, X86_PMC_IDX_FIXED) {
 		if (!__test_and_set_bit(idx, sched->state.used))
-			break;
+			goto done;
 	}
-	sched->state.counter = idx;
 
-	if (idx >= X86_PMC_IDX_MAX)
 	return false;
 
+done:
+	sched->state.counter = idx;
+
 	if (c->overlap)
 		perf_sched_save_state(sched);