Linux Audio

Check our new training course

Real-Time Linux with PREEMPT_RT training

Feb 18-20, 2025
Register
Loading...
v4.17
  1// SPDX-License-Identifier: GPL-2.0
  2#include <stdio.h>
  3#include "evsel.h"
  4#include "stat.h"
  5#include "color.h"
  6#include "pmu.h"
  7#include "rblist.h"
  8#include "evlist.h"
  9#include "expr.h"
 10#include "metricgroup.h"
 
 
 
 
 
 
 
 11
 12/*
 13 * AGGR_GLOBAL: Use CPU 0
 14 * AGGR_SOCKET: Use first CPU of socket
 15 * AGGR_CORE: Use first CPU of core
 16 * AGGR_NONE: Use matching CPU
 17 * AGGR_THREAD: Not supported?
 18 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 19static bool have_frontend_stalled;
 20
 21struct runtime_stat rt_stat;
 22struct stats walltime_nsecs_stats;
 23
 24struct saved_value {
 25	struct rb_node rb_node;
 26	struct perf_evsel *evsel;
 27	enum stat_type type;
 28	int ctx;
 29	int cpu;
 30	struct runtime_stat *stat;
 31	struct stats stats;
 32};
 33
 34static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
 35{
 36	struct saved_value *a = container_of(rb_node,
 37					     struct saved_value,
 38					     rb_node);
 39	const struct saved_value *b = entry;
 40
 41	if (a->cpu != b->cpu)
 42		return a->cpu - b->cpu;
 43
 44	/*
 45	 * Previously the rbtree was used to link generic metrics.
 46	 * The keys were evsel/cpu. Now the rbtree is extended to support
 47	 * per-thread shadow stats. For shadow stats case, the keys
 48	 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
 49	 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
 50	 */
 51	if (a->type != b->type)
 52		return a->type - b->type;
 53
 54	if (a->ctx != b->ctx)
 55		return a->ctx - b->ctx;
 56
 57	if (a->evsel == NULL && b->evsel == NULL) {
 58		if (a->stat == b->stat)
 59			return 0;
 60
 61		if ((char *)a->stat < (char *)b->stat)
 62			return -1;
 63
 64		return 1;
 65	}
 66
 67	if (a->evsel == b->evsel)
 68		return 0;
 69	if ((char *)a->evsel < (char *)b->evsel)
 70		return -1;
 71	return +1;
 72}
 73
 74static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
 75				     const void *entry)
 76{
 77	struct saved_value *nd = malloc(sizeof(struct saved_value));
 78
 79	if (!nd)
 80		return NULL;
 81	memcpy(nd, entry, sizeof(struct saved_value));
 82	return &nd->rb_node;
 83}
 84
 85static void saved_value_delete(struct rblist *rblist __maybe_unused,
 86			       struct rb_node *rb_node)
 87{
 88	struct saved_value *v;
 89
 90	BUG_ON(!rb_node);
 91	v = container_of(rb_node, struct saved_value, rb_node);
 92	free(v);
 93}
 94
 95static struct saved_value *saved_value_lookup(struct perf_evsel *evsel,
 96					      int cpu,
 97					      bool create,
 98					      enum stat_type type,
 99					      int ctx,
100					      struct runtime_stat *st)
101{
102	struct rblist *rblist;
103	struct rb_node *nd;
104	struct saved_value dm = {
105		.cpu = cpu,
106		.evsel = evsel,
107		.type = type,
108		.ctx = ctx,
109		.stat = st,
110	};
111
112	rblist = &st->value_list;
113
114	nd = rblist__find(rblist, &dm);
115	if (nd)
116		return container_of(nd, struct saved_value, rb_node);
117	if (create) {
118		rblist__add_node(rblist, &dm);
119		nd = rblist__find(rblist, &dm);
120		if (nd)
121			return container_of(nd, struct saved_value, rb_node);
122	}
123	return NULL;
124}
125
126void runtime_stat__init(struct runtime_stat *st)
127{
128	struct rblist *rblist = &st->value_list;
129
130	rblist__init(rblist);
131	rblist->node_cmp = saved_value_cmp;
132	rblist->node_new = saved_value_new;
133	rblist->node_delete = saved_value_delete;
134}
135
136void runtime_stat__exit(struct runtime_stat *st)
137{
138	rblist__exit(&st->value_list);
139}
140
141void perf_stat__init_shadow_stats(void)
142{
143	have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
144	runtime_stat__init(&rt_stat);
145}
146
147static int evsel_context(struct perf_evsel *evsel)
148{
149	int ctx = 0;
150
151	if (evsel->attr.exclude_kernel)
152		ctx |= CTX_BIT_KERNEL;
153	if (evsel->attr.exclude_user)
154		ctx |= CTX_BIT_USER;
155	if (evsel->attr.exclude_hv)
156		ctx |= CTX_BIT_HV;
157	if (evsel->attr.exclude_host)
158		ctx |= CTX_BIT_HOST;
159	if (evsel->attr.exclude_idle)
160		ctx |= CTX_BIT_IDLE;
161
162	return ctx;
163}
164
165static void reset_stat(struct runtime_stat *st)
166{
167	struct rblist *rblist;
168	struct rb_node *pos, *next;
169
170	rblist = &st->value_list;
171	next = rb_first(&rblist->entries);
172	while (next) {
173		pos = next;
174		next = rb_next(pos);
175		memset(&container_of(pos, struct saved_value, rb_node)->stats,
176		       0,
177		       sizeof(struct stats));
178	}
179}
180
181void perf_stat__reset_shadow_stats(void)
182{
183	reset_stat(&rt_stat);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
184	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
185}
186
187void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
188{
189	reset_stat(st);
190}
191
192static void update_runtime_stat(struct runtime_stat *st,
193				enum stat_type type,
194				int ctx, int cpu, u64 count)
195{
196	struct saved_value *v = saved_value_lookup(NULL, cpu, true,
197						   type, ctx, st);
198
199	if (v)
200		update_stats(&v->stats, count);
201}
202
203/*
204 * Update various tracking values we maintain to print
205 * more semantic information such as miss/hit ratios,
206 * instruction rates, etc:
207 */
208void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 count,
209				    int cpu, struct runtime_stat *st)
210{
211	int ctx = evsel_context(counter);
212
213	count *= counter->scale;
214
215	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK) ||
216	    perf_evsel__match(counter, SOFTWARE, SW_CPU_CLOCK))
217		update_runtime_stat(st, STAT_NSECS, 0, cpu, count);
218	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
219		update_runtime_stat(st, STAT_CYCLES, ctx, cpu, count);
220	else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
221		update_runtime_stat(st, STAT_CYCLES_IN_TX, ctx, cpu, count);
222	else if (perf_stat_evsel__is(counter, TRANSACTION_START))
223		update_runtime_stat(st, STAT_TRANSACTION, ctx, cpu, count);
224	else if (perf_stat_evsel__is(counter, ELISION_START))
225		update_runtime_stat(st, STAT_ELISION, ctx, cpu, count);
226	else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
227		update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
228				    ctx, cpu, count);
229	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
230		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
231				    ctx, cpu, count);
232	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
233		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
234				    ctx, cpu, count);
235	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
236		update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
237				    ctx, cpu, count);
238	else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
239		update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
240				    ctx, cpu, count);
241	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
242		update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
243				    ctx, cpu, count);
244	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
245		update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
246				    ctx, cpu, count);
247	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
248		update_runtime_stat(st, STAT_BRANCHES, ctx, cpu, count);
249	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
250		update_runtime_stat(st, STAT_CACHEREFS, ctx, cpu, count);
251	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
252		update_runtime_stat(st, STAT_L1_DCACHE, ctx, cpu, count);
253	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
254		update_runtime_stat(st, STAT_L1_ICACHE, ctx, cpu, count);
255	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
256		update_runtime_stat(st, STAT_LL_CACHE, ctx, cpu, count);
257	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
258		update_runtime_stat(st, STAT_DTLB_CACHE, ctx, cpu, count);
259	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
260		update_runtime_stat(st, STAT_ITLB_CACHE, ctx, cpu, count);
261	else if (perf_stat_evsel__is(counter, SMI_NUM))
262		update_runtime_stat(st, STAT_SMI_NUM, ctx, cpu, count);
263	else if (perf_stat_evsel__is(counter, APERF))
264		update_runtime_stat(st, STAT_APERF, ctx, cpu, count);
265
266	if (counter->collect_stat) {
267		struct saved_value *v = saved_value_lookup(counter, cpu, true,
268							   STAT_NONE, 0, st);
269		update_stats(&v->stats, count);
270	}
271}
272
273/* used for get_ratio_color() */
274enum grc_type {
275	GRC_STALLED_CYCLES_FE,
276	GRC_STALLED_CYCLES_BE,
277	GRC_CACHE_MISSES,
278	GRC_MAX_NR
279};
280
281static const char *get_ratio_color(enum grc_type type, double ratio)
282{
283	static const double grc_table[GRC_MAX_NR][3] = {
284		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
285		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
286		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
287	};
288	const char *color = PERF_COLOR_NORMAL;
289
290	if (ratio > grc_table[type][0])
291		color = PERF_COLOR_RED;
292	else if (ratio > grc_table[type][1])
293		color = PERF_COLOR_MAGENTA;
294	else if (ratio > grc_table[type][2])
295		color = PERF_COLOR_YELLOW;
296
297	return color;
298}
299
300static struct perf_evsel *perf_stat__find_event(struct perf_evlist *evsel_list,
301						const char *name)
302{
303	struct perf_evsel *c2;
304
305	evlist__for_each_entry (evsel_list, c2) {
306		if (!strcasecmp(c2->name, name))
307			return c2;
308	}
309	return NULL;
310}
311
312/* Mark MetricExpr target events and link events using them to them. */
313void perf_stat__collect_metric_expr(struct perf_evlist *evsel_list)
314{
315	struct perf_evsel *counter, *leader, **metric_events, *oc;
316	bool found;
317	const char **metric_names;
318	int i;
319	int num_metric_names;
320
321	evlist__for_each_entry(evsel_list, counter) {
322		bool invalid = false;
323
324		leader = counter->leader;
325		if (!counter->metric_expr)
326			continue;
327		metric_events = counter->metric_events;
328		if (!metric_events) {
329			if (expr__find_other(counter->metric_expr, counter->name,
330						&metric_names, &num_metric_names) < 0)
331				continue;
332
333			metric_events = calloc(sizeof(struct perf_evsel *),
334					       num_metric_names + 1);
335			if (!metric_events)
336				return;
337			counter->metric_events = metric_events;
338		}
339
340		for (i = 0; i < num_metric_names; i++) {
341			found = false;
342			if (leader) {
343				/* Search in group */
344				for_each_group_member (oc, leader) {
345					if (!strcasecmp(oc->name, metric_names[i])) {
346						found = true;
347						break;
348					}
349				}
350			}
351			if (!found) {
352				/* Search ignoring groups */
353				oc = perf_stat__find_event(evsel_list, metric_names[i]);
354			}
355			if (!oc) {
356				/* Deduping one is good enough to handle duplicated PMUs. */
357				static char *printed;
358
359				/*
360				 * Adding events automatically would be difficult, because
361				 * it would risk creating groups that are not schedulable.
362				 * perf stat doesn't understand all the scheduling constraints
363				 * of events. So we ask the user instead to add the missing
364				 * events.
365				 */
366				if (!printed || strcasecmp(printed, metric_names[i])) {
367					fprintf(stderr,
368						"Add %s event to groups to get metric expression for %s\n",
369						metric_names[i],
370						counter->name);
371					printed = strdup(metric_names[i]);
372				}
373				invalid = true;
374				continue;
375			}
376			metric_events[i] = oc;
377			oc->collect_stat = true;
378		}
379		metric_events[i] = NULL;
380		free(metric_names);
381		if (invalid) {
382			free(metric_events);
383			counter->metric_events = NULL;
384			counter->metric_expr = NULL;
385		}
386	}
387}
388
389static double runtime_stat_avg(struct runtime_stat *st,
390			       enum stat_type type, int ctx, int cpu)
391{
392	struct saved_value *v;
393
394	v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
395	if (!v)
396		return 0.0;
397
398	return avg_stats(&v->stats);
399}
400
401static double runtime_stat_n(struct runtime_stat *st,
402			     enum stat_type type, int ctx, int cpu)
403{
404	struct saved_value *v;
405
406	v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
407	if (!v)
408		return 0.0;
409
410	return v->stats.n;
411}
412
413static void print_stalled_cycles_frontend(int cpu,
414					  struct perf_evsel *evsel, double avg,
415					  struct perf_stat_output_ctx *out,
416					  struct runtime_stat *st)
417{
418	double total, ratio = 0.0;
419	const char *color;
420	int ctx = evsel_context(evsel);
421
422	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
423
424	if (total)
425		ratio = avg / total * 100.0;
426
427	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
428
429	if (ratio)
430		out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
431				  ratio);
432	else
433		out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
434}
435
436static void print_stalled_cycles_backend(int cpu,
437					 struct perf_evsel *evsel, double avg,
438					 struct perf_stat_output_ctx *out,
439					 struct runtime_stat *st)
440{
441	double total, ratio = 0.0;
442	const char *color;
443	int ctx = evsel_context(evsel);
444
445	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
446
447	if (total)
448		ratio = avg / total * 100.0;
449
450	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
451
452	out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
453}
454
455static void print_branch_misses(int cpu,
456				struct perf_evsel *evsel,
457				double avg,
458				struct perf_stat_output_ctx *out,
459				struct runtime_stat *st)
460{
461	double total, ratio = 0.0;
462	const char *color;
463	int ctx = evsel_context(evsel);
464
465	total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
466
467	if (total)
468		ratio = avg / total * 100.0;
469
470	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
471
472	out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
473}
474
475static void print_l1_dcache_misses(int cpu,
476				   struct perf_evsel *evsel,
477				   double avg,
478				   struct perf_stat_output_ctx *out,
479				   struct runtime_stat *st)
480
481{
482	double total, ratio = 0.0;
483	const char *color;
484	int ctx = evsel_context(evsel);
485
486	total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
487
488	if (total)
489		ratio = avg / total * 100.0;
490
491	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
492
493	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
494}
495
496static void print_l1_icache_misses(int cpu,
497				   struct perf_evsel *evsel,
498				   double avg,
499				   struct perf_stat_output_ctx *out,
500				   struct runtime_stat *st)
501
502{
503	double total, ratio = 0.0;
504	const char *color;
505	int ctx = evsel_context(evsel);
506
507	total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
508
509	if (total)
510		ratio = avg / total * 100.0;
511
512	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
513	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
514}
515
516static void print_dtlb_cache_misses(int cpu,
517				    struct perf_evsel *evsel,
518				    double avg,
519				    struct perf_stat_output_ctx *out,
520				    struct runtime_stat *st)
521{
522	double total, ratio = 0.0;
523	const char *color;
524	int ctx = evsel_context(evsel);
525
526	total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
527
528	if (total)
529		ratio = avg / total * 100.0;
530
531	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
532	out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
533}
534
535static void print_itlb_cache_misses(int cpu,
536				    struct perf_evsel *evsel,
537				    double avg,
538				    struct perf_stat_output_ctx *out,
539				    struct runtime_stat *st)
540{
541	double total, ratio = 0.0;
542	const char *color;
543	int ctx = evsel_context(evsel);
544
545	total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
546
547	if (total)
548		ratio = avg / total * 100.0;
549
550	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
551	out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
552}
553
554static void print_ll_cache_misses(int cpu,
555				  struct perf_evsel *evsel,
556				  double avg,
557				  struct perf_stat_output_ctx *out,
558				  struct runtime_stat *st)
559{
560	double total, ratio = 0.0;
561	const char *color;
562	int ctx = evsel_context(evsel);
563
564	total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
565
566	if (total)
567		ratio = avg / total * 100.0;
568
569	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
570	out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
571}
572
573/*
574 * High level "TopDown" CPU core pipe line bottleneck break down.
575 *
576 * Basic concept following
577 * Yasin, A Top Down Method for Performance analysis and Counter architecture
578 * ISPASS14
579 *
580 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
581 *
582 * Frontend -> Backend -> Retiring
583 * BadSpeculation in addition means out of order execution that is thrown away
584 * (for example branch mispredictions)
585 * Frontend is instruction decoding.
586 * Backend is execution, like computation and accessing data in memory
587 * Retiring is good execution that is not directly bottlenecked
588 *
589 * The formulas are computed in slots.
590 * A slot is an entry in the pipeline each for the pipeline width
591 * (for example a 4-wide pipeline has 4 slots for each cycle)
592 *
593 * Formulas:
594 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
595 *			TotalSlots
596 * Retiring = SlotsRetired / TotalSlots
597 * FrontendBound = FetchBubbles / TotalSlots
598 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
599 *
600 * The kernel provides the mapping to the low level CPU events and any scaling
601 * needed for the CPU pipeline width, for example:
602 *
603 * TotalSlots = Cycles * 4
604 *
605 * The scaling factor is communicated in the sysfs unit.
606 *
607 * In some cases the CPU may not be able to measure all the formulas due to
608 * missing events. In this case multiple formulas are combined, as possible.
609 *
610 * Full TopDown supports more levels to sub-divide each area: for example
611 * BackendBound into computing bound and memory bound. For now we only
612 * support Level 1 TopDown.
613 */
614
615static double sanitize_val(double x)
616{
617	if (x < 0 && x >= -0.02)
618		return 0.0;
619	return x;
620}
621
622static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
623{
624	return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
625}
626
627static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
628{
629	double bad_spec = 0;
630	double total_slots;
631	double total;
632
633	total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
634		runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
635		runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
636
637	total_slots = td_total_slots(ctx, cpu, st);
638	if (total_slots)
639		bad_spec = total / total_slots;
640	return sanitize_val(bad_spec);
641}
642
643static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
644{
645	double retiring = 0;
646	double total_slots = td_total_slots(ctx, cpu, st);
647	double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
648					    ctx, cpu);
649
650	if (total_slots)
651		retiring = ret_slots / total_slots;
652	return retiring;
653}
654
655static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
656{
657	double fe_bound = 0;
658	double total_slots = td_total_slots(ctx, cpu, st);
659	double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
660					    ctx, cpu);
661
662	if (total_slots)
663		fe_bound = fetch_bub / total_slots;
664	return fe_bound;
665}
666
667static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
668{
669	double sum = (td_fe_bound(ctx, cpu, st) +
670		      td_bad_spec(ctx, cpu, st) +
671		      td_retiring(ctx, cpu, st));
672	if (sum == 0)
673		return 0;
674	return sanitize_val(1.0 - sum);
675}
676
677static void print_smi_cost(int cpu, struct perf_evsel *evsel,
678			   struct perf_stat_output_ctx *out,
679			   struct runtime_stat *st)
680{
681	double smi_num, aperf, cycles, cost = 0.0;
682	int ctx = evsel_context(evsel);
683	const char *color = NULL;
684
685	smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
686	aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
687	cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
688
689	if ((cycles == 0) || (aperf == 0))
690		return;
691
692	if (smi_num)
693		cost = (aperf - cycles) / aperf * 100.00;
694
695	if (cost > 10)
696		color = PERF_COLOR_RED;
697	out->print_metric(out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
698	out->print_metric(out->ctx, NULL, "%4.0f", "SMI#", smi_num);
699}
700
701static void generic_metric(const char *metric_expr,
702			   struct perf_evsel **metric_events,
703			   char *name,
704			   const char *metric_name,
705			   double avg,
706			   int cpu,
707			   struct perf_stat_output_ctx *out,
708			   struct runtime_stat *st)
709{
710	print_metric_t print_metric = out->print_metric;
711	struct parse_ctx pctx;
712	double ratio;
713	int i;
714	void *ctxp = out->ctx;
715
716	expr__ctx_init(&pctx);
717	expr__add_id(&pctx, name, avg);
718	for (i = 0; metric_events[i]; i++) {
719		struct saved_value *v;
720		struct stats *stats;
721		double scale;
722
723		if (!strcmp(metric_events[i]->name, "duration_time")) {
724			stats = &walltime_nsecs_stats;
725			scale = 1e-9;
726		} else {
727			v = saved_value_lookup(metric_events[i], cpu, false,
728					       STAT_NONE, 0, st);
729			if (!v)
730				break;
731			stats = &v->stats;
732			scale = 1.0;
733		}
734		expr__add_id(&pctx, metric_events[i]->name, avg_stats(stats)*scale);
735	}
736	if (!metric_events[i]) {
737		const char *p = metric_expr;
738
739		if (expr__parse(&ratio, &pctx, &p) == 0)
740			print_metric(ctxp, NULL, "%8.1f",
741				metric_name ?
742				metric_name :
743				out->force_header ?  name : "",
744				ratio);
745		else
746			print_metric(ctxp, NULL, NULL,
747				     out->force_header ?
748				     (metric_name ? metric_name : name) : "", 0);
749	} else
750		print_metric(ctxp, NULL, NULL, "", 0);
751}
752
753void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
754				   double avg, int cpu,
755				   struct perf_stat_output_ctx *out,
756				   struct rblist *metric_events,
757				   struct runtime_stat *st)
758{
759	void *ctxp = out->ctx;
760	print_metric_t print_metric = out->print_metric;
761	double total, ratio = 0.0, total2;
762	const char *color = NULL;
763	int ctx = evsel_context(evsel);
764	struct metric_event *me;
765	int num = 1;
766
767	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
768		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
769
770		if (total) {
771			ratio = avg / total;
772			print_metric(ctxp, NULL, "%7.2f ",
773					"insn per cycle", ratio);
774		} else {
775			print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
776		}
777
778		total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
779					 ctx, cpu);
780
781		total = max(total, runtime_stat_avg(st,
782						    STAT_STALLED_CYCLES_BACK,
783						    ctx, cpu));
784
785		if (total && avg) {
786			out->new_line(ctxp);
787			ratio = total / avg;
788			print_metric(ctxp, NULL, "%7.2f ",
789					"stalled cycles per insn",
790					ratio);
791		} else if (have_frontend_stalled) {
792			print_metric(ctxp, NULL, NULL,
793				     "stalled cycles per insn", 0);
794		}
795	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
796		if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
797			print_branch_misses(cpu, evsel, avg, out, st);
798		else
799			print_metric(ctxp, NULL, NULL, "of all branches", 0);
800	} else if (
801		evsel->attr.type == PERF_TYPE_HW_CACHE &&
802		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
803					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
804					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
805
806		if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
807			print_l1_dcache_misses(cpu, evsel, avg, out, st);
808		else
809			print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
810	} else if (
811		evsel->attr.type == PERF_TYPE_HW_CACHE &&
812		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
813					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
814					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
815
816		if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
817			print_l1_icache_misses(cpu, evsel, avg, out, st);
818		else
819			print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
820	} else if (
821		evsel->attr.type == PERF_TYPE_HW_CACHE &&
822		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
823					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
824					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
825
826		if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
827			print_dtlb_cache_misses(cpu, evsel, avg, out, st);
828		else
829			print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
830	} else if (
831		evsel->attr.type == PERF_TYPE_HW_CACHE &&
832		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
833					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
834					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
835
836		if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
837			print_itlb_cache_misses(cpu, evsel, avg, out, st);
838		else
839			print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
840	} else if (
841		evsel->attr.type == PERF_TYPE_HW_CACHE &&
842		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
843					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
844					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
845
846		if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
847			print_ll_cache_misses(cpu, evsel, avg, out, st);
848		else
849			print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
850	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
851		total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
852
853		if (total)
854			ratio = avg * 100 / total;
855
856		if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
857			print_metric(ctxp, NULL, "%8.3f %%",
858				     "of all cache refs", ratio);
859		else
860			print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
861	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
862		print_stalled_cycles_frontend(cpu, evsel, avg, out, st);
863	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
864		print_stalled_cycles_backend(cpu, evsel, avg, out, st);
865	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
866		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
867
868		if (total) {
869			ratio = avg / total;
870			print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
871		} else {
872			print_metric(ctxp, NULL, NULL, "Ghz", 0);
873		}
874	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
875		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
876
877		if (total)
878			print_metric(ctxp, NULL,
879					"%7.2f%%", "transactional cycles",
880					100.0 * (avg / total));
881		else
882			print_metric(ctxp, NULL, NULL, "transactional cycles",
883				     0);
884	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
885		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
886		total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
887
888		if (total2 < avg)
889			total2 = avg;
890		if (total)
891			print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
892				100.0 * ((total2-avg) / total));
893		else
894			print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
895	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
896		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
897					 ctx, cpu);
898
899		if (avg)
900			ratio = total / avg;
901
902		if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
903			print_metric(ctxp, NULL, "%8.0f",
904				     "cycles / transaction", ratio);
905		else
906			print_metric(ctxp, NULL, NULL, "cycles / transaction",
907				      0);
908	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
909		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
910					 ctx, cpu);
911
912		if (avg)
913			ratio = total / avg;
914
915		print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
916	} else if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK) ||
917		   perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK)) {
918		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
919			print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
920				     avg / ratio);
921		else
922			print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
923	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
924		double fe_bound = td_fe_bound(ctx, cpu, st);
925
926		if (fe_bound > 0.2)
927			color = PERF_COLOR_RED;
928		print_metric(ctxp, color, "%8.1f%%", "frontend bound",
929				fe_bound * 100.);
930	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
931		double retiring = td_retiring(ctx, cpu, st);
932
933		if (retiring > 0.7)
934			color = PERF_COLOR_GREEN;
935		print_metric(ctxp, color, "%8.1f%%", "retiring",
936				retiring * 100.);
937	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
938		double bad_spec = td_bad_spec(ctx, cpu, st);
939
940		if (bad_spec > 0.1)
941			color = PERF_COLOR_RED;
942		print_metric(ctxp, color, "%8.1f%%", "bad speculation",
943				bad_spec * 100.);
944	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
945		double be_bound = td_be_bound(ctx, cpu, st);
946		const char *name = "backend bound";
947		static int have_recovery_bubbles = -1;
948
949		/* In case the CPU does not support topdown-recovery-bubbles */
950		if (have_recovery_bubbles < 0)
951			have_recovery_bubbles = pmu_have_event("cpu",
952					"topdown-recovery-bubbles");
953		if (!have_recovery_bubbles)
954			name = "backend bound/bad spec";
955
956		if (be_bound > 0.2)
957			color = PERF_COLOR_RED;
958		if (td_total_slots(ctx, cpu, st) > 0)
959			print_metric(ctxp, color, "%8.1f%%", name,
960					be_bound * 100.);
961		else
962			print_metric(ctxp, NULL, NULL, name, 0);
963	} else if (evsel->metric_expr) {
964		generic_metric(evsel->metric_expr, evsel->metric_events, evsel->name,
965				evsel->metric_name, avg, cpu, out, st);
966	} else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) {
967		char unit = 'M';
968		char unit_buf[10];
969
970		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
971
972		if (total)
973			ratio = 1000.0 * avg / total;
974		if (ratio < 0.001) {
975			ratio *= 1000;
976			unit = 'K';
977		}
978		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
979		print_metric(ctxp, NULL, "%8.3f", unit_buf, ratio);
980	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
981		print_smi_cost(cpu, evsel, out, st);
982	} else {
983		num = 0;
984	}
985
986	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
987		struct metric_expr *mexp;
988
989		list_for_each_entry (mexp, &me->head, nd) {
990			if (num++ > 0)
991				out->new_line(ctxp);
992			generic_metric(mexp->metric_expr, mexp->metric_events,
993					evsel->name, mexp->metric_name,
994					avg, cpu, out, st);
995		}
996	}
997	if (num == 0)
998		print_metric(ctxp, NULL, NULL, NULL, 0);
 
999}
v4.10.11
 
  1#include <stdio.h>
  2#include "evsel.h"
  3#include "stat.h"
  4#include "color.h"
  5#include "pmu.h"
  6
  7enum {
  8	CTX_BIT_USER	= 1 << 0,
  9	CTX_BIT_KERNEL	= 1 << 1,
 10	CTX_BIT_HV	= 1 << 2,
 11	CTX_BIT_HOST	= 1 << 3,
 12	CTX_BIT_IDLE	= 1 << 4,
 13	CTX_BIT_MAX	= 1 << 5,
 14};
 15
 16#define NUM_CTX CTX_BIT_MAX
 17
 18/*
 19 * AGGR_GLOBAL: Use CPU 0
 20 * AGGR_SOCKET: Use first CPU of socket
 21 * AGGR_CORE: Use first CPU of core
 22 * AGGR_NONE: Use matching CPU
 23 * AGGR_THREAD: Not supported?
 24 */
 25static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
 26static struct stats runtime_cycles_stats[NUM_CTX][MAX_NR_CPUS];
 27static struct stats runtime_stalled_cycles_front_stats[NUM_CTX][MAX_NR_CPUS];
 28static struct stats runtime_stalled_cycles_back_stats[NUM_CTX][MAX_NR_CPUS];
 29static struct stats runtime_branches_stats[NUM_CTX][MAX_NR_CPUS];
 30static struct stats runtime_cacherefs_stats[NUM_CTX][MAX_NR_CPUS];
 31static struct stats runtime_l1_dcache_stats[NUM_CTX][MAX_NR_CPUS];
 32static struct stats runtime_l1_icache_stats[NUM_CTX][MAX_NR_CPUS];
 33static struct stats runtime_ll_cache_stats[NUM_CTX][MAX_NR_CPUS];
 34static struct stats runtime_itlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
 35static struct stats runtime_dtlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
 36static struct stats runtime_cycles_in_tx_stats[NUM_CTX][MAX_NR_CPUS];
 37static struct stats runtime_transaction_stats[NUM_CTX][MAX_NR_CPUS];
 38static struct stats runtime_elision_stats[NUM_CTX][MAX_NR_CPUS];
 39static struct stats runtime_topdown_total_slots[NUM_CTX][MAX_NR_CPUS];
 40static struct stats runtime_topdown_slots_issued[NUM_CTX][MAX_NR_CPUS];
 41static struct stats runtime_topdown_slots_retired[NUM_CTX][MAX_NR_CPUS];
 42static struct stats runtime_topdown_fetch_bubbles[NUM_CTX][MAX_NR_CPUS];
 43static struct stats runtime_topdown_recovery_bubbles[NUM_CTX][MAX_NR_CPUS];
 44static bool have_frontend_stalled;
 45
 
 46struct stats walltime_nsecs_stats;
 47
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 48void perf_stat__init_shadow_stats(void)
 49{
 50	have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
 
 51}
 52
 53static int evsel_context(struct perf_evsel *evsel)
 54{
 55	int ctx = 0;
 56
 57	if (evsel->attr.exclude_kernel)
 58		ctx |= CTX_BIT_KERNEL;
 59	if (evsel->attr.exclude_user)
 60		ctx |= CTX_BIT_USER;
 61	if (evsel->attr.exclude_hv)
 62		ctx |= CTX_BIT_HV;
 63	if (evsel->attr.exclude_host)
 64		ctx |= CTX_BIT_HOST;
 65	if (evsel->attr.exclude_idle)
 66		ctx |= CTX_BIT_IDLE;
 67
 68	return ctx;
 69}
 70
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 71void perf_stat__reset_shadow_stats(void)
 72{
 73	memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
 74	memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
 75	memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
 76	memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
 77	memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
 78	memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
 79	memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
 80	memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
 81	memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
 82	memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
 83	memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
 84	memset(runtime_cycles_in_tx_stats, 0,
 85			sizeof(runtime_cycles_in_tx_stats));
 86	memset(runtime_transaction_stats, 0,
 87		sizeof(runtime_transaction_stats));
 88	memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
 89	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
 90	memset(runtime_topdown_total_slots, 0, sizeof(runtime_topdown_total_slots));
 91	memset(runtime_topdown_slots_retired, 0, sizeof(runtime_topdown_slots_retired));
 92	memset(runtime_topdown_slots_issued, 0, sizeof(runtime_topdown_slots_issued));
 93	memset(runtime_topdown_fetch_bubbles, 0, sizeof(runtime_topdown_fetch_bubbles));
 94	memset(runtime_topdown_recovery_bubbles, 0, sizeof(runtime_topdown_recovery_bubbles));
 
 
 
 
 
 
 
 
 
 
 
 95}
 96
 97/*
 98 * Update various tracking values we maintain to print
 99 * more semantic information such as miss/hit ratios,
100 * instruction rates, etc:
101 */
102void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 *count,
103				    int cpu)
104{
105	int ctx = evsel_context(counter);
106
 
 
107	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK) ||
108	    perf_evsel__match(counter, SOFTWARE, SW_CPU_CLOCK))
109		update_stats(&runtime_nsecs_stats[cpu], count[0]);
110	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
111		update_stats(&runtime_cycles_stats[ctx][cpu], count[0]);
112	else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
113		update_stats(&runtime_cycles_in_tx_stats[ctx][cpu], count[0]);
114	else if (perf_stat_evsel__is(counter, TRANSACTION_START))
115		update_stats(&runtime_transaction_stats[ctx][cpu], count[0]);
116	else if (perf_stat_evsel__is(counter, ELISION_START))
117		update_stats(&runtime_elision_stats[ctx][cpu], count[0]);
118	else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
119		update_stats(&runtime_topdown_total_slots[ctx][cpu], count[0]);
 
120	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
121		update_stats(&runtime_topdown_slots_issued[ctx][cpu], count[0]);
 
122	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
123		update_stats(&runtime_topdown_slots_retired[ctx][cpu], count[0]);
 
124	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
125		update_stats(&runtime_topdown_fetch_bubbles[ctx][cpu],count[0]);
 
126	else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
127		update_stats(&runtime_topdown_recovery_bubbles[ctx][cpu], count[0]);
 
128	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
129		update_stats(&runtime_stalled_cycles_front_stats[ctx][cpu], count[0]);
 
130	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
131		update_stats(&runtime_stalled_cycles_back_stats[ctx][cpu], count[0]);
 
132	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
133		update_stats(&runtime_branches_stats[ctx][cpu], count[0]);
134	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
135		update_stats(&runtime_cacherefs_stats[ctx][cpu], count[0]);
136	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
137		update_stats(&runtime_l1_dcache_stats[ctx][cpu], count[0]);
138	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
139		update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
140	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
141		update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
142	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
143		update_stats(&runtime_dtlb_cache_stats[ctx][cpu], count[0]);
144	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
145		update_stats(&runtime_itlb_cache_stats[ctx][cpu], count[0]);
 
 
 
 
 
 
 
 
 
 
146}
147
148/* used for get_ratio_color() */
149enum grc_type {
150	GRC_STALLED_CYCLES_FE,
151	GRC_STALLED_CYCLES_BE,
152	GRC_CACHE_MISSES,
153	GRC_MAX_NR
154};
155
156static const char *get_ratio_color(enum grc_type type, double ratio)
157{
158	static const double grc_table[GRC_MAX_NR][3] = {
159		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
160		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
161		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
162	};
163	const char *color = PERF_COLOR_NORMAL;
164
165	if (ratio > grc_table[type][0])
166		color = PERF_COLOR_RED;
167	else if (ratio > grc_table[type][1])
168		color = PERF_COLOR_MAGENTA;
169	else if (ratio > grc_table[type][2])
170		color = PERF_COLOR_YELLOW;
171
172	return color;
173}
174
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
175static void print_stalled_cycles_frontend(int cpu,
176					  struct perf_evsel *evsel, double avg,
177					  struct perf_stat_output_ctx *out)
 
178{
179	double total, ratio = 0.0;
180	const char *color;
181	int ctx = evsel_context(evsel);
182
183	total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
184
185	if (total)
186		ratio = avg / total * 100.0;
187
188	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
189
190	if (ratio)
191		out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
192				  ratio);
193	else
194		out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
195}
196
197static void print_stalled_cycles_backend(int cpu,
198					 struct perf_evsel *evsel, double avg,
199					 struct perf_stat_output_ctx *out)
 
200{
201	double total, ratio = 0.0;
202	const char *color;
203	int ctx = evsel_context(evsel);
204
205	total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
206
207	if (total)
208		ratio = avg / total * 100.0;
209
210	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
211
212	out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
213}
214
215static void print_branch_misses(int cpu,
216				struct perf_evsel *evsel,
217				double avg,
218				struct perf_stat_output_ctx *out)
 
219{
220	double total, ratio = 0.0;
221	const char *color;
222	int ctx = evsel_context(evsel);
223
224	total = avg_stats(&runtime_branches_stats[ctx][cpu]);
225
226	if (total)
227		ratio = avg / total * 100.0;
228
229	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
230
231	out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
232}
233
234static void print_l1_dcache_misses(int cpu,
235				   struct perf_evsel *evsel,
236				   double avg,
237				   struct perf_stat_output_ctx *out)
 
 
238{
239	double total, ratio = 0.0;
240	const char *color;
241	int ctx = evsel_context(evsel);
242
243	total = avg_stats(&runtime_l1_dcache_stats[ctx][cpu]);
244
245	if (total)
246		ratio = avg / total * 100.0;
247
248	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
249
250	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
251}
252
253static void print_l1_icache_misses(int cpu,
254				   struct perf_evsel *evsel,
255				   double avg,
256				   struct perf_stat_output_ctx *out)
 
 
257{
258	double total, ratio = 0.0;
259	const char *color;
260	int ctx = evsel_context(evsel);
261
262	total = avg_stats(&runtime_l1_icache_stats[ctx][cpu]);
263
264	if (total)
265		ratio = avg / total * 100.0;
266
267	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
268	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
269}
270
271static void print_dtlb_cache_misses(int cpu,
272				    struct perf_evsel *evsel,
273				    double avg,
274				    struct perf_stat_output_ctx *out)
 
275{
276	double total, ratio = 0.0;
277	const char *color;
278	int ctx = evsel_context(evsel);
279
280	total = avg_stats(&runtime_dtlb_cache_stats[ctx][cpu]);
281
282	if (total)
283		ratio = avg / total * 100.0;
284
285	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
286	out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
287}
288
289static void print_itlb_cache_misses(int cpu,
290				    struct perf_evsel *evsel,
291				    double avg,
292				    struct perf_stat_output_ctx *out)
 
293{
294	double total, ratio = 0.0;
295	const char *color;
296	int ctx = evsel_context(evsel);
297
298	total = avg_stats(&runtime_itlb_cache_stats[ctx][cpu]);
299
300	if (total)
301		ratio = avg / total * 100.0;
302
303	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
304	out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
305}
306
307static void print_ll_cache_misses(int cpu,
308				  struct perf_evsel *evsel,
309				  double avg,
310				  struct perf_stat_output_ctx *out)
 
311{
312	double total, ratio = 0.0;
313	const char *color;
314	int ctx = evsel_context(evsel);
315
316	total = avg_stats(&runtime_ll_cache_stats[ctx][cpu]);
317
318	if (total)
319		ratio = avg / total * 100.0;
320
321	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
322	out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
323}
324
325/*
326 * High level "TopDown" CPU core pipe line bottleneck break down.
327 *
328 * Basic concept following
329 * Yasin, A Top Down Method for Performance analysis and Counter architecture
330 * ISPASS14
331 *
332 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
333 *
334 * Frontend -> Backend -> Retiring
335 * BadSpeculation in addition means out of order execution that is thrown away
336 * (for example branch mispredictions)
337 * Frontend is instruction decoding.
338 * Backend is execution, like computation and accessing data in memory
339 * Retiring is good execution that is not directly bottlenecked
340 *
341 * The formulas are computed in slots.
342 * A slot is an entry in the pipeline each for the pipeline width
343 * (for example a 4-wide pipeline has 4 slots for each cycle)
344 *
345 * Formulas:
346 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
347 *			TotalSlots
348 * Retiring = SlotsRetired / TotalSlots
349 * FrontendBound = FetchBubbles / TotalSlots
350 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
351 *
352 * The kernel provides the mapping to the low level CPU events and any scaling
353 * needed for the CPU pipeline width, for example:
354 *
355 * TotalSlots = Cycles * 4
356 *
357 * The scaling factor is communicated in the sysfs unit.
358 *
359 * In some cases the CPU may not be able to measure all the formulas due to
360 * missing events. In this case multiple formulas are combined, as possible.
361 *
362 * Full TopDown supports more levels to sub-divide each area: for example
363 * BackendBound into computing bound and memory bound. For now we only
364 * support Level 1 TopDown.
365 */
366
367static double sanitize_val(double x)
368{
369	if (x < 0 && x >= -0.02)
370		return 0.0;
371	return x;
372}
373
374static double td_total_slots(int ctx, int cpu)
375{
376	return avg_stats(&runtime_topdown_total_slots[ctx][cpu]);
377}
378
379static double td_bad_spec(int ctx, int cpu)
380{
381	double bad_spec = 0;
382	double total_slots;
383	double total;
384
385	total = avg_stats(&runtime_topdown_slots_issued[ctx][cpu]) -
386		avg_stats(&runtime_topdown_slots_retired[ctx][cpu]) +
387		avg_stats(&runtime_topdown_recovery_bubbles[ctx][cpu]);
388	total_slots = td_total_slots(ctx, cpu);
 
389	if (total_slots)
390		bad_spec = total / total_slots;
391	return sanitize_val(bad_spec);
392}
393
394static double td_retiring(int ctx, int cpu)
395{
396	double retiring = 0;
397	double total_slots = td_total_slots(ctx, cpu);
398	double ret_slots = avg_stats(&runtime_topdown_slots_retired[ctx][cpu]);
 
399
400	if (total_slots)
401		retiring = ret_slots / total_slots;
402	return retiring;
403}
404
405static double td_fe_bound(int ctx, int cpu)
406{
407	double fe_bound = 0;
408	double total_slots = td_total_slots(ctx, cpu);
409	double fetch_bub = avg_stats(&runtime_topdown_fetch_bubbles[ctx][cpu]);
 
410
411	if (total_slots)
412		fe_bound = fetch_bub / total_slots;
413	return fe_bound;
414}
415
416static double td_be_bound(int ctx, int cpu)
417{
418	double sum = (td_fe_bound(ctx, cpu) +
419		      td_bad_spec(ctx, cpu) +
420		      td_retiring(ctx, cpu));
421	if (sum == 0)
422		return 0;
423	return sanitize_val(1.0 - sum);
424}
425
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
426void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
427				   double avg, int cpu,
428				   struct perf_stat_output_ctx *out)
 
 
429{
430	void *ctxp = out->ctx;
431	print_metric_t print_metric = out->print_metric;
432	double total, ratio = 0.0, total2;
433	const char *color = NULL;
434	int ctx = evsel_context(evsel);
 
 
435
436	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
437		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
 
438		if (total) {
439			ratio = avg / total;
440			print_metric(ctxp, NULL, "%7.2f ",
441					"insn per cycle", ratio);
442		} else {
443			print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
444		}
445		total = avg_stats(&runtime_stalled_cycles_front_stats[ctx][cpu]);
446		total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[ctx][cpu]));
 
 
 
 
 
447
448		if (total && avg) {
449			out->new_line(ctxp);
450			ratio = total / avg;
451			print_metric(ctxp, NULL, "%7.2f ",
452					"stalled cycles per insn",
453					ratio);
454		} else if (have_frontend_stalled) {
455			print_metric(ctxp, NULL, NULL,
456				     "stalled cycles per insn", 0);
457		}
458	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
459		if (runtime_branches_stats[ctx][cpu].n != 0)
460			print_branch_misses(cpu, evsel, avg, out);
461		else
462			print_metric(ctxp, NULL, NULL, "of all branches", 0);
463	} else if (
464		evsel->attr.type == PERF_TYPE_HW_CACHE &&
465		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
466					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
467					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
468		if (runtime_l1_dcache_stats[ctx][cpu].n != 0)
469			print_l1_dcache_misses(cpu, evsel, avg, out);
 
470		else
471			print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
472	} else if (
473		evsel->attr.type == PERF_TYPE_HW_CACHE &&
474		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
475					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
476					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
477		if (runtime_l1_icache_stats[ctx][cpu].n != 0)
478			print_l1_icache_misses(cpu, evsel, avg, out);
 
479		else
480			print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
481	} else if (
482		evsel->attr.type == PERF_TYPE_HW_CACHE &&
483		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
484					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
485					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
486		if (runtime_dtlb_cache_stats[ctx][cpu].n != 0)
487			print_dtlb_cache_misses(cpu, evsel, avg, out);
 
488		else
489			print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
490	} else if (
491		evsel->attr.type == PERF_TYPE_HW_CACHE &&
492		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
493					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
494					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
495		if (runtime_itlb_cache_stats[ctx][cpu].n != 0)
496			print_itlb_cache_misses(cpu, evsel, avg, out);
 
497		else
498			print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
499	} else if (
500		evsel->attr.type == PERF_TYPE_HW_CACHE &&
501		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
502					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
503					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
504		if (runtime_ll_cache_stats[ctx][cpu].n != 0)
505			print_ll_cache_misses(cpu, evsel, avg, out);
 
506		else
507			print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
508	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
509		total = avg_stats(&runtime_cacherefs_stats[ctx][cpu]);
510
511		if (total)
512			ratio = avg * 100 / total;
513
514		if (runtime_cacherefs_stats[ctx][cpu].n != 0)
515			print_metric(ctxp, NULL, "%8.3f %%",
516				     "of all cache refs", ratio);
517		else
518			print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
519	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
520		print_stalled_cycles_frontend(cpu, evsel, avg, out);
521	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
522		print_stalled_cycles_backend(cpu, evsel, avg, out);
523	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
524		total = avg_stats(&runtime_nsecs_stats[cpu]);
525
526		if (total) {
527			ratio = avg / total;
528			print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
529		} else {
530			print_metric(ctxp, NULL, NULL, "Ghz", 0);
531		}
532	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
533		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
 
534		if (total)
535			print_metric(ctxp, NULL,
536					"%7.2f%%", "transactional cycles",
537					100.0 * (avg / total));
538		else
539			print_metric(ctxp, NULL, NULL, "transactional cycles",
540				     0);
541	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
542		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
543		total2 = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
 
544		if (total2 < avg)
545			total2 = avg;
546		if (total)
547			print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
548				100.0 * ((total2-avg) / total));
549		else
550			print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
551	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
552		total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
 
553
554		if (avg)
555			ratio = total / avg;
556
557		if (runtime_cycles_in_tx_stats[ctx][cpu].n != 0)
558			print_metric(ctxp, NULL, "%8.0f",
559				     "cycles / transaction", ratio);
560		else
561			print_metric(ctxp, NULL, NULL, "cycles / transaction",
562				     0);
563	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
564		total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
 
565
566		if (avg)
567			ratio = total / avg;
568
569		print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
570	} else if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK) ||
571		   perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK)) {
572		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
573			print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
574				     avg / ratio);
575		else
576			print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
577	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
578		double fe_bound = td_fe_bound(ctx, cpu);
579
580		if (fe_bound > 0.2)
581			color = PERF_COLOR_RED;
582		print_metric(ctxp, color, "%8.1f%%", "frontend bound",
583				fe_bound * 100.);
584	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
585		double retiring = td_retiring(ctx, cpu);
586
587		if (retiring > 0.7)
588			color = PERF_COLOR_GREEN;
589		print_metric(ctxp, color, "%8.1f%%", "retiring",
590				retiring * 100.);
591	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
592		double bad_spec = td_bad_spec(ctx, cpu);
593
594		if (bad_spec > 0.1)
595			color = PERF_COLOR_RED;
596		print_metric(ctxp, color, "%8.1f%%", "bad speculation",
597				bad_spec * 100.);
598	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
599		double be_bound = td_be_bound(ctx, cpu);
600		const char *name = "backend bound";
601		static int have_recovery_bubbles = -1;
602
603		/* In case the CPU does not support topdown-recovery-bubbles */
604		if (have_recovery_bubbles < 0)
605			have_recovery_bubbles = pmu_have_event("cpu",
606					"topdown-recovery-bubbles");
607		if (!have_recovery_bubbles)
608			name = "backend bound/bad spec";
609
610		if (be_bound > 0.2)
611			color = PERF_COLOR_RED;
612		if (td_total_slots(ctx, cpu) > 0)
613			print_metric(ctxp, color, "%8.1f%%", name,
614					be_bound * 100.);
615		else
616			print_metric(ctxp, NULL, NULL, name, 0);
617	} else if (runtime_nsecs_stats[cpu].n != 0) {
 
 
 
618		char unit = 'M';
619		char unit_buf[10];
620
621		total = avg_stats(&runtime_nsecs_stats[cpu]);
622
623		if (total)
624			ratio = 1000.0 * avg / total;
625		if (ratio < 0.001) {
626			ratio *= 1000;
627			unit = 'K';
628		}
629		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
630		print_metric(ctxp, NULL, "%8.3f", unit_buf, ratio);
 
 
631	} else {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
632		print_metric(ctxp, NULL, NULL, NULL, 0);
633	}
634}