GNU Linux-libre 4.19.286-gnu1
[releases.git] / tools / perf / util / stat-shadow.c
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  */
19 static bool have_frontend_stalled;
20
21 struct runtime_stat rt_stat;
22 struct stats walltime_nsecs_stats;
23
24 struct 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
34 static 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
74 static 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
85 static 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
95 static 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
126 void 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
136 void runtime_stat__exit(struct runtime_stat *st)
137 {
138         rblist__exit(&st->value_list);
139 }
140
141 void 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
147 static 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
165 static 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
181 void perf_stat__reset_shadow_stats(void)
182 {
183         reset_stat(&rt_stat);
184         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
185 }
186
187 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
188 {
189         reset_stat(st);
190 }
191
192 static 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  */
208 void 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         u64 count_ns = count;
213
214         count *= counter->scale;
215
216         if (perf_evsel__is_clock(counter))
217                 update_runtime_stat(st, STAT_NSECS, 0, cpu, count_ns);
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() */
274 enum grc_type {
275         GRC_STALLED_CYCLES_FE,
276         GRC_STALLED_CYCLES_BE,
277         GRC_CACHE_MISSES,
278         GRC_MAX_NR
279 };
280
281 static 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
300 static 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) && !c2->collect_stat)
307                         return c2;
308         }
309         return NULL;
310 }
311
312 /* Mark MetricExpr target events and link events using them to them. */
313 void 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                                                 !oc->collect_stat) {
347                                                 found = true;
348                                                 break;
349                                         }
350                                 }
351                         }
352                         if (!found) {
353                                 /* Search ignoring groups */
354                                 oc = perf_stat__find_event(evsel_list, metric_names[i]);
355                         }
356                         if (!oc) {
357                                 /* Deduping one is good enough to handle duplicated PMUs. */
358                                 static char *printed;
359
360                                 /*
361                                  * Adding events automatically would be difficult, because
362                                  * it would risk creating groups that are not schedulable.
363                                  * perf stat doesn't understand all the scheduling constraints
364                                  * of events. So we ask the user instead to add the missing
365                                  * events.
366                                  */
367                                 if (!printed || strcasecmp(printed, metric_names[i])) {
368                                         fprintf(stderr,
369                                                 "Add %s event to groups to get metric expression for %s\n",
370                                                 metric_names[i],
371                                                 counter->name);
372                                         printed = strdup(metric_names[i]);
373                                 }
374                                 invalid = true;
375                                 continue;
376                         }
377                         metric_events[i] = oc;
378                         oc->collect_stat = true;
379                 }
380                 metric_events[i] = NULL;
381                 free(metric_names);
382                 if (invalid) {
383                         free(metric_events);
384                         counter->metric_events = NULL;
385                         counter->metric_expr = NULL;
386                 }
387         }
388 }
389
390 static double runtime_stat_avg(struct runtime_stat *st,
391                                enum stat_type type, int ctx, int cpu)
392 {
393         struct saved_value *v;
394
395         v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
396         if (!v)
397                 return 0.0;
398
399         return avg_stats(&v->stats);
400 }
401
402 static double runtime_stat_n(struct runtime_stat *st,
403                              enum stat_type type, int ctx, int cpu)
404 {
405         struct saved_value *v;
406
407         v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
408         if (!v)
409                 return 0.0;
410
411         return v->stats.n;
412 }
413
414 static void print_stalled_cycles_frontend(int cpu,
415                                           struct perf_evsel *evsel, double avg,
416                                           struct perf_stat_output_ctx *out,
417                                           struct runtime_stat *st)
418 {
419         double total, ratio = 0.0;
420         const char *color;
421         int ctx = evsel_context(evsel);
422
423         total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
424
425         if (total)
426                 ratio = avg / total * 100.0;
427
428         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
429
430         if (ratio)
431                 out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
432                                   ratio);
433         else
434                 out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
435 }
436
437 static void print_stalled_cycles_backend(int cpu,
438                                          struct perf_evsel *evsel, double avg,
439                                          struct perf_stat_output_ctx *out,
440                                          struct runtime_stat *st)
441 {
442         double total, ratio = 0.0;
443         const char *color;
444         int ctx = evsel_context(evsel);
445
446         total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
447
448         if (total)
449                 ratio = avg / total * 100.0;
450
451         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
452
453         out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
454 }
455
456 static void print_branch_misses(int cpu,
457                                 struct perf_evsel *evsel,
458                                 double avg,
459                                 struct perf_stat_output_ctx *out,
460                                 struct runtime_stat *st)
461 {
462         double total, ratio = 0.0;
463         const char *color;
464         int ctx = evsel_context(evsel);
465
466         total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
467
468         if (total)
469                 ratio = avg / total * 100.0;
470
471         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
472
473         out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
474 }
475
476 static void print_l1_dcache_misses(int cpu,
477                                    struct perf_evsel *evsel,
478                                    double avg,
479                                    struct perf_stat_output_ctx *out,
480                                    struct runtime_stat *st)
481
482 {
483         double total, ratio = 0.0;
484         const char *color;
485         int ctx = evsel_context(evsel);
486
487         total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
488
489         if (total)
490                 ratio = avg / total * 100.0;
491
492         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
493
494         out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
495 }
496
497 static void print_l1_icache_misses(int cpu,
498                                    struct perf_evsel *evsel,
499                                    double avg,
500                                    struct perf_stat_output_ctx *out,
501                                    struct runtime_stat *st)
502
503 {
504         double total, ratio = 0.0;
505         const char *color;
506         int ctx = evsel_context(evsel);
507
508         total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
509
510         if (total)
511                 ratio = avg / total * 100.0;
512
513         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
514         out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
515 }
516
517 static void print_dtlb_cache_misses(int cpu,
518                                     struct perf_evsel *evsel,
519                                     double avg,
520                                     struct perf_stat_output_ctx *out,
521                                     struct runtime_stat *st)
522 {
523         double total, ratio = 0.0;
524         const char *color;
525         int ctx = evsel_context(evsel);
526
527         total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
528
529         if (total)
530                 ratio = avg / total * 100.0;
531
532         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
533         out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
534 }
535
536 static void print_itlb_cache_misses(int cpu,
537                                     struct perf_evsel *evsel,
538                                     double avg,
539                                     struct perf_stat_output_ctx *out,
540                                     struct runtime_stat *st)
541 {
542         double total, ratio = 0.0;
543         const char *color;
544         int ctx = evsel_context(evsel);
545
546         total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
547
548         if (total)
549                 ratio = avg / total * 100.0;
550
551         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
552         out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
553 }
554
555 static void print_ll_cache_misses(int cpu,
556                                   struct perf_evsel *evsel,
557                                   double avg,
558                                   struct perf_stat_output_ctx *out,
559                                   struct runtime_stat *st)
560 {
561         double total, ratio = 0.0;
562         const char *color;
563         int ctx = evsel_context(evsel);
564
565         total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
566
567         if (total)
568                 ratio = avg / total * 100.0;
569
570         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
571         out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
572 }
573
574 /*
575  * High level "TopDown" CPU core pipe line bottleneck break down.
576  *
577  * Basic concept following
578  * Yasin, A Top Down Method for Performance analysis and Counter architecture
579  * ISPASS14
580  *
581  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
582  *
583  * Frontend -> Backend -> Retiring
584  * BadSpeculation in addition means out of order execution that is thrown away
585  * (for example branch mispredictions)
586  * Frontend is instruction decoding.
587  * Backend is execution, like computation and accessing data in memory
588  * Retiring is good execution that is not directly bottlenecked
589  *
590  * The formulas are computed in slots.
591  * A slot is an entry in the pipeline each for the pipeline width
592  * (for example a 4-wide pipeline has 4 slots for each cycle)
593  *
594  * Formulas:
595  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
596  *                      TotalSlots
597  * Retiring = SlotsRetired / TotalSlots
598  * FrontendBound = FetchBubbles / TotalSlots
599  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
600  *
601  * The kernel provides the mapping to the low level CPU events and any scaling
602  * needed for the CPU pipeline width, for example:
603  *
604  * TotalSlots = Cycles * 4
605  *
606  * The scaling factor is communicated in the sysfs unit.
607  *
608  * In some cases the CPU may not be able to measure all the formulas due to
609  * missing events. In this case multiple formulas are combined, as possible.
610  *
611  * Full TopDown supports more levels to sub-divide each area: for example
612  * BackendBound into computing bound and memory bound. For now we only
613  * support Level 1 TopDown.
614  */
615
616 static double sanitize_val(double x)
617 {
618         if (x < 0 && x >= -0.02)
619                 return 0.0;
620         return x;
621 }
622
623 static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
624 {
625         return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
626 }
627
628 static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
629 {
630         double bad_spec = 0;
631         double total_slots;
632         double total;
633
634         total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
635                 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
636                 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
637
638         total_slots = td_total_slots(ctx, cpu, st);
639         if (total_slots)
640                 bad_spec = total / total_slots;
641         return sanitize_val(bad_spec);
642 }
643
644 static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
645 {
646         double retiring = 0;
647         double total_slots = td_total_slots(ctx, cpu, st);
648         double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
649                                             ctx, cpu);
650
651         if (total_slots)
652                 retiring = ret_slots / total_slots;
653         return retiring;
654 }
655
656 static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
657 {
658         double fe_bound = 0;
659         double total_slots = td_total_slots(ctx, cpu, st);
660         double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
661                                             ctx, cpu);
662
663         if (total_slots)
664                 fe_bound = fetch_bub / total_slots;
665         return fe_bound;
666 }
667
668 static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
669 {
670         double sum = (td_fe_bound(ctx, cpu, st) +
671                       td_bad_spec(ctx, cpu, st) +
672                       td_retiring(ctx, cpu, st));
673         if (sum == 0)
674                 return 0;
675         return sanitize_val(1.0 - sum);
676 }
677
678 static void print_smi_cost(int cpu, struct perf_evsel *evsel,
679                            struct perf_stat_output_ctx *out,
680                            struct runtime_stat *st)
681 {
682         double smi_num, aperf, cycles, cost = 0.0;
683         int ctx = evsel_context(evsel);
684         const char *color = NULL;
685
686         smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
687         aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
688         cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
689
690         if ((cycles == 0) || (aperf == 0))
691                 return;
692
693         if (smi_num)
694                 cost = (aperf - cycles) / aperf * 100.00;
695
696         if (cost > 10)
697                 color = PERF_COLOR_RED;
698         out->print_metric(out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
699         out->print_metric(out->ctx, NULL, "%4.0f", "SMI#", smi_num);
700 }
701
702 static void generic_metric(const char *metric_expr,
703                            struct perf_evsel **metric_events,
704                            char *name,
705                            const char *metric_name,
706                            double avg,
707                            int cpu,
708                            struct perf_stat_output_ctx *out,
709                            struct runtime_stat *st)
710 {
711         print_metric_t print_metric = out->print_metric;
712         struct parse_ctx pctx;
713         double ratio;
714         int i;
715         void *ctxp = out->ctx;
716
717         expr__ctx_init(&pctx);
718         expr__add_id(&pctx, name, avg);
719         for (i = 0; metric_events[i]; i++) {
720                 struct saved_value *v;
721                 struct stats *stats;
722                 double scale;
723
724                 if (!strcmp(metric_events[i]->name, "duration_time")) {
725                         stats = &walltime_nsecs_stats;
726                         scale = 1e-9;
727                 } else {
728                         v = saved_value_lookup(metric_events[i], cpu, false,
729                                                STAT_NONE, 0, st);
730                         if (!v)
731                                 break;
732                         stats = &v->stats;
733                         scale = 1.0;
734                 }
735                 expr__add_id(&pctx, metric_events[i]->name, avg_stats(stats)*scale);
736         }
737         if (!metric_events[i]) {
738                 const char *p = metric_expr;
739
740                 if (expr__parse(&ratio, &pctx, &p) == 0)
741                         print_metric(ctxp, NULL, "%8.1f",
742                                 metric_name ?
743                                 metric_name :
744                                 out->force_header ?  name : "",
745                                 ratio);
746                 else
747                         print_metric(ctxp, NULL, NULL,
748                                      out->force_header ?
749                                      (metric_name ? metric_name : name) : "", 0);
750         } else
751                 print_metric(ctxp, NULL, NULL, "", 0);
752 }
753
754 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
755                                    double avg, int cpu,
756                                    struct perf_stat_output_ctx *out,
757                                    struct rblist *metric_events,
758                                    struct runtime_stat *st)
759 {
760         void *ctxp = out->ctx;
761         print_metric_t print_metric = out->print_metric;
762         double total, ratio = 0.0, total2;
763         const char *color = NULL;
764         int ctx = evsel_context(evsel);
765         struct metric_event *me;
766         int num = 1;
767
768         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
769                 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
770
771                 if (total) {
772                         ratio = avg / total;
773                         print_metric(ctxp, NULL, "%7.2f ",
774                                         "insn per cycle", ratio);
775                 } else {
776                         print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
777                 }
778
779                 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
780                                          ctx, cpu);
781
782                 total = max(total, runtime_stat_avg(st,
783                                                     STAT_STALLED_CYCLES_BACK,
784                                                     ctx, cpu));
785
786                 if (total && avg) {
787                         out->new_line(ctxp);
788                         ratio = total / avg;
789                         print_metric(ctxp, NULL, "%7.2f ",
790                                         "stalled cycles per insn",
791                                         ratio);
792                 } else if (have_frontend_stalled) {
793                         print_metric(ctxp, NULL, NULL,
794                                      "stalled cycles per insn", 0);
795                 }
796         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
797                 if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
798                         print_branch_misses(cpu, evsel, avg, out, st);
799                 else
800                         print_metric(ctxp, NULL, NULL, "of all branches", 0);
801         } else if (
802                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
803                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
804                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
805                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
806
807                 if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
808                         print_l1_dcache_misses(cpu, evsel, avg, out, st);
809                 else
810                         print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
811         } else if (
812                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
813                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
814                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
815                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
816
817                 if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
818                         print_l1_icache_misses(cpu, evsel, avg, out, st);
819                 else
820                         print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
821         } else if (
822                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
823                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
824                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
825                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
826
827                 if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
828                         print_dtlb_cache_misses(cpu, evsel, avg, out, st);
829                 else
830                         print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
831         } else if (
832                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
833                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
834                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
835                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
836
837                 if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
838                         print_itlb_cache_misses(cpu, evsel, avg, out, st);
839                 else
840                         print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
841         } else if (
842                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
843                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
844                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
845                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
846
847                 if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
848                         print_ll_cache_misses(cpu, evsel, avg, out, st);
849                 else
850                         print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
851         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
852                 total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
853
854                 if (total)
855                         ratio = avg * 100 / total;
856
857                 if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
858                         print_metric(ctxp, NULL, "%8.3f %%",
859                                      "of all cache refs", ratio);
860                 else
861                         print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
862         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
863                 print_stalled_cycles_frontend(cpu, evsel, avg, out, st);
864         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
865                 print_stalled_cycles_backend(cpu, evsel, avg, out, st);
866         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
867                 total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
868
869                 if (total) {
870                         ratio = avg / total;
871                         print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
872                 } else {
873                         print_metric(ctxp, NULL, NULL, "Ghz", 0);
874                 }
875         } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
876                 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
877
878                 if (total)
879                         print_metric(ctxp, NULL,
880                                         "%7.2f%%", "transactional cycles",
881                                         100.0 * (avg / total));
882                 else
883                         print_metric(ctxp, NULL, NULL, "transactional cycles",
884                                      0);
885         } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
886                 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
887                 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
888
889                 if (total2 < avg)
890                         total2 = avg;
891                 if (total)
892                         print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
893                                 100.0 * ((total2-avg) / total));
894                 else
895                         print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
896         } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
897                 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
898                                          ctx, cpu);
899
900                 if (avg)
901                         ratio = total / avg;
902
903                 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
904                         print_metric(ctxp, NULL, "%8.0f",
905                                      "cycles / transaction", ratio);
906                 else
907                         print_metric(ctxp, NULL, NULL, "cycles / transaction",
908                                       0);
909         } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
910                 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
911                                          ctx, cpu);
912
913                 if (avg)
914                         ratio = total / avg;
915
916                 print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
917         } else if (perf_evsel__is_clock(evsel)) {
918                 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
919                         print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
920                                      avg / (ratio * evsel->scale));
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 }