1 // SPDX-License-Identifier: GPL-2.0
10 #include "metricgroup.h"
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?
19 static bool have_frontend_stalled;
21 struct runtime_stat rt_stat;
22 struct stats walltime_nsecs_stats;
25 struct rb_node rb_node;
26 struct perf_evsel *evsel;
30 struct runtime_stat *stat;
34 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
36 struct saved_value *a = container_of(rb_node,
39 const struct saved_value *b = entry;
42 return a->cpu - b->cpu;
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).
51 if (a->type != b->type)
52 return a->type - b->type;
55 return a->ctx - b->ctx;
57 if (a->evsel == NULL && b->evsel == NULL) {
58 if (a->stat == b->stat)
61 if ((char *)a->stat < (char *)b->stat)
67 if (a->evsel == b->evsel)
69 if ((char *)a->evsel < (char *)b->evsel)
74 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
77 struct saved_value *nd = malloc(sizeof(struct saved_value));
81 memcpy(nd, entry, sizeof(struct saved_value));
85 static void saved_value_delete(struct rblist *rblist __maybe_unused,
86 struct rb_node *rb_node)
88 struct saved_value *v;
91 v = container_of(rb_node, struct saved_value, rb_node);
95 static struct saved_value *saved_value_lookup(struct perf_evsel *evsel,
100 struct runtime_stat *st)
102 struct rblist *rblist;
104 struct saved_value dm = {
112 rblist = &st->value_list;
114 nd = rblist__find(rblist, &dm);
116 return container_of(nd, struct saved_value, rb_node);
118 rblist__add_node(rblist, &dm);
119 nd = rblist__find(rblist, &dm);
121 return container_of(nd, struct saved_value, rb_node);
126 void runtime_stat__init(struct runtime_stat *st)
128 struct rblist *rblist = &st->value_list;
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;
136 void runtime_stat__exit(struct runtime_stat *st)
138 rblist__exit(&st->value_list);
141 void perf_stat__init_shadow_stats(void)
143 have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
144 runtime_stat__init(&rt_stat);
147 static int evsel_context(struct perf_evsel *evsel)
151 if (evsel->attr.exclude_kernel)
152 ctx |= CTX_BIT_KERNEL;
153 if (evsel->attr.exclude_user)
155 if (evsel->attr.exclude_hv)
157 if (evsel->attr.exclude_host)
159 if (evsel->attr.exclude_idle)
165 static void reset_stat(struct runtime_stat *st)
167 struct rblist *rblist;
168 struct rb_node *pos, *next;
170 rblist = &st->value_list;
171 next = rb_first(&rblist->entries);
175 memset(&container_of(pos, struct saved_value, rb_node)->stats,
177 sizeof(struct stats));
181 void perf_stat__reset_shadow_stats(void)
183 reset_stat(&rt_stat);
184 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
187 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
192 static void update_runtime_stat(struct runtime_stat *st,
194 int ctx, int cpu, u64 count)
196 struct saved_value *v = saved_value_lookup(NULL, cpu, true,
200 update_stats(&v->stats, count);
204 * Update various tracking values we maintain to print
205 * more semantic information such as miss/hit ratios,
206 * instruction rates, etc:
208 void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 count,
209 int cpu, struct runtime_stat *st)
211 int ctx = evsel_context(counter);
212 u64 count_ns = count;
214 count *= counter->scale;
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,
229 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
230 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
232 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
233 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
235 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
236 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
238 else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
239 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
241 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
242 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
244 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
245 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
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);
266 if (counter->collect_stat) {
267 struct saved_value *v = saved_value_lookup(counter, cpu, true,
269 update_stats(&v->stats, count);
273 /* used for get_ratio_color() */
275 GRC_STALLED_CYCLES_FE,
276 GRC_STALLED_CYCLES_BE,
281 static const char *get_ratio_color(enum grc_type type, double ratio)
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 },
288 const char *color = PERF_COLOR_NORMAL;
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;
300 static struct perf_evsel *perf_stat__find_event(struct perf_evlist *evsel_list,
303 struct perf_evsel *c2;
305 evlist__for_each_entry (evsel_list, c2) {
306 if (!strcasecmp(c2->name, name) && !c2->collect_stat)
312 /* Mark MetricExpr target events and link events using them to them. */
313 void perf_stat__collect_metric_expr(struct perf_evlist *evsel_list)
315 struct perf_evsel *counter, *leader, **metric_events, *oc;
317 const char **metric_names;
319 int num_metric_names;
321 evlist__for_each_entry(evsel_list, counter) {
322 bool invalid = false;
324 leader = counter->leader;
325 if (!counter->metric_expr)
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)
333 metric_events = calloc(sizeof(struct perf_evsel *),
334 num_metric_names + 1);
337 counter->metric_events = metric_events;
340 for (i = 0; i < num_metric_names; i++) {
343 /* Search in group */
344 for_each_group_member (oc, leader) {
345 if (!strcasecmp(oc->name, metric_names[i]) &&
353 /* Search ignoring groups */
354 oc = perf_stat__find_event(evsel_list, metric_names[i]);
357 /* Deduping one is good enough to handle duplicated PMUs. */
358 static char *printed;
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
367 if (!printed || strcasecmp(printed, metric_names[i])) {
369 "Add %s event to groups to get metric expression for %s\n",
372 printed = strdup(metric_names[i]);
377 metric_events[i] = oc;
378 oc->collect_stat = true;
380 metric_events[i] = NULL;
384 counter->metric_events = NULL;
385 counter->metric_expr = NULL;
390 static double runtime_stat_avg(struct runtime_stat *st,
391 enum stat_type type, int ctx, int cpu)
393 struct saved_value *v;
395 v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
399 return avg_stats(&v->stats);
402 static double runtime_stat_n(struct runtime_stat *st,
403 enum stat_type type, int ctx, int cpu)
405 struct saved_value *v;
407 v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
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)
419 double total, ratio = 0.0;
421 int ctx = evsel_context(evsel);
423 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
426 ratio = avg / total * 100.0;
428 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
431 out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
434 out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
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)
442 double total, ratio = 0.0;
444 int ctx = evsel_context(evsel);
446 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
449 ratio = avg / total * 100.0;
451 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
453 out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
456 static void print_branch_misses(int cpu,
457 struct perf_evsel *evsel,
459 struct perf_stat_output_ctx *out,
460 struct runtime_stat *st)
462 double total, ratio = 0.0;
464 int ctx = evsel_context(evsel);
466 total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
469 ratio = avg / total * 100.0;
471 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
473 out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
476 static void print_l1_dcache_misses(int cpu,
477 struct perf_evsel *evsel,
479 struct perf_stat_output_ctx *out,
480 struct runtime_stat *st)
483 double total, ratio = 0.0;
485 int ctx = evsel_context(evsel);
487 total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
490 ratio = avg / total * 100.0;
492 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
494 out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
497 static void print_l1_icache_misses(int cpu,
498 struct perf_evsel *evsel,
500 struct perf_stat_output_ctx *out,
501 struct runtime_stat *st)
504 double total, ratio = 0.0;
506 int ctx = evsel_context(evsel);
508 total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
511 ratio = avg / total * 100.0;
513 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
514 out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
517 static void print_dtlb_cache_misses(int cpu,
518 struct perf_evsel *evsel,
520 struct perf_stat_output_ctx *out,
521 struct runtime_stat *st)
523 double total, ratio = 0.0;
525 int ctx = evsel_context(evsel);
527 total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
530 ratio = avg / total * 100.0;
532 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
533 out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
536 static void print_itlb_cache_misses(int cpu,
537 struct perf_evsel *evsel,
539 struct perf_stat_output_ctx *out,
540 struct runtime_stat *st)
542 double total, ratio = 0.0;
544 int ctx = evsel_context(evsel);
546 total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
549 ratio = avg / total * 100.0;
551 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
552 out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
555 static void print_ll_cache_misses(int cpu,
556 struct perf_evsel *evsel,
558 struct perf_stat_output_ctx *out,
559 struct runtime_stat *st)
561 double total, ratio = 0.0;
563 int ctx = evsel_context(evsel);
565 total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
568 ratio = avg / total * 100.0;
570 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
571 out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
575 * High level "TopDown" CPU core pipe line bottleneck break down.
577 * Basic concept following
578 * Yasin, A Top Down Method for Performance analysis and Counter architecture
581 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
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
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)
595 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
597 * Retiring = SlotsRetired / TotalSlots
598 * FrontendBound = FetchBubbles / TotalSlots
599 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
601 * The kernel provides the mapping to the low level CPU events and any scaling
602 * needed for the CPU pipeline width, for example:
604 * TotalSlots = Cycles * 4
606 * The scaling factor is communicated in the sysfs unit.
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.
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.
616 static double sanitize_val(double x)
618 if (x < 0 && x >= -0.02)
623 static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
625 return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
628 static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
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);
638 total_slots = td_total_slots(ctx, cpu, st);
640 bad_spec = total / total_slots;
641 return sanitize_val(bad_spec);
644 static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
647 double total_slots = td_total_slots(ctx, cpu, st);
648 double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
652 retiring = ret_slots / total_slots;
656 static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
659 double total_slots = td_total_slots(ctx, cpu, st);
660 double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
664 fe_bound = fetch_bub / total_slots;
668 static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
670 double sum = (td_fe_bound(ctx, cpu, st) +
671 td_bad_spec(ctx, cpu, st) +
672 td_retiring(ctx, cpu, st));
675 return sanitize_val(1.0 - sum);
678 static void print_smi_cost(int cpu, struct perf_evsel *evsel,
679 struct perf_stat_output_ctx *out,
680 struct runtime_stat *st)
682 double smi_num, aperf, cycles, cost = 0.0;
683 int ctx = evsel_context(evsel);
684 const char *color = NULL;
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);
690 if ((cycles == 0) || (aperf == 0))
694 cost = (aperf - cycles) / aperf * 100.00;
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);
702 static void generic_metric(const char *metric_expr,
703 struct perf_evsel **metric_events,
705 const char *metric_name,
708 struct perf_stat_output_ctx *out,
709 struct runtime_stat *st)
711 print_metric_t print_metric = out->print_metric;
712 struct parse_ctx pctx;
715 void *ctxp = out->ctx;
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;
724 if (!strcmp(metric_events[i]->name, "duration_time")) {
725 stats = &walltime_nsecs_stats;
728 v = saved_value_lookup(metric_events[i], cpu, false,
735 expr__add_id(&pctx, metric_events[i]->name, avg_stats(stats)*scale);
737 if (!metric_events[i]) {
738 const char *p = metric_expr;
740 if (expr__parse(&ratio, &pctx, &p) == 0)
741 print_metric(ctxp, NULL, "%8.1f",
744 out->force_header ? name : "",
747 print_metric(ctxp, NULL, NULL,
749 (metric_name ? metric_name : name) : "", 0);
751 print_metric(ctxp, NULL, NULL, "", 0);
754 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
756 struct perf_stat_output_ctx *out,
757 struct rblist *metric_events,
758 struct runtime_stat *st)
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;
768 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
769 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
773 print_metric(ctxp, NULL, "%7.2f ",
774 "insn per cycle", ratio);
776 print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
779 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
782 total = max(total, runtime_stat_avg(st,
783 STAT_STALLED_CYCLES_BACK,
789 print_metric(ctxp, NULL, "%7.2f ",
790 "stalled cycles per insn",
792 } else if (have_frontend_stalled) {
793 print_metric(ctxp, NULL, NULL,
794 "stalled cycles per insn", 0);
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);
800 print_metric(ctxp, NULL, NULL, "of all branches", 0);
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))) {
807 if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
808 print_l1_dcache_misses(cpu, evsel, avg, out, st);
810 print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
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))) {
817 if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
818 print_l1_icache_misses(cpu, evsel, avg, out, st);
820 print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
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))) {
827 if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
828 print_dtlb_cache_misses(cpu, evsel, avg, out, st);
830 print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
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))) {
837 if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
838 print_itlb_cache_misses(cpu, evsel, avg, out, st);
840 print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
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))) {
847 if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
848 print_ll_cache_misses(cpu, evsel, avg, out, st);
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);
855 ratio = avg * 100 / total;
857 if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
858 print_metric(ctxp, NULL, "%8.3f %%",
859 "of all cache refs", ratio);
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);
871 print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
873 print_metric(ctxp, NULL, NULL, "Ghz", 0);
875 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
876 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
879 print_metric(ctxp, NULL,
880 "%7.2f%%", "transactional cycles",
881 100.0 * (avg / total));
883 print_metric(ctxp, NULL, NULL, "transactional cycles",
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);
892 print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
893 100.0 * ((total2-avg) / total));
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,
903 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
904 print_metric(ctxp, NULL, "%8.0f",
905 "cycles / transaction", ratio);
907 print_metric(ctxp, NULL, NULL, "cycles / transaction",
909 } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
910 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
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));
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);
927 color = PERF_COLOR_RED;
928 print_metric(ctxp, color, "%8.1f%%", "frontend bound",
930 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
931 double retiring = td_retiring(ctx, cpu, st);
934 color = PERF_COLOR_GREEN;
935 print_metric(ctxp, color, "%8.1f%%", "retiring",
937 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
938 double bad_spec = td_bad_spec(ctx, cpu, st);
941 color = PERF_COLOR_RED;
942 print_metric(ctxp, color, "%8.1f%%", "bad speculation",
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;
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";
957 color = PERF_COLOR_RED;
958 if (td_total_slots(ctx, cpu, st) > 0)
959 print_metric(ctxp, color, "%8.1f%%", name,
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) {
970 total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
973 ratio = 1000.0 * avg / total;
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);
986 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
987 struct metric_expr *mexp;
989 list_for_each_entry (mexp, &me->head, nd) {
992 generic_metric(mexp->metric_expr, mexp->metric_events,
993 evsel->name, mexp->metric_name,
998 print_metric(ctxp, NULL, NULL, NULL, 0);