1 // SPDX-License-Identifier: GPL-2.0
10 #include "metricgroup.h"
13 #include <linux/zalloc.h>
17 * AGGR_GLOBAL: Use CPU 0
18 * AGGR_SOCKET: Use first CPU of socket
19 * AGGR_DIE: Use first CPU of die
20 * AGGR_CORE: Use first CPU of core
21 * AGGR_NONE: Use matching CPU
22 * AGGR_THREAD: Not supported?
25 struct runtime_stat rt_stat;
26 struct stats walltime_nsecs_stats;
29 struct rb_node rb_node;
35 struct runtime_stat *stat;
41 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
43 struct saved_value *a = container_of(rb_node,
46 const struct saved_value *b = entry;
49 return a->cpu - b->cpu;
52 * Previously the rbtree was used to link generic metrics.
53 * The keys were evsel/cpu. Now the rbtree is extended to support
54 * per-thread shadow stats. For shadow stats case, the keys
55 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
56 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
58 if (a->type != b->type)
59 return a->type - b->type;
62 return a->ctx - b->ctx;
64 if (a->cgrp != b->cgrp)
65 return (char *)a->cgrp < (char *)b->cgrp ? -1 : +1;
67 if (a->evsel == NULL && b->evsel == NULL) {
68 if (a->stat == b->stat)
71 if ((char *)a->stat < (char *)b->stat)
77 if (a->evsel == b->evsel)
79 if ((char *)a->evsel < (char *)b->evsel)
84 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
87 struct saved_value *nd = malloc(sizeof(struct saved_value));
91 memcpy(nd, entry, sizeof(struct saved_value));
95 static void saved_value_delete(struct rblist *rblist __maybe_unused,
96 struct rb_node *rb_node)
98 struct saved_value *v;
101 v = container_of(rb_node, struct saved_value, rb_node);
105 static struct saved_value *saved_value_lookup(struct evsel *evsel,
110 struct runtime_stat *st,
113 struct rblist *rblist;
115 struct saved_value dm = {
124 rblist = &st->value_list;
126 /* don't use context info for clock events */
127 if (type == STAT_NSECS)
130 nd = rblist__find(rblist, &dm);
132 return container_of(nd, struct saved_value, rb_node);
134 rblist__add_node(rblist, &dm);
135 nd = rblist__find(rblist, &dm);
137 return container_of(nd, struct saved_value, rb_node);
142 void runtime_stat__init(struct runtime_stat *st)
144 struct rblist *rblist = &st->value_list;
146 rblist__init(rblist);
147 rblist->node_cmp = saved_value_cmp;
148 rblist->node_new = saved_value_new;
149 rblist->node_delete = saved_value_delete;
152 void runtime_stat__exit(struct runtime_stat *st)
154 rblist__exit(&st->value_list);
157 void perf_stat__init_shadow_stats(void)
159 runtime_stat__init(&rt_stat);
162 static int evsel_context(struct evsel *evsel)
166 if (evsel->core.attr.exclude_kernel)
167 ctx |= CTX_BIT_KERNEL;
168 if (evsel->core.attr.exclude_user)
170 if (evsel->core.attr.exclude_hv)
172 if (evsel->core.attr.exclude_host)
174 if (evsel->core.attr.exclude_idle)
180 static void reset_stat(struct runtime_stat *st)
182 struct rblist *rblist;
183 struct rb_node *pos, *next;
185 rblist = &st->value_list;
186 next = rb_first_cached(&rblist->entries);
190 memset(&container_of(pos, struct saved_value, rb_node)->stats,
192 sizeof(struct stats));
196 void perf_stat__reset_shadow_stats(void)
198 reset_stat(&rt_stat);
199 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
202 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
207 struct runtime_stat_data {
212 static void update_runtime_stat(struct runtime_stat *st,
215 struct runtime_stat_data *rsd)
217 struct saved_value *v = saved_value_lookup(NULL, cpu, true, type,
218 rsd->ctx, st, rsd->cgrp);
221 update_stats(&v->stats, count);
225 * Update various tracking values we maintain to print
226 * more semantic information such as miss/hit ratios,
227 * instruction rates, etc:
229 void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
230 int cpu, struct runtime_stat *st)
232 u64 count_ns = count;
233 struct saved_value *v;
234 struct runtime_stat_data rsd = {
235 .ctx = evsel_context(counter),
236 .cgrp = counter->cgrp,
239 count *= counter->scale;
241 if (evsel__is_clock(counter))
242 update_runtime_stat(st, STAT_NSECS, cpu, count_ns, &rsd);
243 else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
244 update_runtime_stat(st, STAT_CYCLES, cpu, count, &rsd);
245 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
246 update_runtime_stat(st, STAT_CYCLES_IN_TX, cpu, count, &rsd);
247 else if (perf_stat_evsel__is(counter, TRANSACTION_START))
248 update_runtime_stat(st, STAT_TRANSACTION, cpu, count, &rsd);
249 else if (perf_stat_evsel__is(counter, ELISION_START))
250 update_runtime_stat(st, STAT_ELISION, cpu, count, &rsd);
251 else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
252 update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
254 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
255 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
257 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
258 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
260 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
261 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
263 else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
264 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
266 else if (perf_stat_evsel__is(counter, TOPDOWN_RETIRING))
267 update_runtime_stat(st, STAT_TOPDOWN_RETIRING,
269 else if (perf_stat_evsel__is(counter, TOPDOWN_BAD_SPEC))
270 update_runtime_stat(st, STAT_TOPDOWN_BAD_SPEC,
272 else if (perf_stat_evsel__is(counter, TOPDOWN_FE_BOUND))
273 update_runtime_stat(st, STAT_TOPDOWN_FE_BOUND,
275 else if (perf_stat_evsel__is(counter, TOPDOWN_BE_BOUND))
276 update_runtime_stat(st, STAT_TOPDOWN_BE_BOUND,
278 else if (perf_stat_evsel__is(counter, TOPDOWN_HEAVY_OPS))
279 update_runtime_stat(st, STAT_TOPDOWN_HEAVY_OPS,
281 else if (perf_stat_evsel__is(counter, TOPDOWN_BR_MISPREDICT))
282 update_runtime_stat(st, STAT_TOPDOWN_BR_MISPREDICT,
284 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_LAT))
285 update_runtime_stat(st, STAT_TOPDOWN_FETCH_LAT,
287 else if (perf_stat_evsel__is(counter, TOPDOWN_MEM_BOUND))
288 update_runtime_stat(st, STAT_TOPDOWN_MEM_BOUND,
290 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
291 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
293 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
294 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
296 else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
297 update_runtime_stat(st, STAT_BRANCHES, cpu, count, &rsd);
298 else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
299 update_runtime_stat(st, STAT_CACHEREFS, cpu, count, &rsd);
300 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
301 update_runtime_stat(st, STAT_L1_DCACHE, cpu, count, &rsd);
302 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
303 update_runtime_stat(st, STAT_L1_ICACHE, cpu, count, &rsd);
304 else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
305 update_runtime_stat(st, STAT_LL_CACHE, cpu, count, &rsd);
306 else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
307 update_runtime_stat(st, STAT_DTLB_CACHE, cpu, count, &rsd);
308 else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
309 update_runtime_stat(st, STAT_ITLB_CACHE, cpu, count, &rsd);
310 else if (perf_stat_evsel__is(counter, SMI_NUM))
311 update_runtime_stat(st, STAT_SMI_NUM, cpu, count, &rsd);
312 else if (perf_stat_evsel__is(counter, APERF))
313 update_runtime_stat(st, STAT_APERF, cpu, count, &rsd);
315 if (counter->collect_stat) {
316 v = saved_value_lookup(counter, cpu, true, STAT_NONE, 0, st,
318 update_stats(&v->stats, count);
319 if (counter->metric_leader)
320 v->metric_total += count;
321 } else if (counter->metric_leader) {
322 v = saved_value_lookup(counter->metric_leader,
323 cpu, true, STAT_NONE, 0, st, rsd.cgrp);
324 v->metric_total += count;
329 /* used for get_ratio_color() */
331 GRC_STALLED_CYCLES_FE,
332 GRC_STALLED_CYCLES_BE,
337 static const char *get_ratio_color(enum grc_type type, double ratio)
339 static const double grc_table[GRC_MAX_NR][3] = {
340 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
341 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
342 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
344 const char *color = PERF_COLOR_NORMAL;
346 if (ratio > grc_table[type][0])
347 color = PERF_COLOR_RED;
348 else if (ratio > grc_table[type][1])
349 color = PERF_COLOR_MAGENTA;
350 else if (ratio > grc_table[type][2])
351 color = PERF_COLOR_YELLOW;
356 static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
361 evlist__for_each_entry (evsel_list, c2) {
362 if (!strcasecmp(c2->name, name) && !c2->collect_stat)
368 /* Mark MetricExpr target events and link events using them to them. */
369 void perf_stat__collect_metric_expr(struct evlist *evsel_list)
371 struct evsel *counter, *leader, **metric_events, *oc;
373 struct expr_parse_ctx ctx;
374 struct hashmap_entry *cur;
378 expr__ctx_init(&ctx);
379 evlist__for_each_entry(evsel_list, counter) {
380 bool invalid = false;
382 leader = counter->leader;
383 if (!counter->metric_expr)
386 expr__ctx_clear(&ctx);
387 metric_events = counter->metric_events;
388 if (!metric_events) {
389 if (expr__find_other(counter->metric_expr,
394 metric_events = calloc(sizeof(struct evsel *),
395 hashmap__size(&ctx.ids) + 1);
396 if (!metric_events) {
397 expr__ctx_clear(&ctx);
400 counter->metric_events = metric_events;
404 hashmap__for_each_entry((&ctx.ids), cur, bkt) {
405 const char *metric_name = (const char *)cur->key;
409 /* Search in group */
410 for_each_group_member (oc, leader) {
411 if (!strcasecmp(oc->name,
420 /* Search ignoring groups */
421 oc = perf_stat__find_event(evsel_list,
425 /* Deduping one is good enough to handle duplicated PMUs. */
426 static char *printed;
429 * Adding events automatically would be difficult, because
430 * it would risk creating groups that are not schedulable.
431 * perf stat doesn't understand all the scheduling constraints
432 * of events. So we ask the user instead to add the missing
436 strcasecmp(printed, metric_name)) {
438 "Add %s event to groups to get metric expression for %s\n",
441 printed = strdup(metric_name);
446 metric_events[i++] = oc;
447 oc->collect_stat = true;
449 metric_events[i] = NULL;
452 counter->metric_events = NULL;
453 counter->metric_expr = NULL;
456 expr__ctx_clear(&ctx);
459 static double runtime_stat_avg(struct runtime_stat *st,
460 enum stat_type type, int cpu,
461 struct runtime_stat_data *rsd)
463 struct saved_value *v;
465 v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp);
469 return avg_stats(&v->stats);
472 static double runtime_stat_n(struct runtime_stat *st,
473 enum stat_type type, int cpu,
474 struct runtime_stat_data *rsd)
476 struct saved_value *v;
478 v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp);
485 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
487 struct perf_stat_output_ctx *out,
488 struct runtime_stat *st,
489 struct runtime_stat_data *rsd)
491 double total, ratio = 0.0;
494 total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
497 ratio = avg / total * 100.0;
499 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
502 out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
505 out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
508 static void print_stalled_cycles_backend(struct perf_stat_config *config,
510 struct perf_stat_output_ctx *out,
511 struct runtime_stat *st,
512 struct runtime_stat_data *rsd)
514 double total, ratio = 0.0;
517 total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
520 ratio = avg / total * 100.0;
522 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
524 out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
527 static void print_branch_misses(struct perf_stat_config *config,
529 struct perf_stat_output_ctx *out,
530 struct runtime_stat *st,
531 struct runtime_stat_data *rsd)
533 double total, ratio = 0.0;
536 total = runtime_stat_avg(st, STAT_BRANCHES, cpu, rsd);
539 ratio = avg / total * 100.0;
541 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
543 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
546 static void print_l1_dcache_misses(struct perf_stat_config *config,
548 struct perf_stat_output_ctx *out,
549 struct runtime_stat *st,
550 struct runtime_stat_data *rsd)
552 double total, ratio = 0.0;
555 total = runtime_stat_avg(st, STAT_L1_DCACHE, cpu, rsd);
558 ratio = avg / total * 100.0;
560 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
562 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache accesses", ratio);
565 static void print_l1_icache_misses(struct perf_stat_config *config,
567 struct perf_stat_output_ctx *out,
568 struct runtime_stat *st,
569 struct runtime_stat_data *rsd)
571 double total, ratio = 0.0;
574 total = runtime_stat_avg(st, STAT_L1_ICACHE, cpu, rsd);
577 ratio = avg / total * 100.0;
579 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
580 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache accesses", ratio);
583 static void print_dtlb_cache_misses(struct perf_stat_config *config,
585 struct perf_stat_output_ctx *out,
586 struct runtime_stat *st,
587 struct runtime_stat_data *rsd)
589 double total, ratio = 0.0;
592 total = runtime_stat_avg(st, STAT_DTLB_CACHE, cpu, rsd);
595 ratio = avg / total * 100.0;
597 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
598 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache accesses", ratio);
601 static void print_itlb_cache_misses(struct perf_stat_config *config,
603 struct perf_stat_output_ctx *out,
604 struct runtime_stat *st,
605 struct runtime_stat_data *rsd)
607 double total, ratio = 0.0;
610 total = runtime_stat_avg(st, STAT_ITLB_CACHE, cpu, rsd);
613 ratio = avg / total * 100.0;
615 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
616 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache accesses", ratio);
619 static void print_ll_cache_misses(struct perf_stat_config *config,
621 struct perf_stat_output_ctx *out,
622 struct runtime_stat *st,
623 struct runtime_stat_data *rsd)
625 double total, ratio = 0.0;
628 total = runtime_stat_avg(st, STAT_LL_CACHE, cpu, rsd);
631 ratio = avg / total * 100.0;
633 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
634 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache accesses", ratio);
638 * High level "TopDown" CPU core pipe line bottleneck break down.
640 * Basic concept following
641 * Yasin, A Top Down Method for Performance analysis and Counter architecture
644 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
646 * Frontend -> Backend -> Retiring
647 * BadSpeculation in addition means out of order execution that is thrown away
648 * (for example branch mispredictions)
649 * Frontend is instruction decoding.
650 * Backend is execution, like computation and accessing data in memory
651 * Retiring is good execution that is not directly bottlenecked
653 * The formulas are computed in slots.
654 * A slot is an entry in the pipeline each for the pipeline width
655 * (for example a 4-wide pipeline has 4 slots for each cycle)
658 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
660 * Retiring = SlotsRetired / TotalSlots
661 * FrontendBound = FetchBubbles / TotalSlots
662 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
664 * The kernel provides the mapping to the low level CPU events and any scaling
665 * needed for the CPU pipeline width, for example:
667 * TotalSlots = Cycles * 4
669 * The scaling factor is communicated in the sysfs unit.
671 * In some cases the CPU may not be able to measure all the formulas due to
672 * missing events. In this case multiple formulas are combined, as possible.
674 * Full TopDown supports more levels to sub-divide each area: for example
675 * BackendBound into computing bound and memory bound. For now we only
676 * support Level 1 TopDown.
679 static double sanitize_val(double x)
681 if (x < 0 && x >= -0.02)
686 static double td_total_slots(int cpu, struct runtime_stat *st,
687 struct runtime_stat_data *rsd)
689 return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, cpu, rsd);
692 static double td_bad_spec(int cpu, struct runtime_stat *st,
693 struct runtime_stat_data *rsd)
699 total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, cpu, rsd) -
700 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, cpu, rsd) +
701 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, cpu, rsd);
703 total_slots = td_total_slots(cpu, st, rsd);
705 bad_spec = total / total_slots;
706 return sanitize_val(bad_spec);
709 static double td_retiring(int cpu, struct runtime_stat *st,
710 struct runtime_stat_data *rsd)
713 double total_slots = td_total_slots(cpu, st, rsd);
714 double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
718 retiring = ret_slots / total_slots;
722 static double td_fe_bound(int cpu, struct runtime_stat *st,
723 struct runtime_stat_data *rsd)
726 double total_slots = td_total_slots(cpu, st, rsd);
727 double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
731 fe_bound = fetch_bub / total_slots;
735 static double td_be_bound(int cpu, struct runtime_stat *st,
736 struct runtime_stat_data *rsd)
738 double sum = (td_fe_bound(cpu, st, rsd) +
739 td_bad_spec(cpu, st, rsd) +
740 td_retiring(cpu, st, rsd));
743 return sanitize_val(1.0 - sum);
747 * Kernel reports metrics multiplied with slots. To get back
748 * the ratios we need to recreate the sum.
751 static double td_metric_ratio(int cpu, enum stat_type type,
752 struct runtime_stat *stat,
753 struct runtime_stat_data *rsd)
755 double sum = runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) +
756 runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) +
757 runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) +
758 runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd);
759 double d = runtime_stat_avg(stat, type, cpu, rsd);
767 * ... but only if most of the values are actually available.
768 * We allow two missing.
771 static bool full_td(int cpu, struct runtime_stat *stat,
772 struct runtime_stat_data *rsd)
776 if (runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) > 0)
778 if (runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) > 0)
780 if (runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) > 0)
782 if (runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd) > 0)
787 static void print_smi_cost(struct perf_stat_config *config, int cpu,
788 struct perf_stat_output_ctx *out,
789 struct runtime_stat *st,
790 struct runtime_stat_data *rsd)
792 double smi_num, aperf, cycles, cost = 0.0;
793 const char *color = NULL;
795 smi_num = runtime_stat_avg(st, STAT_SMI_NUM, cpu, rsd);
796 aperf = runtime_stat_avg(st, STAT_APERF, cpu, rsd);
797 cycles = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
799 if ((cycles == 0) || (aperf == 0))
803 cost = (aperf - cycles) / aperf * 100.00;
806 color = PERF_COLOR_RED;
807 out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
808 out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
811 static int prepare_metric(struct evsel **metric_events,
812 struct metric_ref *metric_refs,
813 struct expr_parse_ctx *pctx,
815 struct runtime_stat *st)
821 expr__ctx_init(pctx);
822 for (i = 0; metric_events[i]; i++) {
823 struct saved_value *v;
825 u64 metric_total = 0;
827 if (!strcmp(metric_events[i]->name, "duration_time")) {
828 stats = &walltime_nsecs_stats;
831 v = saved_value_lookup(metric_events[i], cpu, false,
833 metric_events[i]->cgrp);
840 metric_total = v->metric_total;
843 n = strdup(metric_events[i]->name);
847 * This display code with --no-merge adds [cpu] postfixes.
848 * These are not supported by the parser. Remove everything
856 expr__add_id_val(pctx, n, metric_total);
858 expr__add_id_val(pctx, n, avg_stats(stats)*scale);
861 for (j = 0; metric_refs && metric_refs[j].metric_name; j++) {
862 ret = expr__add_ref(pctx, &metric_refs[j]);
870 static void generic_metric(struct perf_stat_config *config,
871 const char *metric_expr,
872 struct evsel **metric_events,
873 struct metric_ref *metric_refs,
875 const char *metric_name,
876 const char *metric_unit,
879 struct perf_stat_output_ctx *out,
880 struct runtime_stat *st)
882 print_metric_t print_metric = out->print_metric;
883 struct expr_parse_ctx pctx;
886 void *ctxp = out->ctx;
888 i = prepare_metric(metric_events, metric_refs, &pctx, cpu, st);
892 if (!metric_events[i]) {
893 if (expr__parse(&ratio, &pctx, metric_expr, runtime) == 0) {
897 if (metric_unit && metric_name) {
898 if (perf_pmu__convert_scale(metric_unit,
899 &unit, &scale) >= 0) {
902 if (strstr(metric_expr, "?"))
903 scnprintf(metric_bf, sizeof(metric_bf),
904 "%s %s_%d", unit, metric_name, runtime);
906 scnprintf(metric_bf, sizeof(metric_bf),
907 "%s %s", unit, metric_name);
909 print_metric(config, ctxp, NULL, "%8.1f",
912 print_metric(config, ctxp, NULL, "%8.2f",
915 out->force_header ? name : "",
919 print_metric(config, ctxp, NULL, NULL,
921 (metric_name ? metric_name : name) : "", 0);
924 print_metric(config, ctxp, NULL, NULL,
926 (metric_name ? metric_name : name) : "", 0);
929 expr__ctx_clear(&pctx);
932 double test_generic_metric(struct metric_expr *mexp, int cpu, struct runtime_stat *st)
934 struct expr_parse_ctx pctx;
937 if (prepare_metric(mexp->metric_events, mexp->metric_refs, &pctx, cpu, st) < 0)
940 if (expr__parse(&ratio, &pctx, mexp->metric_expr, 1))
944 expr__ctx_clear(&pctx);
948 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
951 struct perf_stat_output_ctx *out,
952 struct rblist *metric_events,
953 struct runtime_stat *st)
955 void *ctxp = out->ctx;
956 print_metric_t print_metric = out->print_metric;
957 double total, ratio = 0.0, total2;
958 const char *color = NULL;
959 struct runtime_stat_data rsd = {
960 .ctx = evsel_context(evsel),
963 struct metric_event *me;
966 if (config->iostat_run) {
967 iostat_print_metric(config, evsel, out);
968 } else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
969 total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
973 print_metric(config, ctxp, NULL, "%7.2f ",
974 "insn per cycle", ratio);
976 print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
979 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT, cpu, &rsd);
981 total = max(total, runtime_stat_avg(st,
982 STAT_STALLED_CYCLES_BACK,
986 out->new_line(config, ctxp);
988 print_metric(config, ctxp, NULL, "%7.2f ",
989 "stalled cycles per insn",
992 } else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
993 if (runtime_stat_n(st, STAT_BRANCHES, cpu, &rsd) != 0)
994 print_branch_misses(config, cpu, avg, out, st, &rsd);
996 print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
998 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
999 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1000 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1001 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1003 if (runtime_stat_n(st, STAT_L1_DCACHE, cpu, &rsd) != 0)
1004 print_l1_dcache_misses(config, cpu, avg, out, st, &rsd);
1006 print_metric(config, ctxp, NULL, NULL, "of all L1-dcache accesses", 0);
1008 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1009 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1010 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1011 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1013 if (runtime_stat_n(st, STAT_L1_ICACHE, cpu, &rsd) != 0)
1014 print_l1_icache_misses(config, cpu, avg, out, st, &rsd);
1016 print_metric(config, ctxp, NULL, NULL, "of all L1-icache accesses", 0);
1018 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1019 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1020 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1021 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1023 if (runtime_stat_n(st, STAT_DTLB_CACHE, cpu, &rsd) != 0)
1024 print_dtlb_cache_misses(config, cpu, avg, out, st, &rsd);
1026 print_metric(config, ctxp, NULL, NULL, "of all dTLB cache accesses", 0);
1028 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1029 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1030 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1031 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1033 if (runtime_stat_n(st, STAT_ITLB_CACHE, cpu, &rsd) != 0)
1034 print_itlb_cache_misses(config, cpu, avg, out, st, &rsd);
1036 print_metric(config, ctxp, NULL, NULL, "of all iTLB cache accesses", 0);
1038 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1039 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_LL |
1040 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1041 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1043 if (runtime_stat_n(st, STAT_LL_CACHE, cpu, &rsd) != 0)
1044 print_ll_cache_misses(config, cpu, avg, out, st, &rsd);
1046 print_metric(config, ctxp, NULL, NULL, "of all LL-cache accesses", 0);
1047 } else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
1048 total = runtime_stat_avg(st, STAT_CACHEREFS, cpu, &rsd);
1051 ratio = avg * 100 / total;
1053 if (runtime_stat_n(st, STAT_CACHEREFS, cpu, &rsd) != 0)
1054 print_metric(config, ctxp, NULL, "%8.3f %%",
1055 "of all cache refs", ratio);
1057 print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
1058 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1059 print_stalled_cycles_frontend(config, cpu, avg, out, st, &rsd);
1060 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1061 print_stalled_cycles_backend(config, cpu, avg, out, st, &rsd);
1062 } else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1063 total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd);
1066 ratio = avg / total;
1067 print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
1069 print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
1071 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
1072 total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
1075 print_metric(config, ctxp, NULL,
1076 "%7.2f%%", "transactional cycles",
1077 100.0 * (avg / total));
1079 print_metric(config, ctxp, NULL, NULL, "transactional cycles",
1081 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
1082 total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
1083 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1088 print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
1089 100.0 * ((total2-avg) / total));
1091 print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
1092 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
1093 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1096 ratio = total / avg;
1098 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, cpu, &rsd) != 0)
1099 print_metric(config, ctxp, NULL, "%8.0f",
1100 "cycles / transaction", ratio);
1102 print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
1104 } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
1105 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1108 ratio = total / avg;
1110 print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
1111 } else if (evsel__is_clock(evsel)) {
1112 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
1113 print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1114 avg / (ratio * evsel->scale));
1116 print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1117 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1118 double fe_bound = td_fe_bound(cpu, st, &rsd);
1121 color = PERF_COLOR_RED;
1122 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1124 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1125 double retiring = td_retiring(cpu, st, &rsd);
1128 color = PERF_COLOR_GREEN;
1129 print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1131 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1132 double bad_spec = td_bad_spec(cpu, st, &rsd);
1135 color = PERF_COLOR_RED;
1136 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1138 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1139 double be_bound = td_be_bound(cpu, st, &rsd);
1140 const char *name = "backend bound";
1141 static int have_recovery_bubbles = -1;
1143 /* In case the CPU does not support topdown-recovery-bubbles */
1144 if (have_recovery_bubbles < 0)
1145 have_recovery_bubbles = pmu_have_event("cpu",
1146 "topdown-recovery-bubbles");
1147 if (!have_recovery_bubbles)
1148 name = "backend bound/bad spec";
1151 color = PERF_COLOR_RED;
1152 if (td_total_slots(cpu, st, &rsd) > 0)
1153 print_metric(config, ctxp, color, "%8.1f%%", name,
1156 print_metric(config, ctxp, NULL, NULL, name, 0);
1157 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RETIRING) &&
1158 full_td(cpu, st, &rsd)) {
1159 double retiring = td_metric_ratio(cpu,
1160 STAT_TOPDOWN_RETIRING, st,
1163 color = PERF_COLOR_GREEN;
1164 print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1166 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FE_BOUND) &&
1167 full_td(cpu, st, &rsd)) {
1168 double fe_bound = td_metric_ratio(cpu,
1169 STAT_TOPDOWN_FE_BOUND, st,
1172 color = PERF_COLOR_RED;
1173 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1175 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BE_BOUND) &&
1176 full_td(cpu, st, &rsd)) {
1177 double be_bound = td_metric_ratio(cpu,
1178 STAT_TOPDOWN_BE_BOUND, st,
1181 color = PERF_COLOR_RED;
1182 print_metric(config, ctxp, color, "%8.1f%%", "backend bound",
1184 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BAD_SPEC) &&
1185 full_td(cpu, st, &rsd)) {
1186 double bad_spec = td_metric_ratio(cpu,
1187 STAT_TOPDOWN_BAD_SPEC, st,
1190 color = PERF_COLOR_RED;
1191 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1193 } else if (perf_stat_evsel__is(evsel, TOPDOWN_HEAVY_OPS) &&
1194 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1195 double retiring = td_metric_ratio(cpu,
1196 STAT_TOPDOWN_RETIRING, st,
1198 double heavy_ops = td_metric_ratio(cpu,
1199 STAT_TOPDOWN_HEAVY_OPS, st,
1201 double light_ops = retiring - heavy_ops;
1203 if (retiring > 0.7 && heavy_ops > 0.1)
1204 color = PERF_COLOR_GREEN;
1205 print_metric(config, ctxp, color, "%8.1f%%", "heavy operations",
1207 if (retiring > 0.7 && light_ops > 0.6)
1208 color = PERF_COLOR_GREEN;
1211 print_metric(config, ctxp, color, "%8.1f%%", "light operations",
1213 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BR_MISPREDICT) &&
1214 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1215 double bad_spec = td_metric_ratio(cpu,
1216 STAT_TOPDOWN_BAD_SPEC, st,
1218 double br_mis = td_metric_ratio(cpu,
1219 STAT_TOPDOWN_BR_MISPREDICT, st,
1221 double m_clears = bad_spec - br_mis;
1223 if (bad_spec > 0.1 && br_mis > 0.05)
1224 color = PERF_COLOR_RED;
1225 print_metric(config, ctxp, color, "%8.1f%%", "branch mispredict",
1227 if (bad_spec > 0.1 && m_clears > 0.05)
1228 color = PERF_COLOR_RED;
1231 print_metric(config, ctxp, color, "%8.1f%%", "machine clears",
1233 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_LAT) &&
1234 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1235 double fe_bound = td_metric_ratio(cpu,
1236 STAT_TOPDOWN_FE_BOUND, st,
1238 double fetch_lat = td_metric_ratio(cpu,
1239 STAT_TOPDOWN_FETCH_LAT, st,
1241 double fetch_bw = fe_bound - fetch_lat;
1243 if (fe_bound > 0.2 && fetch_lat > 0.15)
1244 color = PERF_COLOR_RED;
1245 print_metric(config, ctxp, color, "%8.1f%%", "fetch latency",
1247 if (fe_bound > 0.2 && fetch_bw > 0.1)
1248 color = PERF_COLOR_RED;
1251 print_metric(config, ctxp, color, "%8.1f%%", "fetch bandwidth",
1253 } else if (perf_stat_evsel__is(evsel, TOPDOWN_MEM_BOUND) &&
1254 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1255 double be_bound = td_metric_ratio(cpu,
1256 STAT_TOPDOWN_BE_BOUND, st,
1258 double mem_bound = td_metric_ratio(cpu,
1259 STAT_TOPDOWN_MEM_BOUND, st,
1261 double core_bound = be_bound - mem_bound;
1263 if (be_bound > 0.2 && mem_bound > 0.2)
1264 color = PERF_COLOR_RED;
1265 print_metric(config, ctxp, color, "%8.1f%%", "memory bound",
1267 if (be_bound > 0.2 && core_bound > 0.1)
1268 color = PERF_COLOR_RED;
1271 print_metric(config, ctxp, color, "%8.1f%%", "Core bound",
1273 } else if (evsel->metric_expr) {
1274 generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1275 evsel->name, evsel->metric_name, NULL, 1, cpu, out, st);
1276 } else if (runtime_stat_n(st, STAT_NSECS, cpu, &rsd) != 0) {
1278 char unit_buf[10] = "/sec";
1280 total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd);
1282 ratio = convert_unit_double(1000000000.0 * avg / total, &unit);
1285 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1286 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1287 } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1288 print_smi_cost(config, cpu, out, st, &rsd);
1293 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1294 struct metric_expr *mexp;
1296 list_for_each_entry (mexp, &me->head, nd) {
1298 out->new_line(config, ctxp);
1299 generic_metric(config, mexp->metric_expr, mexp->metric_events,
1300 mexp->metric_refs, evsel->name, mexp->metric_name,
1301 mexp->metric_unit, mexp->runtime, cpu, out, st);
1305 print_metric(config, ctxp, NULL, NULL, NULL, 0);