1 // SPDX-License-Identifier: GPL-2.0
12 #include "metricgroup.h"
15 #include <linux/zalloc.h>
17 #include "util/hashmap.h"
20 * AGGR_GLOBAL: Use CPU 0
21 * AGGR_SOCKET: Use first CPU of socket
22 * AGGR_DIE: Use first CPU of die
23 * AGGR_CORE: Use first CPU of core
24 * AGGR_NONE: Use matching CPU
25 * AGGR_THREAD: Not supported?
28 struct runtime_stat rt_stat;
29 struct stats walltime_nsecs_stats;
30 struct rusage_stats ru_stats;
33 struct rb_node rb_node;
37 int map_idx; /* cpu or thread map index */
44 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
46 struct saved_value *a = container_of(rb_node,
49 const struct saved_value *b = entry;
51 if (a->map_idx != b->map_idx)
52 return a->map_idx - b->map_idx;
55 * Previously the rbtree was used to link generic metrics.
56 * The keys were evsel/cpu. Now the rbtree is extended to support
57 * per-thread shadow stats. For shadow stats case, the keys
58 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
59 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
61 if (a->type != b->type)
62 return a->type - b->type;
65 return a->ctx - b->ctx;
67 if (a->cgrp != b->cgrp)
68 return (char *)a->cgrp < (char *)b->cgrp ? -1 : +1;
70 if (a->evsel == b->evsel)
72 if ((char *)a->evsel < (char *)b->evsel)
77 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
80 struct saved_value *nd = malloc(sizeof(struct saved_value));
84 memcpy(nd, entry, sizeof(struct saved_value));
88 static void saved_value_delete(struct rblist *rblist __maybe_unused,
89 struct rb_node *rb_node)
91 struct saved_value *v;
94 v = container_of(rb_node, struct saved_value, rb_node);
98 static struct saved_value *saved_value_lookup(struct evsel *evsel,
103 struct runtime_stat *st,
106 struct rblist *rblist;
108 struct saved_value dm = {
116 rblist = &st->value_list;
118 /* don't use context info for clock events */
119 if (type == STAT_NSECS)
122 nd = rblist__find(rblist, &dm);
124 return container_of(nd, struct saved_value, rb_node);
126 rblist__add_node(rblist, &dm);
127 nd = rblist__find(rblist, &dm);
129 return container_of(nd, struct saved_value, rb_node);
134 void runtime_stat__init(struct runtime_stat *st)
136 struct rblist *rblist = &st->value_list;
138 rblist__init(rblist);
139 rblist->node_cmp = saved_value_cmp;
140 rblist->node_new = saved_value_new;
141 rblist->node_delete = saved_value_delete;
144 void runtime_stat__exit(struct runtime_stat *st)
146 rblist__exit(&st->value_list);
149 void perf_stat__init_shadow_stats(void)
151 runtime_stat__init(&rt_stat);
154 static int evsel_context(struct evsel *evsel)
158 if (evsel->core.attr.exclude_kernel)
159 ctx |= CTX_BIT_KERNEL;
160 if (evsel->core.attr.exclude_user)
162 if (evsel->core.attr.exclude_hv)
164 if (evsel->core.attr.exclude_host)
166 if (evsel->core.attr.exclude_idle)
172 static void reset_stat(struct runtime_stat *st)
174 struct rblist *rblist;
175 struct rb_node *pos, *next;
177 rblist = &st->value_list;
178 next = rb_first_cached(&rblist->entries);
182 memset(&container_of(pos, struct saved_value, rb_node)->stats,
184 sizeof(struct stats));
188 void perf_stat__reset_shadow_stats(void)
190 reset_stat(&rt_stat);
191 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
192 memset(&ru_stats, 0, sizeof(ru_stats));
195 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
200 struct runtime_stat_data {
205 static void update_runtime_stat(struct runtime_stat *st,
207 int map_idx, u64 count,
208 struct runtime_stat_data *rsd)
210 struct saved_value *v = saved_value_lookup(NULL, map_idx, true, type,
211 rsd->ctx, st, rsd->cgrp);
214 update_stats(&v->stats, count);
218 * Update various tracking values we maintain to print
219 * more semantic information such as miss/hit ratios,
220 * instruction rates, etc:
222 void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
223 int map_idx, struct runtime_stat *st)
225 u64 count_ns = count;
226 struct saved_value *v;
227 struct runtime_stat_data rsd = {
228 .ctx = evsel_context(counter),
229 .cgrp = counter->cgrp,
232 count *= counter->scale;
234 if (evsel__is_clock(counter))
235 update_runtime_stat(st, STAT_NSECS, map_idx, count_ns, &rsd);
236 else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
237 update_runtime_stat(st, STAT_CYCLES, map_idx, count, &rsd);
238 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
239 update_runtime_stat(st, STAT_CYCLES_IN_TX, map_idx, count, &rsd);
240 else if (perf_stat_evsel__is(counter, TRANSACTION_START))
241 update_runtime_stat(st, STAT_TRANSACTION, map_idx, count, &rsd);
242 else if (perf_stat_evsel__is(counter, ELISION_START))
243 update_runtime_stat(st, STAT_ELISION, map_idx, count, &rsd);
244 else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
245 update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
246 map_idx, count, &rsd);
247 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
248 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
249 map_idx, count, &rsd);
250 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
251 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
252 map_idx, count, &rsd);
253 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
254 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
255 map_idx, count, &rsd);
256 else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
257 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
258 map_idx, count, &rsd);
259 else if (perf_stat_evsel__is(counter, TOPDOWN_RETIRING))
260 update_runtime_stat(st, STAT_TOPDOWN_RETIRING,
261 map_idx, count, &rsd);
262 else if (perf_stat_evsel__is(counter, TOPDOWN_BAD_SPEC))
263 update_runtime_stat(st, STAT_TOPDOWN_BAD_SPEC,
264 map_idx, count, &rsd);
265 else if (perf_stat_evsel__is(counter, TOPDOWN_FE_BOUND))
266 update_runtime_stat(st, STAT_TOPDOWN_FE_BOUND,
267 map_idx, count, &rsd);
268 else if (perf_stat_evsel__is(counter, TOPDOWN_BE_BOUND))
269 update_runtime_stat(st, STAT_TOPDOWN_BE_BOUND,
270 map_idx, count, &rsd);
271 else if (perf_stat_evsel__is(counter, TOPDOWN_HEAVY_OPS))
272 update_runtime_stat(st, STAT_TOPDOWN_HEAVY_OPS,
273 map_idx, count, &rsd);
274 else if (perf_stat_evsel__is(counter, TOPDOWN_BR_MISPREDICT))
275 update_runtime_stat(st, STAT_TOPDOWN_BR_MISPREDICT,
276 map_idx, count, &rsd);
277 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_LAT))
278 update_runtime_stat(st, STAT_TOPDOWN_FETCH_LAT,
279 map_idx, count, &rsd);
280 else if (perf_stat_evsel__is(counter, TOPDOWN_MEM_BOUND))
281 update_runtime_stat(st, STAT_TOPDOWN_MEM_BOUND,
282 map_idx, count, &rsd);
283 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
284 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
285 map_idx, count, &rsd);
286 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
287 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
288 map_idx, count, &rsd);
289 else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
290 update_runtime_stat(st, STAT_BRANCHES, map_idx, count, &rsd);
291 else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
292 update_runtime_stat(st, STAT_CACHEREFS, map_idx, count, &rsd);
293 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
294 update_runtime_stat(st, STAT_L1_DCACHE, map_idx, count, &rsd);
295 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
296 update_runtime_stat(st, STAT_L1_ICACHE, map_idx, count, &rsd);
297 else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
298 update_runtime_stat(st, STAT_LL_CACHE, map_idx, count, &rsd);
299 else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
300 update_runtime_stat(st, STAT_DTLB_CACHE, map_idx, count, &rsd);
301 else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
302 update_runtime_stat(st, STAT_ITLB_CACHE, map_idx, count, &rsd);
303 else if (perf_stat_evsel__is(counter, SMI_NUM))
304 update_runtime_stat(st, STAT_SMI_NUM, map_idx, count, &rsd);
305 else if (perf_stat_evsel__is(counter, APERF))
306 update_runtime_stat(st, STAT_APERF, map_idx, count, &rsd);
308 if (counter->collect_stat) {
309 v = saved_value_lookup(counter, map_idx, true, STAT_NONE, 0, st,
311 update_stats(&v->stats, count);
312 if (counter->metric_leader)
313 v->metric_total += count;
314 } else if (counter->metric_leader) {
315 v = saved_value_lookup(counter->metric_leader,
316 map_idx, true, STAT_NONE, 0, st, rsd.cgrp);
317 v->metric_total += count;
322 /* used for get_ratio_color() */
324 GRC_STALLED_CYCLES_FE,
325 GRC_STALLED_CYCLES_BE,
330 static const char *get_ratio_color(enum grc_type type, double ratio)
332 static const double grc_table[GRC_MAX_NR][3] = {
333 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
334 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
335 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
337 const char *color = PERF_COLOR_NORMAL;
339 if (ratio > grc_table[type][0])
340 color = PERF_COLOR_RED;
341 else if (ratio > grc_table[type][1])
342 color = PERF_COLOR_MAGENTA;
343 else if (ratio > grc_table[type][2])
344 color = PERF_COLOR_YELLOW;
349 static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
354 evlist__for_each_entry (evsel_list, c2) {
355 if (!strcasecmp(c2->name, name) && !c2->collect_stat)
361 /* Mark MetricExpr target events and link events using them to them. */
362 void perf_stat__collect_metric_expr(struct evlist *evsel_list)
364 struct evsel *counter, *leader, **metric_events, *oc;
366 struct expr_parse_ctx *ctx;
367 struct hashmap_entry *cur;
371 ctx = expr__ctx_new();
373 pr_debug("expr__ctx_new failed");
376 evlist__for_each_entry(evsel_list, counter) {
377 bool invalid = false;
379 leader = evsel__leader(counter);
380 if (!counter->metric_expr)
383 expr__ctx_clear(ctx);
384 metric_events = counter->metric_events;
385 if (!metric_events) {
386 if (expr__find_ids(counter->metric_expr,
391 metric_events = calloc(sizeof(struct evsel *),
392 hashmap__size(ctx->ids) + 1);
393 if (!metric_events) {
397 counter->metric_events = metric_events;
401 hashmap__for_each_entry(ctx->ids, cur, bkt) {
402 const char *metric_name = cur->pkey;
406 /* Search in group */
407 for_each_group_member (oc, leader) {
408 if (!strcasecmp(oc->name,
417 /* Search ignoring groups */
418 oc = perf_stat__find_event(evsel_list,
422 /* Deduping one is good enough to handle duplicated PMUs. */
423 static char *printed;
426 * Adding events automatically would be difficult, because
427 * it would risk creating groups that are not schedulable.
428 * perf stat doesn't understand all the scheduling constraints
429 * of events. So we ask the user instead to add the missing
433 strcasecmp(printed, metric_name)) {
435 "Add %s event to groups to get metric expression for %s\n",
439 printed = strdup(metric_name);
444 metric_events[i++] = oc;
445 oc->collect_stat = true;
447 metric_events[i] = NULL;
450 counter->metric_events = NULL;
451 counter->metric_expr = NULL;
457 static double runtime_stat_avg(struct runtime_stat *st,
458 enum stat_type type, int map_idx,
459 struct runtime_stat_data *rsd)
461 struct saved_value *v;
463 v = saved_value_lookup(NULL, map_idx, false, type, rsd->ctx, st, rsd->cgrp);
467 return avg_stats(&v->stats);
470 static double runtime_stat_n(struct runtime_stat *st,
471 enum stat_type type, int map_idx,
472 struct runtime_stat_data *rsd)
474 struct saved_value *v;
476 v = saved_value_lookup(NULL, map_idx, false, type, rsd->ctx, st, rsd->cgrp);
483 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
484 int map_idx, double avg,
485 struct perf_stat_output_ctx *out,
486 struct runtime_stat *st,
487 struct runtime_stat_data *rsd)
489 double total, ratio = 0.0;
492 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, rsd);
495 ratio = avg / total * 100.0;
497 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
500 out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
503 out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
506 static void print_stalled_cycles_backend(struct perf_stat_config *config,
507 int map_idx, double avg,
508 struct perf_stat_output_ctx *out,
509 struct runtime_stat *st,
510 struct runtime_stat_data *rsd)
512 double total, ratio = 0.0;
515 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, rsd);
518 ratio = avg / total * 100.0;
520 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
522 out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
525 static void print_branch_misses(struct perf_stat_config *config,
526 int map_idx, double avg,
527 struct perf_stat_output_ctx *out,
528 struct runtime_stat *st,
529 struct runtime_stat_data *rsd)
531 double total, ratio = 0.0;
534 total = runtime_stat_avg(st, STAT_BRANCHES, map_idx, rsd);
537 ratio = avg / total * 100.0;
539 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
541 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
544 static void print_l1_dcache_misses(struct perf_stat_config *config,
545 int map_idx, double avg,
546 struct perf_stat_output_ctx *out,
547 struct runtime_stat *st,
548 struct runtime_stat_data *rsd)
550 double total, ratio = 0.0;
553 total = runtime_stat_avg(st, STAT_L1_DCACHE, map_idx, rsd);
556 ratio = avg / total * 100.0;
558 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
560 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache accesses", ratio);
563 static void print_l1_icache_misses(struct perf_stat_config *config,
564 int map_idx, double avg,
565 struct perf_stat_output_ctx *out,
566 struct runtime_stat *st,
567 struct runtime_stat_data *rsd)
569 double total, ratio = 0.0;
572 total = runtime_stat_avg(st, STAT_L1_ICACHE, map_idx, rsd);
575 ratio = avg / total * 100.0;
577 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
578 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache accesses", ratio);
581 static void print_dtlb_cache_misses(struct perf_stat_config *config,
582 int map_idx, double avg,
583 struct perf_stat_output_ctx *out,
584 struct runtime_stat *st,
585 struct runtime_stat_data *rsd)
587 double total, ratio = 0.0;
590 total = runtime_stat_avg(st, STAT_DTLB_CACHE, map_idx, rsd);
593 ratio = avg / total * 100.0;
595 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
596 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache accesses", ratio);
599 static void print_itlb_cache_misses(struct perf_stat_config *config,
600 int map_idx, double avg,
601 struct perf_stat_output_ctx *out,
602 struct runtime_stat *st,
603 struct runtime_stat_data *rsd)
605 double total, ratio = 0.0;
608 total = runtime_stat_avg(st, STAT_ITLB_CACHE, map_idx, rsd);
611 ratio = avg / total * 100.0;
613 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
614 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache accesses", ratio);
617 static void print_ll_cache_misses(struct perf_stat_config *config,
618 int map_idx, double avg,
619 struct perf_stat_output_ctx *out,
620 struct runtime_stat *st,
621 struct runtime_stat_data *rsd)
623 double total, ratio = 0.0;
626 total = runtime_stat_avg(st, STAT_LL_CACHE, map_idx, rsd);
629 ratio = avg / total * 100.0;
631 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
632 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache accesses", ratio);
636 * High level "TopDown" CPU core pipe line bottleneck break down.
638 * Basic concept following
639 * Yasin, A Top Down Method for Performance analysis and Counter architecture
642 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
644 * Frontend -> Backend -> Retiring
645 * BadSpeculation in addition means out of order execution that is thrown away
646 * (for example branch mispredictions)
647 * Frontend is instruction decoding.
648 * Backend is execution, like computation and accessing data in memory
649 * Retiring is good execution that is not directly bottlenecked
651 * The formulas are computed in slots.
652 * A slot is an entry in the pipeline each for the pipeline width
653 * (for example a 4-wide pipeline has 4 slots for each cycle)
656 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
658 * Retiring = SlotsRetired / TotalSlots
659 * FrontendBound = FetchBubbles / TotalSlots
660 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
662 * The kernel provides the mapping to the low level CPU events and any scaling
663 * needed for the CPU pipeline width, for example:
665 * TotalSlots = Cycles * 4
667 * The scaling factor is communicated in the sysfs unit.
669 * In some cases the CPU may not be able to measure all the formulas due to
670 * missing events. In this case multiple formulas are combined, as possible.
672 * Full TopDown supports more levels to sub-divide each area: for example
673 * BackendBound into computing bound and memory bound. For now we only
674 * support Level 1 TopDown.
677 static double sanitize_val(double x)
679 if (x < 0 && x >= -0.02)
684 static double td_total_slots(int map_idx, struct runtime_stat *st,
685 struct runtime_stat_data *rsd)
687 return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, map_idx, rsd);
690 static double td_bad_spec(int map_idx, struct runtime_stat *st,
691 struct runtime_stat_data *rsd)
697 total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, map_idx, rsd) -
698 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, map_idx, rsd) +
699 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, map_idx, rsd);
701 total_slots = td_total_slots(map_idx, st, rsd);
703 bad_spec = total / total_slots;
704 return sanitize_val(bad_spec);
707 static double td_retiring(int map_idx, struct runtime_stat *st,
708 struct runtime_stat_data *rsd)
711 double total_slots = td_total_slots(map_idx, st, rsd);
712 double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
716 retiring = ret_slots / total_slots;
720 static double td_fe_bound(int map_idx, struct runtime_stat *st,
721 struct runtime_stat_data *rsd)
724 double total_slots = td_total_slots(map_idx, st, rsd);
725 double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
729 fe_bound = fetch_bub / total_slots;
733 static double td_be_bound(int map_idx, struct runtime_stat *st,
734 struct runtime_stat_data *rsd)
736 double sum = (td_fe_bound(map_idx, st, rsd) +
737 td_bad_spec(map_idx, st, rsd) +
738 td_retiring(map_idx, st, rsd));
741 return sanitize_val(1.0 - sum);
745 * Kernel reports metrics multiplied with slots. To get back
746 * the ratios we need to recreate the sum.
749 static double td_metric_ratio(int map_idx, enum stat_type type,
750 struct runtime_stat *stat,
751 struct runtime_stat_data *rsd)
753 double sum = runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, map_idx, rsd) +
754 runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, map_idx, rsd) +
755 runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, map_idx, rsd) +
756 runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, map_idx, rsd);
757 double d = runtime_stat_avg(stat, type, map_idx, rsd);
765 * ... but only if most of the values are actually available.
766 * We allow two missing.
769 static bool full_td(int map_idx, struct runtime_stat *stat,
770 struct runtime_stat_data *rsd)
774 if (runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, map_idx, rsd) > 0)
776 if (runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, map_idx, rsd) > 0)
778 if (runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, map_idx, rsd) > 0)
780 if (runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, map_idx, rsd) > 0)
785 static void print_smi_cost(struct perf_stat_config *config, int map_idx,
786 struct perf_stat_output_ctx *out,
787 struct runtime_stat *st,
788 struct runtime_stat_data *rsd)
790 double smi_num, aperf, cycles, cost = 0.0;
791 const char *color = NULL;
793 smi_num = runtime_stat_avg(st, STAT_SMI_NUM, map_idx, rsd);
794 aperf = runtime_stat_avg(st, STAT_APERF, map_idx, rsd);
795 cycles = runtime_stat_avg(st, STAT_CYCLES, map_idx, rsd);
797 if ((cycles == 0) || (aperf == 0))
801 cost = (aperf - cycles) / aperf * 100.00;
804 color = PERF_COLOR_RED;
805 out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
806 out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
809 static int prepare_metric(struct evsel **metric_events,
810 struct metric_ref *metric_refs,
811 struct expr_parse_ctx *pctx,
813 struct runtime_stat *st)
819 for (i = 0; metric_events[i]; i++) {
820 struct saved_value *v;
822 u64 metric_total = 0;
825 if (evsel__is_tool(metric_events[i])) {
827 switch (metric_events[i]->tool_event) {
828 case PERF_TOOL_DURATION_TIME:
829 stats = &walltime_nsecs_stats;
832 case PERF_TOOL_USER_TIME:
833 stats = &ru_stats.ru_utime_usec_stat;
836 case PERF_TOOL_SYSTEM_TIME:
837 stats = &ru_stats.ru_stime_usec_stat;
841 pr_err("Invalid tool event 'none'");
844 pr_err("Invalid tool event 'max'");
847 pr_err("Unknown tool event '%s'", evsel__name(metric_events[i]));
851 v = saved_value_lookup(metric_events[i], map_idx, false,
853 metric_events[i]->cgrp);
858 * If an event was scaled during stat gathering, reverse
859 * the scale before computing the metric.
861 scale = 1.0 / metric_events[i]->scale;
863 source_count = evsel__source_count(metric_events[i]);
866 metric_total = v->metric_total * scale;
868 n = strdup(evsel__metric_id(metric_events[i]));
872 expr__add_id_val_source_count(pctx, n,
873 metric_total ? : avg_stats(stats) * scale,
877 for (j = 0; metric_refs && metric_refs[j].metric_name; j++) {
878 ret = expr__add_ref(pctx, &metric_refs[j]);
886 static void generic_metric(struct perf_stat_config *config,
887 const char *metric_expr,
888 struct evsel **metric_events,
889 struct metric_ref *metric_refs,
891 const char *metric_name,
892 const char *metric_unit,
895 struct perf_stat_output_ctx *out,
896 struct runtime_stat *st)
898 print_metric_t print_metric = out->print_metric;
899 struct expr_parse_ctx *pctx;
902 void *ctxp = out->ctx;
904 pctx = expr__ctx_new();
908 if (config->user_requested_cpu_list)
909 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
910 pctx->sctx.runtime = runtime;
911 pctx->sctx.system_wide = config->system_wide;
912 i = prepare_metric(metric_events, metric_refs, pctx, map_idx, st);
914 expr__ctx_free(pctx);
917 if (!metric_events[i]) {
918 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
922 if (metric_unit && metric_name) {
923 if (perf_pmu__convert_scale(metric_unit,
924 &unit, &scale) >= 0) {
927 if (strstr(metric_expr, "?"))
928 scnprintf(metric_bf, sizeof(metric_bf),
929 "%s %s_%d", unit, metric_name, runtime);
931 scnprintf(metric_bf, sizeof(metric_bf),
932 "%s %s", unit, metric_name);
934 print_metric(config, ctxp, NULL, "%8.1f",
937 print_metric(config, ctxp, NULL, "%8.2f",
940 out->force_header ? name : "",
944 print_metric(config, ctxp, NULL, NULL,
946 (metric_name ? metric_name : name) : "", 0);
949 print_metric(config, ctxp, NULL, NULL,
951 (metric_name ? metric_name : name) : "", 0);
954 expr__ctx_free(pctx);
957 double test_generic_metric(struct metric_expr *mexp, int map_idx, struct runtime_stat *st)
959 struct expr_parse_ctx *pctx;
962 pctx = expr__ctx_new();
966 if (prepare_metric(mexp->metric_events, mexp->metric_refs, pctx, map_idx, st) < 0)
969 if (expr__parse(&ratio, pctx, mexp->metric_expr))
973 expr__ctx_free(pctx);
977 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
979 double avg, int map_idx,
980 struct perf_stat_output_ctx *out,
981 struct rblist *metric_events,
982 struct runtime_stat *st)
984 void *ctxp = out->ctx;
985 print_metric_t print_metric = out->print_metric;
986 double total, ratio = 0.0, total2;
987 const char *color = NULL;
988 struct runtime_stat_data rsd = {
989 .ctx = evsel_context(evsel),
992 struct metric_event *me;
995 if (config->iostat_run) {
996 iostat_print_metric(config, evsel, out);
997 } else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
998 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, &rsd);
1001 ratio = avg / total;
1002 print_metric(config, ctxp, NULL, "%7.2f ",
1003 "insn per cycle", ratio);
1005 print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
1008 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT, map_idx, &rsd);
1010 total = max(total, runtime_stat_avg(st,
1011 STAT_STALLED_CYCLES_BACK,
1015 out->new_line(config, ctxp);
1016 ratio = total / avg;
1017 print_metric(config, ctxp, NULL, "%7.2f ",
1018 "stalled cycles per insn",
1021 } else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
1022 if (runtime_stat_n(st, STAT_BRANCHES, map_idx, &rsd) != 0)
1023 print_branch_misses(config, map_idx, avg, out, st, &rsd);
1025 print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
1027 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1028 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1029 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1030 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1032 if (runtime_stat_n(st, STAT_L1_DCACHE, map_idx, &rsd) != 0)
1033 print_l1_dcache_misses(config, map_idx, avg, out, st, &rsd);
1035 print_metric(config, ctxp, NULL, NULL, "of all L1-dcache accesses", 0);
1037 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1038 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1039 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1040 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1042 if (runtime_stat_n(st, STAT_L1_ICACHE, map_idx, &rsd) != 0)
1043 print_l1_icache_misses(config, map_idx, avg, out, st, &rsd);
1045 print_metric(config, ctxp, NULL, NULL, "of all L1-icache accesses", 0);
1047 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1048 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1049 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1050 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1052 if (runtime_stat_n(st, STAT_DTLB_CACHE, map_idx, &rsd) != 0)
1053 print_dtlb_cache_misses(config, map_idx, avg, out, st, &rsd);
1055 print_metric(config, ctxp, NULL, NULL, "of all dTLB cache accesses", 0);
1057 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1058 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1059 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1060 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1062 if (runtime_stat_n(st, STAT_ITLB_CACHE, map_idx, &rsd) != 0)
1063 print_itlb_cache_misses(config, map_idx, avg, out, st, &rsd);
1065 print_metric(config, ctxp, NULL, NULL, "of all iTLB cache accesses", 0);
1067 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1068 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_LL |
1069 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1070 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1072 if (runtime_stat_n(st, STAT_LL_CACHE, map_idx, &rsd) != 0)
1073 print_ll_cache_misses(config, map_idx, avg, out, st, &rsd);
1075 print_metric(config, ctxp, NULL, NULL, "of all LL-cache accesses", 0);
1076 } else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
1077 total = runtime_stat_avg(st, STAT_CACHEREFS, map_idx, &rsd);
1080 ratio = avg * 100 / total;
1082 if (runtime_stat_n(st, STAT_CACHEREFS, map_idx, &rsd) != 0)
1083 print_metric(config, ctxp, NULL, "%8.3f %%",
1084 "of all cache refs", ratio);
1086 print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
1087 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1088 print_stalled_cycles_frontend(config, map_idx, avg, out, st, &rsd);
1089 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1090 print_stalled_cycles_backend(config, map_idx, avg, out, st, &rsd);
1091 } else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1092 total = runtime_stat_avg(st, STAT_NSECS, map_idx, &rsd);
1095 ratio = avg / total;
1096 print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
1098 print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
1100 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
1101 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, &rsd);
1104 print_metric(config, ctxp, NULL,
1105 "%7.2f%%", "transactional cycles",
1106 100.0 * (avg / total));
1108 print_metric(config, ctxp, NULL, NULL, "transactional cycles",
1110 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
1111 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, &rsd);
1112 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, map_idx, &rsd);
1117 print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
1118 100.0 * ((total2-avg) / total));
1120 print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
1121 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
1122 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, map_idx, &rsd);
1125 ratio = total / avg;
1127 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, map_idx, &rsd) != 0)
1128 print_metric(config, ctxp, NULL, "%8.0f",
1129 "cycles / transaction", ratio);
1131 print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
1133 } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
1134 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, map_idx, &rsd);
1137 ratio = total / avg;
1139 print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
1140 } else if (evsel__is_clock(evsel)) {
1141 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
1142 print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1143 avg / (ratio * evsel->scale));
1145 print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1146 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1147 double fe_bound = td_fe_bound(map_idx, st, &rsd);
1150 color = PERF_COLOR_RED;
1151 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1153 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1154 double retiring = td_retiring(map_idx, st, &rsd);
1157 color = PERF_COLOR_GREEN;
1158 print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1160 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1161 double bad_spec = td_bad_spec(map_idx, st, &rsd);
1164 color = PERF_COLOR_RED;
1165 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1167 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1168 double be_bound = td_be_bound(map_idx, st, &rsd);
1169 const char *name = "backend bound";
1170 static int have_recovery_bubbles = -1;
1172 /* In case the CPU does not support topdown-recovery-bubbles */
1173 if (have_recovery_bubbles < 0)
1174 have_recovery_bubbles = pmu_have_event("cpu",
1175 "topdown-recovery-bubbles");
1176 if (!have_recovery_bubbles)
1177 name = "backend bound/bad spec";
1180 color = PERF_COLOR_RED;
1181 if (td_total_slots(map_idx, st, &rsd) > 0)
1182 print_metric(config, ctxp, color, "%8.1f%%", name,
1185 print_metric(config, ctxp, NULL, NULL, name, 0);
1186 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RETIRING) &&
1187 full_td(map_idx, st, &rsd)) {
1188 double retiring = td_metric_ratio(map_idx,
1189 STAT_TOPDOWN_RETIRING, st,
1192 color = PERF_COLOR_GREEN;
1193 print_metric(config, ctxp, color, "%8.1f%%", "Retiring",
1195 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FE_BOUND) &&
1196 full_td(map_idx, st, &rsd)) {
1197 double fe_bound = td_metric_ratio(map_idx,
1198 STAT_TOPDOWN_FE_BOUND, st,
1201 color = PERF_COLOR_RED;
1202 print_metric(config, ctxp, color, "%8.1f%%", "Frontend Bound",
1204 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BE_BOUND) &&
1205 full_td(map_idx, st, &rsd)) {
1206 double be_bound = td_metric_ratio(map_idx,
1207 STAT_TOPDOWN_BE_BOUND, st,
1210 color = PERF_COLOR_RED;
1211 print_metric(config, ctxp, color, "%8.1f%%", "Backend Bound",
1213 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BAD_SPEC) &&
1214 full_td(map_idx, st, &rsd)) {
1215 double bad_spec = td_metric_ratio(map_idx,
1216 STAT_TOPDOWN_BAD_SPEC, st,
1219 color = PERF_COLOR_RED;
1220 print_metric(config, ctxp, color, "%8.1f%%", "Bad Speculation",
1222 } else if (perf_stat_evsel__is(evsel, TOPDOWN_HEAVY_OPS) &&
1223 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1224 double retiring = td_metric_ratio(map_idx,
1225 STAT_TOPDOWN_RETIRING, st,
1227 double heavy_ops = td_metric_ratio(map_idx,
1228 STAT_TOPDOWN_HEAVY_OPS, st,
1230 double light_ops = retiring - heavy_ops;
1232 if (retiring > 0.7 && heavy_ops > 0.1)
1233 color = PERF_COLOR_GREEN;
1234 print_metric(config, ctxp, color, "%8.1f%%", "Heavy Operations",
1236 if (retiring > 0.7 && light_ops > 0.6)
1237 color = PERF_COLOR_GREEN;
1240 print_metric(config, ctxp, color, "%8.1f%%", "Light Operations",
1242 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BR_MISPREDICT) &&
1243 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1244 double bad_spec = td_metric_ratio(map_idx,
1245 STAT_TOPDOWN_BAD_SPEC, st,
1247 double br_mis = td_metric_ratio(map_idx,
1248 STAT_TOPDOWN_BR_MISPREDICT, st,
1250 double m_clears = bad_spec - br_mis;
1252 if (bad_spec > 0.1 && br_mis > 0.05)
1253 color = PERF_COLOR_RED;
1254 print_metric(config, ctxp, color, "%8.1f%%", "Branch Mispredict",
1256 if (bad_spec > 0.1 && m_clears > 0.05)
1257 color = PERF_COLOR_RED;
1260 print_metric(config, ctxp, color, "%8.1f%%", "Machine Clears",
1262 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_LAT) &&
1263 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1264 double fe_bound = td_metric_ratio(map_idx,
1265 STAT_TOPDOWN_FE_BOUND, st,
1267 double fetch_lat = td_metric_ratio(map_idx,
1268 STAT_TOPDOWN_FETCH_LAT, st,
1270 double fetch_bw = fe_bound - fetch_lat;
1272 if (fe_bound > 0.2 && fetch_lat > 0.15)
1273 color = PERF_COLOR_RED;
1274 print_metric(config, ctxp, color, "%8.1f%%", "Fetch Latency",
1276 if (fe_bound > 0.2 && fetch_bw > 0.1)
1277 color = PERF_COLOR_RED;
1280 print_metric(config, ctxp, color, "%8.1f%%", "Fetch Bandwidth",
1282 } else if (perf_stat_evsel__is(evsel, TOPDOWN_MEM_BOUND) &&
1283 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1284 double be_bound = td_metric_ratio(map_idx,
1285 STAT_TOPDOWN_BE_BOUND, st,
1287 double mem_bound = td_metric_ratio(map_idx,
1288 STAT_TOPDOWN_MEM_BOUND, st,
1290 double core_bound = be_bound - mem_bound;
1292 if (be_bound > 0.2 && mem_bound > 0.2)
1293 color = PERF_COLOR_RED;
1294 print_metric(config, ctxp, color, "%8.1f%%", "Memory Bound",
1296 if (be_bound > 0.2 && core_bound > 0.1)
1297 color = PERF_COLOR_RED;
1300 print_metric(config, ctxp, color, "%8.1f%%", "Core Bound",
1302 } else if (evsel->metric_expr) {
1303 generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1304 evsel->name, evsel->metric_name, NULL, 1,
1306 } else if (runtime_stat_n(st, STAT_NSECS, map_idx, &rsd) != 0) {
1308 char unit_buf[10] = "/sec";
1310 total = runtime_stat_avg(st, STAT_NSECS, map_idx, &rsd);
1312 ratio = convert_unit_double(1000000000.0 * avg / total, &unit);
1315 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1316 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1317 } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1318 print_smi_cost(config, map_idx, out, st, &rsd);
1323 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1324 struct metric_expr *mexp;
1326 list_for_each_entry (mexp, &me->head, nd) {
1328 out->new_line(config, ctxp);
1329 generic_metric(config, mexp->metric_expr, mexp->metric_events,
1330 mexp->metric_refs, evsel->name, mexp->metric_name,
1331 mexp->metric_unit, mexp->runtime,
1336 print_metric(config, ctxp, NULL, NULL, NULL, 0);