Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
f87027b9 JO |
2 | #include <stdio.h> |
3 | #include "evsel.h" | |
4 | #include "stat.h" | |
5 | #include "color.h" | |
fb4605ba | 6 | #include "pmu.h" |
37932c18 AK |
7 | #include "rblist.h" |
8 | #include "evlist.h" | |
9 | #include "expr.h" | |
b18f3e36 | 10 | #include "metricgroup.h" |
d8f9da24 | 11 | #include <linux/zalloc.h> |
f87027b9 | 12 | |
44d49a60 AK |
13 | /* |
14 | * AGGR_GLOBAL: Use CPU 0 | |
15 | * AGGR_SOCKET: Use first CPU of socket | |
db5742b6 | 16 | * AGGR_DIE: Use first CPU of die |
44d49a60 AK |
17 | * AGGR_CORE: Use first CPU of core |
18 | * AGGR_NONE: Use matching CPU | |
19 | * AGGR_THREAD: Not supported? | |
20 | */ | |
fb4605ba | 21 | static bool have_frontend_stalled; |
f87027b9 | 22 | |
8efb2df1 | 23 | struct runtime_stat rt_stat; |
f87027b9 JO |
24 | struct stats walltime_nsecs_stats; |
25 | ||
37932c18 AK |
26 | struct saved_value { |
27 | struct rb_node rb_node; | |
32dcd021 | 28 | struct evsel *evsel; |
49cd456a JY |
29 | enum stat_type type; |
30 | int ctx; | |
37932c18 | 31 | int cpu; |
49cd456a | 32 | struct runtime_stat *stat; |
37932c18 AK |
33 | struct stats stats; |
34 | }; | |
35 | ||
36 | static int saved_value_cmp(struct rb_node *rb_node, const void *entry) | |
37 | { | |
38 | struct saved_value *a = container_of(rb_node, | |
39 | struct saved_value, | |
40 | rb_node); | |
41 | const struct saved_value *b = entry; | |
42 | ||
37932c18 AK |
43 | if (a->cpu != b->cpu) |
44 | return a->cpu - b->cpu; | |
49cd456a JY |
45 | |
46 | /* | |
47 | * Previously the rbtree was used to link generic metrics. | |
48 | * The keys were evsel/cpu. Now the rbtree is extended to support | |
49 | * per-thread shadow stats. For shadow stats case, the keys | |
50 | * are cpu/type/ctx/stat (evsel is NULL). For generic metrics | |
51 | * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL). | |
52 | */ | |
53 | if (a->type != b->type) | |
54 | return a->type - b->type; | |
55 | ||
56 | if (a->ctx != b->ctx) | |
57 | return a->ctx - b->ctx; | |
58 | ||
59 | if (a->evsel == NULL && b->evsel == NULL) { | |
60 | if (a->stat == b->stat) | |
61 | return 0; | |
62 | ||
63 | if ((char *)a->stat < (char *)b->stat) | |
64 | return -1; | |
65 | ||
66 | return 1; | |
67 | } | |
68 | ||
5e97665f AK |
69 | if (a->evsel == b->evsel) |
70 | return 0; | |
71 | if ((char *)a->evsel < (char *)b->evsel) | |
72 | return -1; | |
73 | return +1; | |
37932c18 AK |
74 | } |
75 | ||
76 | static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused, | |
77 | const void *entry) | |
78 | { | |
79 | struct saved_value *nd = malloc(sizeof(struct saved_value)); | |
80 | ||
81 | if (!nd) | |
82 | return NULL; | |
83 | memcpy(nd, entry, sizeof(struct saved_value)); | |
84 | return &nd->rb_node; | |
85 | } | |
86 | ||
b984aff7 JY |
87 | static void saved_value_delete(struct rblist *rblist __maybe_unused, |
88 | struct rb_node *rb_node) | |
89 | { | |
90 | struct saved_value *v; | |
91 | ||
92 | BUG_ON(!rb_node); | |
93 | v = container_of(rb_node, struct saved_value, rb_node); | |
94 | free(v); | |
95 | } | |
96 | ||
32dcd021 | 97 | static struct saved_value *saved_value_lookup(struct evsel *evsel, |
4e1a0963 | 98 | int cpu, |
1fcd0394 JY |
99 | bool create, |
100 | enum stat_type type, | |
101 | int ctx, | |
102 | struct runtime_stat *st) | |
37932c18 | 103 | { |
1fcd0394 | 104 | struct rblist *rblist; |
37932c18 AK |
105 | struct rb_node *nd; |
106 | struct saved_value dm = { | |
107 | .cpu = cpu, | |
37932c18 | 108 | .evsel = evsel, |
1fcd0394 JY |
109 | .type = type, |
110 | .ctx = ctx, | |
111 | .stat = st, | |
37932c18 | 112 | }; |
1fcd0394 JY |
113 | |
114 | rblist = &st->value_list; | |
115 | ||
116 | nd = rblist__find(rblist, &dm); | |
37932c18 AK |
117 | if (nd) |
118 | return container_of(nd, struct saved_value, rb_node); | |
119 | if (create) { | |
1fcd0394 JY |
120 | rblist__add_node(rblist, &dm); |
121 | nd = rblist__find(rblist, &dm); | |
37932c18 AK |
122 | if (nd) |
123 | return container_of(nd, struct saved_value, rb_node); | |
124 | } | |
125 | return NULL; | |
126 | } | |
127 | ||
8efb2df1 JY |
128 | void runtime_stat__init(struct runtime_stat *st) |
129 | { | |
130 | struct rblist *rblist = &st->value_list; | |
131 | ||
132 | rblist__init(rblist); | |
133 | rblist->node_cmp = saved_value_cmp; | |
134 | rblist->node_new = saved_value_new; | |
135 | rblist->node_delete = saved_value_delete; | |
136 | } | |
137 | ||
138 | void runtime_stat__exit(struct runtime_stat *st) | |
139 | { | |
140 | rblist__exit(&st->value_list); | |
141 | } | |
142 | ||
fb4605ba AK |
143 | void perf_stat__init_shadow_stats(void) |
144 | { | |
145 | have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend"); | |
8efb2df1 | 146 | runtime_stat__init(&rt_stat); |
fb4605ba AK |
147 | } |
148 | ||
32dcd021 | 149 | static int evsel_context(struct evsel *evsel) |
f87027b9 JO |
150 | { |
151 | int ctx = 0; | |
152 | ||
1fc632ce | 153 | if (evsel->core.attr.exclude_kernel) |
f87027b9 | 154 | ctx |= CTX_BIT_KERNEL; |
1fc632ce | 155 | if (evsel->core.attr.exclude_user) |
f87027b9 | 156 | ctx |= CTX_BIT_USER; |
1fc632ce | 157 | if (evsel->core.attr.exclude_hv) |
f87027b9 | 158 | ctx |= CTX_BIT_HV; |
1fc632ce | 159 | if (evsel->core.attr.exclude_host) |
f87027b9 | 160 | ctx |= CTX_BIT_HOST; |
1fc632ce | 161 | if (evsel->core.attr.exclude_idle) |
f87027b9 JO |
162 | ctx |= CTX_BIT_IDLE; |
163 | ||
164 | return ctx; | |
165 | } | |
166 | ||
6a1e2c5c | 167 | static void reset_stat(struct runtime_stat *st) |
f87027b9 | 168 | { |
6a1e2c5c | 169 | struct rblist *rblist; |
37932c18 AK |
170 | struct rb_node *pos, *next; |
171 | ||
6a1e2c5c | 172 | rblist = &st->value_list; |
ca227029 | 173 | next = rb_first_cached(&rblist->entries); |
37932c18 AK |
174 | while (next) { |
175 | pos = next; | |
176 | next = rb_next(pos); | |
177 | memset(&container_of(pos, struct saved_value, rb_node)->stats, | |
178 | 0, | |
179 | sizeof(struct stats)); | |
180 | } | |
f87027b9 JO |
181 | } |
182 | ||
6a1e2c5c JY |
183 | void perf_stat__reset_shadow_stats(void) |
184 | { | |
185 | reset_stat(&rt_stat); | |
186 | memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats)); | |
187 | } | |
188 | ||
189 | void perf_stat__reset_shadow_per_stat(struct runtime_stat *st) | |
190 | { | |
191 | reset_stat(st); | |
192 | } | |
193 | ||
1fcd0394 JY |
194 | static void update_runtime_stat(struct runtime_stat *st, |
195 | enum stat_type type, | |
196 | int ctx, int cpu, u64 count) | |
197 | { | |
198 | struct saved_value *v = saved_value_lookup(NULL, cpu, true, | |
199 | type, ctx, st); | |
200 | ||
201 | if (v) | |
202 | update_stats(&v->stats, count); | |
203 | } | |
204 | ||
f87027b9 JO |
205 | /* |
206 | * Update various tracking values we maintain to print | |
207 | * more semantic information such as miss/hit ratios, | |
208 | * instruction rates, etc: | |
209 | */ | |
32dcd021 | 210 | void perf_stat__update_shadow_stats(struct evsel *counter, u64 count, |
1fcd0394 | 211 | int cpu, struct runtime_stat *st) |
f87027b9 JO |
212 | { |
213 | int ctx = evsel_context(counter); | |
57ddf091 | 214 | u64 count_ns = count; |
f87027b9 | 215 | |
54830dd0 JO |
216 | count *= counter->scale; |
217 | ||
eb08d006 | 218 | if (perf_evsel__is_clock(counter)) |
57ddf091 | 219 | update_runtime_stat(st, STAT_NSECS, 0, cpu, count_ns); |
f87027b9 | 220 | else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES)) |
1fcd0394 | 221 | update_runtime_stat(st, STAT_CYCLES, ctx, cpu, count); |
f87027b9 | 222 | else if (perf_stat_evsel__is(counter, CYCLES_IN_TX)) |
1fcd0394 | 223 | update_runtime_stat(st, STAT_CYCLES_IN_TX, ctx, cpu, count); |
f87027b9 | 224 | else if (perf_stat_evsel__is(counter, TRANSACTION_START)) |
1fcd0394 | 225 | update_runtime_stat(st, STAT_TRANSACTION, ctx, cpu, count); |
f87027b9 | 226 | else if (perf_stat_evsel__is(counter, ELISION_START)) |
1fcd0394 | 227 | update_runtime_stat(st, STAT_ELISION, ctx, cpu, count); |
239bd47f | 228 | else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS)) |
1fcd0394 JY |
229 | update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS, |
230 | ctx, cpu, count); | |
239bd47f | 231 | else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED)) |
1fcd0394 JY |
232 | update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED, |
233 | ctx, cpu, count); | |
239bd47f | 234 | else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED)) |
1fcd0394 JY |
235 | update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED, |
236 | ctx, cpu, count); | |
239bd47f | 237 | else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES)) |
1fcd0394 JY |
238 | update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES, |
239 | ctx, cpu, count); | |
239bd47f | 240 | else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES)) |
1fcd0394 JY |
241 | update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES, |
242 | ctx, cpu, count); | |
f87027b9 | 243 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) |
1fcd0394 JY |
244 | update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT, |
245 | ctx, cpu, count); | |
f87027b9 | 246 | else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND)) |
1fcd0394 JY |
247 | update_runtime_stat(st, STAT_STALLED_CYCLES_BACK, |
248 | ctx, cpu, count); | |
f87027b9 | 249 | else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS)) |
1fcd0394 | 250 | update_runtime_stat(st, STAT_BRANCHES, ctx, cpu, count); |
f87027b9 | 251 | else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES)) |
1fcd0394 | 252 | update_runtime_stat(st, STAT_CACHEREFS, ctx, cpu, count); |
f87027b9 | 253 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D)) |
1fcd0394 | 254 | update_runtime_stat(st, STAT_L1_DCACHE, ctx, cpu, count); |
f87027b9 | 255 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I)) |
1fcd0394 | 256 | update_runtime_stat(st, STAT_L1_ICACHE, ctx, cpu, count); |
f87027b9 | 257 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL)) |
1fcd0394 | 258 | update_runtime_stat(st, STAT_LL_CACHE, ctx, cpu, count); |
f87027b9 | 259 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB)) |
1fcd0394 | 260 | update_runtime_stat(st, STAT_DTLB_CACHE, ctx, cpu, count); |
f87027b9 | 261 | else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB)) |
1fcd0394 | 262 | update_runtime_stat(st, STAT_ITLB_CACHE, ctx, cpu, count); |
daefd0bc | 263 | else if (perf_stat_evsel__is(counter, SMI_NUM)) |
1fcd0394 | 264 | update_runtime_stat(st, STAT_SMI_NUM, ctx, cpu, count); |
daefd0bc | 265 | else if (perf_stat_evsel__is(counter, APERF)) |
1fcd0394 | 266 | update_runtime_stat(st, STAT_APERF, ctx, cpu, count); |
37932c18 AK |
267 | |
268 | if (counter->collect_stat) { | |
1fcd0394 JY |
269 | struct saved_value *v = saved_value_lookup(counter, cpu, true, |
270 | STAT_NONE, 0, st); | |
54830dd0 | 271 | update_stats(&v->stats, count); |
37932c18 | 272 | } |
f87027b9 JO |
273 | } |
274 | ||
275 | /* used for get_ratio_color() */ | |
276 | enum grc_type { | |
277 | GRC_STALLED_CYCLES_FE, | |
278 | GRC_STALLED_CYCLES_BE, | |
279 | GRC_CACHE_MISSES, | |
280 | GRC_MAX_NR | |
281 | }; | |
282 | ||
283 | static const char *get_ratio_color(enum grc_type type, double ratio) | |
284 | { | |
285 | static const double grc_table[GRC_MAX_NR][3] = { | |
286 | [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 }, | |
287 | [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 }, | |
288 | [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 }, | |
289 | }; | |
290 | const char *color = PERF_COLOR_NORMAL; | |
291 | ||
292 | if (ratio > grc_table[type][0]) | |
293 | color = PERF_COLOR_RED; | |
294 | else if (ratio > grc_table[type][1]) | |
295 | color = PERF_COLOR_MAGENTA; | |
296 | else if (ratio > grc_table[type][2]) | |
297 | color = PERF_COLOR_YELLOW; | |
298 | ||
299 | return color; | |
300 | } | |
301 | ||
63503dba | 302 | static struct evsel *perf_stat__find_event(struct evlist *evsel_list, |
37932c18 AK |
303 | const char *name) |
304 | { | |
32dcd021 | 305 | struct evsel *c2; |
37932c18 AK |
306 | |
307 | evlist__for_each_entry (evsel_list, c2) { | |
145c407c | 308 | if (!strcasecmp(c2->name, name) && !c2->collect_stat) |
37932c18 AK |
309 | return c2; |
310 | } | |
311 | return NULL; | |
312 | } | |
313 | ||
314 | /* Mark MetricExpr target events and link events using them to them. */ | |
63503dba | 315 | void perf_stat__collect_metric_expr(struct evlist *evsel_list) |
37932c18 | 316 | { |
32dcd021 | 317 | struct evsel *counter, *leader, **metric_events, *oc; |
37932c18 AK |
318 | bool found; |
319 | const char **metric_names; | |
320 | int i; | |
321 | int num_metric_names; | |
322 | ||
323 | evlist__for_each_entry(evsel_list, counter) { | |
324 | bool invalid = false; | |
325 | ||
326 | leader = counter->leader; | |
327 | if (!counter->metric_expr) | |
328 | continue; | |
329 | metric_events = counter->metric_events; | |
330 | if (!metric_events) { | |
331 | if (expr__find_other(counter->metric_expr, counter->name, | |
332 | &metric_names, &num_metric_names) < 0) | |
333 | continue; | |
334 | ||
32dcd021 | 335 | metric_events = calloc(sizeof(struct evsel *), |
37932c18 AK |
336 | num_metric_names + 1); |
337 | if (!metric_events) | |
338 | return; | |
339 | counter->metric_events = metric_events; | |
340 | } | |
341 | ||
342 | for (i = 0; i < num_metric_names; i++) { | |
343 | found = false; | |
344 | if (leader) { | |
345 | /* Search in group */ | |
346 | for_each_group_member (oc, leader) { | |
145c407c AK |
347 | if (!strcasecmp(oc->name, metric_names[i]) && |
348 | !oc->collect_stat) { | |
37932c18 AK |
349 | found = true; |
350 | break; | |
351 | } | |
352 | } | |
353 | } | |
354 | if (!found) { | |
355 | /* Search ignoring groups */ | |
356 | oc = perf_stat__find_event(evsel_list, metric_names[i]); | |
357 | } | |
358 | if (!oc) { | |
359 | /* Deduping one is good enough to handle duplicated PMUs. */ | |
360 | static char *printed; | |
361 | ||
362 | /* | |
363 | * Adding events automatically would be difficult, because | |
364 | * it would risk creating groups that are not schedulable. | |
365 | * perf stat doesn't understand all the scheduling constraints | |
366 | * of events. So we ask the user instead to add the missing | |
367 | * events. | |
368 | */ | |
369 | if (!printed || strcasecmp(printed, metric_names[i])) { | |
370 | fprintf(stderr, | |
371 | "Add %s event to groups to get metric expression for %s\n", | |
372 | metric_names[i], | |
373 | counter->name); | |
374 | printed = strdup(metric_names[i]); | |
375 | } | |
376 | invalid = true; | |
377 | continue; | |
378 | } | |
379 | metric_events[i] = oc; | |
380 | oc->collect_stat = true; | |
381 | } | |
382 | metric_events[i] = NULL; | |
383 | free(metric_names); | |
384 | if (invalid) { | |
385 | free(metric_events); | |
386 | counter->metric_events = NULL; | |
387 | counter->metric_expr = NULL; | |
388 | } | |
389 | } | |
390 | } | |
391 | ||
e0128b30 JY |
392 | static double runtime_stat_avg(struct runtime_stat *st, |
393 | enum stat_type type, int ctx, int cpu) | |
394 | { | |
395 | struct saved_value *v; | |
396 | ||
397 | v = saved_value_lookup(NULL, cpu, false, type, ctx, st); | |
398 | if (!v) | |
399 | return 0.0; | |
400 | ||
401 | return avg_stats(&v->stats); | |
402 | } | |
403 | ||
404 | static double runtime_stat_n(struct runtime_stat *st, | |
405 | enum stat_type type, int ctx, int cpu) | |
406 | { | |
407 | struct saved_value *v; | |
408 | ||
409 | v = saved_value_lookup(NULL, cpu, false, type, ctx, st); | |
410 | if (!v) | |
411 | return 0.0; | |
412 | ||
413 | return v->stats.n; | |
414 | } | |
415 | ||
6ca9a082 JO |
416 | static void print_stalled_cycles_frontend(struct perf_stat_config *config, |
417 | int cpu, | |
32dcd021 | 418 | struct evsel *evsel, double avg, |
e0128b30 JY |
419 | struct perf_stat_output_ctx *out, |
420 | struct runtime_stat *st) | |
f87027b9 JO |
421 | { |
422 | double total, ratio = 0.0; | |
423 | const char *color; | |
424 | int ctx = evsel_context(evsel); | |
425 | ||
e0128b30 | 426 | total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu); |
f87027b9 JO |
427 | |
428 | if (total) | |
429 | ratio = avg / total * 100.0; | |
430 | ||
431 | color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio); | |
432 | ||
140aeadc | 433 | if (ratio) |
6ca9a082 | 434 | out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle", |
140aeadc AK |
435 | ratio); |
436 | else | |
6ca9a082 | 437 | out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0); |
f87027b9 JO |
438 | } |
439 | ||
6ca9a082 JO |
440 | static void print_stalled_cycles_backend(struct perf_stat_config *config, |
441 | int cpu, | |
32dcd021 | 442 | struct evsel *evsel, double avg, |
e0128b30 JY |
443 | struct perf_stat_output_ctx *out, |
444 | struct runtime_stat *st) | |
f87027b9 JO |
445 | { |
446 | double total, ratio = 0.0; | |
447 | const char *color; | |
448 | int ctx = evsel_context(evsel); | |
449 | ||
e0128b30 | 450 | total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu); |
f87027b9 JO |
451 | |
452 | if (total) | |
453 | ratio = avg / total * 100.0; | |
454 | ||
455 | color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio); | |
456 | ||
6ca9a082 | 457 | out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio); |
f87027b9 JO |
458 | } |
459 | ||
6ca9a082 JO |
460 | static void print_branch_misses(struct perf_stat_config *config, |
461 | int cpu, | |
32dcd021 | 462 | struct evsel *evsel, |
140aeadc | 463 | double avg, |
e0128b30 JY |
464 | struct perf_stat_output_ctx *out, |
465 | struct runtime_stat *st) | |
f87027b9 JO |
466 | { |
467 | double total, ratio = 0.0; | |
468 | const char *color; | |
469 | int ctx = evsel_context(evsel); | |
470 | ||
e0128b30 | 471 | total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu); |
f87027b9 JO |
472 | |
473 | if (total) | |
474 | ratio = avg / total * 100.0; | |
475 | ||
476 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); | |
477 | ||
6ca9a082 | 478 | out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio); |
f87027b9 JO |
479 | } |
480 | ||
6ca9a082 JO |
481 | static void print_l1_dcache_misses(struct perf_stat_config *config, |
482 | int cpu, | |
32dcd021 | 483 | struct evsel *evsel, |
140aeadc | 484 | double avg, |
e0128b30 JY |
485 | struct perf_stat_output_ctx *out, |
486 | struct runtime_stat *st) | |
487 | ||
f87027b9 JO |
488 | { |
489 | double total, ratio = 0.0; | |
490 | const char *color; | |
491 | int ctx = evsel_context(evsel); | |
492 | ||
e0128b30 | 493 | total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu); |
f87027b9 JO |
494 | |
495 | if (total) | |
496 | ratio = avg / total * 100.0; | |
497 | ||
498 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); | |
499 | ||
6ca9a082 | 500 | out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio); |
f87027b9 JO |
501 | } |
502 | ||
6ca9a082 JO |
503 | static void print_l1_icache_misses(struct perf_stat_config *config, |
504 | int cpu, | |
32dcd021 | 505 | struct evsel *evsel, |
140aeadc | 506 | double avg, |
e0128b30 JY |
507 | struct perf_stat_output_ctx *out, |
508 | struct runtime_stat *st) | |
509 | ||
f87027b9 JO |
510 | { |
511 | double total, ratio = 0.0; | |
512 | const char *color; | |
513 | int ctx = evsel_context(evsel); | |
514 | ||
e0128b30 | 515 | total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu); |
f87027b9 JO |
516 | |
517 | if (total) | |
518 | ratio = avg / total * 100.0; | |
519 | ||
520 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); | |
6ca9a082 | 521 | out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio); |
f87027b9 JO |
522 | } |
523 | ||
6ca9a082 JO |
524 | static void print_dtlb_cache_misses(struct perf_stat_config *config, |
525 | int cpu, | |
32dcd021 | 526 | struct evsel *evsel, |
140aeadc | 527 | double avg, |
e0128b30 JY |
528 | struct perf_stat_output_ctx *out, |
529 | struct runtime_stat *st) | |
f87027b9 JO |
530 | { |
531 | double total, ratio = 0.0; | |
532 | const char *color; | |
533 | int ctx = evsel_context(evsel); | |
534 | ||
e0128b30 | 535 | total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu); |
f87027b9 JO |
536 | |
537 | if (total) | |
538 | ratio = avg / total * 100.0; | |
539 | ||
540 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); | |
6ca9a082 | 541 | out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio); |
f87027b9 JO |
542 | } |
543 | ||
6ca9a082 JO |
544 | static void print_itlb_cache_misses(struct perf_stat_config *config, |
545 | int cpu, | |
32dcd021 | 546 | struct evsel *evsel, |
140aeadc | 547 | double avg, |
e0128b30 JY |
548 | struct perf_stat_output_ctx *out, |
549 | struct runtime_stat *st) | |
f87027b9 JO |
550 | { |
551 | double total, ratio = 0.0; | |
552 | const char *color; | |
553 | int ctx = evsel_context(evsel); | |
554 | ||
e0128b30 | 555 | total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu); |
f87027b9 JO |
556 | |
557 | if (total) | |
558 | ratio = avg / total * 100.0; | |
559 | ||
560 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); | |
6ca9a082 | 561 | out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio); |
f87027b9 JO |
562 | } |
563 | ||
6ca9a082 JO |
564 | static void print_ll_cache_misses(struct perf_stat_config *config, |
565 | int cpu, | |
32dcd021 | 566 | struct evsel *evsel, |
140aeadc | 567 | double avg, |
e0128b30 JY |
568 | struct perf_stat_output_ctx *out, |
569 | struct runtime_stat *st) | |
f87027b9 JO |
570 | { |
571 | double total, ratio = 0.0; | |
572 | const char *color; | |
573 | int ctx = evsel_context(evsel); | |
574 | ||
e0128b30 | 575 | total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu); |
f87027b9 JO |
576 | |
577 | if (total) | |
578 | ratio = avg / total * 100.0; | |
579 | ||
580 | color = get_ratio_color(GRC_CACHE_MISSES, ratio); | |
6ca9a082 | 581 | out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio); |
f87027b9 JO |
582 | } |
583 | ||
239bd47f AK |
584 | /* |
585 | * High level "TopDown" CPU core pipe line bottleneck break down. | |
586 | * | |
587 | * Basic concept following | |
588 | * Yasin, A Top Down Method for Performance analysis and Counter architecture | |
589 | * ISPASS14 | |
590 | * | |
591 | * The CPU pipeline is divided into 4 areas that can be bottlenecks: | |
592 | * | |
593 | * Frontend -> Backend -> Retiring | |
594 | * BadSpeculation in addition means out of order execution that is thrown away | |
595 | * (for example branch mispredictions) | |
596 | * Frontend is instruction decoding. | |
597 | * Backend is execution, like computation and accessing data in memory | |
598 | * Retiring is good execution that is not directly bottlenecked | |
599 | * | |
600 | * The formulas are computed in slots. | |
601 | * A slot is an entry in the pipeline each for the pipeline width | |
602 | * (for example a 4-wide pipeline has 4 slots for each cycle) | |
603 | * | |
604 | * Formulas: | |
605 | * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) / | |
606 | * TotalSlots | |
607 | * Retiring = SlotsRetired / TotalSlots | |
608 | * FrontendBound = FetchBubbles / TotalSlots | |
609 | * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound | |
610 | * | |
611 | * The kernel provides the mapping to the low level CPU events and any scaling | |
612 | * needed for the CPU pipeline width, for example: | |
613 | * | |
614 | * TotalSlots = Cycles * 4 | |
615 | * | |
616 | * The scaling factor is communicated in the sysfs unit. | |
617 | * | |
618 | * In some cases the CPU may not be able to measure all the formulas due to | |
619 | * missing events. In this case multiple formulas are combined, as possible. | |
620 | * | |
621 | * Full TopDown supports more levels to sub-divide each area: for example | |
622 | * BackendBound into computing bound and memory bound. For now we only | |
623 | * support Level 1 TopDown. | |
624 | */ | |
625 | ||
626 | static double sanitize_val(double x) | |
627 | { | |
628 | if (x < 0 && x >= -0.02) | |
629 | return 0.0; | |
630 | return x; | |
631 | } | |
632 | ||
e0128b30 | 633 | static double td_total_slots(int ctx, int cpu, struct runtime_stat *st) |
239bd47f | 634 | { |
e0128b30 | 635 | return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu); |
239bd47f AK |
636 | } |
637 | ||
e0128b30 | 638 | static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st) |
239bd47f AK |
639 | { |
640 | double bad_spec = 0; | |
641 | double total_slots; | |
642 | double total; | |
643 | ||
e0128b30 JY |
644 | total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) - |
645 | runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) + | |
646 | runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu); | |
647 | ||
648 | total_slots = td_total_slots(ctx, cpu, st); | |
239bd47f AK |
649 | if (total_slots) |
650 | bad_spec = total / total_slots; | |
651 | return sanitize_val(bad_spec); | |
652 | } | |
653 | ||
e0128b30 | 654 | static double td_retiring(int ctx, int cpu, struct runtime_stat *st) |
239bd47f AK |
655 | { |
656 | double retiring = 0; | |
e0128b30 JY |
657 | double total_slots = td_total_slots(ctx, cpu, st); |
658 | double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, | |
659 | ctx, cpu); | |
239bd47f AK |
660 | |
661 | if (total_slots) | |
662 | retiring = ret_slots / total_slots; | |
663 | return retiring; | |
664 | } | |
665 | ||
e0128b30 | 666 | static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st) |
239bd47f AK |
667 | { |
668 | double fe_bound = 0; | |
e0128b30 JY |
669 | double total_slots = td_total_slots(ctx, cpu, st); |
670 | double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES, | |
671 | ctx, cpu); | |
239bd47f AK |
672 | |
673 | if (total_slots) | |
674 | fe_bound = fetch_bub / total_slots; | |
675 | return fe_bound; | |
676 | } | |
677 | ||
e0128b30 | 678 | static double td_be_bound(int ctx, int cpu, struct runtime_stat *st) |
239bd47f | 679 | { |
e0128b30 JY |
680 | double sum = (td_fe_bound(ctx, cpu, st) + |
681 | td_bad_spec(ctx, cpu, st) + | |
682 | td_retiring(ctx, cpu, st)); | |
239bd47f AK |
683 | if (sum == 0) |
684 | return 0; | |
685 | return sanitize_val(1.0 - sum); | |
686 | } | |
687 | ||
6ca9a082 | 688 | static void print_smi_cost(struct perf_stat_config *config, |
32dcd021 | 689 | int cpu, struct evsel *evsel, |
e0128b30 JY |
690 | struct perf_stat_output_ctx *out, |
691 | struct runtime_stat *st) | |
daefd0bc KL |
692 | { |
693 | double smi_num, aperf, cycles, cost = 0.0; | |
694 | int ctx = evsel_context(evsel); | |
695 | const char *color = NULL; | |
696 | ||
e0128b30 JY |
697 | smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu); |
698 | aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu); | |
699 | cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu); | |
daefd0bc KL |
700 | |
701 | if ((cycles == 0) || (aperf == 0)) | |
702 | return; | |
703 | ||
704 | if (smi_num) | |
705 | cost = (aperf - cycles) / aperf * 100.00; | |
706 | ||
707 | if (cost > 10) | |
708 | color = PERF_COLOR_RED; | |
6ca9a082 JO |
709 | out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost); |
710 | out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num); | |
daefd0bc KL |
711 | } |
712 | ||
6ca9a082 JO |
713 | static void generic_metric(struct perf_stat_config *config, |
714 | const char *metric_expr, | |
32dcd021 | 715 | struct evsel **metric_events, |
bba49af8 AK |
716 | char *name, |
717 | const char *metric_name, | |
287f2649 | 718 | const char *metric_unit, |
bba49af8 AK |
719 | double avg, |
720 | int cpu, | |
e0128b30 JY |
721 | struct perf_stat_output_ctx *out, |
722 | struct runtime_stat *st) | |
bba49af8 AK |
723 | { |
724 | print_metric_t print_metric = out->print_metric; | |
725 | struct parse_ctx pctx; | |
287f2649 | 726 | double ratio, scale; |
bba49af8 AK |
727 | int i; |
728 | void *ctxp = out->ctx; | |
e3a94273 | 729 | char *n, *pn; |
bba49af8 AK |
730 | |
731 | expr__ctx_init(&pctx); | |
732 | expr__add_id(&pctx, name, avg); | |
733 | for (i = 0; metric_events[i]; i++) { | |
734 | struct saved_value *v; | |
fd48aad9 | 735 | struct stats *stats; |
fd48aad9 AK |
736 | |
737 | if (!strcmp(metric_events[i]->name, "duration_time")) { | |
738 | stats = &walltime_nsecs_stats; | |
739 | scale = 1e-9; | |
740 | } else { | |
1fcd0394 | 741 | v = saved_value_lookup(metric_events[i], cpu, false, |
e0128b30 | 742 | STAT_NONE, 0, st); |
fd48aad9 AK |
743 | if (!v) |
744 | break; | |
745 | stats = &v->stats; | |
746 | scale = 1.0; | |
747 | } | |
e3a94273 AK |
748 | |
749 | n = strdup(metric_events[i]->name); | |
750 | if (!n) | |
751 | return; | |
752 | /* | |
753 | * This display code with --no-merge adds [cpu] postfixes. | |
754 | * These are not supported by the parser. Remove everything | |
755 | * after the space. | |
756 | */ | |
757 | pn = strchr(n, ' '); | |
758 | if (pn) | |
759 | *pn = 0; | |
760 | expr__add_id(&pctx, n, avg_stats(stats)*scale); | |
bba49af8 AK |
761 | } |
762 | if (!metric_events[i]) { | |
763 | const char *p = metric_expr; | |
764 | ||
287f2649 JY |
765 | if (expr__parse(&ratio, &pctx, &p) == 0) { |
766 | char *unit; | |
767 | char metric_bf[64]; | |
768 | ||
769 | if (metric_unit && metric_name) { | |
770 | if (perf_pmu__convert_scale(metric_unit, | |
771 | &unit, &scale) >= 0) { | |
772 | ratio *= scale; | |
773 | } | |
774 | ||
775 | scnprintf(metric_bf, sizeof(metric_bf), | |
776 | "%s %s", unit, metric_name); | |
777 | print_metric(config, ctxp, NULL, "%8.1f", | |
778 | metric_bf, ratio); | |
779 | } else { | |
780 | print_metric(config, ctxp, NULL, "%8.1f", | |
781 | metric_name ? | |
782 | metric_name : | |
783 | out->force_header ? name : "", | |
784 | ratio); | |
785 | } | |
786 | } else { | |
6ca9a082 | 787 | print_metric(config, ctxp, NULL, NULL, |
4ed962eb AK |
788 | out->force_header ? |
789 | (metric_name ? metric_name : name) : "", 0); | |
287f2649 | 790 | } |
bba49af8 | 791 | } else |
6ca9a082 | 792 | print_metric(config, ctxp, NULL, NULL, "", 0); |
e3a94273 AK |
793 | |
794 | for (i = 1; i < pctx.num_ids; i++) | |
d8f9da24 | 795 | zfree(&pctx.ids[i].name); |
bba49af8 AK |
796 | } |
797 | ||
6ca9a082 | 798 | void perf_stat__print_shadow_stats(struct perf_stat_config *config, |
32dcd021 | 799 | struct evsel *evsel, |
140aeadc | 800 | double avg, int cpu, |
b18f3e36 | 801 | struct perf_stat_output_ctx *out, |
e0128b30 JY |
802 | struct rblist *metric_events, |
803 | struct runtime_stat *st) | |
f87027b9 | 804 | { |
140aeadc AK |
805 | void *ctxp = out->ctx; |
806 | print_metric_t print_metric = out->print_metric; | |
f87027b9 | 807 | double total, ratio = 0.0, total2; |
239bd47f | 808 | const char *color = NULL; |
f87027b9 | 809 | int ctx = evsel_context(evsel); |
b18f3e36 AK |
810 | struct metric_event *me; |
811 | int num = 1; | |
f87027b9 JO |
812 | |
813 | if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) { | |
e0128b30 JY |
814 | total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu); |
815 | ||
f87027b9 JO |
816 | if (total) { |
817 | ratio = avg / total; | |
6ca9a082 | 818 | print_metric(config, ctxp, NULL, "%7.2f ", |
140aeadc | 819 | "insn per cycle", ratio); |
f87027b9 | 820 | } else { |
6ca9a082 | 821 | print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0); |
f87027b9 | 822 | } |
e0128b30 JY |
823 | |
824 | total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT, | |
825 | ctx, cpu); | |
826 | ||
827 | total = max(total, runtime_stat_avg(st, | |
828 | STAT_STALLED_CYCLES_BACK, | |
829 | ctx, cpu)); | |
f87027b9 JO |
830 | |
831 | if (total && avg) { | |
6ca9a082 | 832 | out->new_line(config, ctxp); |
f87027b9 | 833 | ratio = total / avg; |
6ca9a082 | 834 | print_metric(config, ctxp, NULL, "%7.2f ", |
140aeadc AK |
835 | "stalled cycles per insn", |
836 | ratio); | |
fb4605ba | 837 | } else if (have_frontend_stalled) { |
146540fb CW |
838 | out->new_line(config, ctxp); |
839 | print_metric(config, ctxp, NULL, "%7.2f ", | |
140aeadc | 840 | "stalled cycles per insn", 0); |
f87027b9 | 841 | } |
140aeadc | 842 | } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) { |
e0128b30 | 843 | if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0) |
6ca9a082 | 844 | print_branch_misses(config, cpu, evsel, avg, out, st); |
140aeadc | 845 | else |
6ca9a082 | 846 | print_metric(config, ctxp, NULL, NULL, "of all branches", 0); |
f87027b9 | 847 | } else if ( |
1fc632ce JO |
848 | evsel->core.attr.type == PERF_TYPE_HW_CACHE && |
849 | evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1D | | |
f87027b9 | 850 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | |
140aeadc | 851 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { |
e0128b30 JY |
852 | |
853 | if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0) | |
6ca9a082 | 854 | print_l1_dcache_misses(config, cpu, evsel, avg, out, st); |
140aeadc | 855 | else |
6ca9a082 | 856 | print_metric(config, ctxp, NULL, NULL, "of all L1-dcache hits", 0); |
f87027b9 | 857 | } else if ( |
1fc632ce JO |
858 | evsel->core.attr.type == PERF_TYPE_HW_CACHE && |
859 | evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1I | | |
f87027b9 | 860 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | |
140aeadc | 861 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { |
e0128b30 JY |
862 | |
863 | if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0) | |
6ca9a082 | 864 | print_l1_icache_misses(config, cpu, evsel, avg, out, st); |
140aeadc | 865 | else |
6ca9a082 | 866 | print_metric(config, ctxp, NULL, NULL, "of all L1-icache hits", 0); |
f87027b9 | 867 | } else if ( |
1fc632ce JO |
868 | evsel->core.attr.type == PERF_TYPE_HW_CACHE && |
869 | evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_DTLB | | |
f87027b9 | 870 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | |
140aeadc | 871 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { |
e0128b30 JY |
872 | |
873 | if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0) | |
6ca9a082 | 874 | print_dtlb_cache_misses(config, cpu, evsel, avg, out, st); |
140aeadc | 875 | else |
6ca9a082 | 876 | print_metric(config, ctxp, NULL, NULL, "of all dTLB cache hits", 0); |
f87027b9 | 877 | } else if ( |
1fc632ce JO |
878 | evsel->core.attr.type == PERF_TYPE_HW_CACHE && |
879 | evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_ITLB | | |
f87027b9 | 880 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | |
140aeadc | 881 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { |
e0128b30 JY |
882 | |
883 | if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0) | |
6ca9a082 | 884 | print_itlb_cache_misses(config, cpu, evsel, avg, out, st); |
140aeadc | 885 | else |
6ca9a082 | 886 | print_metric(config, ctxp, NULL, NULL, "of all iTLB cache hits", 0); |
f87027b9 | 887 | } else if ( |
1fc632ce JO |
888 | evsel->core.attr.type == PERF_TYPE_HW_CACHE && |
889 | evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_LL | | |
f87027b9 | 890 | ((PERF_COUNT_HW_CACHE_OP_READ) << 8) | |
140aeadc | 891 | ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) { |
e0128b30 JY |
892 | |
893 | if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0) | |
6ca9a082 | 894 | print_ll_cache_misses(config, cpu, evsel, avg, out, st); |
140aeadc | 895 | else |
6ca9a082 | 896 | print_metric(config, ctxp, NULL, NULL, "of all LL-cache hits", 0); |
140aeadc | 897 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) { |
e0128b30 | 898 | total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu); |
f87027b9 JO |
899 | |
900 | if (total) | |
901 | ratio = avg * 100 / total; | |
902 | ||
e0128b30 | 903 | if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0) |
6ca9a082 | 904 | print_metric(config, ctxp, NULL, "%8.3f %%", |
140aeadc AK |
905 | "of all cache refs", ratio); |
906 | else | |
6ca9a082 | 907 | print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0); |
f87027b9 | 908 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) { |
6ca9a082 | 909 | print_stalled_cycles_frontend(config, cpu, evsel, avg, out, st); |
f87027b9 | 910 | } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) { |
6ca9a082 | 911 | print_stalled_cycles_backend(config, cpu, evsel, avg, out, st); |
f87027b9 | 912 | } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { |
e0128b30 | 913 | total = runtime_stat_avg(st, STAT_NSECS, 0, cpu); |
f87027b9 JO |
914 | |
915 | if (total) { | |
916 | ratio = avg / total; | |
6ca9a082 | 917 | print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio); |
f87027b9 | 918 | } else { |
6ca9a082 | 919 | print_metric(config, ctxp, NULL, NULL, "Ghz", 0); |
f87027b9 JO |
920 | } |
921 | } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) { | |
e0128b30 JY |
922 | total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu); |
923 | ||
f87027b9 | 924 | if (total) |
6ca9a082 | 925 | print_metric(config, ctxp, NULL, |
140aeadc AK |
926 | "%7.2f%%", "transactional cycles", |
927 | 100.0 * (avg / total)); | |
928 | else | |
6ca9a082 | 929 | print_metric(config, ctxp, NULL, NULL, "transactional cycles", |
140aeadc | 930 | 0); |
f87027b9 | 931 | } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) { |
e0128b30 JY |
932 | total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu); |
933 | total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu); | |
934 | ||
f87027b9 JO |
935 | if (total2 < avg) |
936 | total2 = avg; | |
937 | if (total) | |
6ca9a082 | 938 | print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles", |
f87027b9 | 939 | 100.0 * ((total2-avg) / total)); |
140aeadc | 940 | else |
6ca9a082 | 941 | print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0); |
140aeadc | 942 | } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) { |
e0128b30 JY |
943 | total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, |
944 | ctx, cpu); | |
f87027b9 | 945 | |
54976285 | 946 | if (avg) |
f87027b9 JO |
947 | ratio = total / avg; |
948 | ||
e0128b30 | 949 | if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0) |
6ca9a082 | 950 | print_metric(config, ctxp, NULL, "%8.0f", |
140aeadc AK |
951 | "cycles / transaction", ratio); |
952 | else | |
6ca9a082 | 953 | print_metric(config, ctxp, NULL, NULL, "cycles / transaction", |
e0128b30 | 954 | 0); |
140aeadc | 955 | } else if (perf_stat_evsel__is(evsel, ELISION_START)) { |
e0128b30 JY |
956 | total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, |
957 | ctx, cpu); | |
f87027b9 | 958 | |
54976285 | 959 | if (avg) |
f87027b9 JO |
960 | ratio = total / avg; |
961 | ||
6ca9a082 | 962 | print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio); |
0aa802a7 | 963 | } else if (perf_evsel__is_clock(evsel)) { |
4579ecc8 | 964 | if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0) |
6ca9a082 | 965 | print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized", |
0aa802a7 | 966 | avg / (ratio * evsel->scale)); |
4579ecc8 | 967 | else |
6ca9a082 | 968 | print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0); |
239bd47f | 969 | } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) { |
e0128b30 | 970 | double fe_bound = td_fe_bound(ctx, cpu, st); |
239bd47f AK |
971 | |
972 | if (fe_bound > 0.2) | |
973 | color = PERF_COLOR_RED; | |
6ca9a082 | 974 | print_metric(config, ctxp, color, "%8.1f%%", "frontend bound", |
239bd47f AK |
975 | fe_bound * 100.); |
976 | } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) { | |
e0128b30 | 977 | double retiring = td_retiring(ctx, cpu, st); |
239bd47f AK |
978 | |
979 | if (retiring > 0.7) | |
980 | color = PERF_COLOR_GREEN; | |
6ca9a082 | 981 | print_metric(config, ctxp, color, "%8.1f%%", "retiring", |
239bd47f AK |
982 | retiring * 100.); |
983 | } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) { | |
e0128b30 | 984 | double bad_spec = td_bad_spec(ctx, cpu, st); |
239bd47f AK |
985 | |
986 | if (bad_spec > 0.1) | |
987 | color = PERF_COLOR_RED; | |
6ca9a082 | 988 | print_metric(config, ctxp, color, "%8.1f%%", "bad speculation", |
239bd47f AK |
989 | bad_spec * 100.); |
990 | } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) { | |
e0128b30 | 991 | double be_bound = td_be_bound(ctx, cpu, st); |
239bd47f AK |
992 | const char *name = "backend bound"; |
993 | static int have_recovery_bubbles = -1; | |
994 | ||
995 | /* In case the CPU does not support topdown-recovery-bubbles */ | |
996 | if (have_recovery_bubbles < 0) | |
997 | have_recovery_bubbles = pmu_have_event("cpu", | |
998 | "topdown-recovery-bubbles"); | |
999 | if (!have_recovery_bubbles) | |
1000 | name = "backend bound/bad spec"; | |
1001 | ||
1002 | if (be_bound > 0.2) | |
1003 | color = PERF_COLOR_RED; | |
e0128b30 | 1004 | if (td_total_slots(ctx, cpu, st) > 0) |
6ca9a082 | 1005 | print_metric(config, ctxp, color, "%8.1f%%", name, |
239bd47f AK |
1006 | be_bound * 100.); |
1007 | else | |
6ca9a082 | 1008 | print_metric(config, ctxp, NULL, NULL, name, 0); |
37932c18 | 1009 | } else if (evsel->metric_expr) { |
6ca9a082 | 1010 | generic_metric(config, evsel->metric_expr, evsel->metric_events, evsel->name, |
287f2649 | 1011 | evsel->metric_name, NULL, avg, cpu, out, st); |
e0128b30 | 1012 | } else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) { |
f87027b9 | 1013 | char unit = 'M'; |
140aeadc | 1014 | char unit_buf[10]; |
f87027b9 | 1015 | |
e0128b30 | 1016 | total = runtime_stat_avg(st, STAT_NSECS, 0, cpu); |
f87027b9 JO |
1017 | |
1018 | if (total) | |
1019 | ratio = 1000.0 * avg / total; | |
1020 | if (ratio < 0.001) { | |
1021 | ratio *= 1000; | |
1022 | unit = 'K'; | |
1023 | } | |
140aeadc | 1024 | snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit); |
6ca9a082 | 1025 | print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio); |
daefd0bc | 1026 | } else if (perf_stat_evsel__is(evsel, SMI_NUM)) { |
6ca9a082 | 1027 | print_smi_cost(config, cpu, evsel, out, st); |
f87027b9 | 1028 | } else { |
b18f3e36 | 1029 | num = 0; |
f87027b9 | 1030 | } |
b18f3e36 AK |
1031 | |
1032 | if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) { | |
1033 | struct metric_expr *mexp; | |
1034 | ||
1035 | list_for_each_entry (mexp, &me->head, nd) { | |
1036 | if (num++ > 0) | |
6ca9a082 JO |
1037 | out->new_line(config, ctxp); |
1038 | generic_metric(config, mexp->metric_expr, mexp->metric_events, | |
b18f3e36 | 1039 | evsel->name, mexp->metric_name, |
287f2649 | 1040 | mexp->metric_unit, avg, cpu, out, st); |
b18f3e36 AK |
1041 | } |
1042 | } | |
1043 | if (num == 0) | |
6ca9a082 | 1044 | print_metric(config, ctxp, NULL, NULL, NULL, 0); |
f87027b9 | 1045 | } |