]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
a43783ae | 2 | #include <errno.h> |
fd20e811 | 3 | #include <inttypes.h> |
0007ecea | 4 | #include <math.h> |
0007ecea | 5 | #include "stat.h" |
24e34f68 | 6 | #include "evlist.h" |
e2f56da1 | 7 | #include "evsel.h" |
24e34f68 | 8 | #include "thread_map.h" |
0007ecea XG |
9 | |
10 | void update_stats(struct stats *stats, u64 val) | |
11 | { | |
12 | double delta; | |
13 | ||
14 | stats->n++; | |
15 | delta = val - stats->mean; | |
16 | stats->mean += delta / stats->n; | |
17 | stats->M2 += delta*(val - stats->mean); | |
ffe4f3c0 DA |
18 | |
19 | if (val > stats->max) | |
20 | stats->max = val; | |
21 | ||
22 | if (val < stats->min) | |
23 | stats->min = val; | |
0007ecea XG |
24 | } |
25 | ||
26 | double avg_stats(struct stats *stats) | |
27 | { | |
28 | return stats->mean; | |
29 | } | |
30 | ||
31 | /* | |
32 | * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance | |
33 | * | |
34 | * (\Sum n_i^2) - ((\Sum n_i)^2)/n | |
35 | * s^2 = ------------------------------- | |
36 | * n - 1 | |
37 | * | |
38 | * http://en.wikipedia.org/wiki/Stddev | |
39 | * | |
40 | * The std dev of the mean is related to the std dev by: | |
41 | * | |
42 | * s | |
43 | * s_mean = ------- | |
44 | * sqrt(n) | |
45 | * | |
46 | */ | |
47 | double stddev_stats(struct stats *stats) | |
48 | { | |
49 | double variance, variance_mean; | |
50 | ||
45528f7c | 51 | if (stats->n < 2) |
0007ecea XG |
52 | return 0.0; |
53 | ||
54 | variance = stats->M2 / (stats->n - 1); | |
55 | variance_mean = variance / stats->n; | |
56 | ||
57 | return sqrt(variance_mean); | |
58 | } | |
59 | ||
60 | double rel_stddev_stats(double stddev, double avg) | |
61 | { | |
62 | double pct = 0.0; | |
63 | ||
64 | if (avg) | |
65 | pct = 100.0 * stddev/avg; | |
66 | ||
67 | return pct; | |
68 | } | |
e2f56da1 JO |
69 | |
70 | bool __perf_evsel_stat__is(struct perf_evsel *evsel, | |
71 | enum perf_stat_evsel_id id) | |
72 | { | |
581cc8a2 | 73 | struct perf_stat_evsel *ps = evsel->priv; |
e2f56da1 JO |
74 | |
75 | return ps->id == id; | |
76 | } | |
77 | ||
78 | #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name | |
79 | static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = { | |
4c358d5c JO |
80 | ID(NONE, x), |
81 | ID(CYCLES_IN_TX, cpu/cycles-t/), | |
82 | ID(TRANSACTION_START, cpu/tx-start/), | |
83 | ID(ELISION_START, cpu/el-start/), | |
84 | ID(CYCLES_IN_TX_CP, cpu/cycles-ct/), | |
239bd47f AK |
85 | ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots), |
86 | ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued), | |
87 | ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired), | |
88 | ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles), | |
89 | ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles), | |
daefd0bc KL |
90 | ID(SMI_NUM, msr/smi/), |
91 | ID(APERF, msr/aperf/), | |
e2f56da1 JO |
92 | }; |
93 | #undef ID | |
94 | ||
95 | void perf_stat_evsel_id_init(struct perf_evsel *evsel) | |
96 | { | |
581cc8a2 | 97 | struct perf_stat_evsel *ps = evsel->priv; |
e2f56da1 JO |
98 | int i; |
99 | ||
100 | /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */ | |
101 | ||
102 | for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) { | |
103 | if (!strcmp(perf_evsel__name(evsel), id_str[i])) { | |
104 | ps->id = i; | |
105 | break; | |
106 | } | |
107 | } | |
108 | } | |
a9a3a4d9 | 109 | |
86a2cf31 | 110 | static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel) |
9689edfa JO |
111 | { |
112 | int i; | |
581cc8a2 | 113 | struct perf_stat_evsel *ps = evsel->priv; |
9689edfa JO |
114 | |
115 | for (i = 0; i < 3; i++) | |
116 | init_stats(&ps->res_stats[i]); | |
117 | ||
118 | perf_stat_evsel_id_init(evsel); | |
119 | } | |
120 | ||
86a2cf31 | 121 | static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel) |
9689edfa | 122 | { |
581cc8a2 | 123 | evsel->priv = zalloc(sizeof(struct perf_stat_evsel)); |
9689edfa JO |
124 | if (evsel->priv == NULL) |
125 | return -ENOMEM; | |
126 | perf_evsel__reset_stat_priv(evsel); | |
127 | return 0; | |
128 | } | |
129 | ||
86a2cf31 | 130 | static void perf_evsel__free_stat_priv(struct perf_evsel *evsel) |
9689edfa | 131 | { |
f7794d52 JO |
132 | struct perf_stat_evsel *ps = evsel->priv; |
133 | ||
134 | if (ps) | |
135 | free(ps->group_data); | |
9689edfa JO |
136 | zfree(&evsel->priv); |
137 | } | |
a939512d | 138 | |
86a2cf31 JO |
139 | static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel, |
140 | int ncpus, int nthreads) | |
a939512d JO |
141 | { |
142 | struct perf_counts *counts; | |
143 | ||
144 | counts = perf_counts__new(ncpus, nthreads); | |
145 | if (counts) | |
146 | evsel->prev_raw_counts = counts; | |
147 | ||
148 | return counts ? 0 : -ENOMEM; | |
149 | } | |
150 | ||
86a2cf31 | 151 | static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel) |
a939512d JO |
152 | { |
153 | perf_counts__delete(evsel->prev_raw_counts); | |
154 | evsel->prev_raw_counts = NULL; | |
155 | } | |
24e34f68 | 156 | |
86a2cf31 | 157 | static int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw) |
a7d0a102 JO |
158 | { |
159 | int ncpus = perf_evsel__nr_cpus(evsel); | |
160 | int nthreads = thread_map__nr(evsel->threads); | |
161 | ||
162 | if (perf_evsel__alloc_stat_priv(evsel) < 0 || | |
163 | perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 || | |
164 | (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0)) | |
165 | return -ENOMEM; | |
166 | ||
167 | return 0; | |
168 | } | |
169 | ||
24e34f68 JO |
170 | int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw) |
171 | { | |
172 | struct perf_evsel *evsel; | |
24e34f68 | 173 | |
e5cadb93 | 174 | evlist__for_each_entry(evlist, evsel) { |
a7d0a102 | 175 | if (perf_evsel__alloc_stats(evsel, alloc_raw)) |
24e34f68 JO |
176 | goto out_free; |
177 | } | |
178 | ||
179 | return 0; | |
180 | ||
181 | out_free: | |
182 | perf_evlist__free_stats(evlist); | |
183 | return -1; | |
184 | } | |
185 | ||
186 | void perf_evlist__free_stats(struct perf_evlist *evlist) | |
187 | { | |
188 | struct perf_evsel *evsel; | |
189 | ||
e5cadb93 | 190 | evlist__for_each_entry(evlist, evsel) { |
24e34f68 JO |
191 | perf_evsel__free_stat_priv(evsel); |
192 | perf_evsel__free_counts(evsel); | |
193 | perf_evsel__free_prev_raw_counts(evsel); | |
194 | } | |
195 | } | |
196 | ||
197 | void perf_evlist__reset_stats(struct perf_evlist *evlist) | |
198 | { | |
199 | struct perf_evsel *evsel; | |
200 | ||
e5cadb93 | 201 | evlist__for_each_entry(evlist, evsel) { |
24e34f68 JO |
202 | perf_evsel__reset_stat_priv(evsel); |
203 | perf_evsel__reset_counts(evsel); | |
204 | } | |
205 | } | |
f80010eb JO |
206 | |
207 | static void zero_per_pkg(struct perf_evsel *counter) | |
208 | { | |
209 | if (counter->per_pkg_mask) | |
210 | memset(counter->per_pkg_mask, 0, MAX_NR_CPUS); | |
211 | } | |
212 | ||
02d8dabc SE |
213 | static int check_per_pkg(struct perf_evsel *counter, |
214 | struct perf_counts_values *vals, int cpu, bool *skip) | |
f80010eb JO |
215 | { |
216 | unsigned long *mask = counter->per_pkg_mask; | |
217 | struct cpu_map *cpus = perf_evsel__cpus(counter); | |
218 | int s; | |
219 | ||
220 | *skip = false; | |
221 | ||
222 | if (!counter->per_pkg) | |
223 | return 0; | |
224 | ||
225 | if (cpu_map__empty(cpus)) | |
226 | return 0; | |
227 | ||
228 | if (!mask) { | |
229 | mask = zalloc(MAX_NR_CPUS); | |
230 | if (!mask) | |
231 | return -ENOMEM; | |
232 | ||
233 | counter->per_pkg_mask = mask; | |
234 | } | |
235 | ||
02d8dabc SE |
236 | /* |
237 | * we do not consider an event that has not run as a good | |
238 | * instance to mark a package as used (skip=1). Otherwise | |
239 | * we may run into a situation where the first CPU in a package | |
240 | * is not running anything, yet the second is, and this function | |
241 | * would mark the package as used after the first CPU and would | |
242 | * not read the values from the second CPU. | |
243 | */ | |
244 | if (!(vals->run && vals->ena)) | |
245 | return 0; | |
246 | ||
1fe7a300 | 247 | s = cpu_map__get_socket(cpus, cpu, NULL); |
f80010eb JO |
248 | if (s < 0) |
249 | return -1; | |
250 | ||
251 | *skip = test_and_set_bit(s, mask) == 1; | |
252 | return 0; | |
253 | } | |
254 | ||
255 | static int | |
256 | process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel, | |
257 | int cpu, int thread, | |
258 | struct perf_counts_values *count) | |
259 | { | |
260 | struct perf_counts_values *aggr = &evsel->counts->aggr; | |
261 | static struct perf_counts_values zero; | |
262 | bool skip = false; | |
263 | ||
02d8dabc | 264 | if (check_per_pkg(evsel, count, cpu, &skip)) { |
f80010eb JO |
265 | pr_err("failed to read per-pkg counter\n"); |
266 | return -1; | |
267 | } | |
268 | ||
269 | if (skip) | |
270 | count = &zero; | |
271 | ||
272 | switch (config->aggr_mode) { | |
273 | case AGGR_THREAD: | |
274 | case AGGR_CORE: | |
275 | case AGGR_SOCKET: | |
276 | case AGGR_NONE: | |
277 | if (!evsel->snapshot) | |
278 | perf_evsel__compute_deltas(evsel, cpu, thread, count); | |
279 | perf_counts_values__scale(count, config->scale, NULL); | |
280 | if (config->aggr_mode == AGGR_NONE) | |
281 | perf_stat__update_shadow_stats(evsel, count->values, cpu); | |
282 | break; | |
283 | case AGGR_GLOBAL: | |
284 | aggr->val += count->val; | |
285 | if (config->scale) { | |
286 | aggr->ena += count->ena; | |
287 | aggr->run += count->run; | |
288 | } | |
208df99e | 289 | case AGGR_UNSET: |
f80010eb JO |
290 | default: |
291 | break; | |
292 | } | |
293 | ||
294 | return 0; | |
295 | } | |
296 | ||
297 | static int process_counter_maps(struct perf_stat_config *config, | |
298 | struct perf_evsel *counter) | |
299 | { | |
300 | int nthreads = thread_map__nr(counter->threads); | |
301 | int ncpus = perf_evsel__nr_cpus(counter); | |
302 | int cpu, thread; | |
303 | ||
304 | if (counter->system_wide) | |
305 | nthreads = 1; | |
306 | ||
307 | for (thread = 0; thread < nthreads; thread++) { | |
308 | for (cpu = 0; cpu < ncpus; cpu++) { | |
309 | if (process_counter_values(config, counter, cpu, thread, | |
310 | perf_counts(counter->counts, cpu, thread))) | |
311 | return -1; | |
312 | } | |
313 | } | |
314 | ||
315 | return 0; | |
316 | } | |
317 | ||
318 | int perf_stat_process_counter(struct perf_stat_config *config, | |
319 | struct perf_evsel *counter) | |
320 | { | |
321 | struct perf_counts_values *aggr = &counter->counts->aggr; | |
581cc8a2 | 322 | struct perf_stat_evsel *ps = counter->priv; |
f80010eb | 323 | u64 *count = counter->counts->aggr.values; |
f340c5fc | 324 | u64 val; |
f80010eb JO |
325 | int i, ret; |
326 | ||
327 | aggr->val = aggr->ena = aggr->run = 0; | |
f80010eb | 328 | |
51fd2df1 JO |
329 | /* |
330 | * We calculate counter's data every interval, | |
331 | * and the display code shows ps->res_stats | |
332 | * avg value. We need to zero the stats for | |
333 | * interval mode, otherwise overall avg running | |
334 | * averages will be shown for each interval. | |
335 | */ | |
336 | if (config->interval) | |
337 | init_stats(ps->res_stats); | |
338 | ||
f80010eb JO |
339 | if (counter->per_pkg) |
340 | zero_per_pkg(counter); | |
341 | ||
342 | ret = process_counter_maps(config, counter); | |
343 | if (ret) | |
344 | return ret; | |
345 | ||
346 | if (config->aggr_mode != AGGR_GLOBAL) | |
347 | return 0; | |
348 | ||
349 | if (!counter->snapshot) | |
350 | perf_evsel__compute_deltas(counter, -1, -1, aggr); | |
351 | perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled); | |
352 | ||
353 | for (i = 0; i < 3; i++) | |
354 | update_stats(&ps->res_stats[i], count[i]); | |
355 | ||
bb963e16 | 356 | if (verbose > 0) { |
f80010eb JO |
357 | fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n", |
358 | perf_evsel__name(counter), count[0], count[1], count[2]); | |
359 | } | |
360 | ||
361 | /* | |
362 | * Save the full runtime - to allow normalization during printout: | |
363 | */ | |
f340c5fc AK |
364 | val = counter->scale * *count; |
365 | perf_stat__update_shadow_stats(counter, &val, 0); | |
f80010eb JO |
366 | |
367 | return 0; | |
368 | } | |
0ea0e355 JO |
369 | |
370 | int perf_event__process_stat_event(struct perf_tool *tool __maybe_unused, | |
371 | union perf_event *event, | |
372 | struct perf_session *session) | |
373 | { | |
374 | struct perf_counts_values count; | |
375 | struct stat_event *st = &event->stat; | |
376 | struct perf_evsel *counter; | |
377 | ||
378 | count.val = st->val; | |
379 | count.ena = st->ena; | |
380 | count.run = st->run; | |
381 | ||
382 | counter = perf_evlist__id2evsel(session->evlist, st->id); | |
383 | if (!counter) { | |
384 | pr_err("Failed to resolve counter for stat event.\n"); | |
385 | return -EINVAL; | |
386 | } | |
387 | ||
388 | *perf_counts(counter->counts, st->cpu, st->thread) = count; | |
389 | counter->supported = true; | |
390 | return 0; | |
391 | } | |
e08a4564 JO |
392 | |
393 | size_t perf_event__fprintf_stat(union perf_event *event, FILE *fp) | |
394 | { | |
395 | struct stat_event *st = (struct stat_event *) event; | |
396 | size_t ret; | |
397 | ||
398 | ret = fprintf(fp, "\n... id %" PRIu64 ", cpu %d, thread %d\n", | |
399 | st->id, st->cpu, st->thread); | |
400 | ret += fprintf(fp, "... value %" PRIu64 ", enabled %" PRIu64 ", running %" PRIu64 "\n", | |
401 | st->val, st->ena, st->run); | |
402 | ||
403 | return ret; | |
404 | } | |
405 | ||
406 | size_t perf_event__fprintf_stat_round(union perf_event *event, FILE *fp) | |
407 | { | |
408 | struct stat_round_event *rd = (struct stat_round_event *)event; | |
409 | size_t ret; | |
410 | ||
411 | ret = fprintf(fp, "\n... time %" PRIu64 ", type %s\n", rd->time, | |
412 | rd->type == PERF_STAT_ROUND_TYPE__FINAL ? "FINAL" : "INTERVAL"); | |
413 | ||
414 | return ret; | |
415 | } | |
416 | ||
417 | size_t perf_event__fprintf_stat_config(union perf_event *event, FILE *fp) | |
418 | { | |
419 | struct perf_stat_config sc; | |
420 | size_t ret; | |
421 | ||
422 | perf_event__read_stat_config(&sc, &event->stat_config); | |
423 | ||
424 | ret = fprintf(fp, "\n"); | |
425 | ret += fprintf(fp, "... aggr_mode %d\n", sc.aggr_mode); | |
426 | ret += fprintf(fp, "... scale %d\n", sc.scale); | |
427 | ret += fprintf(fp, "... interval %u\n", sc.interval); | |
428 | ||
429 | return ret; | |
430 | } |