xref: /openbmc/linux/tools/perf/util/pmus.c (revision c1e0230e)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/list.h>
3 #include <linux/zalloc.h>
4 #include <subcmd/pager.h>
5 #include <sys/types.h>
6 #include <dirent.h>
7 #include <pthread.h>
8 #include <string.h>
9 #include <unistd.h>
10 #include "debug.h"
11 #include "evsel.h"
12 #include "pmus.h"
13 #include "pmu.h"
14 #include "print-events.h"
15 
16 /*
17  * core_pmus:  A PMU belongs to core_pmus if it's name is "cpu" or it's sysfs
18  *             directory contains "cpus" file. All PMUs belonging to core_pmus
19  *             must have pmu->is_core=1. If there are more than one PMU in
20  *             this list, perf interprets it as a heterogeneous platform.
21  *             (FWIW, certain ARM platforms having heterogeneous cores uses
22  *             homogeneous PMU, and thus they are treated as homogeneous
23  *             platform by perf because core_pmus will have only one entry)
24  * other_pmus: All other PMUs which are not part of core_pmus list. It doesn't
25  *             matter whether PMU is present per SMT-thread or outside of the
26  *             core in the hw. For e.g., an instance of AMD ibs_fetch// and
27  *             ibs_op// PMUs is present in each hw SMT thread, however they
28  *             are captured under other_pmus. PMUs belonging to other_pmus
29  *             must have pmu->is_core=0 but pmu->is_uncore could be 0 or 1.
30  */
31 static LIST_HEAD(core_pmus);
32 static LIST_HEAD(other_pmus);
33 static bool read_sysfs_core_pmus;
34 static bool read_sysfs_all_pmus;
35 
36 void perf_pmus__destroy(void)
37 {
38 	struct perf_pmu *pmu, *tmp;
39 
40 	list_for_each_entry_safe(pmu, tmp, &core_pmus, list) {
41 		list_del(&pmu->list);
42 
43 		perf_pmu__delete(pmu);
44 	}
45 	list_for_each_entry_safe(pmu, tmp, &other_pmus, list) {
46 		list_del(&pmu->list);
47 
48 		perf_pmu__delete(pmu);
49 	}
50 	read_sysfs_core_pmus = false;
51 	read_sysfs_all_pmus = false;
52 }
53 
54 static struct perf_pmu *pmu_find(const char *name)
55 {
56 	struct perf_pmu *pmu;
57 
58 	list_for_each_entry(pmu, &core_pmus, list) {
59 		if (!strcmp(pmu->name, name) ||
60 		    (pmu->alias_name && !strcmp(pmu->alias_name, name)))
61 			return pmu;
62 	}
63 	list_for_each_entry(pmu, &other_pmus, list) {
64 		if (!strcmp(pmu->name, name) ||
65 		    (pmu->alias_name && !strcmp(pmu->alias_name, name)))
66 			return pmu;
67 	}
68 
69 	return NULL;
70 }
71 
72 struct perf_pmu *perf_pmus__find(const char *name)
73 {
74 	struct perf_pmu *pmu;
75 	int dirfd;
76 	bool core_pmu;
77 
78 	/*
79 	 * Once PMU is loaded it stays in the list,
80 	 * so we keep us from multiple reading/parsing
81 	 * the pmu format definitions.
82 	 */
83 	pmu = pmu_find(name);
84 	if (pmu)
85 		return pmu;
86 
87 	if (read_sysfs_all_pmus)
88 		return NULL;
89 
90 	core_pmu = is_pmu_core(name);
91 	if (core_pmu && read_sysfs_core_pmus)
92 		return NULL;
93 
94 	dirfd = perf_pmu__event_source_devices_fd();
95 	pmu = perf_pmu__lookup(core_pmu ? &core_pmus : &other_pmus, dirfd, name);
96 	close(dirfd);
97 
98 	return pmu;
99 }
100 
101 static struct perf_pmu *perf_pmu__find2(int dirfd, const char *name)
102 {
103 	struct perf_pmu *pmu;
104 	bool core_pmu;
105 
106 	/*
107 	 * Once PMU is loaded it stays in the list,
108 	 * so we keep us from multiple reading/parsing
109 	 * the pmu format definitions.
110 	 */
111 	pmu = pmu_find(name);
112 	if (pmu)
113 		return pmu;
114 
115 	if (read_sysfs_all_pmus)
116 		return NULL;
117 
118 	core_pmu = is_pmu_core(name);
119 	if (core_pmu && read_sysfs_core_pmus)
120 		return NULL;
121 
122 	return perf_pmu__lookup(core_pmu ? &core_pmus : &other_pmus, dirfd, name);
123 }
124 
125 /* Add all pmus in sysfs to pmu list: */
126 static void pmu_read_sysfs(bool core_only)
127 {
128 	int fd;
129 	DIR *dir;
130 	struct dirent *dent;
131 
132 	if (read_sysfs_all_pmus || (core_only && read_sysfs_core_pmus))
133 		return;
134 
135 	fd = perf_pmu__event_source_devices_fd();
136 	if (fd < 0)
137 		return;
138 
139 	dir = fdopendir(fd);
140 	if (!dir) {
141 		close(fd);
142 		return;
143 	}
144 
145 	while ((dent = readdir(dir))) {
146 		if (!strcmp(dent->d_name, ".") || !strcmp(dent->d_name, ".."))
147 			continue;
148 		if (core_only && !is_pmu_core(dent->d_name))
149 			continue;
150 		/* add to static LIST_HEAD(core_pmus) or LIST_HEAD(other_pmus): */
151 		perf_pmu__find2(fd, dent->d_name);
152 	}
153 
154 	closedir(dir);
155 	if (list_empty(&core_pmus)) {
156 		if (!perf_pmu__create_placeholder_core_pmu(&core_pmus))
157 			pr_err("Failure to set up any core PMUs\n");
158 	}
159 	if (!list_empty(&core_pmus)) {
160 		read_sysfs_core_pmus = true;
161 		if (!core_only)
162 			read_sysfs_all_pmus = true;
163 	}
164 }
165 
166 static struct perf_pmu *__perf_pmus__find_by_type(unsigned int type)
167 {
168 	struct perf_pmu *pmu;
169 
170 	list_for_each_entry(pmu, &core_pmus, list) {
171 		if (pmu->type == type)
172 			return pmu;
173 	}
174 
175 	list_for_each_entry(pmu, &other_pmus, list) {
176 		if (pmu->type == type)
177 			return pmu;
178 	}
179 	return NULL;
180 }
181 
182 struct perf_pmu *perf_pmus__find_by_type(unsigned int type)
183 {
184 	struct perf_pmu *pmu = __perf_pmus__find_by_type(type);
185 
186 	if (pmu || read_sysfs_all_pmus)
187 		return pmu;
188 
189 	pmu_read_sysfs(/*core_only=*/false);
190 	pmu = __perf_pmus__find_by_type(type);
191 	return pmu;
192 }
193 
194 /*
195  * pmu iterator: If pmu is NULL, we start at the begin, otherwise return the
196  * next pmu. Returns NULL on end.
197  */
198 struct perf_pmu *perf_pmus__scan(struct perf_pmu *pmu)
199 {
200 	bool use_core_pmus = !pmu || pmu->is_core;
201 
202 	if (!pmu) {
203 		pmu_read_sysfs(/*core_only=*/false);
204 		pmu = list_prepare_entry(pmu, &core_pmus, list);
205 	}
206 	if (use_core_pmus) {
207 		list_for_each_entry_continue(pmu, &core_pmus, list)
208 			return pmu;
209 
210 		pmu = NULL;
211 		pmu = list_prepare_entry(pmu, &other_pmus, list);
212 	}
213 	list_for_each_entry_continue(pmu, &other_pmus, list)
214 		return pmu;
215 	return NULL;
216 }
217 
218 struct perf_pmu *perf_pmus__scan_core(struct perf_pmu *pmu)
219 {
220 	if (!pmu) {
221 		pmu_read_sysfs(/*core_only=*/true);
222 		pmu = list_prepare_entry(pmu, &core_pmus, list);
223 	}
224 	list_for_each_entry_continue(pmu, &core_pmus, list)
225 		return pmu;
226 
227 	return NULL;
228 }
229 
230 const struct perf_pmu *perf_pmus__pmu_for_pmu_filter(const char *str)
231 {
232 	struct perf_pmu *pmu = NULL;
233 
234 	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
235 		if (!strcmp(pmu->name, str))
236 			return pmu;
237 		/* Ignore "uncore_" prefix. */
238 		if (!strncmp(pmu->name, "uncore_", 7)) {
239 			if (!strcmp(pmu->name + 7, str))
240 				return pmu;
241 		}
242 		/* Ignore "cpu_" prefix on Intel hybrid PMUs. */
243 		if (!strncmp(pmu->name, "cpu_", 4)) {
244 			if (!strcmp(pmu->name + 4, str))
245 				return pmu;
246 		}
247 	}
248 	return NULL;
249 }
250 
251 int __weak perf_pmus__num_mem_pmus(void)
252 {
253 	/* All core PMUs are for mem events. */
254 	return perf_pmus__num_core_pmus();
255 }
256 
257 /** Struct for ordering events as output in perf list. */
258 struct sevent {
259 	/** PMU for event. */
260 	const struct perf_pmu *pmu;
261 	/**
262 	 * Optional event for name, desc, etc. If not present then this is a
263 	 * selectable PMU and the event name is shown as "//".
264 	 */
265 	const struct perf_pmu_alias *event;
266 	/** Is the PMU for the CPU? */
267 	bool is_cpu;
268 };
269 
270 static int cmp_sevent(const void *a, const void *b)
271 {
272 	const struct sevent *as = a;
273 	const struct sevent *bs = b;
274 	const char *a_pmu_name = NULL, *b_pmu_name = NULL;
275 	const char *a_name = "//", *a_desc = NULL, *a_topic = "";
276 	const char *b_name = "//", *b_desc = NULL, *b_topic = "";
277 	int ret;
278 
279 	if (as->event) {
280 		a_name = as->event->name;
281 		a_desc = as->event->desc;
282 		a_topic = as->event->topic ?: "";
283 		a_pmu_name = as->event->pmu_name;
284 	}
285 	if (bs->event) {
286 		b_name = bs->event->name;
287 		b_desc = bs->event->desc;
288 		b_topic = bs->event->topic ?: "";
289 		b_pmu_name = bs->event->pmu_name;
290 	}
291 	/* Put extra events last. */
292 	if (!!a_desc != !!b_desc)
293 		return !!a_desc - !!b_desc;
294 
295 	/* Order by topics. */
296 	ret = strcmp(a_topic, b_topic);
297 	if (ret)
298 		return ret;
299 
300 	/* Order CPU core events to be first */
301 	if (as->is_cpu != bs->is_cpu)
302 		return as->is_cpu ? -1 : 1;
303 
304 	/* Order by PMU name. */
305 	if (as->pmu != bs->pmu) {
306 		a_pmu_name = a_pmu_name ?: (as->pmu->name ?: "");
307 		b_pmu_name = b_pmu_name ?: (bs->pmu->name ?: "");
308 		ret = strcmp(a_pmu_name, b_pmu_name);
309 		if (ret)
310 			return ret;
311 	}
312 
313 	/* Order by event name. */
314 	return strcmp(a_name, b_name);
315 }
316 
317 static bool pmu_alias_is_duplicate(struct sevent *alias_a,
318 				   struct sevent *alias_b)
319 {
320 	const char *a_pmu_name = NULL, *b_pmu_name = NULL;
321 	const char *a_name = "//", *b_name = "//";
322 
323 
324 	if (alias_a->event) {
325 		a_name = alias_a->event->name;
326 		a_pmu_name = alias_a->event->pmu_name;
327 	}
328 	if (alias_b->event) {
329 		b_name = alias_b->event->name;
330 		b_pmu_name = alias_b->event->pmu_name;
331 	}
332 
333 	/* Different names -> never duplicates */
334 	if (strcmp(a_name, b_name))
335 		return false;
336 
337 	/* Don't remove duplicates for different PMUs */
338 	a_pmu_name = a_pmu_name ?: (alias_a->pmu->name ?: "");
339 	b_pmu_name = b_pmu_name ?: (alias_b->pmu->name ?: "");
340 	return strcmp(a_pmu_name, b_pmu_name) == 0;
341 }
342 
343 static int sub_non_neg(int a, int b)
344 {
345 	if (b > a)
346 		return 0;
347 	return a - b;
348 }
349 
350 static char *format_alias(char *buf, int len, const struct perf_pmu *pmu,
351 			  const struct perf_pmu_alias *alias)
352 {
353 	struct parse_events_term *term;
354 	int used = snprintf(buf, len, "%s/%s", pmu->name, alias->name);
355 
356 	list_for_each_entry(term, &alias->terms, list) {
357 		if (term->type_val == PARSE_EVENTS__TERM_TYPE_STR)
358 			used += snprintf(buf + used, sub_non_neg(len, used),
359 					",%s=%s", term->config,
360 					term->val.str);
361 	}
362 
363 	if (sub_non_neg(len, used) > 0) {
364 		buf[used] = '/';
365 		used++;
366 	}
367 	if (sub_non_neg(len, used) > 0) {
368 		buf[used] = '\0';
369 		used++;
370 	} else
371 		buf[len - 1] = '\0';
372 
373 	return buf;
374 }
375 
376 void perf_pmus__print_pmu_events(const struct print_callbacks *print_cb, void *print_state)
377 {
378 	struct perf_pmu *pmu;
379 	struct perf_pmu_alias *event;
380 	char buf[1024];
381 	int printed = 0;
382 	int len, j;
383 	struct sevent *aliases;
384 
385 	pmu = NULL;
386 	len = 0;
387 	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
388 		list_for_each_entry(event, &pmu->aliases, list)
389 			len++;
390 		if (pmu->selectable)
391 			len++;
392 	}
393 	aliases = zalloc(sizeof(struct sevent) * len);
394 	if (!aliases) {
395 		pr_err("FATAL: not enough memory to print PMU events\n");
396 		return;
397 	}
398 	pmu = NULL;
399 	j = 0;
400 	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
401 		bool is_cpu = pmu->is_core;
402 
403 		list_for_each_entry(event, &pmu->aliases, list) {
404 			aliases[j].event = event;
405 			aliases[j].pmu = pmu;
406 			aliases[j].is_cpu = is_cpu;
407 			j++;
408 		}
409 		if (pmu->selectable) {
410 			aliases[j].event = NULL;
411 			aliases[j].pmu = pmu;
412 			aliases[j].is_cpu = is_cpu;
413 			j++;
414 		}
415 	}
416 	len = j;
417 	qsort(aliases, len, sizeof(struct sevent), cmp_sevent);
418 	for (j = 0; j < len; j++) {
419 		const char *name, *alias = NULL, *scale_unit = NULL,
420 			*desc = NULL, *long_desc = NULL,
421 			*encoding_desc = NULL, *topic = NULL,
422 			*pmu_name = NULL;
423 		bool deprecated = false;
424 		size_t buf_used;
425 
426 		/* Skip duplicates */
427 		if (j > 0 && pmu_alias_is_duplicate(&aliases[j], &aliases[j - 1]))
428 			continue;
429 
430 		if (!aliases[j].event) {
431 			/* A selectable event. */
432 			pmu_name = aliases[j].pmu->name;
433 			buf_used = snprintf(buf, sizeof(buf), "%s//", pmu_name) + 1;
434 			name = buf;
435 		} else {
436 			if (aliases[j].event->desc) {
437 				name = aliases[j].event->name;
438 				buf_used = 0;
439 			} else {
440 				name = format_alias(buf, sizeof(buf), aliases[j].pmu,
441 						    aliases[j].event);
442 				if (aliases[j].is_cpu) {
443 					alias = name;
444 					name = aliases[j].event->name;
445 				}
446 				buf_used = strlen(buf) + 1;
447 			}
448 			pmu_name = aliases[j].event->pmu_name ?: (aliases[j].pmu->name ?: "");
449 			if (strlen(aliases[j].event->unit) || aliases[j].event->scale != 1.0) {
450 				scale_unit = buf + buf_used;
451 				buf_used += snprintf(buf + buf_used, sizeof(buf) - buf_used,
452 						"%G%s", aliases[j].event->scale,
453 						aliases[j].event->unit) + 1;
454 			}
455 			desc = aliases[j].event->desc;
456 			long_desc = aliases[j].event->long_desc;
457 			topic = aliases[j].event->topic;
458 			encoding_desc = buf + buf_used;
459 			buf_used += snprintf(buf + buf_used, sizeof(buf) - buf_used,
460 					"%s/%s/", pmu_name, aliases[j].event->str) + 1;
461 			deprecated = aliases[j].event->deprecated;
462 		}
463 		print_cb->print_event(print_state,
464 				pmu_name,
465 				topic,
466 				name,
467 				alias,
468 				scale_unit,
469 				deprecated,
470 				"Kernel PMU event",
471 				desc,
472 				long_desc,
473 				encoding_desc);
474 	}
475 	if (printed && pager_in_use())
476 		printf("\n");
477 
478 	zfree(&aliases);
479 }
480 
481 bool perf_pmus__have_event(const char *pname, const char *name)
482 {
483 	struct perf_pmu *pmu = perf_pmus__find(pname);
484 
485 	return pmu && perf_pmu__have_event(pmu, name);
486 }
487 
488 int perf_pmus__num_core_pmus(void)
489 {
490 	static int count;
491 
492 	if (!count) {
493 		struct perf_pmu *pmu = NULL;
494 
495 		while ((pmu = perf_pmus__scan_core(pmu)) != NULL)
496 			count++;
497 	}
498 	return count;
499 }
500 
501 static bool __perf_pmus__supports_extended_type(void)
502 {
503 	struct perf_pmu *pmu = NULL;
504 
505 	if (perf_pmus__num_core_pmus() <= 1)
506 		return false;
507 
508 	while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
509 		if (!is_event_supported(PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES | ((__u64)pmu->type << PERF_PMU_TYPE_SHIFT)))
510 			return false;
511 	}
512 
513 	return true;
514 }
515 
516 static bool perf_pmus__do_support_extended_type;
517 
518 static void perf_pmus__init_supports_extended_type(void)
519 {
520 	perf_pmus__do_support_extended_type = __perf_pmus__supports_extended_type();
521 }
522 
523 bool perf_pmus__supports_extended_type(void)
524 {
525 	static pthread_once_t extended_type_once = PTHREAD_ONCE_INIT;
526 
527 	pthread_once(&extended_type_once, perf_pmus__init_supports_extended_type);
528 
529 	return perf_pmus__do_support_extended_type;
530 }
531 
532 char *perf_pmus__default_pmu_name(void)
533 {
534 	int fd;
535 	DIR *dir;
536 	struct dirent *dent;
537 	char *result = NULL;
538 
539 	if (!list_empty(&core_pmus))
540 		return strdup(list_first_entry(&core_pmus, struct perf_pmu, list)->name);
541 
542 	fd = perf_pmu__event_source_devices_fd();
543 	if (fd < 0)
544 		return strdup("cpu");
545 
546 	dir = fdopendir(fd);
547 	if (!dir) {
548 		close(fd);
549 		return strdup("cpu");
550 	}
551 
552 	while ((dent = readdir(dir))) {
553 		if (!strcmp(dent->d_name, ".") || !strcmp(dent->d_name, ".."))
554 			continue;
555 		if (is_pmu_core(dent->d_name)) {
556 			result = strdup(dent->d_name);
557 			break;
558 		}
559 	}
560 
561 	closedir(dir);
562 	return result ?: strdup("cpu");
563 }
564 
565 struct perf_pmu *evsel__find_pmu(const struct evsel *evsel)
566 {
567 	struct perf_pmu *pmu = evsel->pmu;
568 
569 	if (!pmu) {
570 		pmu = perf_pmus__find_by_type(evsel->core.attr.type);
571 		((struct evsel *)evsel)->pmu = pmu;
572 	}
573 	return pmu;
574 }
575