1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/list.h> 3 #include <linux/list_sort.h> 4 #include <linux/string.h> 5 #include <linux/zalloc.h> 6 #include <subcmd/pager.h> 7 #include <sys/types.h> 8 #include <ctype.h> 9 #include <dirent.h> 10 #include <pthread.h> 11 #include <string.h> 12 #include <unistd.h> 13 #include "cpumap.h" 14 #include "debug.h" 15 #include "evsel.h" 16 #include "pmus.h" 17 #include "pmu.h" 18 #include "print-events.h" 19 20 /* 21 * core_pmus: A PMU belongs to core_pmus if it's name is "cpu" or it's sysfs 22 * directory contains "cpus" file. All PMUs belonging to core_pmus 23 * must have pmu->is_core=1. If there are more than one PMU in 24 * this list, perf interprets it as a heterogeneous platform. 25 * (FWIW, certain ARM platforms having heterogeneous cores uses 26 * homogeneous PMU, and thus they are treated as homogeneous 27 * platform by perf because core_pmus will have only one entry) 28 * other_pmus: All other PMUs which are not part of core_pmus list. It doesn't 29 * matter whether PMU is present per SMT-thread or outside of the 30 * core in the hw. For e.g., an instance of AMD ibs_fetch// and 31 * ibs_op// PMUs is present in each hw SMT thread, however they 32 * are captured under other_pmus. PMUs belonging to other_pmus 33 * must have pmu->is_core=0 but pmu->is_uncore could be 0 or 1. 34 */ 35 static LIST_HEAD(core_pmus); 36 static LIST_HEAD(other_pmus); 37 static bool read_sysfs_core_pmus; 38 static bool read_sysfs_all_pmus; 39 40 int pmu_name_len_no_suffix(const char *str, unsigned long *num) 41 { 42 int orig_len, len; 43 44 orig_len = len = strlen(str); 45 46 /* Non-uncore PMUs have their full length, for example, i915. */ 47 if (!strstarts(str, "uncore_")) 48 return len; 49 50 /* 51 * Count trailing digits and '_', if '_{num}' suffix isn't present use 52 * the full length. 53 */ 54 while (len > 0 && isdigit(str[len - 1])) 55 len--; 56 57 if (len > 0 && len != orig_len && str[len - 1] == '_') { 58 if (num) 59 *num = strtoul(&str[len], NULL, 10); 60 return len - 1; 61 } 62 return orig_len; 63 } 64 65 void perf_pmus__destroy(void) 66 { 67 struct perf_pmu *pmu, *tmp; 68 69 list_for_each_entry_safe(pmu, tmp, &core_pmus, list) { 70 list_del(&pmu->list); 71 72 perf_pmu__delete(pmu); 73 } 74 list_for_each_entry_safe(pmu, tmp, &other_pmus, list) { 75 list_del(&pmu->list); 76 77 perf_pmu__delete(pmu); 78 } 79 read_sysfs_core_pmus = false; 80 read_sysfs_all_pmus = false; 81 } 82 83 static struct perf_pmu *pmu_find(const char *name) 84 { 85 struct perf_pmu *pmu; 86 87 list_for_each_entry(pmu, &core_pmus, list) { 88 if (!strcmp(pmu->name, name) || 89 (pmu->alias_name && !strcmp(pmu->alias_name, name))) 90 return pmu; 91 } 92 list_for_each_entry(pmu, &other_pmus, list) { 93 if (!strcmp(pmu->name, name) || 94 (pmu->alias_name && !strcmp(pmu->alias_name, name))) 95 return pmu; 96 } 97 98 return NULL; 99 } 100 101 struct perf_pmu *perf_pmus__find(const char *name) 102 { 103 struct perf_pmu *pmu; 104 int dirfd; 105 bool core_pmu; 106 107 /* 108 * Once PMU is loaded it stays in the list, 109 * so we keep us from multiple reading/parsing 110 * the pmu format definitions. 111 */ 112 pmu = pmu_find(name); 113 if (pmu) 114 return pmu; 115 116 if (read_sysfs_all_pmus) 117 return NULL; 118 119 core_pmu = is_pmu_core(name); 120 if (core_pmu && read_sysfs_core_pmus) 121 return NULL; 122 123 dirfd = perf_pmu__event_source_devices_fd(); 124 pmu = perf_pmu__lookup(core_pmu ? &core_pmus : &other_pmus, dirfd, name); 125 close(dirfd); 126 127 return pmu; 128 } 129 130 static struct perf_pmu *perf_pmu__find2(int dirfd, const char *name) 131 { 132 struct perf_pmu *pmu; 133 bool core_pmu; 134 135 /* 136 * Once PMU is loaded it stays in the list, 137 * so we keep us from multiple reading/parsing 138 * the pmu format definitions. 139 */ 140 pmu = pmu_find(name); 141 if (pmu) 142 return pmu; 143 144 if (read_sysfs_all_pmus) 145 return NULL; 146 147 core_pmu = is_pmu_core(name); 148 if (core_pmu && read_sysfs_core_pmus) 149 return NULL; 150 151 return perf_pmu__lookup(core_pmu ? &core_pmus : &other_pmus, dirfd, name); 152 } 153 154 static int pmus_cmp(void *priv __maybe_unused, 155 const struct list_head *lhs, const struct list_head *rhs) 156 { 157 unsigned long lhs_num = 0, rhs_num = 0; 158 struct perf_pmu *lhs_pmu = container_of(lhs, struct perf_pmu, list); 159 struct perf_pmu *rhs_pmu = container_of(rhs, struct perf_pmu, list); 160 const char *lhs_pmu_name = lhs_pmu->name ?: ""; 161 const char *rhs_pmu_name = rhs_pmu->name ?: ""; 162 int lhs_pmu_name_len = pmu_name_len_no_suffix(lhs_pmu_name, &lhs_num); 163 int rhs_pmu_name_len = pmu_name_len_no_suffix(rhs_pmu_name, &rhs_num); 164 int ret = strncmp(lhs_pmu_name, rhs_pmu_name, 165 lhs_pmu_name_len < rhs_pmu_name_len ? lhs_pmu_name_len : rhs_pmu_name_len); 166 167 if (lhs_pmu_name_len != rhs_pmu_name_len || ret != 0 || lhs_pmu_name_len == 0) 168 return ret; 169 170 return lhs_num < rhs_num ? -1 : (lhs_num > rhs_num ? 1 : 0); 171 } 172 173 /* Add all pmus in sysfs to pmu list: */ 174 static void pmu_read_sysfs(bool core_only) 175 { 176 int fd; 177 DIR *dir; 178 struct dirent *dent; 179 180 if (read_sysfs_all_pmus || (core_only && read_sysfs_core_pmus)) 181 return; 182 183 fd = perf_pmu__event_source_devices_fd(); 184 if (fd < 0) 185 return; 186 187 dir = fdopendir(fd); 188 if (!dir) { 189 close(fd); 190 return; 191 } 192 193 while ((dent = readdir(dir))) { 194 if (!strcmp(dent->d_name, ".") || !strcmp(dent->d_name, "..")) 195 continue; 196 if (core_only && !is_pmu_core(dent->d_name)) 197 continue; 198 /* add to static LIST_HEAD(core_pmus) or LIST_HEAD(other_pmus): */ 199 perf_pmu__find2(fd, dent->d_name); 200 } 201 202 closedir(dir); 203 if (list_empty(&core_pmus)) { 204 if (!perf_pmu__create_placeholder_core_pmu(&core_pmus)) 205 pr_err("Failure to set up any core PMUs\n"); 206 } 207 list_sort(NULL, &core_pmus, pmus_cmp); 208 list_sort(NULL, &other_pmus, pmus_cmp); 209 if (!list_empty(&core_pmus)) { 210 read_sysfs_core_pmus = true; 211 if (!core_only) 212 read_sysfs_all_pmus = true; 213 } 214 } 215 216 static struct perf_pmu *__perf_pmus__find_by_type(unsigned int type) 217 { 218 struct perf_pmu *pmu; 219 220 list_for_each_entry(pmu, &core_pmus, list) { 221 if (pmu->type == type) 222 return pmu; 223 } 224 225 list_for_each_entry(pmu, &other_pmus, list) { 226 if (pmu->type == type) 227 return pmu; 228 } 229 return NULL; 230 } 231 232 struct perf_pmu *perf_pmus__find_by_type(unsigned int type) 233 { 234 struct perf_pmu *pmu = __perf_pmus__find_by_type(type); 235 236 if (pmu || read_sysfs_all_pmus) 237 return pmu; 238 239 pmu_read_sysfs(/*core_only=*/false); 240 pmu = __perf_pmus__find_by_type(type); 241 return pmu; 242 } 243 244 /* 245 * pmu iterator: If pmu is NULL, we start at the begin, otherwise return the 246 * next pmu. Returns NULL on end. 247 */ 248 struct perf_pmu *perf_pmus__scan(struct perf_pmu *pmu) 249 { 250 bool use_core_pmus = !pmu || pmu->is_core; 251 252 if (!pmu) { 253 pmu_read_sysfs(/*core_only=*/false); 254 pmu = list_prepare_entry(pmu, &core_pmus, list); 255 } 256 if (use_core_pmus) { 257 list_for_each_entry_continue(pmu, &core_pmus, list) 258 return pmu; 259 260 pmu = NULL; 261 pmu = list_prepare_entry(pmu, &other_pmus, list); 262 } 263 list_for_each_entry_continue(pmu, &other_pmus, list) 264 return pmu; 265 return NULL; 266 } 267 268 struct perf_pmu *perf_pmus__scan_core(struct perf_pmu *pmu) 269 { 270 if (!pmu) { 271 pmu_read_sysfs(/*core_only=*/true); 272 return list_first_entry_or_null(&core_pmus, typeof(*pmu), list); 273 } 274 list_for_each_entry_continue(pmu, &core_pmus, list) 275 return pmu; 276 277 return NULL; 278 } 279 280 static struct perf_pmu *perf_pmus__scan_skip_duplicates(struct perf_pmu *pmu) 281 { 282 bool use_core_pmus = !pmu || pmu->is_core; 283 int last_pmu_name_len = 0; 284 const char *last_pmu_name = (pmu && pmu->name) ? pmu->name : ""; 285 286 if (!pmu) { 287 pmu_read_sysfs(/*core_only=*/false); 288 pmu = list_prepare_entry(pmu, &core_pmus, list); 289 } else 290 last_pmu_name_len = pmu_name_len_no_suffix(pmu->name ?: "", NULL); 291 292 if (use_core_pmus) { 293 list_for_each_entry_continue(pmu, &core_pmus, list) { 294 int pmu_name_len = pmu_name_len_no_suffix(pmu->name ?: "", /*num=*/NULL); 295 296 if (last_pmu_name_len == pmu_name_len && 297 !strncmp(last_pmu_name, pmu->name ?: "", pmu_name_len)) 298 continue; 299 300 return pmu; 301 } 302 pmu = NULL; 303 pmu = list_prepare_entry(pmu, &other_pmus, list); 304 } 305 list_for_each_entry_continue(pmu, &other_pmus, list) { 306 int pmu_name_len = pmu_name_len_no_suffix(pmu->name ?: "", /*num=*/NULL); 307 308 if (last_pmu_name_len == pmu_name_len && 309 !strncmp(last_pmu_name, pmu->name ?: "", pmu_name_len)) 310 continue; 311 312 return pmu; 313 } 314 return NULL; 315 } 316 317 const struct perf_pmu *perf_pmus__pmu_for_pmu_filter(const char *str) 318 { 319 struct perf_pmu *pmu = NULL; 320 321 while ((pmu = perf_pmus__scan(pmu)) != NULL) { 322 if (!strcmp(pmu->name, str)) 323 return pmu; 324 /* Ignore "uncore_" prefix. */ 325 if (!strncmp(pmu->name, "uncore_", 7)) { 326 if (!strcmp(pmu->name + 7, str)) 327 return pmu; 328 } 329 /* Ignore "cpu_" prefix on Intel hybrid PMUs. */ 330 if (!strncmp(pmu->name, "cpu_", 4)) { 331 if (!strcmp(pmu->name + 4, str)) 332 return pmu; 333 } 334 } 335 return NULL; 336 } 337 338 int __weak perf_pmus__num_mem_pmus(void) 339 { 340 /* All core PMUs are for mem events. */ 341 return perf_pmus__num_core_pmus(); 342 } 343 344 /** Struct for ordering events as output in perf list. */ 345 struct sevent { 346 /** PMU for event. */ 347 const struct perf_pmu *pmu; 348 const char *name; 349 const char* alias; 350 const char *scale_unit; 351 const char *desc; 352 const char *long_desc; 353 const char *encoding_desc; 354 const char *topic; 355 const char *pmu_name; 356 bool deprecated; 357 }; 358 359 static int cmp_sevent(const void *a, const void *b) 360 { 361 const struct sevent *as = a; 362 const struct sevent *bs = b; 363 bool a_iscpu, b_iscpu; 364 int ret; 365 366 /* Put extra events last. */ 367 if (!!as->desc != !!bs->desc) 368 return !!as->desc - !!bs->desc; 369 370 /* Order by topics. */ 371 ret = strcmp(as->topic ?: "", bs->topic ?: ""); 372 if (ret) 373 return ret; 374 375 /* Order CPU core events to be first */ 376 a_iscpu = as->pmu ? as->pmu->is_core : true; 377 b_iscpu = bs->pmu ? bs->pmu->is_core : true; 378 if (a_iscpu != b_iscpu) 379 return a_iscpu ? -1 : 1; 380 381 /* Order by PMU name. */ 382 if (as->pmu != bs->pmu) { 383 ret = strcmp(as->pmu_name ?: "", bs->pmu_name ?: ""); 384 if (ret) 385 return ret; 386 } 387 388 /* Order by event name. */ 389 return strcmp(as->name, bs->name); 390 } 391 392 static bool pmu_alias_is_duplicate(struct sevent *a, struct sevent *b) 393 { 394 /* Different names -> never duplicates */ 395 if (strcmp(a->name ?: "//", b->name ?: "//")) 396 return false; 397 398 /* Don't remove duplicates for different PMUs */ 399 return strcmp(a->pmu_name, b->pmu_name) == 0; 400 } 401 402 struct events_callback_state { 403 struct sevent *aliases; 404 size_t aliases_len; 405 size_t index; 406 }; 407 408 static int perf_pmus__print_pmu_events__callback(void *vstate, 409 struct pmu_event_info *info) 410 { 411 struct events_callback_state *state = vstate; 412 struct sevent *s; 413 414 if (state->index >= state->aliases_len) { 415 pr_err("Unexpected event %s/%s/\n", info->pmu->name, info->name); 416 return 1; 417 } 418 s = &state->aliases[state->index]; 419 s->pmu = info->pmu; 420 #define COPY_STR(str) s->str = info->str ? strdup(info->str) : NULL 421 COPY_STR(name); 422 COPY_STR(alias); 423 COPY_STR(scale_unit); 424 COPY_STR(desc); 425 COPY_STR(long_desc); 426 COPY_STR(encoding_desc); 427 COPY_STR(topic); 428 COPY_STR(pmu_name); 429 #undef COPY_STR 430 s->deprecated = info->deprecated; 431 state->index++; 432 return 0; 433 } 434 435 void perf_pmus__print_pmu_events(const struct print_callbacks *print_cb, void *print_state) 436 { 437 struct perf_pmu *pmu; 438 int printed = 0; 439 int len; 440 struct sevent *aliases; 441 struct events_callback_state state; 442 bool skip_duplicate_pmus = print_cb->skip_duplicate_pmus(print_state); 443 struct perf_pmu *(*scan_fn)(struct perf_pmu *); 444 445 if (skip_duplicate_pmus) 446 scan_fn = perf_pmus__scan_skip_duplicates; 447 else 448 scan_fn = perf_pmus__scan; 449 450 pmu = NULL; 451 len = 0; 452 while ((pmu = scan_fn(pmu)) != NULL) 453 len += perf_pmu__num_events(pmu); 454 455 aliases = zalloc(sizeof(struct sevent) * len); 456 if (!aliases) { 457 pr_err("FATAL: not enough memory to print PMU events\n"); 458 return; 459 } 460 pmu = NULL; 461 state = (struct events_callback_state) { 462 .aliases = aliases, 463 .aliases_len = len, 464 .index = 0, 465 }; 466 while ((pmu = scan_fn(pmu)) != NULL) { 467 perf_pmu__for_each_event(pmu, skip_duplicate_pmus, &state, 468 perf_pmus__print_pmu_events__callback); 469 } 470 qsort(aliases, len, sizeof(struct sevent), cmp_sevent); 471 for (int j = 0; j < len; j++) { 472 /* Skip duplicates */ 473 if (j > 0 && pmu_alias_is_duplicate(&aliases[j], &aliases[j - 1])) 474 continue; 475 476 print_cb->print_event(print_state, 477 aliases[j].pmu_name, 478 aliases[j].topic, 479 aliases[j].name, 480 aliases[j].alias, 481 aliases[j].scale_unit, 482 aliases[j].deprecated, 483 "Kernel PMU event", 484 aliases[j].desc, 485 aliases[j].long_desc, 486 aliases[j].encoding_desc); 487 zfree(&aliases[j].name); 488 zfree(&aliases[j].alias); 489 zfree(&aliases[j].scale_unit); 490 zfree(&aliases[j].desc); 491 zfree(&aliases[j].long_desc); 492 zfree(&aliases[j].encoding_desc); 493 zfree(&aliases[j].topic); 494 zfree(&aliases[j].pmu_name); 495 } 496 if (printed && pager_in_use()) 497 printf("\n"); 498 499 zfree(&aliases); 500 } 501 502 bool perf_pmus__have_event(const char *pname, const char *name) 503 { 504 struct perf_pmu *pmu = perf_pmus__find(pname); 505 506 return pmu && perf_pmu__have_event(pmu, name); 507 } 508 509 int perf_pmus__num_core_pmus(void) 510 { 511 static int count; 512 513 if (!count) { 514 struct perf_pmu *pmu = NULL; 515 516 while ((pmu = perf_pmus__scan_core(pmu)) != NULL) 517 count++; 518 } 519 return count; 520 } 521 522 static bool __perf_pmus__supports_extended_type(void) 523 { 524 struct perf_pmu *pmu = NULL; 525 526 if (perf_pmus__num_core_pmus() <= 1) 527 return false; 528 529 while ((pmu = perf_pmus__scan_core(pmu)) != NULL) { 530 if (!is_event_supported(PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES | ((__u64)pmu->type << PERF_PMU_TYPE_SHIFT))) 531 return false; 532 } 533 534 return true; 535 } 536 537 static bool perf_pmus__do_support_extended_type; 538 539 static void perf_pmus__init_supports_extended_type(void) 540 { 541 perf_pmus__do_support_extended_type = __perf_pmus__supports_extended_type(); 542 } 543 544 bool perf_pmus__supports_extended_type(void) 545 { 546 static pthread_once_t extended_type_once = PTHREAD_ONCE_INIT; 547 548 pthread_once(&extended_type_once, perf_pmus__init_supports_extended_type); 549 550 return perf_pmus__do_support_extended_type; 551 } 552 553 char *perf_pmus__default_pmu_name(void) 554 { 555 int fd; 556 DIR *dir; 557 struct dirent *dent; 558 char *result = NULL; 559 560 if (!list_empty(&core_pmus)) 561 return strdup(list_first_entry(&core_pmus, struct perf_pmu, list)->name); 562 563 fd = perf_pmu__event_source_devices_fd(); 564 if (fd < 0) 565 return strdup("cpu"); 566 567 dir = fdopendir(fd); 568 if (!dir) { 569 close(fd); 570 return strdup("cpu"); 571 } 572 573 while ((dent = readdir(dir))) { 574 if (!strcmp(dent->d_name, ".") || !strcmp(dent->d_name, "..")) 575 continue; 576 if (is_pmu_core(dent->d_name)) { 577 result = strdup(dent->d_name); 578 break; 579 } 580 } 581 582 closedir(dir); 583 return result ?: strdup("cpu"); 584 } 585 586 struct perf_pmu *evsel__find_pmu(const struct evsel *evsel) 587 { 588 struct perf_pmu *pmu = evsel->pmu; 589 590 if (!pmu) { 591 pmu = perf_pmus__find_by_type(evsel->core.attr.type); 592 ((struct evsel *)evsel)->pmu = pmu; 593 } 594 return pmu; 595 } 596 597 struct perf_pmu *perf_pmus__find_core_pmu(void) 598 { 599 struct perf_pmu *pmu = NULL; 600 601 while ((pmu = perf_pmus__scan_core(pmu))) { 602 /* 603 * The cpumap should cover all CPUs. Otherwise, some CPUs may 604 * not support some events or have different event IDs. 605 */ 606 if (RC_CHK_ACCESS(pmu->cpus)->nr != cpu__max_cpu().cpu) 607 return NULL; 608 609 return pmu; 610 } 611 return NULL; 612 } 613