1 // SPDX-License-Identifier: GPL-2.0 2 #include <api/fs/fs.h> 3 #include "cpumap.h" 4 #include "debug.h" 5 #include "event.h" 6 #include <assert.h> 7 #include <dirent.h> 8 #include <stdio.h> 9 #include <stdlib.h> 10 #include <linux/bitmap.h> 11 #include "asm/bug.h" 12 13 #include <linux/ctype.h> 14 #include <linux/zalloc.h> 15 #include <internal/cpumap.h> 16 17 static struct perf_cpu max_cpu_num; 18 static struct perf_cpu max_present_cpu_num; 19 static int max_node_num; 20 /** 21 * The numa node X as read from /sys/devices/system/node/nodeX indexed by the 22 * CPU number. 23 */ 24 static int *cpunode_map; 25 26 bool perf_record_cpu_map_data__test_bit(int i, 27 const struct perf_record_cpu_map_data *data) 28 { 29 int bit_word32 = i / 32; 30 __u32 bit_mask32 = 1U << (i & 31); 31 int bit_word64 = i / 64; 32 __u64 bit_mask64 = ((__u64)1) << (i & 63); 33 34 return (data->mask32_data.long_size == 4) 35 ? (bit_word32 < data->mask32_data.nr) && 36 (data->mask32_data.mask[bit_word32] & bit_mask32) != 0 37 : (bit_word64 < data->mask64_data.nr) && 38 (data->mask64_data.mask[bit_word64] & bit_mask64) != 0; 39 } 40 41 /* Read ith mask value from data into the given 64-bit sized bitmap */ 42 static void perf_record_cpu_map_data__read_one_mask(const struct perf_record_cpu_map_data *data, 43 int i, unsigned long *bitmap) 44 { 45 #if __SIZEOF_LONG__ == 8 46 if (data->mask32_data.long_size == 4) 47 bitmap[0] = data->mask32_data.mask[i]; 48 else 49 bitmap[0] = data->mask64_data.mask[i]; 50 #else 51 if (data->mask32_data.long_size == 4) { 52 bitmap[0] = data->mask32_data.mask[i]; 53 bitmap[1] = 0; 54 } else { 55 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 56 bitmap[0] = (unsigned long)(data->mask64_data.mask[i] >> 32); 57 bitmap[1] = (unsigned long)data->mask64_data.mask[i]; 58 #else 59 bitmap[0] = (unsigned long)data->mask64_data.mask[i]; 60 bitmap[1] = (unsigned long)(data->mask64_data.mask[i] >> 32); 61 #endif 62 } 63 #endif 64 } 65 static struct perf_cpu_map *cpu_map__from_entries(const struct perf_record_cpu_map_data *data) 66 { 67 struct perf_cpu_map *map; 68 69 map = perf_cpu_map__empty_new(data->cpus_data.nr); 70 if (map) { 71 unsigned i; 72 73 for (i = 0; i < data->cpus_data.nr; i++) { 74 /* 75 * Special treatment for -1, which is not real cpu number, 76 * and we need to use (int) -1 to initialize map[i], 77 * otherwise it would become 65535. 78 */ 79 if (data->cpus_data.cpu[i] == (u16) -1) 80 map->map[i].cpu = -1; 81 else 82 map->map[i].cpu = (int) data->cpus_data.cpu[i]; 83 } 84 } 85 86 return map; 87 } 88 89 static struct perf_cpu_map *cpu_map__from_mask(const struct perf_record_cpu_map_data *data) 90 { 91 DECLARE_BITMAP(local_copy, 64); 92 int weight = 0, mask_nr = data->mask32_data.nr; 93 struct perf_cpu_map *map; 94 95 for (int i = 0; i < mask_nr; i++) { 96 perf_record_cpu_map_data__read_one_mask(data, i, local_copy); 97 weight += bitmap_weight(local_copy, 64); 98 } 99 100 map = perf_cpu_map__empty_new(weight); 101 if (!map) 102 return NULL; 103 104 for (int i = 0, j = 0; i < mask_nr; i++) { 105 int cpus_per_i = (i * data->mask32_data.long_size * BITS_PER_BYTE); 106 int cpu; 107 108 perf_record_cpu_map_data__read_one_mask(data, i, local_copy); 109 for_each_set_bit(cpu, local_copy, 64) 110 map->map[j++].cpu = cpu + cpus_per_i; 111 } 112 return map; 113 114 } 115 116 static struct perf_cpu_map *cpu_map__from_range(const struct perf_record_cpu_map_data *data) 117 { 118 struct perf_cpu_map *map; 119 unsigned int i = 0; 120 121 map = perf_cpu_map__empty_new(data->range_cpu_data.end_cpu - 122 data->range_cpu_data.start_cpu + 1 + data->range_cpu_data.any_cpu); 123 if (!map) 124 return NULL; 125 126 if (data->range_cpu_data.any_cpu) 127 map->map[i++].cpu = -1; 128 129 for (int cpu = data->range_cpu_data.start_cpu; cpu <= data->range_cpu_data.end_cpu; 130 i++, cpu++) 131 map->map[i].cpu = cpu; 132 133 return map; 134 } 135 136 struct perf_cpu_map *cpu_map__new_data(const struct perf_record_cpu_map_data *data) 137 { 138 switch (data->type) { 139 case PERF_CPU_MAP__CPUS: 140 return cpu_map__from_entries(data); 141 case PERF_CPU_MAP__MASK: 142 return cpu_map__from_mask(data); 143 case PERF_CPU_MAP__RANGE_CPUS: 144 return cpu_map__from_range(data); 145 default: 146 pr_err("cpu_map__new_data unknown type %d\n", data->type); 147 return NULL; 148 } 149 } 150 151 size_t cpu_map__fprintf(struct perf_cpu_map *map, FILE *fp) 152 { 153 #define BUFSIZE 1024 154 char buf[BUFSIZE]; 155 156 cpu_map__snprint(map, buf, sizeof(buf)); 157 return fprintf(fp, "%s\n", buf); 158 #undef BUFSIZE 159 } 160 161 struct perf_cpu_map *perf_cpu_map__empty_new(int nr) 162 { 163 struct perf_cpu_map *cpus = malloc(sizeof(*cpus) + sizeof(int) * nr); 164 165 if (cpus != NULL) { 166 int i; 167 168 cpus->nr = nr; 169 for (i = 0; i < nr; i++) 170 cpus->map[i].cpu = -1; 171 172 refcount_set(&cpus->refcnt, 1); 173 } 174 175 return cpus; 176 } 177 178 struct cpu_aggr_map *cpu_aggr_map__empty_new(int nr) 179 { 180 struct cpu_aggr_map *cpus = malloc(sizeof(*cpus) + sizeof(struct aggr_cpu_id) * nr); 181 182 if (cpus != NULL) { 183 int i; 184 185 cpus->nr = nr; 186 for (i = 0; i < nr; i++) 187 cpus->map[i] = aggr_cpu_id__empty(); 188 189 refcount_set(&cpus->refcnt, 1); 190 } 191 192 return cpus; 193 } 194 195 static int cpu__get_topology_int(int cpu, const char *name, int *value) 196 { 197 char path[PATH_MAX]; 198 199 snprintf(path, PATH_MAX, 200 "devices/system/cpu/cpu%d/topology/%s", cpu, name); 201 202 return sysfs__read_int(path, value); 203 } 204 205 int cpu__get_socket_id(struct perf_cpu cpu) 206 { 207 int value, ret = cpu__get_topology_int(cpu.cpu, "physical_package_id", &value); 208 return ret ?: value; 209 } 210 211 struct aggr_cpu_id aggr_cpu_id__socket(struct perf_cpu cpu, void *data __maybe_unused) 212 { 213 struct aggr_cpu_id id = aggr_cpu_id__empty(); 214 215 id.socket = cpu__get_socket_id(cpu); 216 return id; 217 } 218 219 static int aggr_cpu_id__cmp(const void *a_pointer, const void *b_pointer) 220 { 221 struct aggr_cpu_id *a = (struct aggr_cpu_id *)a_pointer; 222 struct aggr_cpu_id *b = (struct aggr_cpu_id *)b_pointer; 223 224 if (a->node != b->node) 225 return a->node - b->node; 226 else if (a->socket != b->socket) 227 return a->socket - b->socket; 228 else if (a->die != b->die) 229 return a->die - b->die; 230 else if (a->core != b->core) 231 return a->core - b->core; 232 else 233 return a->thread_idx - b->thread_idx; 234 } 235 236 struct cpu_aggr_map *cpu_aggr_map__new(const struct perf_cpu_map *cpus, 237 aggr_cpu_id_get_t get_id, 238 void *data, bool needs_sort) 239 { 240 int idx; 241 struct perf_cpu cpu; 242 struct cpu_aggr_map *c = cpu_aggr_map__empty_new(cpus->nr); 243 244 if (!c) 245 return NULL; 246 247 /* Reset size as it may only be partially filled */ 248 c->nr = 0; 249 250 perf_cpu_map__for_each_cpu(cpu, idx, cpus) { 251 bool duplicate = false; 252 struct aggr_cpu_id cpu_id = get_id(cpu, data); 253 254 for (int j = 0; j < c->nr; j++) { 255 if (aggr_cpu_id__equal(&cpu_id, &c->map[j])) { 256 duplicate = true; 257 break; 258 } 259 } 260 if (!duplicate) { 261 c->map[c->nr] = cpu_id; 262 c->nr++; 263 } 264 } 265 /* Trim. */ 266 if (c->nr != cpus->nr) { 267 struct cpu_aggr_map *trimmed_c = 268 realloc(c, 269 sizeof(struct cpu_aggr_map) + sizeof(struct aggr_cpu_id) * c->nr); 270 271 if (trimmed_c) 272 c = trimmed_c; 273 } 274 275 /* ensure we process id in increasing order */ 276 if (needs_sort) 277 qsort(c->map, c->nr, sizeof(struct aggr_cpu_id), aggr_cpu_id__cmp); 278 279 return c; 280 281 } 282 283 int cpu__get_die_id(struct perf_cpu cpu) 284 { 285 int value, ret = cpu__get_topology_int(cpu.cpu, "die_id", &value); 286 287 return ret ?: value; 288 } 289 290 struct aggr_cpu_id aggr_cpu_id__die(struct perf_cpu cpu, void *data) 291 { 292 struct aggr_cpu_id id; 293 int die; 294 295 die = cpu__get_die_id(cpu); 296 /* There is no die_id on legacy system. */ 297 if (die == -1) 298 die = 0; 299 300 /* 301 * die_id is relative to socket, so start 302 * with the socket ID and then add die to 303 * make a unique ID. 304 */ 305 id = aggr_cpu_id__socket(cpu, data); 306 if (aggr_cpu_id__is_empty(&id)) 307 return id; 308 309 id.die = die; 310 return id; 311 } 312 313 int cpu__get_core_id(struct perf_cpu cpu) 314 { 315 int value, ret = cpu__get_topology_int(cpu.cpu, "core_id", &value); 316 return ret ?: value; 317 } 318 319 struct aggr_cpu_id aggr_cpu_id__core(struct perf_cpu cpu, void *data) 320 { 321 struct aggr_cpu_id id; 322 int core = cpu__get_core_id(cpu); 323 324 /* aggr_cpu_id__die returns a struct with socket and die set. */ 325 id = aggr_cpu_id__die(cpu, data); 326 if (aggr_cpu_id__is_empty(&id)) 327 return id; 328 329 /* 330 * core_id is relative to socket and die, we need a global id. 331 * So we combine the result from cpu_map__get_die with the core id 332 */ 333 id.core = core; 334 return id; 335 336 } 337 338 struct aggr_cpu_id aggr_cpu_id__cpu(struct perf_cpu cpu, void *data) 339 { 340 struct aggr_cpu_id id; 341 342 /* aggr_cpu_id__core returns a struct with socket, die and core set. */ 343 id = aggr_cpu_id__core(cpu, data); 344 if (aggr_cpu_id__is_empty(&id)) 345 return id; 346 347 id.cpu = cpu; 348 return id; 349 350 } 351 352 struct aggr_cpu_id aggr_cpu_id__node(struct perf_cpu cpu, void *data __maybe_unused) 353 { 354 struct aggr_cpu_id id = aggr_cpu_id__empty(); 355 356 id.node = cpu__get_node(cpu); 357 return id; 358 } 359 360 struct aggr_cpu_id aggr_cpu_id__global(struct perf_cpu cpu, void *data __maybe_unused) 361 { 362 struct aggr_cpu_id id = aggr_cpu_id__empty(); 363 364 /* it always aggregates to the cpu 0 */ 365 cpu.cpu = 0; 366 id.cpu = cpu; 367 return id; 368 } 369 370 /* setup simple routines to easily access node numbers given a cpu number */ 371 static int get_max_num(char *path, int *max) 372 { 373 size_t num; 374 char *buf; 375 int err = 0; 376 377 if (filename__read_str(path, &buf, &num)) 378 return -1; 379 380 buf[num] = '\0'; 381 382 /* start on the right, to find highest node num */ 383 while (--num) { 384 if ((buf[num] == ',') || (buf[num] == '-')) { 385 num++; 386 break; 387 } 388 } 389 if (sscanf(&buf[num], "%d", max) < 1) { 390 err = -1; 391 goto out; 392 } 393 394 /* convert from 0-based to 1-based */ 395 (*max)++; 396 397 out: 398 free(buf); 399 return err; 400 } 401 402 /* Determine highest possible cpu in the system for sparse allocation */ 403 static void set_max_cpu_num(void) 404 { 405 const char *mnt; 406 char path[PATH_MAX]; 407 int ret = -1; 408 409 /* set up default */ 410 max_cpu_num.cpu = 4096; 411 max_present_cpu_num.cpu = 4096; 412 413 mnt = sysfs__mountpoint(); 414 if (!mnt) 415 goto out; 416 417 /* get the highest possible cpu number for a sparse allocation */ 418 ret = snprintf(path, PATH_MAX, "%s/devices/system/cpu/possible", mnt); 419 if (ret >= PATH_MAX) { 420 pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX); 421 goto out; 422 } 423 424 ret = get_max_num(path, &max_cpu_num.cpu); 425 if (ret) 426 goto out; 427 428 /* get the highest present cpu number for a sparse allocation */ 429 ret = snprintf(path, PATH_MAX, "%s/devices/system/cpu/present", mnt); 430 if (ret >= PATH_MAX) { 431 pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX); 432 goto out; 433 } 434 435 ret = get_max_num(path, &max_present_cpu_num.cpu); 436 437 out: 438 if (ret) 439 pr_err("Failed to read max cpus, using default of %d\n", max_cpu_num.cpu); 440 } 441 442 /* Determine highest possible node in the system for sparse allocation */ 443 static void set_max_node_num(void) 444 { 445 const char *mnt; 446 char path[PATH_MAX]; 447 int ret = -1; 448 449 /* set up default */ 450 max_node_num = 8; 451 452 mnt = sysfs__mountpoint(); 453 if (!mnt) 454 goto out; 455 456 /* get the highest possible cpu number for a sparse allocation */ 457 ret = snprintf(path, PATH_MAX, "%s/devices/system/node/possible", mnt); 458 if (ret >= PATH_MAX) { 459 pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX); 460 goto out; 461 } 462 463 ret = get_max_num(path, &max_node_num); 464 465 out: 466 if (ret) 467 pr_err("Failed to read max nodes, using default of %d\n", max_node_num); 468 } 469 470 int cpu__max_node(void) 471 { 472 if (unlikely(!max_node_num)) 473 set_max_node_num(); 474 475 return max_node_num; 476 } 477 478 struct perf_cpu cpu__max_cpu(void) 479 { 480 if (unlikely(!max_cpu_num.cpu)) 481 set_max_cpu_num(); 482 483 return max_cpu_num; 484 } 485 486 struct perf_cpu cpu__max_present_cpu(void) 487 { 488 if (unlikely(!max_present_cpu_num.cpu)) 489 set_max_cpu_num(); 490 491 return max_present_cpu_num; 492 } 493 494 495 int cpu__get_node(struct perf_cpu cpu) 496 { 497 if (unlikely(cpunode_map == NULL)) { 498 pr_debug("cpu_map not initialized\n"); 499 return -1; 500 } 501 502 return cpunode_map[cpu.cpu]; 503 } 504 505 static int init_cpunode_map(void) 506 { 507 int i; 508 509 set_max_cpu_num(); 510 set_max_node_num(); 511 512 cpunode_map = calloc(max_cpu_num.cpu, sizeof(int)); 513 if (!cpunode_map) { 514 pr_err("%s: calloc failed\n", __func__); 515 return -1; 516 } 517 518 for (i = 0; i < max_cpu_num.cpu; i++) 519 cpunode_map[i] = -1; 520 521 return 0; 522 } 523 524 int cpu__setup_cpunode_map(void) 525 { 526 struct dirent *dent1, *dent2; 527 DIR *dir1, *dir2; 528 unsigned int cpu, mem; 529 char buf[PATH_MAX]; 530 char path[PATH_MAX]; 531 const char *mnt; 532 int n; 533 534 /* initialize globals */ 535 if (init_cpunode_map()) 536 return -1; 537 538 mnt = sysfs__mountpoint(); 539 if (!mnt) 540 return 0; 541 542 n = snprintf(path, PATH_MAX, "%s/devices/system/node", mnt); 543 if (n >= PATH_MAX) { 544 pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX); 545 return -1; 546 } 547 548 dir1 = opendir(path); 549 if (!dir1) 550 return 0; 551 552 /* walk tree and setup map */ 553 while ((dent1 = readdir(dir1)) != NULL) { 554 if (dent1->d_type != DT_DIR || sscanf(dent1->d_name, "node%u", &mem) < 1) 555 continue; 556 557 n = snprintf(buf, PATH_MAX, "%s/%s", path, dent1->d_name); 558 if (n >= PATH_MAX) { 559 pr_err("sysfs path crossed PATH_MAX(%d) size\n", PATH_MAX); 560 continue; 561 } 562 563 dir2 = opendir(buf); 564 if (!dir2) 565 continue; 566 while ((dent2 = readdir(dir2)) != NULL) { 567 if (dent2->d_type != DT_LNK || sscanf(dent2->d_name, "cpu%u", &cpu) < 1) 568 continue; 569 cpunode_map[cpu] = mem; 570 } 571 closedir(dir2); 572 } 573 closedir(dir1); 574 return 0; 575 } 576 577 size_t cpu_map__snprint(struct perf_cpu_map *map, char *buf, size_t size) 578 { 579 int i, start = -1; 580 bool first = true; 581 size_t ret = 0; 582 583 #define COMMA first ? "" : "," 584 585 for (i = 0; i < map->nr + 1; i++) { 586 struct perf_cpu cpu = { .cpu = INT_MAX }; 587 bool last = i == map->nr; 588 589 if (!last) 590 cpu = map->map[i]; 591 592 if (start == -1) { 593 start = i; 594 if (last) { 595 ret += snprintf(buf + ret, size - ret, 596 "%s%d", COMMA, 597 map->map[i].cpu); 598 } 599 } else if (((i - start) != (cpu.cpu - map->map[start].cpu)) || last) { 600 int end = i - 1; 601 602 if (start == end) { 603 ret += snprintf(buf + ret, size - ret, 604 "%s%d", COMMA, 605 map->map[start].cpu); 606 } else { 607 ret += snprintf(buf + ret, size - ret, 608 "%s%d-%d", COMMA, 609 map->map[start].cpu, map->map[end].cpu); 610 } 611 first = false; 612 start = i; 613 } 614 } 615 616 #undef COMMA 617 618 pr_debug2("cpumask list: %s\n", buf); 619 return ret; 620 } 621 622 static char hex_char(unsigned char val) 623 { 624 if (val < 10) 625 return val + '0'; 626 if (val < 16) 627 return val - 10 + 'a'; 628 return '?'; 629 } 630 631 size_t cpu_map__snprint_mask(struct perf_cpu_map *map, char *buf, size_t size) 632 { 633 int i, cpu; 634 char *ptr = buf; 635 unsigned char *bitmap; 636 struct perf_cpu last_cpu = perf_cpu_map__cpu(map, map->nr - 1); 637 638 if (buf == NULL) 639 return 0; 640 641 bitmap = zalloc(last_cpu.cpu / 8 + 1); 642 if (bitmap == NULL) { 643 buf[0] = '\0'; 644 return 0; 645 } 646 647 for (i = 0; i < map->nr; i++) { 648 cpu = perf_cpu_map__cpu(map, i).cpu; 649 bitmap[cpu / 8] |= 1 << (cpu % 8); 650 } 651 652 for (cpu = last_cpu.cpu / 4 * 4; cpu >= 0; cpu -= 4) { 653 unsigned char bits = bitmap[cpu / 8]; 654 655 if (cpu % 8) 656 bits >>= 4; 657 else 658 bits &= 0xf; 659 660 *ptr++ = hex_char(bits); 661 if ((cpu % 32) == 0 && cpu > 0) 662 *ptr++ = ','; 663 } 664 *ptr = '\0'; 665 free(bitmap); 666 667 buf[size - 1] = '\0'; 668 return ptr - buf; 669 } 670 671 const struct perf_cpu_map *cpu_map__online(void) /* thread unsafe */ 672 { 673 static const struct perf_cpu_map *online = NULL; 674 675 if (!online) 676 online = perf_cpu_map__new(NULL); /* from /sys/devices/system/cpu/online */ 677 678 return online; 679 } 680 681 bool aggr_cpu_id__equal(const struct aggr_cpu_id *a, const struct aggr_cpu_id *b) 682 { 683 return a->thread_idx == b->thread_idx && 684 a->node == b->node && 685 a->socket == b->socket && 686 a->die == b->die && 687 a->core == b->core && 688 a->cpu.cpu == b->cpu.cpu; 689 } 690 691 bool aggr_cpu_id__is_empty(const struct aggr_cpu_id *a) 692 { 693 return a->thread_idx == -1 && 694 a->node == -1 && 695 a->socket == -1 && 696 a->die == -1 && 697 a->core == -1 && 698 a->cpu.cpu == -1; 699 } 700 701 struct aggr_cpu_id aggr_cpu_id__empty(void) 702 { 703 struct aggr_cpu_id ret = { 704 .thread_idx = -1, 705 .node = -1, 706 .socket = -1, 707 .die = -1, 708 .core = -1, 709 .cpu = (struct perf_cpu){ .cpu = -1 }, 710 }; 711 return ret; 712 } 713