1 // SPDX-License-Identifier: GPL-2.0 2 #include <stddef.h> 3 #include <stdlib.h> 4 #include <string.h> 5 #include <errno.h> 6 #include <sys/types.h> 7 #include <sys/stat.h> 8 #include <unistd.h> 9 #include <api/fs/fs.h> 10 #include <linux/kernel.h> 11 #include "map_symbol.h" 12 #include "mem-events.h" 13 #include "debug.h" 14 #include "symbol.h" 15 #include "pmu.h" 16 #include "pmus.h" 17 18 unsigned int perf_mem_events__loads_ldlat = 30; 19 20 #define E(t, n, s) { .tag = t, .name = n, .sysfs_name = s } 21 22 static struct perf_mem_event perf_mem_events[PERF_MEM_EVENTS__MAX] = { 23 E("ldlat-loads", "cpu/mem-loads,ldlat=%u/P", "cpu/events/mem-loads"), 24 E("ldlat-stores", "cpu/mem-stores/P", "cpu/events/mem-stores"), 25 E(NULL, NULL, NULL), 26 }; 27 #undef E 28 29 static char mem_loads_name[100]; 30 static bool mem_loads_name__init; 31 32 struct perf_mem_event * __weak perf_mem_events__ptr(int i) 33 { 34 if (i >= PERF_MEM_EVENTS__MAX) 35 return NULL; 36 37 return &perf_mem_events[i]; 38 } 39 40 const char * __weak perf_mem_events__name(int i, const char *pmu_name __maybe_unused) 41 { 42 struct perf_mem_event *e = perf_mem_events__ptr(i); 43 44 if (!e) 45 return NULL; 46 47 if (i == PERF_MEM_EVENTS__LOAD) { 48 if (!mem_loads_name__init) { 49 mem_loads_name__init = true; 50 scnprintf(mem_loads_name, sizeof(mem_loads_name), 51 e->name, perf_mem_events__loads_ldlat); 52 } 53 return mem_loads_name; 54 } 55 56 return e->name; 57 } 58 59 __weak bool is_mem_loads_aux_event(struct evsel *leader __maybe_unused) 60 { 61 return false; 62 } 63 64 int perf_mem_events__parse(const char *str) 65 { 66 char *tok, *saveptr = NULL; 67 bool found = false; 68 char *buf; 69 int j; 70 71 /* We need buffer that we know we can write to. */ 72 buf = malloc(strlen(str) + 1); 73 if (!buf) 74 return -ENOMEM; 75 76 strcpy(buf, str); 77 78 tok = strtok_r((char *)buf, ",", &saveptr); 79 80 while (tok) { 81 for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) { 82 struct perf_mem_event *e = perf_mem_events__ptr(j); 83 84 if (!e->tag) 85 continue; 86 87 if (strstr(e->tag, tok)) 88 e->record = found = true; 89 } 90 91 tok = strtok_r(NULL, ",", &saveptr); 92 } 93 94 free(buf); 95 96 if (found) 97 return 0; 98 99 pr_err("failed: event '%s' not found, use '-e list' to get list of available events\n", str); 100 return -1; 101 } 102 103 static bool perf_mem_event__supported(const char *mnt, struct perf_pmu *pmu, 104 struct perf_mem_event *e) 105 { 106 char sysfs_name[100]; 107 char path[PATH_MAX]; 108 struct stat st; 109 110 scnprintf(sysfs_name, sizeof(sysfs_name), e->sysfs_name, pmu->name); 111 scnprintf(path, PATH_MAX, "%s/devices/%s", mnt, sysfs_name); 112 return !stat(path, &st); 113 } 114 115 int perf_mem_events__init(void) 116 { 117 const char *mnt = sysfs__mount(); 118 bool found = false; 119 int j; 120 121 if (!mnt) 122 return -ENOENT; 123 124 for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) { 125 struct perf_mem_event *e = perf_mem_events__ptr(j); 126 struct perf_pmu *pmu = NULL; 127 128 /* 129 * If the event entry isn't valid, skip initialization 130 * and "e->supported" will keep false. 131 */ 132 if (!e->tag) 133 continue; 134 135 /* 136 * Scan all PMUs not just core ones, since perf mem/c2c on 137 * platforms like AMD uses IBS OP PMU which is independent 138 * of core PMU. 139 */ 140 while ((pmu = perf_pmus__scan(pmu)) != NULL) { 141 e->supported |= perf_mem_event__supported(mnt, pmu, e); 142 if (e->supported) { 143 found = true; 144 break; 145 } 146 } 147 } 148 149 return found ? 0 : -ENOENT; 150 } 151 152 void perf_mem_events__list(void) 153 { 154 int j; 155 156 for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) { 157 struct perf_mem_event *e = perf_mem_events__ptr(j); 158 159 fprintf(stderr, "%-*s%-*s%s", 160 e->tag ? 13 : 0, 161 e->tag ? : "", 162 e->tag && verbose > 0 ? 25 : 0, 163 e->tag && verbose > 0 ? perf_mem_events__name(j, NULL) : "", 164 e->supported ? ": available\n" : ""); 165 } 166 } 167 168 static void perf_mem_events__print_unsupport_hybrid(struct perf_mem_event *e, 169 int idx) 170 { 171 const char *mnt = sysfs__mount(); 172 struct perf_pmu *pmu = NULL; 173 174 while ((pmu = perf_pmus__scan(pmu)) != NULL) { 175 if (!perf_mem_event__supported(mnt, pmu, e)) { 176 pr_err("failed: event '%s' not supported\n", 177 perf_mem_events__name(idx, pmu->name)); 178 } 179 } 180 } 181 182 int perf_mem_events__record_args(const char **rec_argv, int *argv_nr, 183 char **rec_tmp, int *tmp_nr) 184 { 185 const char *mnt = sysfs__mount(); 186 int i = *argv_nr, k = 0; 187 struct perf_mem_event *e; 188 189 for (int j = 0; j < PERF_MEM_EVENTS__MAX; j++) { 190 e = perf_mem_events__ptr(j); 191 if (!e->record) 192 continue; 193 194 if (perf_pmus__num_mem_pmus() == 1) { 195 if (!e->supported) { 196 pr_err("failed: event '%s' not supported\n", 197 perf_mem_events__name(j, NULL)); 198 return -1; 199 } 200 201 rec_argv[i++] = "-e"; 202 rec_argv[i++] = perf_mem_events__name(j, NULL); 203 } else { 204 struct perf_pmu *pmu = NULL; 205 206 if (!e->supported) { 207 perf_mem_events__print_unsupport_hybrid(e, j); 208 return -1; 209 } 210 211 while ((pmu = perf_pmus__scan(pmu)) != NULL) { 212 const char *s = perf_mem_events__name(j, pmu->name); 213 214 if (!perf_mem_event__supported(mnt, pmu, e)) 215 continue; 216 217 rec_argv[i++] = "-e"; 218 if (s) { 219 char *copy = strdup(s); 220 if (!copy) 221 return -1; 222 223 rec_argv[i++] = copy; 224 rec_tmp[k++] = copy; 225 } 226 } 227 } 228 } 229 230 *argv_nr = i; 231 *tmp_nr = k; 232 return 0; 233 } 234 235 static const char * const tlb_access[] = { 236 "N/A", 237 "HIT", 238 "MISS", 239 "L1", 240 "L2", 241 "Walker", 242 "Fault", 243 }; 244 245 int perf_mem__tlb_scnprintf(char *out, size_t sz, struct mem_info *mem_info) 246 { 247 size_t l = 0, i; 248 u64 m = PERF_MEM_TLB_NA; 249 u64 hit, miss; 250 251 sz -= 1; /* -1 for null termination */ 252 out[0] = '\0'; 253 254 if (mem_info) 255 m = mem_info->data_src.mem_dtlb; 256 257 hit = m & PERF_MEM_TLB_HIT; 258 miss = m & PERF_MEM_TLB_MISS; 259 260 /* already taken care of */ 261 m &= ~(PERF_MEM_TLB_HIT|PERF_MEM_TLB_MISS); 262 263 for (i = 0; m && i < ARRAY_SIZE(tlb_access); i++, m >>= 1) { 264 if (!(m & 0x1)) 265 continue; 266 if (l) { 267 strcat(out, " or "); 268 l += 4; 269 } 270 l += scnprintf(out + l, sz - l, tlb_access[i]); 271 } 272 if (*out == '\0') 273 l += scnprintf(out, sz - l, "N/A"); 274 if (hit) 275 l += scnprintf(out + l, sz - l, " hit"); 276 if (miss) 277 l += scnprintf(out + l, sz - l, " miss"); 278 279 return l; 280 } 281 282 static const char * const mem_lvl[] = { 283 "N/A", 284 "HIT", 285 "MISS", 286 "L1", 287 "LFB/MAB", 288 "L2", 289 "L3", 290 "Local RAM", 291 "Remote RAM (1 hop)", 292 "Remote RAM (2 hops)", 293 "Remote Cache (1 hop)", 294 "Remote Cache (2 hops)", 295 "I/O", 296 "Uncached", 297 }; 298 299 static const char * const mem_lvlnum[] = { 300 [PERF_MEM_LVLNUM_UNC] = "Uncached", 301 [PERF_MEM_LVLNUM_CXL] = "CXL", 302 [PERF_MEM_LVLNUM_IO] = "I/O", 303 [PERF_MEM_LVLNUM_ANY_CACHE] = "Any cache", 304 [PERF_MEM_LVLNUM_LFB] = "LFB/MAB", 305 [PERF_MEM_LVLNUM_RAM] = "RAM", 306 [PERF_MEM_LVLNUM_PMEM] = "PMEM", 307 [PERF_MEM_LVLNUM_NA] = "N/A", 308 }; 309 310 static const char * const mem_hops[] = { 311 "N/A", 312 /* 313 * While printing, 'Remote' will be added to represent 314 * 'Remote core, same node' accesses as remote field need 315 * to be set with mem_hops field. 316 */ 317 "core, same node", 318 "node, same socket", 319 "socket, same board", 320 "board", 321 }; 322 323 static int perf_mem__op_scnprintf(char *out, size_t sz, struct mem_info *mem_info) 324 { 325 u64 op = PERF_MEM_LOCK_NA; 326 int l; 327 328 if (mem_info) 329 op = mem_info->data_src.mem_op; 330 331 if (op & PERF_MEM_OP_NA) 332 l = scnprintf(out, sz, "N/A"); 333 else if (op & PERF_MEM_OP_LOAD) 334 l = scnprintf(out, sz, "LOAD"); 335 else if (op & PERF_MEM_OP_STORE) 336 l = scnprintf(out, sz, "STORE"); 337 else if (op & PERF_MEM_OP_PFETCH) 338 l = scnprintf(out, sz, "PFETCH"); 339 else if (op & PERF_MEM_OP_EXEC) 340 l = scnprintf(out, sz, "EXEC"); 341 else 342 l = scnprintf(out, sz, "No"); 343 344 return l; 345 } 346 347 int perf_mem__lvl_scnprintf(char *out, size_t sz, struct mem_info *mem_info) 348 { 349 union perf_mem_data_src data_src; 350 int printed = 0; 351 size_t l = 0; 352 size_t i; 353 int lvl; 354 char hit_miss[5] = {0}; 355 356 sz -= 1; /* -1 for null termination */ 357 out[0] = '\0'; 358 359 if (!mem_info) 360 goto na; 361 362 data_src = mem_info->data_src; 363 364 if (data_src.mem_lvl & PERF_MEM_LVL_HIT) 365 memcpy(hit_miss, "hit", 3); 366 else if (data_src.mem_lvl & PERF_MEM_LVL_MISS) 367 memcpy(hit_miss, "miss", 4); 368 369 lvl = data_src.mem_lvl_num; 370 if (lvl && lvl != PERF_MEM_LVLNUM_NA) { 371 if (data_src.mem_remote) { 372 strcat(out, "Remote "); 373 l += 7; 374 } 375 376 if (data_src.mem_hops) 377 l += scnprintf(out + l, sz - l, "%s ", mem_hops[data_src.mem_hops]); 378 379 if (mem_lvlnum[lvl]) 380 l += scnprintf(out + l, sz - l, mem_lvlnum[lvl]); 381 else 382 l += scnprintf(out + l, sz - l, "L%d", lvl); 383 384 l += scnprintf(out + l, sz - l, " %s", hit_miss); 385 return l; 386 } 387 388 lvl = data_src.mem_lvl; 389 if (!lvl) 390 goto na; 391 392 lvl &= ~(PERF_MEM_LVL_NA | PERF_MEM_LVL_HIT | PERF_MEM_LVL_MISS); 393 if (!lvl) 394 goto na; 395 396 for (i = 0; lvl && i < ARRAY_SIZE(mem_lvl); i++, lvl >>= 1) { 397 if (!(lvl & 0x1)) 398 continue; 399 if (printed++) { 400 strcat(out, " or "); 401 l += 4; 402 } 403 l += scnprintf(out + l, sz - l, mem_lvl[i]); 404 } 405 406 if (printed) { 407 l += scnprintf(out + l, sz - l, " %s", hit_miss); 408 return l; 409 } 410 411 na: 412 strcat(out, "N/A"); 413 return 3; 414 } 415 416 static const char * const snoop_access[] = { 417 "N/A", 418 "None", 419 "Hit", 420 "Miss", 421 "HitM", 422 }; 423 424 static const char * const snoopx_access[] = { 425 "Fwd", 426 "Peer", 427 }; 428 429 int perf_mem__snp_scnprintf(char *out, size_t sz, struct mem_info *mem_info) 430 { 431 size_t i, l = 0; 432 u64 m = PERF_MEM_SNOOP_NA; 433 434 sz -= 1; /* -1 for null termination */ 435 out[0] = '\0'; 436 437 if (mem_info) 438 m = mem_info->data_src.mem_snoop; 439 440 for (i = 0; m && i < ARRAY_SIZE(snoop_access); i++, m >>= 1) { 441 if (!(m & 0x1)) 442 continue; 443 if (l) { 444 strcat(out, " or "); 445 l += 4; 446 } 447 l += scnprintf(out + l, sz - l, snoop_access[i]); 448 } 449 450 m = 0; 451 if (mem_info) 452 m = mem_info->data_src.mem_snoopx; 453 454 for (i = 0; m && i < ARRAY_SIZE(snoopx_access); i++, m >>= 1) { 455 if (!(m & 0x1)) 456 continue; 457 458 if (l) { 459 strcat(out, " or "); 460 l += 4; 461 } 462 l += scnprintf(out + l, sz - l, snoopx_access[i]); 463 } 464 465 if (*out == '\0') 466 l += scnprintf(out, sz - l, "N/A"); 467 468 return l; 469 } 470 471 int perf_mem__lck_scnprintf(char *out, size_t sz, struct mem_info *mem_info) 472 { 473 u64 mask = PERF_MEM_LOCK_NA; 474 int l; 475 476 if (mem_info) 477 mask = mem_info->data_src.mem_lock; 478 479 if (mask & PERF_MEM_LOCK_NA) 480 l = scnprintf(out, sz, "N/A"); 481 else if (mask & PERF_MEM_LOCK_LOCKED) 482 l = scnprintf(out, sz, "Yes"); 483 else 484 l = scnprintf(out, sz, "No"); 485 486 return l; 487 } 488 489 int perf_mem__blk_scnprintf(char *out, size_t sz, struct mem_info *mem_info) 490 { 491 size_t l = 0; 492 u64 mask = PERF_MEM_BLK_NA; 493 494 sz -= 1; /* -1 for null termination */ 495 out[0] = '\0'; 496 497 if (mem_info) 498 mask = mem_info->data_src.mem_blk; 499 500 if (!mask || (mask & PERF_MEM_BLK_NA)) { 501 l += scnprintf(out + l, sz - l, " N/A"); 502 return l; 503 } 504 if (mask & PERF_MEM_BLK_DATA) 505 l += scnprintf(out + l, sz - l, " Data"); 506 if (mask & PERF_MEM_BLK_ADDR) 507 l += scnprintf(out + l, sz - l, " Addr"); 508 509 return l; 510 } 511 512 int perf_script__meminfo_scnprintf(char *out, size_t sz, struct mem_info *mem_info) 513 { 514 int i = 0; 515 516 i += scnprintf(out, sz, "|OP "); 517 i += perf_mem__op_scnprintf(out + i, sz - i, mem_info); 518 i += scnprintf(out + i, sz - i, "|LVL "); 519 i += perf_mem__lvl_scnprintf(out + i, sz, mem_info); 520 i += scnprintf(out + i, sz - i, "|SNP "); 521 i += perf_mem__snp_scnprintf(out + i, sz - i, mem_info); 522 i += scnprintf(out + i, sz - i, "|TLB "); 523 i += perf_mem__tlb_scnprintf(out + i, sz - i, mem_info); 524 i += scnprintf(out + i, sz - i, "|LCK "); 525 i += perf_mem__lck_scnprintf(out + i, sz - i, mem_info); 526 i += scnprintf(out + i, sz - i, "|BLK "); 527 i += perf_mem__blk_scnprintf(out + i, sz - i, mem_info); 528 529 return i; 530 } 531 532 int c2c_decode_stats(struct c2c_stats *stats, struct mem_info *mi) 533 { 534 union perf_mem_data_src *data_src = &mi->data_src; 535 u64 daddr = mi->daddr.addr; 536 u64 op = data_src->mem_op; 537 u64 lvl = data_src->mem_lvl; 538 u64 snoop = data_src->mem_snoop; 539 u64 snoopx = data_src->mem_snoopx; 540 u64 lock = data_src->mem_lock; 541 u64 blk = data_src->mem_blk; 542 /* 543 * Skylake might report unknown remote level via this 544 * bit, consider it when evaluating remote HITMs. 545 * 546 * Incase of power, remote field can also be used to denote cache 547 * accesses from the another core of same node. Hence, setting 548 * mrem only when HOPS is zero along with set remote field. 549 */ 550 bool mrem = (data_src->mem_remote && !data_src->mem_hops); 551 int err = 0; 552 553 #define HITM_INC(__f) \ 554 do { \ 555 stats->__f++; \ 556 stats->tot_hitm++; \ 557 } while (0) 558 559 #define PEER_INC(__f) \ 560 do { \ 561 stats->__f++; \ 562 stats->tot_peer++; \ 563 } while (0) 564 565 #define P(a, b) PERF_MEM_##a##_##b 566 567 stats->nr_entries++; 568 569 if (lock & P(LOCK, LOCKED)) stats->locks++; 570 571 if (blk & P(BLK, DATA)) stats->blk_data++; 572 if (blk & P(BLK, ADDR)) stats->blk_addr++; 573 574 if (op & P(OP, LOAD)) { 575 /* load */ 576 stats->load++; 577 578 if (!daddr) { 579 stats->ld_noadrs++; 580 return -1; 581 } 582 583 if (lvl & P(LVL, HIT)) { 584 if (lvl & P(LVL, UNC)) stats->ld_uncache++; 585 if (lvl & P(LVL, IO)) stats->ld_io++; 586 if (lvl & P(LVL, LFB)) stats->ld_fbhit++; 587 if (lvl & P(LVL, L1 )) stats->ld_l1hit++; 588 if (lvl & P(LVL, L2)) { 589 stats->ld_l2hit++; 590 591 if (snoopx & P(SNOOPX, PEER)) 592 PEER_INC(lcl_peer); 593 } 594 if (lvl & P(LVL, L3 )) { 595 if (snoop & P(SNOOP, HITM)) 596 HITM_INC(lcl_hitm); 597 else 598 stats->ld_llchit++; 599 600 if (snoopx & P(SNOOPX, PEER)) 601 PEER_INC(lcl_peer); 602 } 603 604 if (lvl & P(LVL, LOC_RAM)) { 605 stats->lcl_dram++; 606 if (snoop & P(SNOOP, HIT)) 607 stats->ld_shared++; 608 else 609 stats->ld_excl++; 610 } 611 612 if ((lvl & P(LVL, REM_RAM1)) || 613 (lvl & P(LVL, REM_RAM2)) || 614 mrem) { 615 stats->rmt_dram++; 616 if (snoop & P(SNOOP, HIT)) 617 stats->ld_shared++; 618 else 619 stats->ld_excl++; 620 } 621 } 622 623 if ((lvl & P(LVL, REM_CCE1)) || 624 (lvl & P(LVL, REM_CCE2)) || 625 mrem) { 626 if (snoop & P(SNOOP, HIT)) { 627 stats->rmt_hit++; 628 } else if (snoop & P(SNOOP, HITM)) { 629 HITM_INC(rmt_hitm); 630 } else if (snoopx & P(SNOOPX, PEER)) { 631 stats->rmt_hit++; 632 PEER_INC(rmt_peer); 633 } 634 } 635 636 if ((lvl & P(LVL, MISS))) 637 stats->ld_miss++; 638 639 } else if (op & P(OP, STORE)) { 640 /* store */ 641 stats->store++; 642 643 if (!daddr) { 644 stats->st_noadrs++; 645 return -1; 646 } 647 648 if (lvl & P(LVL, HIT)) { 649 if (lvl & P(LVL, UNC)) stats->st_uncache++; 650 if (lvl & P(LVL, L1 )) stats->st_l1hit++; 651 } 652 if (lvl & P(LVL, MISS)) 653 if (lvl & P(LVL, L1)) stats->st_l1miss++; 654 if (lvl & P(LVL, NA)) 655 stats->st_na++; 656 } else { 657 /* unparsable data_src? */ 658 stats->noparse++; 659 return -1; 660 } 661 662 if (!mi->daddr.ms.map || !mi->iaddr.ms.map) { 663 stats->nomap++; 664 return -1; 665 } 666 667 #undef P 668 #undef HITM_INC 669 return err; 670 } 671 672 void c2c_add_stats(struct c2c_stats *stats, struct c2c_stats *add) 673 { 674 stats->nr_entries += add->nr_entries; 675 676 stats->locks += add->locks; 677 stats->store += add->store; 678 stats->st_uncache += add->st_uncache; 679 stats->st_noadrs += add->st_noadrs; 680 stats->st_l1hit += add->st_l1hit; 681 stats->st_l1miss += add->st_l1miss; 682 stats->st_na += add->st_na; 683 stats->load += add->load; 684 stats->ld_excl += add->ld_excl; 685 stats->ld_shared += add->ld_shared; 686 stats->ld_uncache += add->ld_uncache; 687 stats->ld_io += add->ld_io; 688 stats->ld_miss += add->ld_miss; 689 stats->ld_noadrs += add->ld_noadrs; 690 stats->ld_fbhit += add->ld_fbhit; 691 stats->ld_l1hit += add->ld_l1hit; 692 stats->ld_l2hit += add->ld_l2hit; 693 stats->ld_llchit += add->ld_llchit; 694 stats->lcl_hitm += add->lcl_hitm; 695 stats->rmt_hitm += add->rmt_hitm; 696 stats->tot_hitm += add->tot_hitm; 697 stats->lcl_peer += add->lcl_peer; 698 stats->rmt_peer += add->rmt_peer; 699 stats->tot_peer += add->tot_peer; 700 stats->rmt_hit += add->rmt_hit; 701 stats->lcl_dram += add->lcl_dram; 702 stats->rmt_dram += add->rmt_dram; 703 stats->blk_data += add->blk_data; 704 stats->blk_addr += add->blk_addr; 705 stats->nomap += add->nomap; 706 stats->noparse += add->noparse; 707 } 708