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