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