1 // SPDX-License-Identifier: GPL-2.0 2 #include "builtin.h" 3 4 #include "util/dso.h" 5 #include "util/evlist.h" 6 #include "util/evsel.h" 7 #include "util/config.h" 8 #include "util/map.h" 9 #include "util/symbol.h" 10 #include "util/thread.h" 11 #include "util/header.h" 12 #include "util/session.h" 13 #include "util/tool.h" 14 #include "util/callchain.h" 15 #include "util/time-utils.h" 16 #include <linux/err.h> 17 18 #include <subcmd/pager.h> 19 #include <subcmd/parse-options.h> 20 #include "util/trace-event.h" 21 #include "util/data.h" 22 #include "util/cpumap.h" 23 24 #include "util/debug.h" 25 #include "util/event.h" 26 #include "util/string2.h" 27 #include "util/util.h" 28 29 #include <linux/kernel.h> 30 #include <linux/numa.h> 31 #include <linux/rbtree.h> 32 #include <linux/string.h> 33 #include <linux/zalloc.h> 34 #include <errno.h> 35 #include <inttypes.h> 36 #include <locale.h> 37 #include <regex.h> 38 39 #include <linux/ctype.h> 40 #include <event-parse.h> 41 42 static int kmem_slab; 43 static int kmem_page; 44 45 static long kmem_page_size; 46 static enum { 47 KMEM_SLAB, 48 KMEM_PAGE, 49 } kmem_default = KMEM_SLAB; /* for backward compatibility */ 50 51 struct alloc_stat; 52 typedef int (*sort_fn_t)(void *, void *); 53 54 static int alloc_flag; 55 static int caller_flag; 56 57 static int alloc_lines = -1; 58 static int caller_lines = -1; 59 60 static bool raw_ip; 61 62 struct alloc_stat { 63 u64 call_site; 64 u64 ptr; 65 u64 bytes_req; 66 u64 bytes_alloc; 67 u64 last_alloc; 68 u32 hit; 69 u32 pingpong; 70 71 short alloc_cpu; 72 73 struct rb_node node; 74 }; 75 76 static struct rb_root root_alloc_stat; 77 static struct rb_root root_alloc_sorted; 78 static struct rb_root root_caller_stat; 79 static struct rb_root root_caller_sorted; 80 81 static unsigned long total_requested, total_allocated, total_freed; 82 static unsigned long nr_allocs, nr_cross_allocs; 83 84 /* filters for controlling start and stop of time of analysis */ 85 static struct perf_time_interval ptime; 86 static const char *time_str; 87 88 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr, 89 int bytes_req, int bytes_alloc, int cpu) 90 { 91 struct rb_node **node = &root_alloc_stat.rb_node; 92 struct rb_node *parent = NULL; 93 struct alloc_stat *data = NULL; 94 95 while (*node) { 96 parent = *node; 97 data = rb_entry(*node, struct alloc_stat, node); 98 99 if (ptr > data->ptr) 100 node = &(*node)->rb_right; 101 else if (ptr < data->ptr) 102 node = &(*node)->rb_left; 103 else 104 break; 105 } 106 107 if (data && data->ptr == ptr) { 108 data->hit++; 109 data->bytes_req += bytes_req; 110 data->bytes_alloc += bytes_alloc; 111 } else { 112 data = malloc(sizeof(*data)); 113 if (!data) { 114 pr_err("%s: malloc failed\n", __func__); 115 return -1; 116 } 117 data->ptr = ptr; 118 data->pingpong = 0; 119 data->hit = 1; 120 data->bytes_req = bytes_req; 121 data->bytes_alloc = bytes_alloc; 122 123 rb_link_node(&data->node, parent, node); 124 rb_insert_color(&data->node, &root_alloc_stat); 125 } 126 data->call_site = call_site; 127 data->alloc_cpu = cpu; 128 data->last_alloc = bytes_alloc; 129 130 return 0; 131 } 132 133 static int insert_caller_stat(unsigned long call_site, 134 int bytes_req, int bytes_alloc) 135 { 136 struct rb_node **node = &root_caller_stat.rb_node; 137 struct rb_node *parent = NULL; 138 struct alloc_stat *data = NULL; 139 140 while (*node) { 141 parent = *node; 142 data = rb_entry(*node, struct alloc_stat, node); 143 144 if (call_site > data->call_site) 145 node = &(*node)->rb_right; 146 else if (call_site < data->call_site) 147 node = &(*node)->rb_left; 148 else 149 break; 150 } 151 152 if (data && data->call_site == call_site) { 153 data->hit++; 154 data->bytes_req += bytes_req; 155 data->bytes_alloc += bytes_alloc; 156 } else { 157 data = malloc(sizeof(*data)); 158 if (!data) { 159 pr_err("%s: malloc failed\n", __func__); 160 return -1; 161 } 162 data->call_site = call_site; 163 data->pingpong = 0; 164 data->hit = 1; 165 data->bytes_req = bytes_req; 166 data->bytes_alloc = bytes_alloc; 167 168 rb_link_node(&data->node, parent, node); 169 rb_insert_color(&data->node, &root_caller_stat); 170 } 171 172 return 0; 173 } 174 175 static int evsel__process_alloc_event(struct perf_sample *sample) 176 { 177 unsigned long ptr = perf_sample__intval(sample, "ptr"), 178 call_site = perf_sample__intval(sample, "call_site"); 179 int bytes_req = perf_sample__intval(sample, "bytes_req"), 180 bytes_alloc = perf_sample__intval(sample, "bytes_alloc"); 181 182 if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) || 183 insert_caller_stat(call_site, bytes_req, bytes_alloc)) 184 return -1; 185 186 total_requested += bytes_req; 187 total_allocated += bytes_alloc; 188 189 nr_allocs++; 190 191 /* 192 * Commit 11e9734bcb6a ("mm/slab_common: unify NUMA and UMA 193 * version of tracepoints") adds the field "node" into the 194 * tracepoints 'kmalloc' and 'kmem_cache_alloc'. 195 * 196 * The legacy tracepoints 'kmalloc_node' and 'kmem_cache_alloc_node' 197 * also contain the field "node". 198 * 199 * If the tracepoint contains the field "node" the tool stats the 200 * cross allocation. 201 */ 202 if (evsel__field(sample->evsel, "node")) { 203 int node1, node2; 204 205 node1 = cpu__get_node((struct perf_cpu){.cpu = sample->cpu}); 206 node2 = perf_sample__intval(sample, "node"); 207 208 /* 209 * If the field "node" is NUMA_NO_NODE (-1), we don't take it 210 * as a cross allocation. 211 */ 212 if ((node2 != NUMA_NO_NODE) && (node1 != node2)) 213 nr_cross_allocs++; 214 } 215 216 return 0; 217 } 218 219 static int ptr_cmp(void *, void *); 220 static int slab_callsite_cmp(void *, void *); 221 222 static struct alloc_stat *search_alloc_stat(unsigned long ptr, 223 unsigned long call_site, 224 struct rb_root *root, 225 sort_fn_t sort_fn) 226 { 227 struct rb_node *node = root->rb_node; 228 struct alloc_stat key = { .ptr = ptr, .call_site = call_site }; 229 230 while (node) { 231 struct alloc_stat *data; 232 int cmp; 233 234 data = rb_entry(node, struct alloc_stat, node); 235 236 cmp = sort_fn(&key, data); 237 if (cmp < 0) 238 node = node->rb_left; 239 else if (cmp > 0) 240 node = node->rb_right; 241 else 242 return data; 243 } 244 return NULL; 245 } 246 247 static int evsel__process_free_event(struct perf_sample *sample) 248 { 249 unsigned long ptr = perf_sample__intval(sample, "ptr"); 250 struct alloc_stat *s_alloc, *s_caller; 251 252 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp); 253 if (!s_alloc) 254 return 0; 255 256 total_freed += s_alloc->last_alloc; 257 258 if ((short)sample->cpu != s_alloc->alloc_cpu) { 259 s_alloc->pingpong++; 260 261 s_caller = search_alloc_stat(0, s_alloc->call_site, 262 &root_caller_stat, 263 slab_callsite_cmp); 264 if (!s_caller) 265 return -1; 266 s_caller->pingpong++; 267 } 268 s_alloc->alloc_cpu = -1; 269 270 return 0; 271 } 272 273 static u64 total_page_alloc_bytes; 274 static u64 total_page_free_bytes; 275 static u64 total_page_nomatch_bytes; 276 static u64 total_page_fail_bytes; 277 static unsigned long nr_page_allocs; 278 static unsigned long nr_page_frees; 279 static unsigned long nr_page_fails; 280 static unsigned long nr_page_nomatch; 281 282 static bool use_pfn; 283 static bool live_page; 284 static struct perf_session *kmem_session; 285 286 #define MAX_MIGRATE_TYPES 6 287 #define MAX_PAGE_ORDER 11 288 289 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES]; 290 291 struct page_stat { 292 struct rb_node node; 293 u64 page; 294 u64 callsite; 295 int order; 296 unsigned gfp_flags; 297 unsigned migrate_type; 298 u64 alloc_bytes; 299 u64 free_bytes; 300 int nr_alloc; 301 int nr_free; 302 }; 303 304 static struct rb_root page_live_tree; 305 static struct rb_root page_alloc_tree; 306 static struct rb_root page_alloc_sorted; 307 static struct rb_root page_caller_tree; 308 static struct rb_root page_caller_sorted; 309 310 struct alloc_func { 311 u64 start; 312 u64 end; 313 char *name; 314 }; 315 316 static int nr_alloc_funcs; 317 static struct alloc_func *alloc_func_list; 318 319 static int funcmp(const void *a, const void *b) 320 { 321 const struct alloc_func *fa = a; 322 const struct alloc_func *fb = b; 323 324 if (fa->start > fb->start) 325 return 1; 326 else 327 return -1; 328 } 329 330 static int callcmp(const void *a, const void *b) 331 { 332 const struct alloc_func *fa = a; 333 const struct alloc_func *fb = b; 334 335 if (fb->start <= fa->start && fa->end < fb->end) 336 return 0; 337 338 if (fa->start > fb->start) 339 return 1; 340 else 341 return -1; 342 } 343 344 static int build_alloc_func_list(void) 345 { 346 int ret; 347 struct map *kernel_map; 348 struct symbol *sym; 349 struct rb_node *node; 350 struct alloc_func *func; 351 struct machine *machine = &kmem_session->machines.host; 352 regex_t alloc_func_regex; 353 static const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?"; 354 355 ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED); 356 if (ret) { 357 char err[BUFSIZ]; 358 359 regerror(ret, &alloc_func_regex, err, sizeof(err)); 360 pr_err("Invalid regex: %s\n%s", pattern, err); 361 return -EINVAL; 362 } 363 364 kernel_map = machine__kernel_map(machine); 365 if (map__load(kernel_map) < 0) { 366 pr_err("cannot load kernel map\n"); 367 return -ENOENT; 368 } 369 370 map__for_each_symbol(kernel_map, sym, node) { 371 if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0)) 372 continue; 373 374 func = realloc(alloc_func_list, 375 (nr_alloc_funcs + 1) * sizeof(*func)); 376 if (func == NULL) 377 return -ENOMEM; 378 379 pr_debug("alloc func: %s\n", sym->name); 380 func[nr_alloc_funcs].start = sym->start; 381 func[nr_alloc_funcs].end = sym->end; 382 func[nr_alloc_funcs].name = sym->name; 383 384 alloc_func_list = func; 385 nr_alloc_funcs++; 386 } 387 388 qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp); 389 390 regfree(&alloc_func_regex); 391 return 0; 392 } 393 394 /* 395 * Find first non-memory allocation function from callchain. 396 * The allocation functions are in the 'alloc_func_list'. 397 */ 398 static u64 find_callsite(struct perf_sample *sample) 399 { 400 struct addr_location al; 401 struct machine *machine = &kmem_session->machines.host; 402 struct callchain_cursor_node *node; 403 struct callchain_cursor *cursor; 404 u64 result = sample->ip; 405 406 addr_location__init(&al); 407 if (alloc_func_list == NULL) { 408 if (build_alloc_func_list() < 0) 409 goto out; 410 } 411 412 al.thread = machine__findnew_thread(machine, sample->pid, sample->tid); 413 414 cursor = get_tls_callchain_cursor(); 415 if (cursor == NULL) 416 goto out; 417 418 sample__resolve_callchain(sample, cursor, /*parent=*/NULL, &al, 16); 419 420 callchain_cursor_commit(cursor); 421 while (true) { 422 struct alloc_func key, *caller; 423 u64 addr; 424 425 node = callchain_cursor_current(cursor); 426 if (node == NULL) 427 break; 428 429 key.start = key.end = node->ip; 430 caller = bsearch(&key, alloc_func_list, nr_alloc_funcs, 431 sizeof(key), callcmp); 432 if (!caller) { 433 /* found */ 434 if (node->ms.map) 435 addr = map__dso_unmap_ip(node->ms.map, node->ip); 436 else 437 addr = node->ip; 438 439 result = addr; 440 goto out; 441 } else 442 pr_debug3("skipping alloc function: %s\n", caller->name); 443 444 callchain_cursor_advance(cursor); 445 } 446 447 pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip); 448 out: 449 addr_location__exit(&al); 450 return result; 451 } 452 453 struct sort_dimension { 454 const char name[20]; 455 sort_fn_t cmp; 456 struct list_head list; 457 }; 458 459 static LIST_HEAD(page_alloc_sort_input); 460 static LIST_HEAD(page_caller_sort_input); 461 462 static struct page_stat * 463 __page_stat__findnew_page(struct page_stat *pstat, bool create) 464 { 465 struct rb_node **node = &page_live_tree.rb_node; 466 struct rb_node *parent = NULL; 467 struct page_stat *data; 468 469 while (*node) { 470 s64 cmp; 471 472 parent = *node; 473 data = rb_entry(*node, struct page_stat, node); 474 475 cmp = data->page - pstat->page; 476 if (cmp < 0) 477 node = &parent->rb_left; 478 else if (cmp > 0) 479 node = &parent->rb_right; 480 else 481 return data; 482 } 483 484 if (!create) 485 return NULL; 486 487 data = zalloc(sizeof(*data)); 488 if (data != NULL) { 489 data->page = pstat->page; 490 data->order = pstat->order; 491 data->gfp_flags = pstat->gfp_flags; 492 data->migrate_type = pstat->migrate_type; 493 494 rb_link_node(&data->node, parent, node); 495 rb_insert_color(&data->node, &page_live_tree); 496 } 497 498 return data; 499 } 500 501 static struct page_stat *page_stat__find_page(struct page_stat *pstat) 502 { 503 return __page_stat__findnew_page(pstat, false); 504 } 505 506 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat) 507 { 508 return __page_stat__findnew_page(pstat, true); 509 } 510 511 static struct page_stat * 512 __page_stat__findnew_alloc(struct page_stat *pstat, bool create) 513 { 514 struct rb_node **node = &page_alloc_tree.rb_node; 515 struct rb_node *parent = NULL; 516 struct page_stat *data; 517 struct sort_dimension *sort; 518 519 while (*node) { 520 int cmp = 0; 521 522 parent = *node; 523 data = rb_entry(*node, struct page_stat, node); 524 525 list_for_each_entry(sort, &page_alloc_sort_input, list) { 526 cmp = sort->cmp(pstat, data); 527 if (cmp) 528 break; 529 } 530 531 if (cmp < 0) 532 node = &parent->rb_left; 533 else if (cmp > 0) 534 node = &parent->rb_right; 535 else 536 return data; 537 } 538 539 if (!create) 540 return NULL; 541 542 data = zalloc(sizeof(*data)); 543 if (data != NULL) { 544 data->page = pstat->page; 545 data->order = pstat->order; 546 data->gfp_flags = pstat->gfp_flags; 547 data->migrate_type = pstat->migrate_type; 548 549 rb_link_node(&data->node, parent, node); 550 rb_insert_color(&data->node, &page_alloc_tree); 551 } 552 553 return data; 554 } 555 556 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat) 557 { 558 return __page_stat__findnew_alloc(pstat, false); 559 } 560 561 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat) 562 { 563 return __page_stat__findnew_alloc(pstat, true); 564 } 565 566 static struct page_stat * 567 __page_stat__findnew_caller(struct page_stat *pstat, bool create) 568 { 569 struct rb_node **node = &page_caller_tree.rb_node; 570 struct rb_node *parent = NULL; 571 struct page_stat *data; 572 struct sort_dimension *sort; 573 574 while (*node) { 575 int cmp = 0; 576 577 parent = *node; 578 data = rb_entry(*node, struct page_stat, node); 579 580 list_for_each_entry(sort, &page_caller_sort_input, list) { 581 cmp = sort->cmp(pstat, data); 582 if (cmp) 583 break; 584 } 585 586 if (cmp < 0) 587 node = &parent->rb_left; 588 else if (cmp > 0) 589 node = &parent->rb_right; 590 else 591 return data; 592 } 593 594 if (!create) 595 return NULL; 596 597 data = zalloc(sizeof(*data)); 598 if (data != NULL) { 599 data->callsite = pstat->callsite; 600 data->order = pstat->order; 601 data->gfp_flags = pstat->gfp_flags; 602 data->migrate_type = pstat->migrate_type; 603 604 rb_link_node(&data->node, parent, node); 605 rb_insert_color(&data->node, &page_caller_tree); 606 } 607 608 return data; 609 } 610 611 static struct page_stat *page_stat__find_caller(struct page_stat *pstat) 612 { 613 return __page_stat__findnew_caller(pstat, false); 614 } 615 616 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat) 617 { 618 return __page_stat__findnew_caller(pstat, true); 619 } 620 621 static bool valid_page(u64 pfn_or_page) 622 { 623 if (use_pfn && pfn_or_page == -1UL) 624 return false; 625 if (!use_pfn && pfn_or_page == 0) 626 return false; 627 return true; 628 } 629 630 struct gfp_flag { 631 unsigned int flags; 632 char *compact_str; 633 char *human_readable; 634 }; 635 636 static struct gfp_flag *gfps; 637 static int nr_gfps; 638 639 static int gfpcmp(const void *a, const void *b) 640 { 641 const struct gfp_flag *fa = a; 642 const struct gfp_flag *fb = b; 643 644 return fa->flags - fb->flags; 645 } 646 647 /* see include/trace/events/mmflags.h */ 648 static const struct { 649 const char *original; 650 const char *compact; 651 } gfp_compact_table[] = { 652 { "GFP_TRANSHUGE", "THP" }, 653 { "GFP_TRANSHUGE_LIGHT", "THL" }, 654 { "GFP_HIGHUSER_MOVABLE", "HUM" }, 655 { "GFP_HIGHUSER", "HU" }, 656 { "GFP_USER", "U" }, 657 { "GFP_KERNEL_ACCOUNT", "KAC" }, 658 { "GFP_KERNEL", "K" }, 659 { "GFP_NOFS", "NF" }, 660 { "GFP_ATOMIC", "A" }, 661 { "GFP_NOIO", "NI" }, 662 { "GFP_NOWAIT", "NW" }, 663 { "GFP_DMA", "D" }, 664 { "__GFP_HIGHMEM", "HM" }, 665 { "GFP_DMA32", "D32" }, 666 { "__GFP_HIGH", "H" }, 667 { "__GFP_IO", "I" }, 668 { "__GFP_FS", "F" }, 669 { "__GFP_NOWARN", "NWR" }, 670 { "__GFP_RETRY_MAYFAIL", "R" }, 671 { "__GFP_NOFAIL", "NF" }, 672 { "__GFP_NORETRY", "NR" }, 673 { "__GFP_COMP", "C" }, 674 { "__GFP_ZERO", "Z" }, 675 { "__GFP_NOMEMALLOC", "NMA" }, 676 { "__GFP_MEMALLOC", "MA" }, 677 { "__GFP_HARDWALL", "HW" }, 678 { "__GFP_THISNODE", "TN" }, 679 { "__GFP_RECLAIMABLE", "RC" }, 680 { "__GFP_MOVABLE", "M" }, 681 { "__GFP_ACCOUNT", "AC" }, 682 { "__GFP_WRITE", "WR" }, 683 { "__GFP_RECLAIM", "R" }, 684 { "__GFP_DIRECT_RECLAIM", "DR" }, 685 { "__GFP_KSWAPD_RECLAIM", "KR" }, 686 }; 687 688 static size_t max_gfp_len; 689 690 static char *compact_gfp_flags(char *gfp_flags) 691 { 692 char *orig_flags = strdup(gfp_flags); 693 char *new_flags = NULL; 694 char *str, *pos = NULL; 695 size_t len = 0; 696 697 if (orig_flags == NULL) 698 return NULL; 699 700 str = strtok_r(orig_flags, "|", &pos); 701 while (str) { 702 size_t i; 703 char *new; 704 const char *cpt; 705 706 for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) { 707 if (strcmp(gfp_compact_table[i].original, str)) 708 continue; 709 710 cpt = gfp_compact_table[i].compact; 711 new = realloc(new_flags, len + strlen(cpt) + 2); 712 if (new == NULL) { 713 free(new_flags); 714 free(orig_flags); 715 return NULL; 716 } 717 718 new_flags = new; 719 720 if (!len) { 721 strcpy(new_flags, cpt); 722 } else { 723 strcat(new_flags, "|"); 724 strcat(new_flags, cpt); 725 len++; 726 } 727 728 len += strlen(cpt); 729 } 730 731 str = strtok_r(NULL, "|", &pos); 732 } 733 734 if (max_gfp_len < len) 735 max_gfp_len = len; 736 737 free(orig_flags); 738 return new_flags; 739 } 740 741 static char *compact_gfp_string(unsigned long gfp_flags) 742 { 743 struct gfp_flag key = { 744 .flags = gfp_flags, 745 }; 746 struct gfp_flag *gfp; 747 748 gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp); 749 if (gfp) 750 return gfp->compact_str; 751 752 return NULL; 753 } 754 755 static int parse_gfp_flags(struct perf_sample *sample, unsigned int gfp_flags) 756 { 757 struct tep_record record = { 758 .cpu = sample->cpu, 759 .data = sample->raw_data, 760 .size = sample->raw_size, 761 }; 762 struct trace_seq seq; 763 char *str, *pos = NULL; 764 const struct tep_event *tp_format; 765 766 if (nr_gfps) { 767 struct gfp_flag key = { 768 .flags = gfp_flags, 769 }; 770 771 if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp)) 772 return 0; 773 } 774 775 trace_seq_init(&seq); 776 tp_format = evsel__tp_format(sample->evsel); 777 if (tp_format) 778 tep_print_event(tp_format->tep, &seq, &record, "%s", TEP_PRINT_INFO); 779 780 str = strtok_r(seq.buffer, " ", &pos); 781 while (str) { 782 if (!strncmp(str, "gfp_flags=", 10)) { 783 struct gfp_flag *new; 784 785 new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps)); 786 if (new == NULL) 787 goto err_out; 788 789 gfps = new; 790 new += nr_gfps; 791 792 new->flags = gfp_flags; 793 new->human_readable = strdup(str + 10); 794 if (!new->human_readable) 795 goto err_out; 796 new->compact_str = compact_gfp_flags(str + 10); 797 if (!new->compact_str) { 798 free(new->human_readable); 799 goto err_out; 800 } 801 nr_gfps++; 802 qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp); 803 } 804 805 str = strtok_r(NULL, " ", &pos); 806 } 807 808 trace_seq_destroy(&seq); 809 return 0; 810 err_out: 811 trace_seq_destroy(&seq); 812 return -ENOMEM; 813 } 814 815 static int evsel__process_page_alloc_event(struct perf_sample *sample) 816 { 817 u64 page; 818 unsigned int order = perf_sample__intval(sample, "order"); 819 unsigned int gfp_flags = perf_sample__intval(sample, "gfp_flags"); 820 unsigned int migrate_type = perf_sample__intval(sample, "migratetype"); 821 u64 bytes = kmem_page_size << order; 822 u64 callsite; 823 struct page_stat *pstat; 824 struct page_stat this = { 825 .order = order, 826 .gfp_flags = gfp_flags, 827 .migrate_type = migrate_type, 828 }; 829 830 if (order >= MAX_PAGE_ORDER) { 831 pr_debug("Out-of-bounds order %u\n", order); 832 return -1; 833 } 834 835 if (migrate_type >= MAX_MIGRATE_TYPES) { 836 pr_debug("Out-of-bounds migratetype %u\n", migrate_type); 837 return -1; 838 } 839 840 if (use_pfn) 841 page = perf_sample__intval(sample, "pfn"); 842 else 843 page = perf_sample__intval(sample, "page"); 844 845 nr_page_allocs++; 846 total_page_alloc_bytes += bytes; 847 848 if (!valid_page(page)) { 849 nr_page_fails++; 850 total_page_fail_bytes += bytes; 851 852 return 0; 853 } 854 855 if (parse_gfp_flags(sample, gfp_flags) < 0) 856 return -1; 857 858 callsite = find_callsite(sample); 859 860 /* 861 * This is to find the current page (with correct gfp flags and 862 * migrate type) at free event. 863 */ 864 this.page = page; 865 pstat = page_stat__findnew_page(&this); 866 if (pstat == NULL) 867 return -ENOMEM; 868 869 pstat->nr_alloc++; 870 pstat->alloc_bytes += bytes; 871 pstat->callsite = callsite; 872 873 if (!live_page) { 874 pstat = page_stat__findnew_alloc(&this); 875 if (pstat == NULL) 876 return -ENOMEM; 877 878 pstat->nr_alloc++; 879 pstat->alloc_bytes += bytes; 880 pstat->callsite = callsite; 881 } 882 883 this.callsite = callsite; 884 pstat = page_stat__findnew_caller(&this); 885 if (pstat == NULL) 886 return -ENOMEM; 887 888 pstat->nr_alloc++; 889 pstat->alloc_bytes += bytes; 890 891 order_stats[order][migrate_type]++; 892 893 return 0; 894 } 895 896 static int evsel__process_page_free_event(struct perf_sample *sample) 897 { 898 u64 page; 899 unsigned int order = perf_sample__intval(sample, "order"); 900 u64 bytes = kmem_page_size << order; 901 struct page_stat *pstat; 902 struct page_stat this = { 903 .order = order, 904 }; 905 906 if (order >= MAX_PAGE_ORDER) { 907 pr_debug("Out-of-bounds order %u\n", order); 908 return -1; 909 } 910 911 if (use_pfn) 912 page = perf_sample__intval(sample, "pfn"); 913 else 914 page = perf_sample__intval(sample, "page"); 915 916 nr_page_frees++; 917 total_page_free_bytes += bytes; 918 919 this.page = page; 920 pstat = page_stat__find_page(&this); 921 if (pstat == NULL) { 922 pr_debug2("missing free at page %"PRIx64" (order: %d)\n", 923 page, order); 924 925 nr_page_nomatch++; 926 total_page_nomatch_bytes += bytes; 927 928 return 0; 929 } 930 931 this.gfp_flags = pstat->gfp_flags; 932 this.migrate_type = pstat->migrate_type; 933 this.callsite = pstat->callsite; 934 935 rb_erase(&pstat->node, &page_live_tree); 936 free(pstat); 937 938 if (live_page) { 939 order_stats[this.order][this.migrate_type]--; 940 } else { 941 pstat = page_stat__find_alloc(&this); 942 if (pstat == NULL) 943 return -ENOMEM; 944 945 pstat->nr_free++; 946 pstat->free_bytes += bytes; 947 } 948 949 pstat = page_stat__find_caller(&this); 950 if (pstat == NULL) 951 return -ENOENT; 952 953 pstat->nr_free++; 954 pstat->free_bytes += bytes; 955 956 if (live_page) { 957 pstat->nr_alloc--; 958 pstat->alloc_bytes -= bytes; 959 960 if (pstat->nr_alloc == 0) { 961 rb_erase(&pstat->node, &page_caller_tree); 962 free(pstat); 963 } 964 } 965 966 return 0; 967 } 968 969 static bool perf_kmem__skip_sample(struct perf_sample *sample) 970 { 971 /* skip sample based on time? */ 972 if (perf_time__skip_sample(&ptime, sample->time)) 973 return true; 974 975 return false; 976 } 977 978 typedef int (*tracepoint_handler)(struct perf_sample *sample); 979 980 static int process_sample_event(const struct perf_tool *tool __maybe_unused, 981 union perf_event *event, 982 struct perf_sample *sample, 983 struct machine *machine) 984 { 985 struct evsel *evsel = sample->evsel; 986 int err = 0; 987 struct thread *thread = machine__findnew_thread(machine, sample->pid, 988 sample->tid); 989 990 if (thread == NULL) { 991 pr_debug("problem processing %s (%u) event at offset %#" PRIx64 ", skipping it.\n", 992 perf_event__name(event->header.type), event->header.type, 993 sample->file_offset); 994 return -1; 995 } 996 997 if (perf_kmem__skip_sample(sample)) { 998 thread__put(thread); 999 return 0; 1000 } 1001 1002 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread__tid(thread)); 1003 1004 if (evsel->handler != NULL) { 1005 tracepoint_handler f = evsel->handler; 1006 err = f(sample); 1007 } 1008 1009 thread__put(thread); 1010 1011 return err; 1012 } 1013 1014 static double fragmentation(unsigned long n_req, unsigned long n_alloc) 1015 { 1016 if (n_alloc == 0) 1017 return 0.0; 1018 else 1019 return 100.0 - (100.0 * n_req / n_alloc); 1020 } 1021 1022 static void __print_slab_result(struct rb_root *root, 1023 struct perf_session *session, 1024 int n_lines, int is_caller) 1025 { 1026 struct rb_node *next; 1027 struct machine *machine = &session->machines.host; 1028 1029 printf("%.105s\n", graph_dotted_line); 1030 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr"); 1031 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n"); 1032 printf("%.105s\n", graph_dotted_line); 1033 1034 next = rb_first(root); 1035 1036 while (next && n_lines--) { 1037 struct alloc_stat *data = rb_entry(next, struct alloc_stat, 1038 node); 1039 struct symbol *sym = NULL; 1040 struct map *map; 1041 char buf[BUFSIZ]; 1042 u64 addr; 1043 1044 if (is_caller) { 1045 addr = data->call_site; 1046 if (!raw_ip) 1047 sym = machine__find_kernel_symbol(machine, addr, &map); 1048 } else 1049 addr = data->ptr; 1050 1051 if (sym != NULL) 1052 snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name, 1053 addr - map__unmap_ip(map, sym->start)); 1054 else 1055 snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr); 1056 printf(" %-34s |", buf); 1057 1058 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n", 1059 (unsigned long long)data->bytes_alloc, 1060 (unsigned long)data->bytes_alloc / data->hit, 1061 (unsigned long long)data->bytes_req, 1062 (unsigned long)data->bytes_req / data->hit, 1063 (unsigned long)data->hit, 1064 (unsigned long)data->pingpong, 1065 fragmentation(data->bytes_req, data->bytes_alloc)); 1066 1067 next = rb_next(next); 1068 } 1069 1070 if (n_lines == -1) 1071 printf(" ... | ... | ... | ... | ... | ... \n"); 1072 1073 printf("%.105s\n", graph_dotted_line); 1074 } 1075 1076 static const char * const migrate_type_str[] = { 1077 "UNMOVABL", 1078 "RECLAIM", 1079 "MOVABLE", 1080 "RESERVED", 1081 "CMA/ISLT", 1082 "UNKNOWN", 1083 }; 1084 1085 static void __print_page_alloc_result(struct perf_session *session, int n_lines) 1086 { 1087 struct rb_node *next = rb_first(&page_alloc_sorted); 1088 struct machine *machine = &session->machines.host; 1089 const char *format; 1090 int gfp_len = max(strlen("GFP flags"), max_gfp_len); 1091 1092 printf("\n%.105s\n", graph_dotted_line); 1093 printf(" %-16s | %5s alloc (KB) | Hits | Order | Mig.type | %-*s | Callsite\n", 1094 use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total", 1095 gfp_len, "GFP flags"); 1096 printf("%.105s\n", graph_dotted_line); 1097 1098 if (use_pfn) 1099 format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n"; 1100 else 1101 format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n"; 1102 1103 while (next && n_lines--) { 1104 struct page_stat *data; 1105 struct symbol *sym; 1106 struct map *map; 1107 char buf[32]; 1108 char *caller = buf; 1109 1110 data = rb_entry(next, struct page_stat, node); 1111 sym = machine__find_kernel_symbol(machine, data->callsite, &map); 1112 if (sym) 1113 caller = sym->name; 1114 else 1115 scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite); 1116 1117 printf(format, (unsigned long long)data->page, 1118 (unsigned long long)data->alloc_bytes / 1024, 1119 data->nr_alloc, data->order, 1120 migrate_type_str[data->migrate_type], 1121 gfp_len, compact_gfp_string(data->gfp_flags), caller); 1122 1123 next = rb_next(next); 1124 } 1125 1126 if (n_lines == -1) { 1127 printf(" ... | ... | ... | ... | ... | %-*s | ...\n", 1128 gfp_len, "..."); 1129 } 1130 1131 printf("%.105s\n", graph_dotted_line); 1132 } 1133 1134 static void __print_page_caller_result(struct perf_session *session, int n_lines) 1135 { 1136 struct rb_node *next = rb_first(&page_caller_sorted); 1137 struct machine *machine = &session->machines.host; 1138 int gfp_len = max(strlen("GFP flags"), max_gfp_len); 1139 1140 printf("\n%.105s\n", graph_dotted_line); 1141 printf(" %5s alloc (KB) | Hits | Order | Mig.type | %-*s | Callsite\n", 1142 live_page ? "Live" : "Total", gfp_len, "GFP flags"); 1143 printf("%.105s\n", graph_dotted_line); 1144 1145 while (next && n_lines--) { 1146 struct page_stat *data; 1147 struct symbol *sym; 1148 struct map *map; 1149 char buf[32]; 1150 char *caller = buf; 1151 1152 data = rb_entry(next, struct page_stat, node); 1153 sym = machine__find_kernel_symbol(machine, data->callsite, &map); 1154 if (sym) 1155 caller = sym->name; 1156 else 1157 scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite); 1158 1159 printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n", 1160 (unsigned long long)data->alloc_bytes / 1024, 1161 data->nr_alloc, data->order, 1162 migrate_type_str[data->migrate_type], 1163 gfp_len, compact_gfp_string(data->gfp_flags), caller); 1164 1165 next = rb_next(next); 1166 } 1167 1168 if (n_lines == -1) { 1169 printf(" ... | ... | ... | ... | %-*s | ...\n", 1170 gfp_len, "..."); 1171 } 1172 1173 printf("%.105s\n", graph_dotted_line); 1174 } 1175 1176 static void print_gfp_flags(void) 1177 { 1178 int i; 1179 1180 printf("#\n"); 1181 printf("# GFP flags\n"); 1182 printf("# ---------\n"); 1183 for (i = 0; i < nr_gfps; i++) { 1184 printf("# %08x: %*s: %s\n", gfps[i].flags, 1185 (int) max_gfp_len, gfps[i].compact_str, 1186 gfps[i].human_readable); 1187 } 1188 } 1189 1190 static void print_slab_summary(void) 1191 { 1192 printf("\nSUMMARY (SLAB allocator)"); 1193 printf("\n========================\n"); 1194 printf("Total bytes requested: %'lu\n", total_requested); 1195 printf("Total bytes allocated: %'lu\n", total_allocated); 1196 printf("Total bytes freed: %'lu\n", total_freed); 1197 if (total_allocated > total_freed) { 1198 printf("Net total bytes allocated: %'lu\n", 1199 total_allocated - total_freed); 1200 } 1201 printf("Total bytes wasted on internal fragmentation: %'lu\n", 1202 total_allocated - total_requested); 1203 printf("Internal fragmentation: %f%%\n", 1204 fragmentation(total_requested, total_allocated)); 1205 printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs); 1206 } 1207 1208 static void print_page_summary(void) 1209 { 1210 int o, m; 1211 u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch; 1212 u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes; 1213 1214 printf("\nSUMMARY (page allocator)"); 1215 printf("\n========================\n"); 1216 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total allocation requests", 1217 nr_page_allocs, total_page_alloc_bytes / 1024); 1218 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total free requests", 1219 nr_page_frees, total_page_free_bytes / 1024); 1220 printf("\n"); 1221 1222 printf("%-30s: %'16"PRIu64" [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests", 1223 nr_alloc_freed, (total_alloc_freed_bytes) / 1024); 1224 printf("%-30s: %'16"PRIu64" [ %'16"PRIu64" KB ]\n", "Total alloc-only requests", 1225 nr_page_allocs - nr_alloc_freed, 1226 (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024); 1227 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total free-only requests", 1228 nr_page_nomatch, total_page_nomatch_bytes / 1024); 1229 printf("\n"); 1230 1231 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total allocation failures", 1232 nr_page_fails, total_page_fail_bytes / 1024); 1233 printf("\n"); 1234 1235 printf("%5s %12s %12s %12s %12s %12s\n", "Order", "Unmovable", 1236 "Reclaimable", "Movable", "Reserved", "CMA/Isolated"); 1237 printf("%.5s %.12s %.12s %.12s %.12s %.12s\n", graph_dotted_line, 1238 graph_dotted_line, graph_dotted_line, graph_dotted_line, 1239 graph_dotted_line, graph_dotted_line); 1240 1241 for (o = 0; o < MAX_PAGE_ORDER; o++) { 1242 printf("%5d", o); 1243 for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) { 1244 if (order_stats[o][m]) 1245 printf(" %'12d", order_stats[o][m]); 1246 else 1247 printf(" %12c", '.'); 1248 } 1249 printf("\n"); 1250 } 1251 } 1252 1253 static void print_slab_result(struct perf_session *session) 1254 { 1255 if (caller_flag) 1256 __print_slab_result(&root_caller_sorted, session, caller_lines, 1); 1257 if (alloc_flag) 1258 __print_slab_result(&root_alloc_sorted, session, alloc_lines, 0); 1259 print_slab_summary(); 1260 } 1261 1262 static void print_page_result(struct perf_session *session) 1263 { 1264 if (caller_flag || alloc_flag) 1265 print_gfp_flags(); 1266 if (caller_flag) 1267 __print_page_caller_result(session, caller_lines); 1268 if (alloc_flag) 1269 __print_page_alloc_result(session, alloc_lines); 1270 print_page_summary(); 1271 } 1272 1273 static void print_result(struct perf_session *session) 1274 { 1275 if (kmem_slab) 1276 print_slab_result(session); 1277 if (kmem_page) 1278 print_page_result(session); 1279 } 1280 1281 static LIST_HEAD(slab_caller_sort); 1282 static LIST_HEAD(slab_alloc_sort); 1283 static LIST_HEAD(page_caller_sort); 1284 static LIST_HEAD(page_alloc_sort); 1285 1286 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data, 1287 struct list_head *sort_list) 1288 { 1289 struct rb_node **new = &(root->rb_node); 1290 struct rb_node *parent = NULL; 1291 struct sort_dimension *sort; 1292 1293 while (*new) { 1294 struct alloc_stat *this; 1295 int cmp = 0; 1296 1297 this = rb_entry(*new, struct alloc_stat, node); 1298 parent = *new; 1299 1300 list_for_each_entry(sort, sort_list, list) { 1301 cmp = sort->cmp(data, this); 1302 if (cmp) 1303 break; 1304 } 1305 1306 if (cmp > 0) 1307 new = &((*new)->rb_left); 1308 else 1309 new = &((*new)->rb_right); 1310 } 1311 1312 rb_link_node(&data->node, parent, new); 1313 rb_insert_color(&data->node, root); 1314 } 1315 1316 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted, 1317 struct list_head *sort_list) 1318 { 1319 struct rb_node *node; 1320 struct alloc_stat *data; 1321 1322 for (;;) { 1323 node = rb_first(root); 1324 if (!node) 1325 break; 1326 1327 rb_erase(node, root); 1328 data = rb_entry(node, struct alloc_stat, node); 1329 sort_slab_insert(root_sorted, data, sort_list); 1330 } 1331 } 1332 1333 static void sort_page_insert(struct rb_root *root, struct page_stat *data, 1334 struct list_head *sort_list) 1335 { 1336 struct rb_node **new = &root->rb_node; 1337 struct rb_node *parent = NULL; 1338 struct sort_dimension *sort; 1339 1340 while (*new) { 1341 struct page_stat *this; 1342 int cmp = 0; 1343 1344 this = rb_entry(*new, struct page_stat, node); 1345 parent = *new; 1346 1347 list_for_each_entry(sort, sort_list, list) { 1348 cmp = sort->cmp(data, this); 1349 if (cmp) 1350 break; 1351 } 1352 1353 if (cmp > 0) 1354 new = &parent->rb_left; 1355 else 1356 new = &parent->rb_right; 1357 } 1358 1359 rb_link_node(&data->node, parent, new); 1360 rb_insert_color(&data->node, root); 1361 } 1362 1363 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted, 1364 struct list_head *sort_list) 1365 { 1366 struct rb_node *node; 1367 struct page_stat *data; 1368 1369 for (;;) { 1370 node = rb_first(root); 1371 if (!node) 1372 break; 1373 1374 rb_erase(node, root); 1375 data = rb_entry(node, struct page_stat, node); 1376 sort_page_insert(root_sorted, data, sort_list); 1377 } 1378 } 1379 1380 static void sort_result(void) 1381 { 1382 if (kmem_slab) { 1383 __sort_slab_result(&root_alloc_stat, &root_alloc_sorted, 1384 &slab_alloc_sort); 1385 __sort_slab_result(&root_caller_stat, &root_caller_sorted, 1386 &slab_caller_sort); 1387 } 1388 if (kmem_page) { 1389 if (live_page) 1390 __sort_page_result(&page_live_tree, &page_alloc_sorted, 1391 &page_alloc_sort); 1392 else 1393 __sort_page_result(&page_alloc_tree, &page_alloc_sorted, 1394 &page_alloc_sort); 1395 1396 __sort_page_result(&page_caller_tree, &page_caller_sorted, 1397 &page_caller_sort); 1398 } 1399 } 1400 1401 static int __cmd_kmem(struct perf_session *session) 1402 { 1403 int err = -EINVAL; 1404 struct evsel *evsel; 1405 const struct evsel_str_handler kmem_tracepoints[] = { 1406 /* slab allocator */ 1407 { "kmem:kmalloc", evsel__process_alloc_event, }, 1408 { "kmem:kmem_cache_alloc", evsel__process_alloc_event, }, 1409 { "kmem:kmalloc_node", evsel__process_alloc_event, }, 1410 { "kmem:kmem_cache_alloc_node", evsel__process_alloc_event, }, 1411 { "kmem:kfree", evsel__process_free_event, }, 1412 { "kmem:kmem_cache_free", evsel__process_free_event, }, 1413 /* page allocator */ 1414 { "kmem:mm_page_alloc", evsel__process_page_alloc_event, }, 1415 { "kmem:mm_page_free", evsel__process_page_free_event, }, 1416 }; 1417 1418 if (!perf_session__has_traces(session, "kmem record")) 1419 goto out; 1420 1421 if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) { 1422 pr_err("Initializing perf session tracepoint handlers failed\n"); 1423 goto out; 1424 } 1425 1426 evlist__for_each_entry(session->evlist, evsel) { 1427 if (evsel__name_is(evsel, "kmem:mm_page_alloc") && 1428 evsel__field(evsel, "pfn")) { 1429 use_pfn = true; 1430 break; 1431 } 1432 } 1433 1434 setup_pager(); 1435 err = perf_session__process_events(session); 1436 if (err != 0) { 1437 pr_err("error during process events: %d\n", err); 1438 goto out; 1439 } 1440 sort_result(); 1441 print_result(session); 1442 out: 1443 return err; 1444 } 1445 1446 /* slab sort keys */ 1447 static int ptr_cmp(void *a, void *b) 1448 { 1449 struct alloc_stat *l = a; 1450 struct alloc_stat *r = b; 1451 1452 if (l->ptr < r->ptr) 1453 return -1; 1454 else if (l->ptr > r->ptr) 1455 return 1; 1456 return 0; 1457 } 1458 1459 static struct sort_dimension ptr_sort_dimension = { 1460 .name = "ptr", 1461 .cmp = ptr_cmp, 1462 }; 1463 1464 static int slab_callsite_cmp(void *a, void *b) 1465 { 1466 struct alloc_stat *l = a; 1467 struct alloc_stat *r = b; 1468 1469 if (l->call_site < r->call_site) 1470 return -1; 1471 else if (l->call_site > r->call_site) 1472 return 1; 1473 return 0; 1474 } 1475 1476 static struct sort_dimension callsite_sort_dimension = { 1477 .name = "callsite", 1478 .cmp = slab_callsite_cmp, 1479 }; 1480 1481 static int hit_cmp(void *a, void *b) 1482 { 1483 struct alloc_stat *l = a; 1484 struct alloc_stat *r = b; 1485 1486 if (l->hit < r->hit) 1487 return -1; 1488 else if (l->hit > r->hit) 1489 return 1; 1490 return 0; 1491 } 1492 1493 static struct sort_dimension hit_sort_dimension = { 1494 .name = "hit", 1495 .cmp = hit_cmp, 1496 }; 1497 1498 static int bytes_cmp(void *a, void *b) 1499 { 1500 struct alloc_stat *l = a; 1501 struct alloc_stat *r = b; 1502 1503 if (l->bytes_alloc < r->bytes_alloc) 1504 return -1; 1505 else if (l->bytes_alloc > r->bytes_alloc) 1506 return 1; 1507 return 0; 1508 } 1509 1510 static struct sort_dimension bytes_sort_dimension = { 1511 .name = "bytes", 1512 .cmp = bytes_cmp, 1513 }; 1514 1515 static int frag_cmp(void *a, void *b) 1516 { 1517 double x, y; 1518 struct alloc_stat *l = a; 1519 struct alloc_stat *r = b; 1520 1521 x = fragmentation(l->bytes_req, l->bytes_alloc); 1522 y = fragmentation(r->bytes_req, r->bytes_alloc); 1523 1524 if (x < y) 1525 return -1; 1526 else if (x > y) 1527 return 1; 1528 return 0; 1529 } 1530 1531 static struct sort_dimension frag_sort_dimension = { 1532 .name = "frag", 1533 .cmp = frag_cmp, 1534 }; 1535 1536 static int pingpong_cmp(void *a, void *b) 1537 { 1538 struct alloc_stat *l = a; 1539 struct alloc_stat *r = b; 1540 1541 if (l->pingpong < r->pingpong) 1542 return -1; 1543 else if (l->pingpong > r->pingpong) 1544 return 1; 1545 return 0; 1546 } 1547 1548 static struct sort_dimension pingpong_sort_dimension = { 1549 .name = "pingpong", 1550 .cmp = pingpong_cmp, 1551 }; 1552 1553 /* page sort keys */ 1554 static int page_cmp(void *a, void *b) 1555 { 1556 struct page_stat *l = a; 1557 struct page_stat *r = b; 1558 1559 if (l->page < r->page) 1560 return -1; 1561 else if (l->page > r->page) 1562 return 1; 1563 return 0; 1564 } 1565 1566 static struct sort_dimension page_sort_dimension = { 1567 .name = "page", 1568 .cmp = page_cmp, 1569 }; 1570 1571 static int page_callsite_cmp(void *a, void *b) 1572 { 1573 struct page_stat *l = a; 1574 struct page_stat *r = b; 1575 1576 if (l->callsite < r->callsite) 1577 return -1; 1578 else if (l->callsite > r->callsite) 1579 return 1; 1580 return 0; 1581 } 1582 1583 static struct sort_dimension page_callsite_sort_dimension = { 1584 .name = "callsite", 1585 .cmp = page_callsite_cmp, 1586 }; 1587 1588 static int page_hit_cmp(void *a, void *b) 1589 { 1590 struct page_stat *l = a; 1591 struct page_stat *r = b; 1592 1593 if (l->nr_alloc < r->nr_alloc) 1594 return -1; 1595 else if (l->nr_alloc > r->nr_alloc) 1596 return 1; 1597 return 0; 1598 } 1599 1600 static struct sort_dimension page_hit_sort_dimension = { 1601 .name = "hit", 1602 .cmp = page_hit_cmp, 1603 }; 1604 1605 static int page_bytes_cmp(void *a, void *b) 1606 { 1607 struct page_stat *l = a; 1608 struct page_stat *r = b; 1609 1610 if (l->alloc_bytes < r->alloc_bytes) 1611 return -1; 1612 else if (l->alloc_bytes > r->alloc_bytes) 1613 return 1; 1614 return 0; 1615 } 1616 1617 static struct sort_dimension page_bytes_sort_dimension = { 1618 .name = "bytes", 1619 .cmp = page_bytes_cmp, 1620 }; 1621 1622 static int page_order_cmp(void *a, void *b) 1623 { 1624 struct page_stat *l = a; 1625 struct page_stat *r = b; 1626 1627 if (l->order < r->order) 1628 return -1; 1629 else if (l->order > r->order) 1630 return 1; 1631 return 0; 1632 } 1633 1634 static struct sort_dimension page_order_sort_dimension = { 1635 .name = "order", 1636 .cmp = page_order_cmp, 1637 }; 1638 1639 static int migrate_type_cmp(void *a, void *b) 1640 { 1641 struct page_stat *l = a; 1642 struct page_stat *r = b; 1643 1644 /* for internal use to find free'd page */ 1645 if (l->migrate_type == -1U) 1646 return 0; 1647 1648 if (l->migrate_type < r->migrate_type) 1649 return -1; 1650 else if (l->migrate_type > r->migrate_type) 1651 return 1; 1652 return 0; 1653 } 1654 1655 static struct sort_dimension migrate_type_sort_dimension = { 1656 .name = "migtype", 1657 .cmp = migrate_type_cmp, 1658 }; 1659 1660 static int gfp_flags_cmp(void *a, void *b) 1661 { 1662 struct page_stat *l = a; 1663 struct page_stat *r = b; 1664 1665 /* for internal use to find free'd page */ 1666 if (l->gfp_flags == -1U) 1667 return 0; 1668 1669 if (l->gfp_flags < r->gfp_flags) 1670 return -1; 1671 else if (l->gfp_flags > r->gfp_flags) 1672 return 1; 1673 return 0; 1674 } 1675 1676 static struct sort_dimension gfp_flags_sort_dimension = { 1677 .name = "gfp", 1678 .cmp = gfp_flags_cmp, 1679 }; 1680 1681 static struct sort_dimension *slab_sorts[] = { 1682 &ptr_sort_dimension, 1683 &callsite_sort_dimension, 1684 &hit_sort_dimension, 1685 &bytes_sort_dimension, 1686 &frag_sort_dimension, 1687 &pingpong_sort_dimension, 1688 }; 1689 1690 static struct sort_dimension *page_sorts[] = { 1691 &page_sort_dimension, 1692 &page_callsite_sort_dimension, 1693 &page_hit_sort_dimension, 1694 &page_bytes_sort_dimension, 1695 &page_order_sort_dimension, 1696 &migrate_type_sort_dimension, 1697 &gfp_flags_sort_dimension, 1698 }; 1699 1700 static int slab_sort_dimension__add(const char *tok, struct list_head *list) 1701 { 1702 struct sort_dimension *sort; 1703 int i; 1704 1705 for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) { 1706 if (!strcmp(slab_sorts[i]->name, tok)) { 1707 sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i])); 1708 if (!sort) { 1709 pr_err("%s: memdup failed\n", __func__); 1710 return -1; 1711 } 1712 list_add_tail(&sort->list, list); 1713 return 0; 1714 } 1715 } 1716 1717 return -1; 1718 } 1719 1720 static int page_sort_dimension__add(const char *tok, struct list_head *list) 1721 { 1722 struct sort_dimension *sort; 1723 int i; 1724 1725 for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) { 1726 if (!strcmp(page_sorts[i]->name, tok)) { 1727 sort = memdup(page_sorts[i], sizeof(*page_sorts[i])); 1728 if (!sort) { 1729 pr_err("%s: memdup failed\n", __func__); 1730 return -1; 1731 } 1732 list_add_tail(&sort->list, list); 1733 return 0; 1734 } 1735 } 1736 1737 return -1; 1738 } 1739 1740 static int setup_slab_sorting(struct list_head *sort_list, const char *arg) 1741 { 1742 char *tok; 1743 char *str = strdup(arg); 1744 char *pos = str; 1745 1746 if (!str) { 1747 pr_err("%s: strdup failed\n", __func__); 1748 return -1; 1749 } 1750 1751 while (true) { 1752 tok = strsep(&pos, ","); 1753 if (!tok) 1754 break; 1755 if (slab_sort_dimension__add(tok, sort_list) < 0) { 1756 pr_err("Unknown slab --sort key: '%s'", tok); 1757 free(str); 1758 return -1; 1759 } 1760 } 1761 1762 free(str); 1763 return 0; 1764 } 1765 1766 static int setup_page_sorting(struct list_head *sort_list, const char *arg) 1767 { 1768 char *tok; 1769 char *str = strdup(arg); 1770 char *pos = str; 1771 1772 if (!str) { 1773 pr_err("%s: strdup failed\n", __func__); 1774 return -1; 1775 } 1776 1777 while (true) { 1778 tok = strsep(&pos, ","); 1779 if (!tok) 1780 break; 1781 if (page_sort_dimension__add(tok, sort_list) < 0) { 1782 pr_err("Unknown page --sort key: '%s'", tok); 1783 free(str); 1784 return -1; 1785 } 1786 } 1787 1788 free(str); 1789 return 0; 1790 } 1791 1792 static int parse_sort_opt(const struct option *opt __maybe_unused, 1793 const char *arg, int unset __maybe_unused) 1794 { 1795 if (!arg) 1796 return -1; 1797 1798 if (kmem_page > kmem_slab || 1799 (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) { 1800 if (caller_flag > alloc_flag) 1801 return setup_page_sorting(&page_caller_sort, arg); 1802 else 1803 return setup_page_sorting(&page_alloc_sort, arg); 1804 } else { 1805 if (caller_flag > alloc_flag) 1806 return setup_slab_sorting(&slab_caller_sort, arg); 1807 else 1808 return setup_slab_sorting(&slab_alloc_sort, arg); 1809 } 1810 1811 return 0; 1812 } 1813 1814 static int parse_caller_opt(const struct option *opt __maybe_unused, 1815 const char *arg __maybe_unused, 1816 int unset __maybe_unused) 1817 { 1818 caller_flag = (alloc_flag + 1); 1819 return 0; 1820 } 1821 1822 static int parse_alloc_opt(const struct option *opt __maybe_unused, 1823 const char *arg __maybe_unused, 1824 int unset __maybe_unused) 1825 { 1826 alloc_flag = (caller_flag + 1); 1827 return 0; 1828 } 1829 1830 static int parse_slab_opt(const struct option *opt __maybe_unused, 1831 const char *arg __maybe_unused, 1832 int unset __maybe_unused) 1833 { 1834 kmem_slab = (kmem_page + 1); 1835 return 0; 1836 } 1837 1838 static int parse_page_opt(const struct option *opt __maybe_unused, 1839 const char *arg __maybe_unused, 1840 int unset __maybe_unused) 1841 { 1842 kmem_page = (kmem_slab + 1); 1843 return 0; 1844 } 1845 1846 static int parse_line_opt(const struct option *opt __maybe_unused, 1847 const char *arg, int unset __maybe_unused) 1848 { 1849 int lines; 1850 1851 if (!arg) 1852 return -1; 1853 1854 lines = strtoul(arg, NULL, 10); 1855 1856 if (caller_flag > alloc_flag) 1857 caller_lines = lines; 1858 else 1859 alloc_lines = lines; 1860 1861 return 0; 1862 } 1863 1864 static bool slab_legacy_tp_is_exposed(void) 1865 { 1866 /* 1867 * The tracepoints "kmem:kmalloc_node" and 1868 * "kmem:kmem_cache_alloc_node" have been removed on the latest 1869 * kernel, if the tracepoint "kmem:kmalloc_node" is existed it 1870 * means the tool is running on an old kernel, we need to 1871 * rollback to support these legacy tracepoints. 1872 */ 1873 return IS_ERR(trace_event__tp_format("kmem", "kmalloc_node")) ? 1874 false : true; 1875 } 1876 1877 static int __cmd_record(int argc, const char **argv) 1878 { 1879 const char * const record_args[] = { 1880 "record", "-a", "-R", "-c", "1", 1881 }; 1882 const char * const slab_events[] = { 1883 "-e", "kmem:kmalloc", 1884 "-e", "kmem:kfree", 1885 "-e", "kmem:kmem_cache_alloc", 1886 "-e", "kmem:kmem_cache_free", 1887 }; 1888 const char * const slab_legacy_events[] = { 1889 "-e", "kmem:kmalloc_node", 1890 "-e", "kmem:kmem_cache_alloc_node", 1891 }; 1892 const char * const page_events[] = { 1893 "-e", "kmem:mm_page_alloc", 1894 "-e", "kmem:mm_page_free", 1895 }; 1896 unsigned int rec_argc, i, j; 1897 const char **rec_argv; 1898 unsigned int slab_legacy_tp_exposed = slab_legacy_tp_is_exposed(); 1899 1900 rec_argc = ARRAY_SIZE(record_args) + argc - 1; 1901 if (kmem_slab) { 1902 rec_argc += ARRAY_SIZE(slab_events); 1903 if (slab_legacy_tp_exposed) 1904 rec_argc += ARRAY_SIZE(slab_legacy_events); 1905 } 1906 if (kmem_page) 1907 rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */ 1908 1909 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 1910 1911 if (rec_argv == NULL) 1912 return -ENOMEM; 1913 1914 for (i = 0; i < ARRAY_SIZE(record_args); i++) 1915 rec_argv[i] = strdup(record_args[i]); 1916 1917 if (kmem_slab) { 1918 for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++) 1919 rec_argv[i] = strdup(slab_events[j]); 1920 if (slab_legacy_tp_exposed) { 1921 for (j = 0; j < ARRAY_SIZE(slab_legacy_events); j++, i++) 1922 rec_argv[i] = strdup(slab_legacy_events[j]); 1923 } 1924 } 1925 if (kmem_page) { 1926 rec_argv[i++] = strdup("-g"); 1927 1928 for (j = 0; j < ARRAY_SIZE(page_events); j++, i++) 1929 rec_argv[i] = strdup(page_events[j]); 1930 } 1931 1932 for (j = 1; j < (unsigned int)argc; j++, i++) 1933 rec_argv[i] = argv[j]; 1934 1935 return cmd_record(i, rec_argv); 1936 } 1937 1938 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused) 1939 { 1940 if (!strcmp(var, "kmem.default")) { 1941 if (!strcmp(value, "slab")) 1942 kmem_default = KMEM_SLAB; 1943 else if (!strcmp(value, "page")) 1944 kmem_default = KMEM_PAGE; 1945 else 1946 pr_err("invalid default value ('slab' or 'page' required): %s\n", 1947 value); 1948 return 0; 1949 } 1950 1951 return 0; 1952 } 1953 1954 int cmd_kmem(int argc, const char **argv) 1955 { 1956 const char * const default_slab_sort = "frag,hit,bytes"; 1957 const char * const default_page_sort = "bytes,hit"; 1958 struct perf_data data = { 1959 .mode = PERF_DATA_MODE_READ, 1960 }; 1961 const struct option kmem_options[] = { 1962 OPT_STRING('i', "input", &input_name, "file", "input file name"), 1963 OPT_INCR('v', "verbose", &verbose, 1964 "be more verbose (show symbol address, etc)"), 1965 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL, 1966 "show per-callsite statistics", parse_caller_opt), 1967 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL, 1968 "show per-allocation statistics", parse_alloc_opt), 1969 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]", 1970 "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, " 1971 "page, order, migtype, gfp", parse_sort_opt), 1972 OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt), 1973 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), 1974 OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"), 1975 OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator", 1976 parse_slab_opt), 1977 OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator", 1978 parse_page_opt), 1979 OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"), 1980 OPT_STRING(0, "time", &time_str, "str", 1981 "Time span of interest (start,stop)"), 1982 OPT_END() 1983 }; 1984 const char *const kmem_subcommands[] = { "record", "stat", NULL }; 1985 const char *kmem_usage[] = { 1986 NULL, 1987 NULL 1988 }; 1989 struct perf_session *session; 1990 struct perf_tool perf_kmem; 1991 static const char errmsg[] = "No %s allocation events found. Have you run 'perf kmem record --%s'?\n"; 1992 int ret = perf_config(kmem_config, NULL); 1993 1994 if (ret) 1995 return ret; 1996 1997 argc = parse_options_subcommand(argc, argv, kmem_options, 1998 kmem_subcommands, kmem_usage, 1999 PARSE_OPT_STOP_AT_NON_OPTION); 2000 2001 if (!argc) 2002 usage_with_options(kmem_usage, kmem_options); 2003 2004 if (kmem_slab == 0 && kmem_page == 0) { 2005 if (kmem_default == KMEM_SLAB) 2006 kmem_slab = 1; 2007 else 2008 kmem_page = 1; 2009 } 2010 2011 if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) { 2012 symbol__init(NULL); 2013 return __cmd_record(argc, argv); 2014 } 2015 2016 data.path = input_name; 2017 2018 perf_tool__init(&perf_kmem, /*ordered_events=*/true); 2019 perf_kmem.sample = process_sample_event; 2020 perf_kmem.comm = perf_event__process_comm; 2021 perf_kmem.mmap = perf_event__process_mmap; 2022 perf_kmem.mmap2 = perf_event__process_mmap2; 2023 perf_kmem.namespaces = perf_event__process_namespaces; 2024 2025 kmem_session = session = perf_session__new(&data, &perf_kmem); 2026 if (IS_ERR(session)) 2027 return PTR_ERR(session); 2028 2029 ret = -1; 2030 2031 if (kmem_slab) { 2032 if (!evlist__find_tracepoint_by_name(session->evlist, "kmem:kmalloc")) { 2033 pr_err(errmsg, "slab", "slab"); 2034 goto out_delete; 2035 } 2036 } 2037 2038 if (kmem_page) { 2039 struct evsel *evsel = evlist__find_tracepoint_by_name(session->evlist, "kmem:mm_page_alloc"); 2040 const struct tep_event *tp_format = evsel ? evsel__tp_format(evsel) : NULL; 2041 2042 if (tp_format == NULL) { 2043 pr_err(errmsg, "page", "page"); 2044 goto out_delete; 2045 } 2046 kmem_page_size = tep_get_page_size(tp_format->tep); 2047 symbol_conf.use_callchain = true; 2048 } 2049 2050 symbol__init(perf_session__env(session)); 2051 2052 if (perf_time__parse_str(&ptime, time_str) != 0) { 2053 pr_err("Invalid time string\n"); 2054 ret = -EINVAL; 2055 goto out_delete; 2056 } 2057 2058 if (!strcmp(argv[0], "stat")) { 2059 setlocale(LC_ALL, ""); 2060 2061 if (cpu__setup_cpunode_map()) 2062 goto out_delete; 2063 2064 if (list_empty(&slab_caller_sort)) 2065 setup_slab_sorting(&slab_caller_sort, default_slab_sort); 2066 if (list_empty(&slab_alloc_sort)) 2067 setup_slab_sorting(&slab_alloc_sort, default_slab_sort); 2068 if (list_empty(&page_caller_sort)) 2069 setup_page_sorting(&page_caller_sort, default_page_sort); 2070 if (list_empty(&page_alloc_sort)) 2071 setup_page_sorting(&page_alloc_sort, default_page_sort); 2072 2073 if (kmem_page) { 2074 setup_page_sorting(&page_alloc_sort_input, 2075 "page,order,migtype,gfp"); 2076 setup_page_sorting(&page_caller_sort_input, 2077 "callsite,order,migtype,gfp"); 2078 } 2079 ret = __cmd_kmem(session); 2080 } else 2081 usage_with_options(kmem_usage, kmem_options); 2082 2083 out_delete: 2084 perf_session__delete(session); 2085 /* free usage string allocated by parse_options_subcommand */ 2086 free((void *)kmem_usage[0]); 2087 2088 return ret; 2089 } 2090