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