1#include "builtin.h" 2#include "perf.h" 3 4#include "util/evlist.h" 5#include "util/evsel.h" 6#include "util/util.h" 7#include "util/cache.h" 8#include "util/symbol.h" 9#include "util/thread.h" 10#include "util/header.h" 11#include "util/session.h" 12#include "util/tool.h" 13 14#include "util/parse-options.h" 15#include "util/trace-event.h" 16 17#include "util/debug.h" 18 19#include <linux/rbtree.h> 20#include <linux/string.h> 21 22struct alloc_stat; 23typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *); 24 25static int alloc_flag; 26static int caller_flag; 27 28static int alloc_lines = -1; 29static int caller_lines = -1; 30 31static bool raw_ip; 32 33static int *cpunode_map; 34static int max_cpu_num; 35 36struct alloc_stat { 37 u64 call_site; 38 u64 ptr; 39 u64 bytes_req; 40 u64 bytes_alloc; 41 u32 hit; 42 u32 pingpong; 43 44 short alloc_cpu; 45 46 struct rb_node node; 47}; 48 49static struct rb_root root_alloc_stat; 50static struct rb_root root_alloc_sorted; 51static struct rb_root root_caller_stat; 52static struct rb_root root_caller_sorted; 53 54static unsigned long total_requested, total_allocated; 55static unsigned long nr_allocs, nr_cross_allocs; 56 57#define PATH_SYS_NODE "/sys/devices/system/node" 58 59static int init_cpunode_map(void) 60{ 61 FILE *fp; 62 int i, err = -1; 63 64 fp = fopen("/sys/devices/system/cpu/kernel_max", "r"); 65 if (!fp) { 66 max_cpu_num = 4096; 67 return 0; 68 } 69 70 if (fscanf(fp, "%d", &max_cpu_num) < 1) { 71 pr_err("Failed to read 'kernel_max' from sysfs"); 72 goto out_close; 73 } 74 75 max_cpu_num++; 76 77 cpunode_map = calloc(max_cpu_num, sizeof(int)); 78 if (!cpunode_map) { 79 pr_err("%s: calloc failed\n", __func__); 80 goto out_close; 81 } 82 83 for (i = 0; i < max_cpu_num; i++) 84 cpunode_map[i] = -1; 85 86 err = 0; 87out_close: 88 fclose(fp); 89 return err; 90} 91 92static int setup_cpunode_map(void) 93{ 94 struct dirent *dent1, *dent2; 95 DIR *dir1, *dir2; 96 unsigned int cpu, mem; 97 char buf[PATH_MAX]; 98 99 if (init_cpunode_map()) 100 return -1; 101 102 dir1 = opendir(PATH_SYS_NODE); 103 if (!dir1) 104 return -1; 105 106 while ((dent1 = readdir(dir1)) != NULL) { 107 if (dent1->d_type != DT_DIR || 108 sscanf(dent1->d_name, "node%u", &mem) < 1) 109 continue; 110 111 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name); 112 dir2 = opendir(buf); 113 if (!dir2) 114 continue; 115 while ((dent2 = readdir(dir2)) != NULL) { 116 if (dent2->d_type != DT_LNK || 117 sscanf(dent2->d_name, "cpu%u", &cpu) < 1) 118 continue; 119 cpunode_map[cpu] = mem; 120 } 121 closedir(dir2); 122 } 123 closedir(dir1); 124 return 0; 125} 126 127static int insert_alloc_stat(unsigned long call_site, unsigned long ptr, 128 int bytes_req, int bytes_alloc, int cpu) 129{ 130 struct rb_node **node = &root_alloc_stat.rb_node; 131 struct rb_node *parent = NULL; 132 struct alloc_stat *data = NULL; 133 134 while (*node) { 135 parent = *node; 136 data = rb_entry(*node, struct alloc_stat, node); 137 138 if (ptr > data->ptr) 139 node = &(*node)->rb_right; 140 else if (ptr < data->ptr) 141 node = &(*node)->rb_left; 142 else 143 break; 144 } 145 146 if (data && data->ptr == ptr) { 147 data->hit++; 148 data->bytes_req += bytes_req; 149 data->bytes_alloc += bytes_alloc; 150 } else { 151 data = malloc(sizeof(*data)); 152 if (!data) { 153 pr_err("%s: malloc failed\n", __func__); 154 return -1; 155 } 156 data->ptr = ptr; 157 data->pingpong = 0; 158 data->hit = 1; 159 data->bytes_req = bytes_req; 160 data->bytes_alloc = bytes_alloc; 161 162 rb_link_node(&data->node, parent, node); 163 rb_insert_color(&data->node, &root_alloc_stat); 164 } 165 data->call_site = call_site; 166 data->alloc_cpu = cpu; 167 return 0; 168} 169 170static int insert_caller_stat(unsigned long call_site, 171 int bytes_req, int bytes_alloc) 172{ 173 struct rb_node **node = &root_caller_stat.rb_node; 174 struct rb_node *parent = NULL; 175 struct alloc_stat *data = NULL; 176 177 while (*node) { 178 parent = *node; 179 data = rb_entry(*node, struct alloc_stat, node); 180 181 if (call_site > data->call_site) 182 node = &(*node)->rb_right; 183 else if (call_site < data->call_site) 184 node = &(*node)->rb_left; 185 else 186 break; 187 } 188 189 if (data && data->call_site == call_site) { 190 data->hit++; 191 data->bytes_req += bytes_req; 192 data->bytes_alloc += bytes_alloc; 193 } else { 194 data = malloc(sizeof(*data)); 195 if (!data) { 196 pr_err("%s: malloc failed\n", __func__); 197 return -1; 198 } 199 data->call_site = call_site; 200 data->pingpong = 0; 201 data->hit = 1; 202 data->bytes_req = bytes_req; 203 data->bytes_alloc = bytes_alloc; 204 205 rb_link_node(&data->node, parent, node); 206 rb_insert_color(&data->node, &root_caller_stat); 207 } 208 209 return 0; 210} 211 212static int perf_evsel__process_alloc_event(struct perf_evsel *evsel, 213 struct perf_sample *sample) 214{ 215 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"), 216 call_site = perf_evsel__intval(evsel, sample, "call_site"); 217 int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"), 218 bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc"); 219 220 if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) || 221 insert_caller_stat(call_site, bytes_req, bytes_alloc)) 222 return -1; 223 224 total_requested += bytes_req; 225 total_allocated += bytes_alloc; 226 227 nr_allocs++; 228 return 0; 229} 230 231static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel, 232 struct perf_sample *sample) 233{ 234 int ret = perf_evsel__process_alloc_event(evsel, sample); 235 236 if (!ret) { 237 int node1 = cpunode_map[sample->cpu], 238 node2 = perf_evsel__intval(evsel, sample, "node"); 239 240 if (node1 != node2) 241 nr_cross_allocs++; 242 } 243 244 return ret; 245} 246 247static int ptr_cmp(struct alloc_stat *, struct alloc_stat *); 248static int callsite_cmp(struct alloc_stat *, struct alloc_stat *); 249 250static struct alloc_stat *search_alloc_stat(unsigned long ptr, 251 unsigned long call_site, 252 struct rb_root *root, 253 sort_fn_t sort_fn) 254{ 255 struct rb_node *node = root->rb_node; 256 struct alloc_stat key = { .ptr = ptr, .call_site = call_site }; 257 258 while (node) { 259 struct alloc_stat *data; 260 int cmp; 261 262 data = rb_entry(node, struct alloc_stat, node); 263 264 cmp = sort_fn(&key, data); 265 if (cmp < 0) 266 node = node->rb_left; 267 else if (cmp > 0) 268 node = node->rb_right; 269 else 270 return data; 271 } 272 return NULL; 273} 274 275static int perf_evsel__process_free_event(struct perf_evsel *evsel, 276 struct perf_sample *sample) 277{ 278 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"); 279 struct alloc_stat *s_alloc, *s_caller; 280 281 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp); 282 if (!s_alloc) 283 return 0; 284 285 if ((short)sample->cpu != s_alloc->alloc_cpu) { 286 s_alloc->pingpong++; 287 288 s_caller = search_alloc_stat(0, s_alloc->call_site, 289 &root_caller_stat, callsite_cmp); 290 if (!s_caller) 291 return -1; 292 s_caller->pingpong++; 293 } 294 s_alloc->alloc_cpu = -1; 295 296 return 0; 297} 298 299typedef int (*tracepoint_handler)(struct perf_evsel *evsel, 300 struct perf_sample *sample); 301 302static int process_sample_event(struct perf_tool *tool __maybe_unused, 303 union perf_event *event, 304 struct perf_sample *sample, 305 struct perf_evsel *evsel, 306 struct machine *machine) 307{ 308 struct thread *thread = machine__findnew_thread(machine, event->ip.pid); 309 310 if (thread == NULL) { 311 pr_debug("problem processing %d event, skipping it.\n", 312 event->header.type); 313 return -1; 314 } 315 316 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid); 317 318 if (evsel->handler.func != NULL) { 319 tracepoint_handler f = evsel->handler.func; 320 return f(evsel, sample); 321 } 322 323 return 0; 324} 325 326static struct perf_tool perf_kmem = { 327 .sample = process_sample_event, 328 .comm = perf_event__process_comm, 329 .ordered_samples = true, 330}; 331 332static double fragmentation(unsigned long n_req, unsigned long n_alloc) 333{ 334 if (n_alloc == 0) 335 return 0.0; 336 else 337 return 100.0 - (100.0 * n_req / n_alloc); 338} 339 340static void __print_result(struct rb_root *root, struct perf_session *session, 341 int n_lines, int is_caller) 342{ 343 struct rb_node *next; 344 struct machine *machine = &session->machines.host; 345 346 printf("%.102s\n", graph_dotted_line); 347 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr"); 348 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n"); 349 printf("%.102s\n", graph_dotted_line); 350 351 next = rb_first(root); 352 353 while (next && n_lines--) { 354 struct alloc_stat *data = rb_entry(next, struct alloc_stat, 355 node); 356 struct symbol *sym = NULL; 357 struct map *map; 358 char buf[BUFSIZ]; 359 u64 addr; 360 361 if (is_caller) { 362 addr = data->call_site; 363 if (!raw_ip) 364 sym = machine__find_kernel_function(machine, addr, &map, NULL); 365 } else 366 addr = data->ptr; 367 368 if (sym != NULL) 369 snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name, 370 addr - map->unmap_ip(map, sym->start)); 371 else 372 snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr); 373 printf(" %-34s |", buf); 374 375 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n", 376 (unsigned long long)data->bytes_alloc, 377 (unsigned long)data->bytes_alloc / data->hit, 378 (unsigned long long)data->bytes_req, 379 (unsigned long)data->bytes_req / data->hit, 380 (unsigned long)data->hit, 381 (unsigned long)data->pingpong, 382 fragmentation(data->bytes_req, data->bytes_alloc)); 383 384 next = rb_next(next); 385 } 386 387 if (n_lines == -1) 388 printf(" ... | ... | ... | ... | ... | ... \n"); 389 390 printf("%.102s\n", graph_dotted_line); 391} 392 393static void print_summary(void) 394{ 395 printf("\nSUMMARY\n=======\n"); 396 printf("Total bytes requested: %lu\n", total_requested); 397 printf("Total bytes allocated: %lu\n", total_allocated); 398 printf("Total bytes wasted on internal fragmentation: %lu\n", 399 total_allocated - total_requested); 400 printf("Internal fragmentation: %f%%\n", 401 fragmentation(total_requested, total_allocated)); 402 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs); 403} 404 405static void print_result(struct perf_session *session) 406{ 407 if (caller_flag) 408 __print_result(&root_caller_sorted, session, caller_lines, 1); 409 if (alloc_flag) 410 __print_result(&root_alloc_sorted, session, alloc_lines, 0); 411 print_summary(); 412} 413 414struct sort_dimension { 415 const char name[20]; 416 sort_fn_t cmp; 417 struct list_head list; 418}; 419 420static LIST_HEAD(caller_sort); 421static LIST_HEAD(alloc_sort); 422 423static void sort_insert(struct rb_root *root, struct alloc_stat *data, 424 struct list_head *sort_list) 425{ 426 struct rb_node **new = &(root->rb_node); 427 struct rb_node *parent = NULL; 428 struct sort_dimension *sort; 429 430 while (*new) { 431 struct alloc_stat *this; 432 int cmp = 0; 433 434 this = rb_entry(*new, struct alloc_stat, node); 435 parent = *new; 436 437 list_for_each_entry(sort, sort_list, list) { 438 cmp = sort->cmp(data, this); 439 if (cmp) 440 break; 441 } 442 443 if (cmp > 0) 444 new = &((*new)->rb_left); 445 else 446 new = &((*new)->rb_right); 447 } 448 449 rb_link_node(&data->node, parent, new); 450 rb_insert_color(&data->node, root); 451} 452 453static void __sort_result(struct rb_root *root, struct rb_root *root_sorted, 454 struct list_head *sort_list) 455{ 456 struct rb_node *node; 457 struct alloc_stat *data; 458 459 for (;;) { 460 node = rb_first(root); 461 if (!node) 462 break; 463 464 rb_erase(node, root); 465 data = rb_entry(node, struct alloc_stat, node); 466 sort_insert(root_sorted, data, sort_list); 467 } 468} 469 470static void sort_result(void) 471{ 472 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort); 473 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort); 474} 475 476static int __cmd_kmem(void) 477{ 478 int err = -EINVAL; 479 struct perf_session *session; 480 const struct perf_evsel_str_handler kmem_tracepoints[] = { 481 { "kmem:kmalloc", perf_evsel__process_alloc_event, }, 482 { "kmem:kmem_cache_alloc", perf_evsel__process_alloc_event, }, 483 { "kmem:kmalloc_node", perf_evsel__process_alloc_node_event, }, 484 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, }, 485 { "kmem:kfree", perf_evsel__process_free_event, }, 486 { "kmem:kmem_cache_free", perf_evsel__process_free_event, }, 487 }; 488 489 session = perf_session__new(input_name, O_RDONLY, 0, false, &perf_kmem); 490 if (session == NULL) 491 return -ENOMEM; 492 493 if (perf_session__create_kernel_maps(session) < 0) 494 goto out_delete; 495 496 if (!perf_session__has_traces(session, "kmem record")) 497 goto out_delete; 498 499 if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) { 500 pr_err("Initializing perf session tracepoint handlers failed\n"); 501 return -1; 502 } 503 504 setup_pager(); 505 err = perf_session__process_events(session, &perf_kmem); 506 if (err != 0) 507 goto out_delete; 508 sort_result(); 509 print_result(session); 510out_delete: 511 perf_session__delete(session); 512 return err; 513} 514 515static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r) 516{ 517 if (l->ptr < r->ptr) 518 return -1; 519 else if (l->ptr > r->ptr) 520 return 1; 521 return 0; 522} 523 524static struct sort_dimension ptr_sort_dimension = { 525 .name = "ptr", 526 .cmp = ptr_cmp, 527}; 528 529static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r) 530{ 531 if (l->call_site < r->call_site) 532 return -1; 533 else if (l->call_site > r->call_site) 534 return 1; 535 return 0; 536} 537 538static struct sort_dimension callsite_sort_dimension = { 539 .name = "callsite", 540 .cmp = callsite_cmp, 541}; 542 543static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r) 544{ 545 if (l->hit < r->hit) 546 return -1; 547 else if (l->hit > r->hit) 548 return 1; 549 return 0; 550} 551 552static struct sort_dimension hit_sort_dimension = { 553 .name = "hit", 554 .cmp = hit_cmp, 555}; 556 557static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r) 558{ 559 if (l->bytes_alloc < r->bytes_alloc) 560 return -1; 561 else if (l->bytes_alloc > r->bytes_alloc) 562 return 1; 563 return 0; 564} 565 566static struct sort_dimension bytes_sort_dimension = { 567 .name = "bytes", 568 .cmp = bytes_cmp, 569}; 570 571static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r) 572{ 573 double x, y; 574 575 x = fragmentation(l->bytes_req, l->bytes_alloc); 576 y = fragmentation(r->bytes_req, r->bytes_alloc); 577 578 if (x < y) 579 return -1; 580 else if (x > y) 581 return 1; 582 return 0; 583} 584 585static struct sort_dimension frag_sort_dimension = { 586 .name = "frag", 587 .cmp = frag_cmp, 588}; 589 590static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r) 591{ 592 if (l->pingpong < r->pingpong) 593 return -1; 594 else if (l->pingpong > r->pingpong) 595 return 1; 596 return 0; 597} 598 599static struct sort_dimension pingpong_sort_dimension = { 600 .name = "pingpong", 601 .cmp = pingpong_cmp, 602}; 603 604static struct sort_dimension *avail_sorts[] = { 605 &ptr_sort_dimension, 606 &callsite_sort_dimension, 607 &hit_sort_dimension, 608 &bytes_sort_dimension, 609 &frag_sort_dimension, 610 &pingpong_sort_dimension, 611}; 612 613#define NUM_AVAIL_SORTS ((int)ARRAY_SIZE(avail_sorts)) 614 615static int sort_dimension__add(const char *tok, struct list_head *list) 616{ 617 struct sort_dimension *sort; 618 int i; 619 620 for (i = 0; i < NUM_AVAIL_SORTS; i++) { 621 if (!strcmp(avail_sorts[i]->name, tok)) { 622 sort = memdup(avail_sorts[i], sizeof(*avail_sorts[i])); 623 if (!sort) { 624 pr_err("%s: memdup failed\n", __func__); 625 return -1; 626 } 627 list_add_tail(&sort->list, list); 628 return 0; 629 } 630 } 631 632 return -1; 633} 634 635static int setup_sorting(struct list_head *sort_list, const char *arg) 636{ 637 char *tok; 638 char *str = strdup(arg); 639 640 if (!str) { 641 pr_err("%s: strdup failed\n", __func__); 642 return -1; 643 } 644 645 while (true) { 646 tok = strsep(&str, ","); 647 if (!tok) 648 break; 649 if (sort_dimension__add(tok, sort_list) < 0) { 650 error("Unknown --sort key: '%s'", tok); 651 free(str); 652 return -1; 653 } 654 } 655 656 free(str); 657 return 0; 658} 659 660static int parse_sort_opt(const struct option *opt __maybe_unused, 661 const char *arg, int unset __maybe_unused) 662{ 663 if (!arg) 664 return -1; 665 666 if (caller_flag > alloc_flag) 667 return setup_sorting(&caller_sort, arg); 668 else 669 return setup_sorting(&alloc_sort, arg); 670 671 return 0; 672} 673 674static int parse_caller_opt(const struct option *opt __maybe_unused, 675 const char *arg __maybe_unused, 676 int unset __maybe_unused) 677{ 678 caller_flag = (alloc_flag + 1); 679 return 0; 680} 681 682static int parse_alloc_opt(const struct option *opt __maybe_unused, 683 const char *arg __maybe_unused, 684 int unset __maybe_unused) 685{ 686 alloc_flag = (caller_flag + 1); 687 return 0; 688} 689 690static int parse_line_opt(const struct option *opt __maybe_unused, 691 const char *arg, int unset __maybe_unused) 692{ 693 int lines; 694 695 if (!arg) 696 return -1; 697 698 lines = strtoul(arg, NULL, 10); 699 700 if (caller_flag > alloc_flag) 701 caller_lines = lines; 702 else 703 alloc_lines = lines; 704 705 return 0; 706} 707 708static int __cmd_record(int argc, const char **argv) 709{ 710 const char * const record_args[] = { 711 "record", "-a", "-R", "-c", "1", 712 "-e", "kmem:kmalloc", 713 "-e", "kmem:kmalloc_node", 714 "-e", "kmem:kfree", 715 "-e", "kmem:kmem_cache_alloc", 716 "-e", "kmem:kmem_cache_alloc_node", 717 "-e", "kmem:kmem_cache_free", 718 }; 719 unsigned int rec_argc, i, j; 720 const char **rec_argv; 721 722 rec_argc = ARRAY_SIZE(record_args) + argc - 1; 723 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 724 725 if (rec_argv == NULL) 726 return -ENOMEM; 727 728 for (i = 0; i < ARRAY_SIZE(record_args); i++) 729 rec_argv[i] = strdup(record_args[i]); 730 731 for (j = 1; j < (unsigned int)argc; j++, i++) 732 rec_argv[i] = argv[j]; 733 734 return cmd_record(i, rec_argv, NULL); 735} 736 737int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused) 738{ 739 const char * const default_sort_order = "frag,hit,bytes"; 740 const struct option kmem_options[] = { 741 OPT_STRING('i', "input", &input_name, "file", "input file name"), 742 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL, 743 "show per-callsite statistics", parse_caller_opt), 744 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL, 745 "show per-allocation statistics", parse_alloc_opt), 746 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]", 747 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag", 748 parse_sort_opt), 749 OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt), 750 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), 751 OPT_END() 752 }; 753 const char * const kmem_usage[] = { 754 "perf kmem [<options>] {record|stat}", 755 NULL 756 }; 757 argc = parse_options(argc, argv, kmem_options, kmem_usage, 0); 758 759 if (!argc) 760 usage_with_options(kmem_usage, kmem_options); 761 762 symbol__init(); 763 764 if (!strncmp(argv[0], "rec", 3)) { 765 return __cmd_record(argc, argv); 766 } else if (!strcmp(argv[0], "stat")) { 767 if (setup_cpunode_map()) 768 return -1; 769 770 if (list_empty(&caller_sort)) 771 setup_sorting(&caller_sort, default_sort_order); 772 if (list_empty(&alloc_sort)) 773 setup_sorting(&alloc_sort, default_sort_order); 774 775 return __cmd_kmem(); 776 } else 777 usage_with_options(kmem_usage, kmem_options); 778 779 return 0; 780} 781 782