tools/perf/builtin-kmem.c

   1 #include "builtin.h"
   2 #include "perf.h"
   3
   4 #include "util/util.h"
   5 #include "util/cache.h"
   6 #include "util/symbol.h"
   7 #include "util/thread.h"
   8 #include "util/header.h"
   9
  10 #include "util/parse-options.h"
  11 #include "util/trace-event.h"
  12
  13 #include "util/debug.h"
  14 #include "util/data_map.h"
  15
  16 #include <linux/rbtree.h>
  17
  18 struct alloc_stat;
  19 typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
  20
  21 static char const               *input_name = "perf.data";
  22
  23 static struct perf_header       *header;
  24 static u64                      sample_type;
  25
  26 static int                      alloc_flag;
  27 static int                      caller_flag;
  28
  29 static int                      alloc_lines = -1;
  30 static int                      caller_lines = -1;
  31
  32 static bool                     raw_ip;
  33
  34 static char                     default_sort_order[] = "frag,hit,bytes";
  35
  36 static int                      *cpunode_map;
  37 static int                      max_cpu_num;
  38
  39 struct alloc_stat {
  40         u64     call_site;
  41         u64     ptr;
  42         u64     bytes_req;
  43         u64     bytes_alloc;
  44         u32     hit;
  45         u32     pingpong;
  46
  47         short   alloc_cpu;
  48
  49         struct rb_node node;
  50 };
  51
  52 static struct rb_root root_alloc_stat;
  53 static struct rb_root root_alloc_sorted;
  54 static struct rb_root root_caller_stat;
  55 static struct rb_root root_caller_sorted;
  56
  57 static unsigned long total_requested, total_allocated;
  58 static unsigned long nr_allocs, nr_cross_allocs;
  59
  60 struct raw_event_sample {
  61         u32 size;
  62         char data[0];
  63 };
  64
  65 #define PATH_SYS_NODE   "/sys/devices/system/node"
  66
  67 static void init_cpunode_map(void)
  68 {
  69         FILE *fp;
  70         int i;
  71
  72         fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
  73         if (!fp) {
  74                 max_cpu_num = 4096;
  75                 return;
  76         }
  77
  78         if (fscanf(fp, "%d", &max_cpu_num) < 1)
  79                 die("Failed to read 'kernel_max' from sysfs");
  80         max_cpu_num++;
  81
  82         cpunode_map = calloc(max_cpu_num, sizeof(int));
  83         if (!cpunode_map)
  84                 die("calloc");
  85         for (i = 0; i < max_cpu_num; i++)
  86                 cpunode_map[i] = -1;
  87         fclose(fp);
  88 }
  89
  90 static void setup_cpunode_map(void)
  91 {
  92         struct dirent *dent1, *dent2;
  93         DIR *dir1, *dir2;
  94         unsigned int cpu, mem;
  95         char buf[PATH_MAX];
  96
  97         init_cpunode_map();
  98
  99         dir1 = opendir(PATH_SYS_NODE);
 100         if (!dir1)
 101                 return;
 102
 103         while (true) {
 104                 dent1 = readdir(dir1);
 105                 if (!dent1)
 106                         break;
 107
 108                 if (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 (true) {
 116                         dent2 = readdir(dir2);
 117                         if (!dent2)
 118                                 break;
 119                         if (sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
 120                                 continue;
 121                         cpunode_map[cpu] = mem;
 122                 }
 123         }
 124 }
 125
 126 static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
 127                               int bytes_req, int bytes_alloc, int cpu)
 128 {
 129         struct rb_node **node = &root_alloc_stat.rb_node;
 130         struct rb_node *parent = NULL;
 131         struct alloc_stat *data = NULL;
 132
 133         while (*node) {
 134                 parent = *node;
 135                 data = rb_entry(*node, struct alloc_stat, node);
 136
 137                 if (ptr > data->ptr)
 138                         node = &(*node)->rb_right;
 139                 else if (ptr < data->ptr)
 140                         node = &(*node)->rb_left;
 141                 else
 142                         break;
 143         }
 144
 145         if (data && data->ptr == ptr) {
 146                 data->hit++;
 147                 data->bytes_req += bytes_req;
 148                 data->bytes_alloc += bytes_req;
 149         } else {
 150                 data = malloc(sizeof(*data));
 151                 if (!data)
 152                         die("malloc");
 153                 data->ptr = ptr;
 154                 data->pingpong = 0;
 155                 data->hit = 1;
 156                 data->bytes_req = bytes_req;
 157                 data->bytes_alloc = bytes_alloc;
 158
 159                 rb_link_node(&data->node, parent, node);
 160                 rb_insert_color(&data->node, &root_alloc_stat);
 161         }
 162         data->call_site = call_site;
 163         data->alloc_cpu = cpu;
 164 }
 165
 166 static void insert_caller_stat(unsigned long call_site,
 167                               int bytes_req, int bytes_alloc)
 168 {
 169         struct rb_node **node = &root_caller_stat.rb_node;
 170         struct rb_node *parent = NULL;
 171         struct alloc_stat *data = NULL;
 172
 173         while (*node) {
 174                 parent = *node;
 175                 data = rb_entry(*node, struct alloc_stat, node);
 176
 177                 if (call_site > data->call_site)
 178                         node = &(*node)->rb_right;
 179                 else if (call_site < data->call_site)
 180                         node = &(*node)->rb_left;
 181                 else
 182                         break;
 183         }
 184
 185         if (data && data->call_site == call_site) {
 186                 data->hit++;
 187                 data->bytes_req += bytes_req;
 188                 data->bytes_alloc += bytes_req;
 189         } else {
 190                 data = malloc(sizeof(*data));
 191                 if (!data)
 192                         die("malloc");
 193                 data->call_site = call_site;
 194                 data->pingpong = 0;
 195                 data->hit = 1;
 196                 data->bytes_req = bytes_req;
 197                 data->bytes_alloc = bytes_alloc;
 198
 199                 rb_link_node(&data->node, parent, node);
 200                 rb_insert_color(&data->node, &root_caller_stat);
 201         }
 202 }
 203
 204 static void process_alloc_event(struct raw_event_sample *raw,
 205                                 struct event *event,
 206                                 int cpu,
 207                                 u64 timestamp __used,
 208                                 struct thread *thread __used,
 209                                 int node)
 210 {
 211         unsigned long call_site;
 212         unsigned long ptr;
 213         int bytes_req;
 214         int bytes_alloc;
 215         int node1, node2;
 216
 217         ptr = raw_field_value(event, "ptr", raw->data);
 218         call_site = raw_field_value(event, "call_site", raw->data);
 219         bytes_req = raw_field_value(event, "bytes_req", raw->data);
 220         bytes_alloc = raw_field_value(event, "bytes_alloc", raw->data);
 221
 222         insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
 223         insert_caller_stat(call_site, bytes_req, bytes_alloc);
 224
 225         total_requested += bytes_req;
 226         total_allocated += bytes_alloc;
 227
 228         if (node) {
 229                 node1 = cpunode_map[cpu];
 230                 node2 = raw_field_value(event, "node", raw->data);
 231                 if (node1 != node2)
 232                         nr_cross_allocs++;
 233         }
 234         nr_allocs++;
 235 }
 236
 237 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
 238 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
 239
 240 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
 241                                             unsigned long call_site,
 242                                             struct rb_root *root,
 243                                             sort_fn_t sort_fn)
 244 {
 245         struct rb_node *node = root->rb_node;
 246         struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
 247
 248         while (node) {
 249                 struct alloc_stat *data;
 250                 int cmp;
 251
 252                 data = rb_entry(node, struct alloc_stat, node);
 253
 254                 cmp = sort_fn(&key, data);
 255                 if (cmp < 0)
 256                         node = node->rb_left;
 257                 else if (cmp > 0)
 258                         node = node->rb_right;
 259                 else
 260                         return data;
 261         }
 262         return NULL;
 263 }
 264
 265 static void process_free_event(struct raw_event_sample *raw,
 266                                struct event *event,
 267                                int cpu,
 268                                u64 timestamp __used,
 269                                struct thread *thread __used)
 270 {
 271         unsigned long ptr;
 272         struct alloc_stat *s_alloc, *s_caller;
 273
 274         ptr = raw_field_value(event, "ptr", raw->data);
 275
 276         s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
 277         if (!s_alloc)
 278                 return;
 279
 280         if (cpu != s_alloc->alloc_cpu) {
 281                 s_alloc->pingpong++;
 282
 283                 s_caller = search_alloc_stat(0, s_alloc->call_site,
 284                                              &root_caller_stat, callsite_cmp);
 285                 assert(s_caller);
 286                 s_caller->pingpong++;
 287         }
 288         s_alloc->alloc_cpu = -1;
 289 }
 290
 291 static void
 292 process_raw_event(event_t *raw_event __used, u32 size, void *data,
 293                   int cpu, u64 timestamp, struct thread *thread)
 294 {
 295         struct raw_event_sample *raw;
 296         struct event *event;
 297         int type;
 298
 299         raw = malloc_or_die(sizeof(*raw)+size);
 300         raw->size = size;
 301         memcpy(raw->data, data, size);
 302
 303         type = trace_parse_common_type(raw->data);
 304         event = trace_find_event(type);
 305
 306         if (!strcmp(event->name, "kmalloc") ||
 307             !strcmp(event->name, "kmem_cache_alloc")) {
 308                 process_alloc_event(raw, event, cpu, timestamp, thread, 0);
 309                 return;
 310         }
 311
 312         if (!strcmp(event->name, "kmalloc_node") ||
 313             !strcmp(event->name, "kmem_cache_alloc_node")) {
 314                 process_alloc_event(raw, event, cpu, timestamp, thread, 1);
 315                 return;
 316         }
 317
 318         if (!strcmp(event->name, "kfree") ||
 319             !strcmp(event->name, "kmem_cache_free")) {
 320                 process_free_event(raw, event, cpu, timestamp, thread);
 321                 return;
 322         }
 323 }
 324
 325 static int process_sample_event(event_t *event)
 326 {
 327         struct sample_data data;
 328         struct thread *thread;
 329
 330         memset(&data, 0, sizeof(data));
 331         data.time = -1;
 332         data.cpu = -1;
 333         data.period = 1;
 334
 335         event__parse_sample(event, sample_type, &data);
 336
 337         dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
 338                 event->header.misc,
 339                 data.pid, data.tid,
 340                 (void *)(long)data.ip,
 341                 (long long)data.period);
 342
 343         thread = threads__findnew(event->ip.pid);
 344         if (thread == NULL) {
 345                 pr_debug("problem processing %d event, skipping it.\n",
 346                          event->header.type);
 347                 return -1;
 348         }
 349
 350         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 351
 352         process_raw_event(event, data.raw_size, data.raw_data, data.cpu,
 353                           data.time, thread);
 354
 355         return 0;
 356 }
 357
 358 static int sample_type_check(u64 type)
 359 {
 360         sample_type = type;
 361
 362         if (!(sample_type & PERF_SAMPLE_RAW)) {
 363                 fprintf(stderr,
 364                         "No trace sample to read. Did you call perf record "
 365                         "without -R?");
 366                 return -1;
 367         }
 368
 369         return 0;
 370 }
 371
 372 static struct perf_file_handler file_handler = {
 373         .process_sample_event   = process_sample_event,
 374         .process_comm_event     = event__process_comm,
 375         .sample_type_check      = sample_type_check,
 376 };
 377
 378 static int read_events(void)
 379 {
 380         register_idle_thread();
 381         register_perf_file_handler(&file_handler);
 382
 383         return mmap_dispatch_perf_file(&header, input_name, 0, 0,
 384                                        &event__cwdlen, &event__cwd);
 385 }
 386
 387 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
 388 {
 389         if (n_alloc == 0)
 390                 return 0.0;
 391         else
 392                 return 100.0 - (100.0 * n_req / n_alloc);
 393 }
 394
 395 static void __print_result(struct rb_root *root, int n_lines, int is_caller)
 396 {
 397         struct rb_node *next;
 398
 399         printf("%.102s\n", graph_dotted_line);
 400         printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
 401         printf(" Total_alloc/Per | Total_req/Per   | Hit   | Ping-pong | Frag\n");
 402         printf("%.102s\n", graph_dotted_line);
 403
 404         next = rb_first(root);
 405
 406         while (next && n_lines--) {
 407                 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
 408                                                    node);
 409                 struct symbol *sym = NULL;
 410                 char buf[BUFSIZ];
 411                 u64 addr;
 412
 413                 if (is_caller) {
 414                         addr = data->call_site;
 415                         if (!raw_ip)
 416                                 sym = thread__find_function(kthread, addr, NULL);
 417                 } else
 418                         addr = data->ptr;
 419
 420                 if (sym != NULL)
 421                         snprintf(buf, sizeof(buf), "%s+%Lx", sym->name,
 422                                  addr - sym->start);
 423                 else
 424                         snprintf(buf, sizeof(buf), "%#Lx", addr);
 425                 printf(" %-34s |", buf);
 426
 427                 printf(" %9llu/%-5lu | %9llu/%-5lu | %6lu | %8lu | %6.3f%%\n",
 428                        (unsigned long long)data->bytes_alloc,
 429                        (unsigned long)data->bytes_alloc / data->hit,
 430                        (unsigned long long)data->bytes_req,
 431                        (unsigned long)data->bytes_req / data->hit,
 432                        (unsigned long)data->hit,
 433                        (unsigned long)data->pingpong,
 434                        fragmentation(data->bytes_req, data->bytes_alloc));
 435
 436                 next = rb_next(next);
 437         }
 438
 439         if (n_lines == -1)
 440                 printf(" ...                                | ...             | ...             | ...    | ...      | ...   \n");
 441
 442         printf("%.102s\n", graph_dotted_line);
 443 }
 444
 445 static void print_summary(void)
 446 {
 447         printf("\nSUMMARY\n=======\n");
 448         printf("Total bytes requested: %lu\n", total_requested);
 449         printf("Total bytes allocated: %lu\n", total_allocated);
 450         printf("Total bytes wasted on internal fragmentation: %lu\n",
 451                total_allocated - total_requested);
 452         printf("Internal fragmentation: %f%%\n",
 453                fragmentation(total_requested, total_allocated));
 454         printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
 455 }
 456
 457 static void print_result(void)
 458 {
 459         if (caller_flag)
 460                 __print_result(&root_caller_sorted, caller_lines, 1);
 461         if (alloc_flag)
 462                 __print_result(&root_alloc_sorted, alloc_lines, 0);
 463         print_summary();
 464 }
 465
 466 struct sort_dimension {
 467         const char              name[20];
 468         sort_fn_t               cmp;
 469         struct list_head        list;
 470 };
 471
 472 static LIST_HEAD(caller_sort);
 473 static LIST_HEAD(alloc_sort);
 474
 475 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
 476                         struct list_head *sort_list)
 477 {
 478         struct rb_node **new = &(root->rb_node);
 479         struct rb_node *parent = NULL;
 480         struct sort_dimension *sort;
 481
 482         while (*new) {
 483                 struct alloc_stat *this;
 484                 int cmp = 0;
 485
 486                 this = rb_entry(*new, struct alloc_stat, node);
 487                 parent = *new;
 488
 489                 list_for_each_entry(sort, sort_list, list) {
 490                         cmp = sort->cmp(data, this);
 491                         if (cmp)
 492                                 break;
 493                 }
 494
 495                 if (cmp > 0)
 496                         new = &((*new)->rb_left);
 497                 else
 498                         new = &((*new)->rb_right);
 499         }
 500
 501         rb_link_node(&data->node, parent, new);
 502         rb_insert_color(&data->node, root);
 503 }
 504
 505 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
 506                           struct list_head *sort_list)
 507 {
 508         struct rb_node *node;
 509         struct alloc_stat *data;
 510
 511         for (;;) {
 512                 node = rb_first(root);
 513                 if (!node)
 514                         break;
 515
 516                 rb_erase(node, root);
 517                 data = rb_entry(node, struct alloc_stat, node);
 518                 sort_insert(root_sorted, data, sort_list);
 519         }
 520 }
 521
 522 static void sort_result(void)
 523 {
 524         __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
 525         __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
 526 }
 527
 528 static int __cmd_kmem(void)
 529 {
 530         setup_pager();
 531         read_events();
 532         sort_result();
 533         print_result();
 534
 535         return 0;
 536 }
 537
 538 static const char * const kmem_usage[] = {
 539         "perf kmem [<options>] {record}",
 540         NULL
 541 };
 542
 543 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
 544 {
 545         if (l->ptr < r->ptr)
 546                 return -1;
 547         else if (l->ptr > r->ptr)
 548                 return 1;
 549         return 0;
 550 }
 551
 552 static struct sort_dimension ptr_sort_dimension = {
 553         .name   = "ptr",
 554         .cmp    = ptr_cmp,
 555 };
 556
 557 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
 558 {
 559         if (l->call_site < r->call_site)
 560                 return -1;
 561         else if (l->call_site > r->call_site)
 562                 return 1;
 563         return 0;
 564 }
 565
 566 static struct sort_dimension callsite_sort_dimension = {
 567         .name   = "callsite",
 568         .cmp    = callsite_cmp,
 569 };
 570
 571 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
 572 {
 573         if (l->hit < r->hit)
 574                 return -1;
 575         else if (l->hit > r->hit)
 576                 return 1;
 577         return 0;
 578 }
 579
 580 static struct sort_dimension hit_sort_dimension = {
 581         .name   = "hit",
 582         .cmp    = hit_cmp,
 583 };
 584
 585 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
 586 {
 587         if (l->bytes_alloc < r->bytes_alloc)
 588                 return -1;
 589         else if (l->bytes_alloc > r->bytes_alloc)
 590                 return 1;
 591         return 0;
 592 }
 593
 594 static struct sort_dimension bytes_sort_dimension = {
 595         .name   = "bytes",
 596         .cmp    = bytes_cmp,
 597 };
 598
 599 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
 600 {
 601         double x, y;
 602
 603         x = fragmentation(l->bytes_req, l->bytes_alloc);
 604         y = fragmentation(r->bytes_req, r->bytes_alloc);
 605
 606         if (x < y)
 607                 return -1;
 608         else if (x > y)
 609                 return 1;
 610         return 0;
 611 }
 612
 613 static struct sort_dimension frag_sort_dimension = {
 614         .name   = "frag",
 615         .cmp    = frag_cmp,
 616 };
 617
 618 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
 619 {
 620         if (l->pingpong < r->pingpong)
 621                 return -1;
 622         else if (l->pingpong > r->pingpong)
 623                 return 1;
 624         return 0;
 625 }
 626
 627 static struct sort_dimension pingpong_sort_dimension = {
 628         .name   = "pingpong",
 629         .cmp    = pingpong_cmp,
 630 };
 631
 632 static struct sort_dimension *avail_sorts[] = {
 633         &ptr_sort_dimension,
 634         &callsite_sort_dimension,
 635         &hit_sort_dimension,
 636         &bytes_sort_dimension,
 637         &frag_sort_dimension,
 638         &pingpong_sort_dimension,
 639 };
 640
 641 #define NUM_AVAIL_SORTS \
 642         (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
 643
 644 static int sort_dimension__add(const char *tok, struct list_head *list)
 645 {
 646         struct sort_dimension *sort;
 647         int i;
 648
 649         for (i = 0; i < NUM_AVAIL_SORTS; i++) {
 650                 if (!strcmp(avail_sorts[i]->name, tok)) {
 651                         sort = malloc(sizeof(*sort));
 652                         if (!sort)
 653                                 die("malloc");
 654                         memcpy(sort, avail_sorts[i], sizeof(*sort));
 655                         list_add_tail(&sort->list, list);
 656                         return 0;
 657                 }
 658         }
 659
 660         return -1;
 661 }
 662
 663 static int setup_sorting(struct list_head *sort_list, const char *arg)
 664 {
 665         char *tok;
 666         char *str = strdup(arg);
 667
 668         if (!str)
 669                 die("strdup");
 670
 671         while (true) {
 672                 tok = strsep(&str, ",");
 673                 if (!tok)
 674                         break;
 675                 if (sort_dimension__add(tok, sort_list) < 0) {
 676                         error("Unknown --sort key: '%s'", tok);
 677                         return -1;
 678                 }
 679         }
 680
 681         free(str);
 682         return 0;
 683 }
 684
 685 static int parse_sort_opt(const struct option *opt __used,
 686                           const char *arg, int unset __used)
 687 {
 688         if (!arg)
 689                 return -1;
 690
 691         if (caller_flag > alloc_flag)
 692                 return setup_sorting(&caller_sort, arg);
 693         else
 694                 return setup_sorting(&alloc_sort, arg);
 695
 696         return 0;
 697 }
 698
 699 static int parse_stat_opt(const struct option *opt __used,
 700                           const char *arg, int unset __used)
 701 {
 702         if (!arg)
 703                 return -1;
 704
 705         if (strcmp(arg, "alloc") == 0)
 706                 alloc_flag = (caller_flag + 1);
 707         else if (strcmp(arg, "caller") == 0)
 708                 caller_flag = (alloc_flag + 1);
 709         else
 710                 return -1;
 711         return 0;
 712 }
 713
 714 static int parse_line_opt(const struct option *opt __used,
 715                           const char *arg, int unset __used)
 716 {
 717         int lines;
 718
 719         if (!arg)
 720                 return -1;
 721
 722         lines = strtoul(arg, NULL, 10);
 723
 724         if (caller_flag > alloc_flag)
 725                 caller_lines = lines;
 726         else
 727                 alloc_lines = lines;
 728
 729         return 0;
 730 }
 731
 732 static const struct option kmem_options[] = {
 733         OPT_STRING('i', "input", &input_name, "file",
 734                    "input file name"),
 735         OPT_CALLBACK(0, "stat", NULL, "<alloc>|<caller>",
 736                      "stat selector, Pass 'alloc' or 'caller'.",
 737                      parse_stat_opt),
 738         OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
 739                      "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
 740                      parse_sort_opt),
 741         OPT_CALLBACK('l', "line", NULL, "num",
 742                      "show n lins",
 743                      parse_line_opt),
 744         OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
 745         OPT_END()
 746 };
 747
 748 static const char *record_args[] = {
 749         "record",
 750         "-a",
 751         "-R",
 752         "-M",
 753         "-f",
 754         "-c", "1",
 755         "-e", "kmem:kmalloc",
 756         "-e", "kmem:kmalloc_node",
 757         "-e", "kmem:kfree",
 758         "-e", "kmem:kmem_cache_alloc",
 759         "-e", "kmem:kmem_cache_alloc_node",
 760         "-e", "kmem:kmem_cache_free",
 761 };
 762
 763 static int __cmd_record(int argc, const char **argv)
 764 {
 765         unsigned int rec_argc, i, j;
 766         const char **rec_argv;
 767
 768         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
 769         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 770
 771         for (i = 0; i < ARRAY_SIZE(record_args); i++)
 772                 rec_argv[i] = strdup(record_args[i]);
 773
 774         for (j = 1; j < (unsigned int)argc; j++, i++)
 775                 rec_argv[i] = argv[j];
 776
 777         return cmd_record(i, rec_argv, NULL);
 778 }
 779
 780 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
 781 {
 782         symbol__init(0);
 783
 784         argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
 785
 786         if (argc && !strncmp(argv[0], "rec", 3))
 787                 return __cmd_record(argc, argv);
 788         else if (argc)
 789                 usage_with_options(kmem_usage, kmem_options);
 790
 791         if (list_empty(&caller_sort))
 792                 setup_sorting(&caller_sort, default_sort_order);
 793         if (list_empty(&alloc_sort))
 794                 setup_sorting(&alloc_sort, default_sort_order);
 795
 796         setup_cpunode_map();
 797
 798         return __cmd_kmem();
 799 }
 800