tools/perf/builtin-kmem.c

   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
  21 struct alloc_stat;
  22 typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
  23
  24 static const char               *input_name;
  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 #define PATH_SYS_NODE   "/sys/devices/system/node"
  61
  62 static int init_cpunode_map(void)
  63 {
  64         FILE *fp;
  65         int i, err = -1;
  66
  67         fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
  68         if (!fp) {
  69                 max_cpu_num = 4096;
  70                 return 0;
  71         }
  72
  73         if (fscanf(fp, "%d", &max_cpu_num) < 1) {
  74                 pr_err("Failed to read 'kernel_max' from sysfs");
  75                 goto out_close;
  76         }
  77
  78         max_cpu_num++;
  79
  80         cpunode_map = calloc(max_cpu_num, sizeof(int));
  81         if (!cpunode_map) {
  82                 pr_err("%s: calloc failed\n", __func__);
  83                 goto out_close;
  84         }
  85
  86         for (i = 0; i < max_cpu_num; i++)
  87                 cpunode_map[i] = -1;
  88
  89         err = 0;
  90 out_close:
  91         fclose(fp);
  92         return err;
  93 }
  94
  95 static int setup_cpunode_map(void)
  96 {
  97         struct dirent *dent1, *dent2;
  98         DIR *dir1, *dir2;
  99         unsigned int cpu, mem;
 100         char buf[PATH_MAX];
 101
 102         if (init_cpunode_map())
 103                 return -1;
 104
 105         dir1 = opendir(PATH_SYS_NODE);
 106         if (!dir1)
 107                 return -1;
 108
 109         while ((dent1 = readdir(dir1)) != NULL) {
 110                 if (dent1->d_type != DT_DIR ||
 111                     sscanf(dent1->d_name, "node%u", &mem) < 1)
 112                         continue;
 113
 114                 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
 115                 dir2 = opendir(buf);
 116                 if (!dir2)
 117                         continue;
 118                 while ((dent2 = readdir(dir2)) != NULL) {
 119                         if (dent2->d_type != DT_LNK ||
 120                             sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
 121                                 continue;
 122                         cpunode_map[cpu] = mem;
 123                 }
 124                 closedir(dir2);
 125         }
 126         closedir(dir1);
 127         return 0;
 128 }
 129
 130 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
 131                              int bytes_req, int bytes_alloc, int cpu)
 132 {
 133         struct rb_node **node = &root_alloc_stat.rb_node;
 134         struct rb_node *parent = NULL;
 135         struct alloc_stat *data = NULL;
 136
 137         while (*node) {
 138                 parent = *node;
 139                 data = rb_entry(*node, struct alloc_stat, node);
 140
 141                 if (ptr > data->ptr)
 142                         node = &(*node)->rb_right;
 143                 else if (ptr < data->ptr)
 144                         node = &(*node)->rb_left;
 145                 else
 146                         break;
 147         }
 148
 149         if (data && data->ptr == ptr) {
 150                 data->hit++;
 151                 data->bytes_req += bytes_req;
 152                 data->bytes_alloc += bytes_alloc;
 153         } else {
 154                 data = malloc(sizeof(*data));
 155                 if (!data) {
 156                         pr_err("%s: malloc failed\n", __func__);
 157                         return -1;
 158                 }
 159                 data->ptr = ptr;
 160                 data->pingpong = 0;
 161                 data->hit = 1;
 162                 data->bytes_req = bytes_req;
 163                 data->bytes_alloc = bytes_alloc;
 164
 165                 rb_link_node(&data->node, parent, node);
 166                 rb_insert_color(&data->node, &root_alloc_stat);
 167         }
 168         data->call_site = call_site;
 169         data->alloc_cpu = cpu;
 170         return 0;
 171 }
 172
 173 static int insert_caller_stat(unsigned long call_site,
 174                               int bytes_req, int bytes_alloc)
 175 {
 176         struct rb_node **node = &root_caller_stat.rb_node;
 177         struct rb_node *parent = NULL;
 178         struct alloc_stat *data = NULL;
 179
 180         while (*node) {
 181                 parent = *node;
 182                 data = rb_entry(*node, struct alloc_stat, node);
 183
 184                 if (call_site > data->call_site)
 185                         node = &(*node)->rb_right;
 186                 else if (call_site < data->call_site)
 187                         node = &(*node)->rb_left;
 188                 else
 189                         break;
 190         }
 191
 192         if (data && data->call_site == call_site) {
 193                 data->hit++;
 194                 data->bytes_req += bytes_req;
 195                 data->bytes_alloc += bytes_alloc;
 196         } else {
 197                 data = malloc(sizeof(*data));
 198                 if (!data) {
 199                         pr_err("%s: malloc failed\n", __func__);
 200                         return -1;
 201                 }
 202                 data->call_site = call_site;
 203                 data->pingpong = 0;
 204                 data->hit = 1;
 205                 data->bytes_req = bytes_req;
 206                 data->bytes_alloc = bytes_alloc;
 207
 208                 rb_link_node(&data->node, parent, node);
 209                 rb_insert_color(&data->node, &root_caller_stat);
 210         }
 211
 212         return 0;
 213 }
 214
 215 static int perf_evsel__process_alloc_event(struct perf_evsel *evsel,
 216                                            struct perf_sample *sample)
 217 {
 218         unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
 219                       call_site = perf_evsel__intval(evsel, sample, "call_site");
 220         int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
 221             bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
 222
 223         if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
 224             insert_caller_stat(call_site, bytes_req, bytes_alloc))
 225                 return -1;
 226
 227         total_requested += bytes_req;
 228         total_allocated += bytes_alloc;
 229
 230         nr_allocs++;
 231         return 0;
 232 }
 233
 234 static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,
 235                                                 struct perf_sample *sample)
 236 {
 237         int ret = perf_evsel__process_alloc_event(evsel, sample);
 238
 239         if (!ret) {
 240                 int node1 = cpunode_map[sample->cpu],
 241                     node2 = perf_evsel__intval(evsel, sample, "node");
 242
 243                 if (node1 != node2)
 244                         nr_cross_allocs++;
 245         }
 246
 247         return ret;
 248 }
 249
 250 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
 251 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
 252
 253 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
 254                                             unsigned long call_site,
 255                                             struct rb_root *root,
 256                                             sort_fn_t sort_fn)
 257 {
 258         struct rb_node *node = root->rb_node;
 259         struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
 260
 261         while (node) {
 262                 struct alloc_stat *data;
 263                 int cmp;
 264
 265                 data = rb_entry(node, struct alloc_stat, node);
 266
 267                 cmp = sort_fn(&key, data);
 268                 if (cmp < 0)
 269                         node = node->rb_left;
 270                 else if (cmp > 0)
 271                         node = node->rb_right;
 272                 else
 273                         return data;
 274         }
 275         return NULL;
 276 }
 277
 278 static int perf_evsel__process_free_event(struct perf_evsel *evsel,
 279                                           struct perf_sample *sample)
 280 {
 281         unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
 282         struct alloc_stat *s_alloc, *s_caller;
 283
 284         s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
 285         if (!s_alloc)
 286                 return 0;
 287
 288         if ((short)sample->cpu != s_alloc->alloc_cpu) {
 289                 s_alloc->pingpong++;
 290
 291                 s_caller = search_alloc_stat(0, s_alloc->call_site,
 292                                              &root_caller_stat, callsite_cmp);
 293                 if (!s_caller)
 294                         return -1;
 295                 s_caller->pingpong++;
 296         }
 297         s_alloc->alloc_cpu = -1;
 298
 299         return 0;
 300 }
 301
 302 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
 303                                   struct perf_sample *sample);
 304
 305 static int process_sample_event(struct perf_tool *tool __maybe_unused,
 306                                 union perf_event *event,
 307                                 struct perf_sample *sample,
 308                                 struct perf_evsel *evsel,
 309                                 struct machine *machine)
 310 {
 311         struct thread *thread = machine__findnew_thread(machine, event->ip.pid);
 312
 313         if (thread == NULL) {
 314                 pr_debug("problem processing %d event, skipping it.\n",
 315                          event->header.type);
 316                 return -1;
 317         }
 318
 319         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 320
 321         if (evsel->handler.func != NULL) {
 322                 tracepoint_handler f = evsel->handler.func;
 323                 return f(evsel, sample);
 324         }
 325
 326         return 0;
 327 }
 328
 329 static struct perf_tool perf_kmem = {
 330         .sample          = process_sample_event,
 331         .comm            = perf_event__process_comm,
 332         .ordered_samples = true,
 333 };
 334
 335 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
 336 {
 337         if (n_alloc == 0)
 338                 return 0.0;
 339         else
 340                 return 100.0 - (100.0 * n_req / n_alloc);
 341 }
 342
 343 static void __print_result(struct rb_root *root, struct perf_session *session,
 344                            int n_lines, int is_caller)
 345 {
 346         struct rb_node *next;
 347         struct machine *machine;
 348
 349         printf("%.102s\n", graph_dotted_line);
 350         printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
 351         printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
 352         printf("%.102s\n", graph_dotted_line);
 353
 354         next = rb_first(root);
 355
 356         machine = perf_session__find_host_machine(session);
 357         if (!machine) {
 358                 pr_err("__print_result: couldn't find kernel information\n");
 359                 return;
 360         }
 361         while (next && n_lines--) {
 362                 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
 363                                                    node);
 364                 struct symbol *sym = NULL;
 365                 struct map *map;
 366                 char buf[BUFSIZ];
 367                 u64 addr;
 368
 369                 if (is_caller) {
 370                         addr = data->call_site;
 371                         if (!raw_ip)
 372                                 sym = machine__find_kernel_function(machine, addr, &map, NULL);
 373                 } else
 374                         addr = data->ptr;
 375
 376                 if (sym != NULL)
 377                         snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
 378                                  addr - map->unmap_ip(map, sym->start));
 379                 else
 380                         snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
 381                 printf(" %-34s |", buf);
 382
 383                 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
 384                        (unsigned long long)data->bytes_alloc,
 385                        (unsigned long)data->bytes_alloc / data->hit,
 386                        (unsigned long long)data->bytes_req,
 387                        (unsigned long)data->bytes_req / data->hit,
 388                        (unsigned long)data->hit,
 389                        (unsigned long)data->pingpong,
 390                        fragmentation(data->bytes_req, data->bytes_alloc));
 391
 392                 next = rb_next(next);
 393         }
 394
 395         if (n_lines == -1)
 396                 printf(" ...                                | ...             | ...             | ...    | ...      | ...   \n");
 397
 398         printf("%.102s\n", graph_dotted_line);
 399 }
 400
 401 static void print_summary(void)
 402 {
 403         printf("\nSUMMARY\n=======\n");
 404         printf("Total bytes requested: %lu\n", total_requested);
 405         printf("Total bytes allocated: %lu\n", total_allocated);
 406         printf("Total bytes wasted on internal fragmentation: %lu\n",
 407                total_allocated - total_requested);
 408         printf("Internal fragmentation: %f%%\n",
 409                fragmentation(total_requested, total_allocated));
 410         printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
 411 }
 412
 413 static void print_result(struct perf_session *session)
 414 {
 415         if (caller_flag)
 416                 __print_result(&root_caller_sorted, session, caller_lines, 1);
 417         if (alloc_flag)
 418                 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
 419         print_summary();
 420 }
 421
 422 struct sort_dimension {
 423         const char              name[20];
 424         sort_fn_t               cmp;
 425         struct list_head        list;
 426 };
 427
 428 static LIST_HEAD(caller_sort);
 429 static LIST_HEAD(alloc_sort);
 430
 431 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
 432                         struct list_head *sort_list)
 433 {
 434         struct rb_node **new = &(root->rb_node);
 435         struct rb_node *parent = NULL;
 436         struct sort_dimension *sort;
 437
 438         while (*new) {
 439                 struct alloc_stat *this;
 440                 int cmp = 0;
 441
 442                 this = rb_entry(*new, struct alloc_stat, node);
 443                 parent = *new;
 444
 445                 list_for_each_entry(sort, sort_list, list) {
 446                         cmp = sort->cmp(data, this);
 447                         if (cmp)
 448                                 break;
 449                 }
 450
 451                 if (cmp > 0)
 452                         new = &((*new)->rb_left);
 453                 else
 454                         new = &((*new)->rb_right);
 455         }
 456
 457         rb_link_node(&data->node, parent, new);
 458         rb_insert_color(&data->node, root);
 459 }
 460
 461 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
 462                           struct list_head *sort_list)
 463 {
 464         struct rb_node *node;
 465         struct alloc_stat *data;
 466
 467         for (;;) {
 468                 node = rb_first(root);
 469                 if (!node)
 470                         break;
 471
 472                 rb_erase(node, root);
 473                 data = rb_entry(node, struct alloc_stat, node);
 474                 sort_insert(root_sorted, data, sort_list);
 475         }
 476 }
 477
 478 static void sort_result(void)
 479 {
 480         __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
 481         __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
 482 }
 483
 484 static int __cmd_kmem(void)
 485 {
 486         int err = -EINVAL;
 487         struct perf_session *session;
 488         const struct perf_evsel_str_handler kmem_tracepoints[] = {
 489                 { "kmem:kmalloc",               perf_evsel__process_alloc_event, },
 490                 { "kmem:kmem_cache_alloc",      perf_evsel__process_alloc_event, },
 491                 { "kmem:kmalloc_node",          perf_evsel__process_alloc_node_event, },
 492                 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
 493                 { "kmem:kfree",                 perf_evsel__process_free_event, },
 494                 { "kmem:kmem_cache_free",       perf_evsel__process_free_event, },
 495         };
 496
 497         session = perf_session__new(input_name, O_RDONLY, 0, false, &perf_kmem);
 498         if (session == NULL)
 499                 return -ENOMEM;
 500
 501         if (perf_session__create_kernel_maps(session) < 0)
 502                 goto out_delete;
 503
 504         if (!perf_session__has_traces(session, "kmem record"))
 505                 goto out_delete;
 506
 507         if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
 508                 pr_err("Initializing perf session tracepoint handlers failed\n");
 509                 return -1;
 510         }
 511
 512         setup_pager();
 513         err = perf_session__process_events(session, &perf_kmem);
 514         if (err != 0)
 515                 goto out_delete;
 516         sort_result();
 517         print_result(session);
 518 out_delete:
 519         perf_session__delete(session);
 520         return err;
 521 }
 522
 523 static const char * const kmem_usage[] = {
 524         "perf kmem [<options>] {record|stat}",
 525         NULL
 526 };
 527
 528 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
 529 {
 530         if (l->ptr < r->ptr)
 531                 return -1;
 532         else if (l->ptr > r->ptr)
 533                 return 1;
 534         return 0;
 535 }
 536
 537 static struct sort_dimension ptr_sort_dimension = {
 538         .name   = "ptr",
 539         .cmp    = ptr_cmp,
 540 };
 541
 542 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
 543 {
 544         if (l->call_site < r->call_site)
 545                 return -1;
 546         else if (l->call_site > r->call_site)
 547                 return 1;
 548         return 0;
 549 }
 550
 551 static struct sort_dimension callsite_sort_dimension = {
 552         .name   = "callsite",
 553         .cmp    = callsite_cmp,
 554 };
 555
 556 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
 557 {
 558         if (l->hit < r->hit)
 559                 return -1;
 560         else if (l->hit > r->hit)
 561                 return 1;
 562         return 0;
 563 }
 564
 565 static struct sort_dimension hit_sort_dimension = {
 566         .name   = "hit",
 567         .cmp    = hit_cmp,
 568 };
 569
 570 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
 571 {
 572         if (l->bytes_alloc < r->bytes_alloc)
 573                 return -1;
 574         else if (l->bytes_alloc > r->bytes_alloc)
 575                 return 1;
 576         return 0;
 577 }
 578
 579 static struct sort_dimension bytes_sort_dimension = {
 580         .name   = "bytes",
 581         .cmp    = bytes_cmp,
 582 };
 583
 584 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
 585 {
 586         double x, y;
 587
 588         x = fragmentation(l->bytes_req, l->bytes_alloc);
 589         y = fragmentation(r->bytes_req, r->bytes_alloc);
 590
 591         if (x < y)
 592                 return -1;
 593         else if (x > y)
 594                 return 1;
 595         return 0;
 596 }
 597
 598 static struct sort_dimension frag_sort_dimension = {
 599         .name   = "frag",
 600         .cmp    = frag_cmp,
 601 };
 602
 603 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
 604 {
 605         if (l->pingpong < r->pingpong)
 606                 return -1;
 607         else if (l->pingpong > r->pingpong)
 608                 return 1;
 609         return 0;
 610 }
 611
 612 static struct sort_dimension pingpong_sort_dimension = {
 613         .name   = "pingpong",
 614         .cmp    = pingpong_cmp,
 615 };
 616
 617 static struct sort_dimension *avail_sorts[] = {
 618         &ptr_sort_dimension,
 619         &callsite_sort_dimension,
 620         &hit_sort_dimension,
 621         &bytes_sort_dimension,
 622         &frag_sort_dimension,
 623         &pingpong_sort_dimension,
 624 };
 625
 626 #define NUM_AVAIL_SORTS \
 627         (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
 628
 629 static int sort_dimension__add(const char *tok, struct list_head *list)
 630 {
 631         struct sort_dimension *sort;
 632         int i;
 633
 634         for (i = 0; i < NUM_AVAIL_SORTS; i++) {
 635                 if (!strcmp(avail_sorts[i]->name, tok)) {
 636                         sort = malloc(sizeof(*sort));
 637                         if (!sort) {
 638                                 pr_err("%s: malloc failed\n", __func__);
 639                                 return -1;
 640                         }
 641                         memcpy(sort, avail_sorts[i], sizeof(*sort));
 642                         list_add_tail(&sort->list, list);
 643                         return 0;
 644                 }
 645         }
 646
 647         return -1;
 648 }
 649
 650 static int setup_sorting(struct list_head *sort_list, const char *arg)
 651 {
 652         char *tok;
 653         char *str = strdup(arg);
 654
 655         if (!str) {
 656                 pr_err("%s: strdup failed\n", __func__);
 657                 return -1;
 658         }
 659
 660         while (true) {
 661                 tok = strsep(&str, ",");
 662                 if (!tok)
 663                         break;
 664                 if (sort_dimension__add(tok, sort_list) < 0) {
 665                         error("Unknown --sort key: '%s'", tok);
 666                         free(str);
 667                         return -1;
 668                 }
 669         }
 670
 671         free(str);
 672         return 0;
 673 }
 674
 675 static int parse_sort_opt(const struct option *opt __maybe_unused,
 676                           const char *arg, int unset __maybe_unused)
 677 {
 678         if (!arg)
 679                 return -1;
 680
 681         if (caller_flag > alloc_flag)
 682                 return setup_sorting(&caller_sort, arg);
 683         else
 684                 return setup_sorting(&alloc_sort, arg);
 685
 686         return 0;
 687 }
 688
 689 static int parse_caller_opt(const struct option *opt __maybe_unused,
 690                             const char *arg __maybe_unused,
 691                             int unset __maybe_unused)
 692 {
 693         caller_flag = (alloc_flag + 1);
 694         return 0;
 695 }
 696
 697 static int parse_alloc_opt(const struct option *opt __maybe_unused,
 698                            const char *arg __maybe_unused,
 699                            int unset __maybe_unused)
 700 {
 701         alloc_flag = (caller_flag + 1);
 702         return 0;
 703 }
 704
 705 static int parse_line_opt(const struct option *opt __maybe_unused,
 706                           const char *arg, int unset __maybe_unused)
 707 {
 708         int lines;
 709
 710         if (!arg)
 711                 return -1;
 712
 713         lines = strtoul(arg, NULL, 10);
 714
 715         if (caller_flag > alloc_flag)
 716                 caller_lines = lines;
 717         else
 718                 alloc_lines = lines;
 719
 720         return 0;
 721 }
 722
 723 static const struct option kmem_options[] = {
 724         OPT_STRING('i', "input", &input_name, "file",
 725                    "input file name"),
 726         OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
 727                            "show per-callsite statistics",
 728                            parse_caller_opt),
 729         OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
 730                            "show per-allocation statistics",
 731                            parse_alloc_opt),
 732         OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
 733                      "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
 734                      parse_sort_opt),
 735         OPT_CALLBACK('l', "line", NULL, "num",
 736                      "show n lines",
 737                      parse_line_opt),
 738         OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
 739         OPT_END()
 740 };
 741
 742 static const char *record_args[] = {
 743         "record",
 744         "-a",
 745         "-R",
 746         "-f",
 747         "-c", "1",
 748         "-e", "kmem:kmalloc",
 749         "-e", "kmem:kmalloc_node",
 750         "-e", "kmem:kfree",
 751         "-e", "kmem:kmem_cache_alloc",
 752         "-e", "kmem:kmem_cache_alloc_node",
 753         "-e", "kmem:kmem_cache_free",
 754 };
 755
 756 static int __cmd_record(int argc, const char **argv)
 757 {
 758         unsigned int rec_argc, i, j;
 759         const char **rec_argv;
 760
 761         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
 762         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 763
 764         if (rec_argv == NULL)
 765                 return -ENOMEM;
 766
 767         for (i = 0; i < ARRAY_SIZE(record_args); i++)
 768                 rec_argv[i] = strdup(record_args[i]);
 769
 770         for (j = 1; j < (unsigned int)argc; j++, i++)
 771                 rec_argv[i] = argv[j];
 772
 773         return cmd_record(i, rec_argv, NULL);
 774 }
 775
 776 int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused)
 777 {
 778         argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
 779
 780         if (!argc)
 781                 usage_with_options(kmem_usage, kmem_options);
 782
 783         symbol__init();
 784
 785         if (!strncmp(argv[0], "rec", 3)) {
 786                 return __cmd_record(argc, argv);
 787         } else if (!strcmp(argv[0], "stat")) {
 788                 if (setup_cpunode_map())
 789                         return -1;
 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                 return __cmd_kmem();
 797         } else
 798                 usage_with_options(kmem_usage, kmem_options);
 799
 800         return 0;
 801 }
 802