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 char                     *cwd;
  37 static int                      cwdlen;
  38
  39 static int                      *cpunode_map;
  40 static int                      max_cpu_num;
  41
  42 struct alloc_stat {
  43         u64     call_site;
  44         u64     ptr;
  45         u64     bytes_req;
  46         u64     bytes_alloc;
  47         u32     hit;
  48         u32     pingpong;
  49
  50         short   alloc_cpu;
  51
  52         struct rb_node node;
  53 };
  54
  55 static struct rb_root root_alloc_stat;
  56 static struct rb_root root_alloc_sorted;
  57 static struct rb_root root_caller_stat;
  58 static struct rb_root root_caller_sorted;
  59
  60 static unsigned long total_requested, total_allocated;
  61 static unsigned long nr_allocs, nr_cross_allocs;
  62
  63 struct raw_event_sample {
  64         u32 size;
  65         char data[0];
  66 };
  67
  68 #define PATH_SYS_NODE   "/sys/devices/system/node"
  69
  70 static void init_cpunode_map(void)
  71 {
  72         FILE *fp;
  73         int i;
  74
  75         fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
  76         if (!fp) {
  77                 max_cpu_num = 4096;
  78                 return;
  79         }
  80
  81         if (fscanf(fp, "%d", &max_cpu_num) < 1)
  82                 die("Failed to read 'kernel_max' from sysfs");
  83         max_cpu_num++;
  84
  85         cpunode_map = calloc(max_cpu_num, sizeof(int));
  86         if (!cpunode_map)
  87                 die("calloc");
  88         for (i = 0; i < max_cpu_num; i++)
  89                 cpunode_map[i] = -1;
  90         fclose(fp);
  91 }
  92
  93 static void setup_cpunode_map(void)
  94 {
  95         struct dirent *dent1, *dent2;
  96         DIR *dir1, *dir2;
  97         unsigned int cpu, mem;
  98         char buf[PATH_MAX];
  99
 100         init_cpunode_map();
 101
 102         dir1 = opendir(PATH_SYS_NODE);
 103         if (!dir1)
 104                 return;
 105
 106         while (true) {
 107                 dent1 = readdir(dir1);
 108                 if (!dent1)
 109                         break;
 110
 111                 if (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 (true) {
 119                         dent2 = readdir(dir2);
 120                         if (!dent2)
 121                                 break;
 122                         if (sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
 123                                 continue;
 124                         cpunode_map[cpu] = mem;
 125                 }
 126         }
 127 }
 128
 129 static int
 130 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
 131 {
 132         struct thread *thread = threads__findnew(event->comm.pid);
 133
 134         dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
 135                 (void *)(offset + head),
 136                 (void *)(long)(event->header.size),
 137                 event->comm.comm, event->comm.pid);
 138
 139         if (thread == NULL ||
 140             thread__set_comm(thread, event->comm.comm)) {
 141                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 142                 return -1;
 143         }
 144
 145         return 0;
 146 }
 147
 148 static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
 149                               int bytes_req, int bytes_alloc, int cpu)
 150 {
 151         struct rb_node **node = &root_alloc_stat.rb_node;
 152         struct rb_node *parent = NULL;
 153         struct alloc_stat *data = NULL;
 154
 155         while (*node) {
 156                 parent = *node;
 157                 data = rb_entry(*node, struct alloc_stat, node);
 158
 159                 if (ptr > data->ptr)
 160                         node = &(*node)->rb_right;
 161                 else if (ptr < data->ptr)
 162                         node = &(*node)->rb_left;
 163                 else
 164                         break;
 165         }
 166
 167         if (data && data->ptr == ptr) {
 168                 data->hit++;
 169                 data->bytes_req += bytes_req;
 170                 data->bytes_alloc += bytes_req;
 171         } else {
 172                 data = malloc(sizeof(*data));
 173                 if (!data)
 174                         die("malloc");
 175                 data->ptr = ptr;
 176                 data->pingpong = 0;
 177                 data->hit = 1;
 178                 data->bytes_req = bytes_req;
 179                 data->bytes_alloc = bytes_alloc;
 180
 181                 rb_link_node(&data->node, parent, node);
 182                 rb_insert_color(&data->node, &root_alloc_stat);
 183         }
 184         data->call_site = call_site;
 185         data->alloc_cpu = cpu;
 186 }
 187
 188 static void insert_caller_stat(unsigned long call_site,
 189                               int bytes_req, int bytes_alloc)
 190 {
 191         struct rb_node **node = &root_caller_stat.rb_node;
 192         struct rb_node *parent = NULL;
 193         struct alloc_stat *data = NULL;
 194
 195         while (*node) {
 196                 parent = *node;
 197                 data = rb_entry(*node, struct alloc_stat, node);
 198
 199                 if (call_site > data->call_site)
 200                         node = &(*node)->rb_right;
 201                 else if (call_site < data->call_site)
 202                         node = &(*node)->rb_left;
 203                 else
 204                         break;
 205         }
 206
 207         if (data && data->call_site == call_site) {
 208                 data->hit++;
 209                 data->bytes_req += bytes_req;
 210                 data->bytes_alloc += bytes_req;
 211         } else {
 212                 data = malloc(sizeof(*data));
 213                 if (!data)
 214                         die("malloc");
 215                 data->call_site = call_site;
 216                 data->pingpong = 0;
 217                 data->hit = 1;
 218                 data->bytes_req = bytes_req;
 219                 data->bytes_alloc = bytes_alloc;
 220
 221                 rb_link_node(&data->node, parent, node);
 222                 rb_insert_color(&data->node, &root_caller_stat);
 223         }
 224 }
 225
 226 static void process_alloc_event(struct raw_event_sample *raw,
 227                                 struct event *event,
 228                                 int cpu,
 229                                 u64 timestamp __used,
 230                                 struct thread *thread __used,
 231                                 int node)
 232 {
 233         unsigned long call_site;
 234         unsigned long ptr;
 235         int bytes_req;
 236         int bytes_alloc;
 237         int node1, node2;
 238
 239         ptr = raw_field_value(event, "ptr", raw->data);
 240         call_site = raw_field_value(event, "call_site", raw->data);
 241         bytes_req = raw_field_value(event, "bytes_req", raw->data);
 242         bytes_alloc = raw_field_value(event, "bytes_alloc", raw->data);
 243
 244         insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
 245         insert_caller_stat(call_site, bytes_req, bytes_alloc);
 246
 247         total_requested += bytes_req;
 248         total_allocated += bytes_alloc;
 249
 250         if (node) {
 251                 node1 = cpunode_map[cpu];
 252                 node2 = raw_field_value(event, "node", raw->data);
 253                 if (node1 != node2)
 254                         nr_cross_allocs++;
 255         }
 256         nr_allocs++;
 257 }
 258
 259 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
 260 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
 261
 262 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
 263                                             unsigned long call_site,
 264                                             struct rb_root *root,
 265                                             sort_fn_t sort_fn)
 266 {
 267         struct rb_node *node = root->rb_node;
 268         struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
 269
 270         while (node) {
 271                 struct alloc_stat *data;
 272                 int cmp;
 273
 274                 data = rb_entry(node, struct alloc_stat, node);
 275
 276                 cmp = sort_fn(&key, data);
 277                 if (cmp < 0)
 278                         node = node->rb_left;
 279                 else if (cmp > 0)
 280                         node = node->rb_right;
 281                 else
 282                         return data;
 283         }
 284         return NULL;
 285 }
 286
 287 static void process_free_event(struct raw_event_sample *raw,
 288                                struct event *event,
 289                                int cpu,
 290                                u64 timestamp __used,
 291                                struct thread *thread __used)
 292 {
 293         unsigned long ptr;
 294         struct alloc_stat *s_alloc, *s_caller;
 295
 296         ptr = raw_field_value(event, "ptr", raw->data);
 297
 298         s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
 299         if (!s_alloc)
 300                 return;
 301
 302         if (cpu != s_alloc->alloc_cpu) {
 303                 s_alloc->pingpong++;
 304
 305                 s_caller = search_alloc_stat(0, s_alloc->call_site,
 306                                              &root_caller_stat, callsite_cmp);
 307                 assert(s_caller);
 308                 s_caller->pingpong++;
 309         }
 310         s_alloc->alloc_cpu = -1;
 311 }
 312
 313 static void
 314 process_raw_event(event_t *raw_event __used, void *more_data,
 315                   int cpu, u64 timestamp, struct thread *thread)
 316 {
 317         struct raw_event_sample *raw = more_data;
 318         struct event *event;
 319         int type;
 320
 321         type = trace_parse_common_type(raw->data);
 322         event = trace_find_event(type);
 323
 324         if (!strcmp(event->name, "kmalloc") ||
 325             !strcmp(event->name, "kmem_cache_alloc")) {
 326                 process_alloc_event(raw, event, cpu, timestamp, thread, 0);
 327                 return;
 328         }
 329
 330         if (!strcmp(event->name, "kmalloc_node") ||
 331             !strcmp(event->name, "kmem_cache_alloc_node")) {
 332                 process_alloc_event(raw, event, cpu, timestamp, thread, 1);
 333                 return;
 334         }
 335
 336         if (!strcmp(event->name, "kfree") ||
 337             !strcmp(event->name, "kmem_cache_free")) {
 338                 process_free_event(raw, event, cpu, timestamp, thread);
 339                 return;
 340         }
 341 }
 342
 343 static int
 344 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
 345 {
 346         u64 ip = event->ip.ip;
 347         u64 timestamp = -1;
 348         u32 cpu = -1;
 349         u64 period = 1;
 350         void *more_data = event->ip.__more_data;
 351         struct thread *thread = threads__findnew(event->ip.pid);
 352
 353         if (sample_type & PERF_SAMPLE_TIME) {
 354                 timestamp = *(u64 *)more_data;
 355                 more_data += sizeof(u64);
 356         }
 357
 358         if (sample_type & PERF_SAMPLE_CPU) {
 359                 cpu = *(u32 *)more_data;
 360                 more_data += sizeof(u32);
 361                 more_data += sizeof(u32); /* reserved */
 362         }
 363
 364         if (sample_type & PERF_SAMPLE_PERIOD) {
 365                 period = *(u64 *)more_data;
 366                 more_data += sizeof(u64);
 367         }
 368
 369         dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
 370                 (void *)(offset + head),
 371                 (void *)(long)(event->header.size),
 372                 event->header.misc,
 373                 event->ip.pid, event->ip.tid,
 374                 (void *)(long)ip,
 375                 (long long)period);
 376
 377         if (thread == NULL) {
 378                 pr_debug("problem processing %d event, skipping it.\n",
 379                          event->header.type);
 380                 return -1;
 381         }
 382
 383         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 384
 385         process_raw_event(event, more_data, cpu, timestamp, thread);
 386
 387         return 0;
 388 }
 389
 390 static int sample_type_check(u64 type)
 391 {
 392         sample_type = type;
 393
 394         if (!(sample_type & PERF_SAMPLE_RAW)) {
 395                 fprintf(stderr,
 396                         "No trace sample to read. Did you call perf record "
 397                         "without -R?");
 398                 return -1;
 399         }
 400
 401         return 0;
 402 }
 403
 404 static struct perf_file_handler file_handler = {
 405         .process_sample_event   = process_sample_event,
 406         .process_comm_event     = process_comm_event,
 407         .sample_type_check      = sample_type_check,
 408 };
 409
 410 static int read_events(void)
 411 {
 412         register_idle_thread();
 413         register_perf_file_handler(&file_handler);
 414
 415         return mmap_dispatch_perf_file(&header, input_name, 0, 0,
 416                                        &cwdlen, &cwd);
 417 }
 418
 419 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
 420 {
 421         if (n_alloc == 0)
 422                 return 0.0;
 423         else
 424                 return 100.0 - (100.0 * n_req / n_alloc);
 425 }
 426
 427 static void __print_result(struct rb_root *root, int n_lines, int is_caller)
 428 {
 429         struct rb_node *next;
 430
 431         printf("%.102s\n", graph_dotted_line);
 432         printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
 433         printf(" Total_alloc/Per | Total_req/Per   | Hit   | Ping-pong | Frag\n");
 434         printf("%.102s\n", graph_dotted_line);
 435
 436         next = rb_first(root);
 437
 438         while (next && n_lines--) {
 439                 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
 440                                                    node);
 441                 struct symbol *sym = NULL;
 442                 char buf[BUFSIZ];
 443                 u64 addr;
 444
 445                 if (is_caller) {
 446                         addr = data->call_site;
 447                         if (!raw_ip)
 448                                 sym = kernel_maps__find_function(addr, NULL, NULL);
 449                 } else
 450                         addr = data->ptr;
 451
 452                 if (sym != NULL)
 453                         snprintf(buf, sizeof(buf), "%s+%Lx", sym->name,
 454                                  addr - sym->start);
 455                 else
 456                         snprintf(buf, sizeof(buf), "%#Lx", addr);
 457                 printf(" %-34s |", buf);
 458
 459                 printf(" %9llu/%-5lu | %9llu/%-5lu | %6lu | %8lu | %6.3f%%\n",
 460                        (unsigned long long)data->bytes_alloc,
 461                        (unsigned long)data->bytes_alloc / data->hit,
 462                        (unsigned long long)data->bytes_req,
 463                        (unsigned long)data->bytes_req / data->hit,
 464                        (unsigned long)data->hit,
 465                        (unsigned long)data->pingpong,
 466                        fragmentation(data->bytes_req, data->bytes_alloc));
 467
 468                 next = rb_next(next);
 469         }
 470
 471         if (n_lines == -1)
 472                 printf(" ...                                | ...             | ...             | ...    | ...      | ...   \n");
 473
 474         printf("%.102s\n", graph_dotted_line);
 475 }
 476
 477 static void print_summary(void)
 478 {
 479         printf("\nSUMMARY\n=======\n");
 480         printf("Total bytes requested: %lu\n", total_requested);
 481         printf("Total bytes allocated: %lu\n", total_allocated);
 482         printf("Total bytes wasted on internal fragmentation: %lu\n",
 483                total_allocated - total_requested);
 484         printf("Internal fragmentation: %f%%\n",
 485                fragmentation(total_requested, total_allocated));
 486         printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
 487 }
 488
 489 static void print_result(void)
 490 {
 491         if (caller_flag)
 492                 __print_result(&root_caller_sorted, caller_lines, 1);
 493         if (alloc_flag)
 494                 __print_result(&root_alloc_sorted, alloc_lines, 0);
 495         print_summary();
 496 }
 497
 498 struct sort_dimension {
 499         const char              name[20];
 500         sort_fn_t               cmp;
 501         struct list_head        list;
 502 };
 503
 504 static LIST_HEAD(caller_sort);
 505 static LIST_HEAD(alloc_sort);
 506
 507 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
 508                         struct list_head *sort_list)
 509 {
 510         struct rb_node **new = &(root->rb_node);
 511         struct rb_node *parent = NULL;
 512         struct sort_dimension *sort;
 513
 514         while (*new) {
 515                 struct alloc_stat *this;
 516                 int cmp = 0;
 517
 518                 this = rb_entry(*new, struct alloc_stat, node);
 519                 parent = *new;
 520
 521                 list_for_each_entry(sort, sort_list, list) {
 522                         cmp = sort->cmp(data, this);
 523                         if (cmp)
 524                                 break;
 525                 }
 526
 527                 if (cmp > 0)
 528                         new = &((*new)->rb_left);
 529                 else
 530                         new = &((*new)->rb_right);
 531         }
 532
 533         rb_link_node(&data->node, parent, new);
 534         rb_insert_color(&data->node, root);
 535 }
 536
 537 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
 538                           struct list_head *sort_list)
 539 {
 540         struct rb_node *node;
 541         struct alloc_stat *data;
 542
 543         for (;;) {
 544                 node = rb_first(root);
 545                 if (!node)
 546                         break;
 547
 548                 rb_erase(node, root);
 549                 data = rb_entry(node, struct alloc_stat, node);
 550                 sort_insert(root_sorted, data, sort_list);
 551         }
 552 }
 553
 554 static void sort_result(void)
 555 {
 556         __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
 557         __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
 558 }
 559
 560 static int __cmd_kmem(void)
 561 {
 562         setup_pager();
 563         read_events();
 564         sort_result();
 565         print_result();
 566
 567         return 0;
 568 }
 569
 570 static const char * const kmem_usage[] = {
 571         "perf kmem [<options>] {record}",
 572         NULL
 573 };
 574
 575 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
 576 {
 577         if (l->ptr < r->ptr)
 578                 return -1;
 579         else if (l->ptr > r->ptr)
 580                 return 1;
 581         return 0;
 582 }
 583
 584 static struct sort_dimension ptr_sort_dimension = {
 585         .name   = "ptr",
 586         .cmp    = ptr_cmp,
 587 };
 588
 589 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
 590 {
 591         if (l->call_site < r->call_site)
 592                 return -1;
 593         else if (l->call_site > r->call_site)
 594                 return 1;
 595         return 0;
 596 }
 597
 598 static struct sort_dimension callsite_sort_dimension = {
 599         .name   = "callsite",
 600         .cmp    = callsite_cmp,
 601 };
 602
 603 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
 604 {
 605         if (l->hit < r->hit)
 606                 return -1;
 607         else if (l->hit > r->hit)
 608                 return 1;
 609         return 0;
 610 }
 611
 612 static struct sort_dimension hit_sort_dimension = {
 613         .name   = "hit",
 614         .cmp    = hit_cmp,
 615 };
 616
 617 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
 618 {
 619         if (l->bytes_alloc < r->bytes_alloc)
 620                 return -1;
 621         else if (l->bytes_alloc > r->bytes_alloc)
 622                 return 1;
 623         return 0;
 624 }
 625
 626 static struct sort_dimension bytes_sort_dimension = {
 627         .name   = "bytes",
 628         .cmp    = bytes_cmp,
 629 };
 630
 631 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
 632 {
 633         double x, y;
 634
 635         x = fragmentation(l->bytes_req, l->bytes_alloc);
 636         y = fragmentation(r->bytes_req, r->bytes_alloc);
 637
 638         if (x < y)
 639                 return -1;
 640         else if (x > y)
 641                 return 1;
 642         return 0;
 643 }
 644
 645 static struct sort_dimension frag_sort_dimension = {
 646         .name   = "frag",
 647         .cmp    = frag_cmp,
 648 };
 649
 650 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
 651 {
 652         if (l->pingpong < r->pingpong)
 653                 return -1;
 654         else if (l->pingpong > r->pingpong)
 655                 return 1;
 656         return 0;
 657 }
 658
 659 static struct sort_dimension pingpong_sort_dimension = {
 660         .name   = "pingpong",
 661         .cmp    = pingpong_cmp,
 662 };
 663
 664 static struct sort_dimension *avail_sorts[] = {
 665         &ptr_sort_dimension,
 666         &callsite_sort_dimension,
 667         &hit_sort_dimension,
 668         &bytes_sort_dimension,
 669         &frag_sort_dimension,
 670         &pingpong_sort_dimension,
 671 };
 672
 673 #define NUM_AVAIL_SORTS \
 674         (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
 675
 676 static int sort_dimension__add(const char *tok, struct list_head *list)
 677 {
 678         struct sort_dimension *sort;
 679         int i;
 680
 681         for (i = 0; i < NUM_AVAIL_SORTS; i++) {
 682                 if (!strcmp(avail_sorts[i]->name, tok)) {
 683                         sort = malloc(sizeof(*sort));
 684                         if (!sort)
 685                                 die("malloc");
 686                         memcpy(sort, avail_sorts[i], sizeof(*sort));
 687                         list_add_tail(&sort->list, list);
 688                         return 0;
 689                 }
 690         }
 691
 692         return -1;
 693 }
 694
 695 static int setup_sorting(struct list_head *sort_list, const char *arg)
 696 {
 697         char *tok;
 698         char *str = strdup(arg);
 699
 700         if (!str)
 701                 die("strdup");
 702
 703         while (true) {
 704                 tok = strsep(&str, ",");
 705                 if (!tok)
 706                         break;
 707                 if (sort_dimension__add(tok, sort_list) < 0) {
 708                         error("Unknown --sort key: '%s'", tok);
 709                         return -1;
 710                 }
 711         }
 712
 713         free(str);
 714         return 0;
 715 }
 716
 717 static int parse_sort_opt(const struct option *opt __used,
 718                           const char *arg, int unset __used)
 719 {
 720         if (!arg)
 721                 return -1;
 722
 723         if (caller_flag > alloc_flag)
 724                 return setup_sorting(&caller_sort, arg);
 725         else
 726                 return setup_sorting(&alloc_sort, arg);
 727
 728         return 0;
 729 }
 730
 731 static int parse_stat_opt(const struct option *opt __used,
 732                           const char *arg, int unset __used)
 733 {
 734         if (!arg)
 735                 return -1;
 736
 737         if (strcmp(arg, "alloc") == 0)
 738                 alloc_flag = (caller_flag + 1);
 739         else if (strcmp(arg, "caller") == 0)
 740                 caller_flag = (alloc_flag + 1);
 741         else
 742                 return -1;
 743         return 0;
 744 }
 745
 746 static int parse_line_opt(const struct option *opt __used,
 747                           const char *arg, int unset __used)
 748 {
 749         int lines;
 750
 751         if (!arg)
 752                 return -1;
 753
 754         lines = strtoul(arg, NULL, 10);
 755
 756         if (caller_flag > alloc_flag)
 757                 caller_lines = lines;
 758         else
 759                 alloc_lines = lines;
 760
 761         return 0;
 762 }
 763
 764 static const struct option kmem_options[] = {
 765         OPT_STRING('i', "input", &input_name, "file",
 766                    "input file name"),
 767         OPT_CALLBACK(0, "stat", NULL, "<alloc>|<caller>",
 768                      "stat selector, Pass 'alloc' or 'caller'.",
 769                      parse_stat_opt),
 770         OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
 771                      "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
 772                      parse_sort_opt),
 773         OPT_CALLBACK('l', "line", NULL, "num",
 774                      "show n lins",
 775                      parse_line_opt),
 776         OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
 777         OPT_END()
 778 };
 779
 780 static const char *record_args[] = {
 781         "record",
 782         "-a",
 783         "-R",
 784         "-M",
 785         "-f",
 786         "-c", "1",
 787         "-e", "kmem:kmalloc",
 788         "-e", "kmem:kmalloc_node",
 789         "-e", "kmem:kfree",
 790         "-e", "kmem:kmem_cache_alloc",
 791         "-e", "kmem:kmem_cache_alloc_node",
 792         "-e", "kmem:kmem_cache_free",
 793 };
 794
 795 static int __cmd_record(int argc, const char **argv)
 796 {
 797         unsigned int rec_argc, i, j;
 798         const char **rec_argv;
 799
 800         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
 801         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 802
 803         for (i = 0; i < ARRAY_SIZE(record_args); i++)
 804                 rec_argv[i] = strdup(record_args[i]);
 805
 806         for (j = 1; j < (unsigned int)argc; j++, i++)
 807                 rec_argv[i] = argv[j];
 808
 809         return cmd_record(i, rec_argv, NULL);
 810 }
 811
 812 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
 813 {
 814         symbol__init(0);
 815
 816         argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
 817
 818         if (argc && !strncmp(argv[0], "rec", 3))
 819                 return __cmd_record(argc, argv);
 820         else if (argc)
 821                 usage_with_options(kmem_usage, kmem_options);
 822
 823         if (list_empty(&caller_sort))
 824                 setup_sorting(&caller_sort, default_sort_order);
 825         if (list_empty(&alloc_sort))
 826                 setup_sorting(&alloc_sort, default_sort_order);
 827
 828         setup_cpunode_map();
 829
 830         return __cmd_kmem();
 831 }
 832