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