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/config.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 #include "util/time-utils.h"
  15
  16 #include <subcmd/parse-options.h>
  17 #include "util/trace-event.h"
  18 #include "util/data.h"
  19 #include "util/cpumap.h"
  20
  21 #include "util/debug.h"
  22
  23 #include <linux/rbtree.h>
  24 #include <linux/string.h>
  25 #include <locale.h>
  26 #include <regex.h>
  27
  28 static int      kmem_slab;
  29 static int      kmem_page;
  30
  31 static long     kmem_page_size;
  32 static enum {
  33         KMEM_SLAB,
  34         KMEM_PAGE,
  35 } kmem_default = KMEM_SLAB;  /* for backward compatibility */
  36
  37 struct alloc_stat;
  38 typedef int (*sort_fn_t)(void *, void *);
  39
  40 static int                      alloc_flag;
  41 static int                      caller_flag;
  42
  43 static int                      alloc_lines = -1;
  44 static int                      caller_lines = -1;
  45
  46 static bool                     raw_ip;
  47
  48 struct alloc_stat {
  49         u64     call_site;
  50         u64     ptr;
  51         u64     bytes_req;
  52         u64     bytes_alloc;
  53         u64     last_alloc;
  54         u32     hit;
  55         u32     pingpong;
  56
  57         short   alloc_cpu;
  58
  59         struct rb_node node;
  60 };
  61
  62 static struct rb_root root_alloc_stat;
  63 static struct rb_root root_alloc_sorted;
  64 static struct rb_root root_caller_stat;
  65 static struct rb_root root_caller_sorted;
  66
  67 static unsigned long total_requested, total_allocated, total_freed;
  68 static unsigned long nr_allocs, nr_cross_allocs;
  69
  70 /* filters for controlling start and stop of time of analysis */
  71 static struct perf_time_interval ptime;
  72 const char *time_str;
  73
  74 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
  75                              int bytes_req, int bytes_alloc, int cpu)
  76 {
  77         struct rb_node **node = &root_alloc_stat.rb_node;
  78         struct rb_node *parent = NULL;
  79         struct alloc_stat *data = NULL;
  80
  81         while (*node) {
  82                 parent = *node;
  83                 data = rb_entry(*node, struct alloc_stat, node);
  84
  85                 if (ptr > data->ptr)
  86                         node = &(*node)->rb_right;
  87                 else if (ptr < data->ptr)
  88                         node = &(*node)->rb_left;
  89                 else
  90                         break;
  91         }
  92
  93         if (data && data->ptr == ptr) {
  94                 data->hit++;
  95                 data->bytes_req += bytes_req;
  96                 data->bytes_alloc += bytes_alloc;
  97         } else {
  98                 data = malloc(sizeof(*data));
  99                 if (!data) {
 100                         pr_err("%s: malloc failed\n", __func__);
 101                         return -1;
 102                 }
 103                 data->ptr = ptr;
 104                 data->pingpong = 0;
 105                 data->hit = 1;
 106                 data->bytes_req = bytes_req;
 107                 data->bytes_alloc = bytes_alloc;
 108
 109                 rb_link_node(&data->node, parent, node);
 110                 rb_insert_color(&data->node, &root_alloc_stat);
 111         }
 112         data->call_site = call_site;
 113         data->alloc_cpu = cpu;
 114         data->last_alloc = bytes_alloc;
 115
 116         return 0;
 117 }
 118
 119 static int insert_caller_stat(unsigned long call_site,
 120                               int bytes_req, int bytes_alloc)
 121 {
 122         struct rb_node **node = &root_caller_stat.rb_node;
 123         struct rb_node *parent = NULL;
 124         struct alloc_stat *data = NULL;
 125
 126         while (*node) {
 127                 parent = *node;
 128                 data = rb_entry(*node, struct alloc_stat, node);
 129
 130                 if (call_site > data->call_site)
 131                         node = &(*node)->rb_right;
 132                 else if (call_site < data->call_site)
 133                         node = &(*node)->rb_left;
 134                 else
 135                         break;
 136         }
 137
 138         if (data && data->call_site == call_site) {
 139                 data->hit++;
 140                 data->bytes_req += bytes_req;
 141                 data->bytes_alloc += bytes_alloc;
 142         } else {
 143                 data = malloc(sizeof(*data));
 144                 if (!data) {
 145                         pr_err("%s: malloc failed\n", __func__);
 146                         return -1;
 147                 }
 148                 data->call_site = call_site;
 149                 data->pingpong = 0;
 150                 data->hit = 1;
 151                 data->bytes_req = bytes_req;
 152                 data->bytes_alloc = bytes_alloc;
 153
 154                 rb_link_node(&data->node, parent, node);
 155                 rb_insert_color(&data->node, &root_caller_stat);
 156         }
 157
 158         return 0;
 159 }
 160
 161 static int perf_evsel__process_alloc_event(struct perf_evsel *evsel,
 162                                            struct perf_sample *sample)
 163 {
 164         unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
 165                       call_site = perf_evsel__intval(evsel, sample, "call_site");
 166         int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
 167             bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
 168
 169         if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
 170             insert_caller_stat(call_site, bytes_req, bytes_alloc))
 171                 return -1;
 172
 173         total_requested += bytes_req;
 174         total_allocated += bytes_alloc;
 175
 176         nr_allocs++;
 177         return 0;
 178 }
 179
 180 static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,
 181                                                 struct perf_sample *sample)
 182 {
 183         int ret = perf_evsel__process_alloc_event(evsel, sample);
 184
 185         if (!ret) {
 186                 int node1 = cpu__get_node(sample->cpu),
 187                     node2 = perf_evsel__intval(evsel, sample, "node");
 188
 189                 if (node1 != node2)
 190                         nr_cross_allocs++;
 191         }
 192
 193         return ret;
 194 }
 195
 196 static int ptr_cmp(void *, void *);
 197 static int slab_callsite_cmp(void *, void *);
 198
 199 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
 200                                             unsigned long call_site,
 201                                             struct rb_root *root,
 202                                             sort_fn_t sort_fn)
 203 {
 204         struct rb_node *node = root->rb_node;
 205         struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
 206
 207         while (node) {
 208                 struct alloc_stat *data;
 209                 int cmp;
 210
 211                 data = rb_entry(node, struct alloc_stat, node);
 212
 213                 cmp = sort_fn(&key, data);
 214                 if (cmp < 0)
 215                         node = node->rb_left;
 216                 else if (cmp > 0)
 217                         node = node->rb_right;
 218                 else
 219                         return data;
 220         }
 221         return NULL;
 222 }
 223
 224 static int perf_evsel__process_free_event(struct perf_evsel *evsel,
 225                                           struct perf_sample *sample)
 226 {
 227         unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
 228         struct alloc_stat *s_alloc, *s_caller;
 229
 230         s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
 231         if (!s_alloc)
 232                 return 0;
 233
 234         total_freed += s_alloc->last_alloc;
 235
 236         if ((short)sample->cpu != s_alloc->alloc_cpu) {
 237                 s_alloc->pingpong++;
 238
 239                 s_caller = search_alloc_stat(0, s_alloc->call_site,
 240                                              &root_caller_stat,
 241                                              slab_callsite_cmp);
 242                 if (!s_caller)
 243                         return -1;
 244                 s_caller->pingpong++;
 245         }
 246         s_alloc->alloc_cpu = -1;
 247
 248         return 0;
 249 }
 250
 251 static u64 total_page_alloc_bytes;
 252 static u64 total_page_free_bytes;
 253 static u64 total_page_nomatch_bytes;
 254 static u64 total_page_fail_bytes;
 255 static unsigned long nr_page_allocs;
 256 static unsigned long nr_page_frees;
 257 static unsigned long nr_page_fails;
 258 static unsigned long nr_page_nomatch;
 259
 260 static bool use_pfn;
 261 static bool live_page;
 262 static struct perf_session *kmem_session;
 263
 264 #define MAX_MIGRATE_TYPES  6
 265 #define MAX_PAGE_ORDER     11
 266
 267 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
 268
 269 struct page_stat {
 270         struct rb_node  node;
 271         u64             page;
 272         u64             callsite;
 273         int             order;
 274         unsigned        gfp_flags;
 275         unsigned        migrate_type;
 276         u64             alloc_bytes;
 277         u64             free_bytes;
 278         int             nr_alloc;
 279         int             nr_free;
 280 };
 281
 282 static struct rb_root page_live_tree;
 283 static struct rb_root page_alloc_tree;
 284 static struct rb_root page_alloc_sorted;
 285 static struct rb_root page_caller_tree;
 286 static struct rb_root page_caller_sorted;
 287
 288 struct alloc_func {
 289         u64 start;
 290         u64 end;
 291         char *name;
 292 };
 293
 294 static int nr_alloc_funcs;
 295 static struct alloc_func *alloc_func_list;
 296
 297 static int funcmp(const void *a, const void *b)
 298 {
 299         const struct alloc_func *fa = a;
 300         const struct alloc_func *fb = b;
 301
 302         if (fa->start > fb->start)
 303                 return 1;
 304         else
 305                 return -1;
 306 }
 307
 308 static int callcmp(const void *a, const void *b)
 309 {
 310         const struct alloc_func *fa = a;
 311         const struct alloc_func *fb = b;
 312
 313         if (fb->start <= fa->start && fa->end < fb->end)
 314                 return 0;
 315
 316         if (fa->start > fb->start)
 317                 return 1;
 318         else
 319                 return -1;
 320 }
 321
 322 static int build_alloc_func_list(void)
 323 {
 324         int ret;
 325         struct map *kernel_map;
 326         struct symbol *sym;
 327         struct rb_node *node;
 328         struct alloc_func *func;
 329         struct machine *machine = &kmem_session->machines.host;
 330         regex_t alloc_func_regex;
 331         const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
 332
 333         ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
 334         if (ret) {
 335                 char err[BUFSIZ];
 336
 337                 regerror(ret, &alloc_func_regex, err, sizeof(err));
 338                 pr_err("Invalid regex: %s\n%s", pattern, err);
 339                 return -EINVAL;
 340         }
 341
 342         kernel_map = machine__kernel_map(machine);
 343         if (map__load(kernel_map) < 0) {
 344                 pr_err("cannot load kernel map\n");
 345                 return -ENOENT;
 346         }
 347
 348         map__for_each_symbol(kernel_map, sym, node) {
 349                 if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
 350                         continue;
 351
 352                 func = realloc(alloc_func_list,
 353                                (nr_alloc_funcs + 1) * sizeof(*func));
 354                 if (func == NULL)
 355                         return -ENOMEM;
 356
 357                 pr_debug("alloc func: %s\n", sym->name);
 358                 func[nr_alloc_funcs].start = sym->start;
 359                 func[nr_alloc_funcs].end   = sym->end;
 360                 func[nr_alloc_funcs].name  = sym->name;
 361
 362                 alloc_func_list = func;
 363                 nr_alloc_funcs++;
 364         }
 365
 366         qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
 367
 368         regfree(&alloc_func_regex);
 369         return 0;
 370 }
 371
 372 /*
 373  * Find first non-memory allocation function from callchain.
 374  * The allocation functions are in the 'alloc_func_list'.
 375  */
 376 static u64 find_callsite(struct perf_evsel *evsel, struct perf_sample *sample)
 377 {
 378         struct addr_location al;
 379         struct machine *machine = &kmem_session->machines.host;
 380         struct callchain_cursor_node *node;
 381
 382         if (alloc_func_list == NULL) {
 383                 if (build_alloc_func_list() < 0)
 384                         goto out;
 385         }
 386
 387         al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
 388         sample__resolve_callchain(sample, &callchain_cursor, NULL, evsel, &al, 16);
 389
 390         callchain_cursor_commit(&callchain_cursor);
 391         while (true) {
 392                 struct alloc_func key, *caller;
 393                 u64 addr;
 394
 395                 node = callchain_cursor_current(&callchain_cursor);
 396                 if (node == NULL)
 397                         break;
 398
 399                 key.start = key.end = node->ip;
 400                 caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
 401                                  sizeof(key), callcmp);
 402                 if (!caller) {
 403                         /* found */
 404                         if (node->map)
 405                                 addr = map__unmap_ip(node->map, node->ip);
 406                         else
 407                                 addr = node->ip;
 408
 409                         return addr;
 410                 } else
 411                         pr_debug3("skipping alloc function: %s\n", caller->name);
 412
 413                 callchain_cursor_advance(&callchain_cursor);
 414         }
 415
 416 out:
 417         pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
 418         return sample->ip;
 419 }
 420
 421 struct sort_dimension {
 422         const char              name[20];
 423         sort_fn_t               cmp;
 424         struct list_head        list;
 425 };
 426
 427 static LIST_HEAD(page_alloc_sort_input);
 428 static LIST_HEAD(page_caller_sort_input);
 429
 430 static struct page_stat *
 431 __page_stat__findnew_page(struct page_stat *pstat, bool create)
 432 {
 433         struct rb_node **node = &page_live_tree.rb_node;
 434         struct rb_node *parent = NULL;
 435         struct page_stat *data;
 436
 437         while (*node) {
 438                 s64 cmp;
 439
 440                 parent = *node;
 441                 data = rb_entry(*node, struct page_stat, node);
 442
 443                 cmp = data->page - pstat->page;
 444                 if (cmp < 0)
 445                         node = &parent->rb_left;
 446                 else if (cmp > 0)
 447                         node = &parent->rb_right;
 448                 else
 449                         return data;
 450         }
 451
 452         if (!create)
 453                 return NULL;
 454
 455         data = zalloc(sizeof(*data));
 456         if (data != NULL) {
 457                 data->page = pstat->page;
 458                 data->order = pstat->order;
 459                 data->gfp_flags = pstat->gfp_flags;
 460                 data->migrate_type = pstat->migrate_type;
 461
 462                 rb_link_node(&data->node, parent, node);
 463                 rb_insert_color(&data->node, &page_live_tree);
 464         }
 465
 466         return data;
 467 }
 468
 469 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
 470 {
 471         return __page_stat__findnew_page(pstat, false);
 472 }
 473
 474 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
 475 {
 476         return __page_stat__findnew_page(pstat, true);
 477 }
 478
 479 static struct page_stat *
 480 __page_stat__findnew_alloc(struct page_stat *pstat, bool create)
 481 {
 482         struct rb_node **node = &page_alloc_tree.rb_node;
 483         struct rb_node *parent = NULL;
 484         struct page_stat *data;
 485         struct sort_dimension *sort;
 486
 487         while (*node) {
 488                 int cmp = 0;
 489
 490                 parent = *node;
 491                 data = rb_entry(*node, struct page_stat, node);
 492
 493                 list_for_each_entry(sort, &page_alloc_sort_input, list) {
 494                         cmp = sort->cmp(pstat, data);
 495                         if (cmp)
 496                                 break;
 497                 }
 498
 499                 if (cmp < 0)
 500                         node = &parent->rb_left;
 501                 else if (cmp > 0)
 502                         node = &parent->rb_right;
 503                 else
 504                         return data;
 505         }
 506
 507         if (!create)
 508                 return NULL;
 509
 510         data = zalloc(sizeof(*data));
 511         if (data != NULL) {
 512                 data->page = pstat->page;
 513                 data->order = pstat->order;
 514                 data->gfp_flags = pstat->gfp_flags;
 515                 data->migrate_type = pstat->migrate_type;
 516
 517                 rb_link_node(&data->node, parent, node);
 518                 rb_insert_color(&data->node, &page_alloc_tree);
 519         }
 520
 521         return data;
 522 }
 523
 524 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
 525 {
 526         return __page_stat__findnew_alloc(pstat, false);
 527 }
 528
 529 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
 530 {
 531         return __page_stat__findnew_alloc(pstat, true);
 532 }
 533
 534 static struct page_stat *
 535 __page_stat__findnew_caller(struct page_stat *pstat, bool create)
 536 {
 537         struct rb_node **node = &page_caller_tree.rb_node;
 538         struct rb_node *parent = NULL;
 539         struct page_stat *data;
 540         struct sort_dimension *sort;
 541
 542         while (*node) {
 543                 int cmp = 0;
 544
 545                 parent = *node;
 546                 data = rb_entry(*node, struct page_stat, node);
 547
 548                 list_for_each_entry(sort, &page_caller_sort_input, list) {
 549                         cmp = sort->cmp(pstat, data);
 550                         if (cmp)
 551                                 break;
 552                 }
 553
 554                 if (cmp < 0)
 555                         node = &parent->rb_left;
 556                 else if (cmp > 0)
 557                         node = &parent->rb_right;
 558                 else
 559                         return data;
 560         }
 561
 562         if (!create)
 563                 return NULL;
 564
 565         data = zalloc(sizeof(*data));
 566         if (data != NULL) {
 567                 data->callsite = pstat->callsite;
 568                 data->order = pstat->order;
 569                 data->gfp_flags = pstat->gfp_flags;
 570                 data->migrate_type = pstat->migrate_type;
 571
 572                 rb_link_node(&data->node, parent, node);
 573                 rb_insert_color(&data->node, &page_caller_tree);
 574         }
 575
 576         return data;
 577 }
 578
 579 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
 580 {
 581         return __page_stat__findnew_caller(pstat, false);
 582 }
 583
 584 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
 585 {
 586         return __page_stat__findnew_caller(pstat, true);
 587 }
 588
 589 static bool valid_page(u64 pfn_or_page)
 590 {
 591         if (use_pfn && pfn_or_page == -1UL)
 592                 return false;
 593         if (!use_pfn && pfn_or_page == 0)
 594                 return false;
 595         return true;
 596 }
 597
 598 struct gfp_flag {
 599         unsigned int flags;
 600         char *compact_str;
 601         char *human_readable;
 602 };
 603
 604 static struct gfp_flag *gfps;
 605 static int nr_gfps;
 606
 607 static int gfpcmp(const void *a, const void *b)
 608 {
 609         const struct gfp_flag *fa = a;
 610         const struct gfp_flag *fb = b;
 611
 612         return fa->flags - fb->flags;
 613 }
 614
 615 /* see include/trace/events/mmflags.h */
 616 static const struct {
 617         const char *original;
 618         const char *compact;
 619 } gfp_compact_table[] = {
 620         { "GFP_TRANSHUGE",              "THP" },
 621         { "GFP_TRANSHUGE_LIGHT",        "THL" },
 622         { "GFP_HIGHUSER_MOVABLE",       "HUM" },
 623         { "GFP_HIGHUSER",               "HU" },
 624         { "GFP_USER",                   "U" },
 625         { "GFP_TEMPORARY",              "TMP" },
 626         { "GFP_KERNEL_ACCOUNT",         "KAC" },
 627         { "GFP_KERNEL",                 "K" },
 628         { "GFP_NOFS",                   "NF" },
 629         { "GFP_ATOMIC",                 "A" },
 630         { "GFP_NOIO",                   "NI" },
 631         { "GFP_NOWAIT",                 "NW" },
 632         { "GFP_DMA",                    "D" },
 633         { "__GFP_HIGHMEM",              "HM" },
 634         { "GFP_DMA32",                  "D32" },
 635         { "__GFP_HIGH",                 "H" },
 636         { "__GFP_ATOMIC",               "_A" },
 637         { "__GFP_IO",                   "I" },
 638         { "__GFP_FS",                   "F" },
 639         { "__GFP_COLD",                 "CO" },
 640         { "__GFP_NOWARN",               "NWR" },
 641         { "__GFP_REPEAT",               "R" },
 642         { "__GFP_NOFAIL",               "NF" },
 643         { "__GFP_NORETRY",              "NR" },
 644         { "__GFP_COMP",                 "C" },
 645         { "__GFP_ZERO",                 "Z" },
 646         { "__GFP_NOMEMALLOC",           "NMA" },
 647         { "__GFP_MEMALLOC",             "MA" },
 648         { "__GFP_HARDWALL",             "HW" },
 649         { "__GFP_THISNODE",             "TN" },
 650         { "__GFP_RECLAIMABLE",          "RC" },
 651         { "__GFP_MOVABLE",              "M" },
 652         { "__GFP_ACCOUNT",              "AC" },
 653         { "__GFP_NOTRACK",              "NT" },
 654         { "__GFP_WRITE",                "WR" },
 655         { "__GFP_RECLAIM",              "R" },
 656         { "__GFP_DIRECT_RECLAIM",       "DR" },
 657         { "__GFP_KSWAPD_RECLAIM",       "KR" },
 658 };
 659
 660 static size_t max_gfp_len;
 661
 662 static char *compact_gfp_flags(char *gfp_flags)
 663 {
 664         char *orig_flags = strdup(gfp_flags);
 665         char *new_flags = NULL;
 666         char *str, *pos = NULL;
 667         size_t len = 0;
 668
 669         if (orig_flags == NULL)
 670                 return NULL;
 671
 672         str = strtok_r(orig_flags, "|", &pos);
 673         while (str) {
 674                 size_t i;
 675                 char *new;
 676                 const char *cpt;
 677
 678                 for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
 679                         if (strcmp(gfp_compact_table[i].original, str))
 680                                 continue;
 681
 682                         cpt = gfp_compact_table[i].compact;
 683                         new = realloc(new_flags, len + strlen(cpt) + 2);
 684                         if (new == NULL) {
 685                                 free(new_flags);
 686                                 return NULL;
 687                         }
 688
 689                         new_flags = new;
 690
 691                         if (!len) {
 692                                 strcpy(new_flags, cpt);
 693                         } else {
 694                                 strcat(new_flags, "|");
 695                                 strcat(new_flags, cpt);
 696                                 len++;
 697                         }
 698
 699                         len += strlen(cpt);
 700                 }
 701
 702                 str = strtok_r(NULL, "|", &pos);
 703         }
 704
 705         if (max_gfp_len < len)
 706                 max_gfp_len = len;
 707
 708         free(orig_flags);
 709         return new_flags;
 710 }
 711
 712 static char *compact_gfp_string(unsigned long gfp_flags)
 713 {
 714         struct gfp_flag key = {
 715                 .flags = gfp_flags,
 716         };
 717         struct gfp_flag *gfp;
 718
 719         gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
 720         if (gfp)
 721                 return gfp->compact_str;
 722
 723         return NULL;
 724 }
 725
 726 static int parse_gfp_flags(struct perf_evsel *evsel, struct perf_sample *sample,
 727                            unsigned int gfp_flags)
 728 {
 729         struct pevent_record record = {
 730                 .cpu = sample->cpu,
 731                 .data = sample->raw_data,
 732                 .size = sample->raw_size,
 733         };
 734         struct trace_seq seq;
 735         char *str, *pos = NULL;
 736
 737         if (nr_gfps) {
 738                 struct gfp_flag key = {
 739                         .flags = gfp_flags,
 740                 };
 741
 742                 if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
 743                         return 0;
 744         }
 745
 746         trace_seq_init(&seq);
 747         pevent_event_info(&seq, evsel->tp_format, &record);
 748
 749         str = strtok_r(seq.buffer, " ", &pos);
 750         while (str) {
 751                 if (!strncmp(str, "gfp_flags=", 10)) {
 752                         struct gfp_flag *new;
 753
 754                         new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
 755                         if (new == NULL)
 756                                 return -ENOMEM;
 757
 758                         gfps = new;
 759                         new += nr_gfps++;
 760
 761                         new->flags = gfp_flags;
 762                         new->human_readable = strdup(str + 10);
 763                         new->compact_str = compact_gfp_flags(str + 10);
 764                         if (!new->human_readable || !new->compact_str)
 765                                 return -ENOMEM;
 766
 767                         qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
 768                 }
 769
 770                 str = strtok_r(NULL, " ", &pos);
 771         }
 772
 773         trace_seq_destroy(&seq);
 774         return 0;
 775 }
 776
 777 static int perf_evsel__process_page_alloc_event(struct perf_evsel *evsel,
 778                                                 struct perf_sample *sample)
 779 {
 780         u64 page;
 781         unsigned int order = perf_evsel__intval(evsel, sample, "order");
 782         unsigned int gfp_flags = perf_evsel__intval(evsel, sample, "gfp_flags");
 783         unsigned int migrate_type = perf_evsel__intval(evsel, sample,
 784                                                        "migratetype");
 785         u64 bytes = kmem_page_size << order;
 786         u64 callsite;
 787         struct page_stat *pstat;
 788         struct page_stat this = {
 789                 .order = order,
 790                 .gfp_flags = gfp_flags,
 791                 .migrate_type = migrate_type,
 792         };
 793
 794         if (use_pfn)
 795                 page = perf_evsel__intval(evsel, sample, "pfn");
 796         else
 797                 page = perf_evsel__intval(evsel, sample, "page");
 798
 799         nr_page_allocs++;
 800         total_page_alloc_bytes += bytes;
 801
 802         if (!valid_page(page)) {
 803                 nr_page_fails++;
 804                 total_page_fail_bytes += bytes;
 805
 806                 return 0;
 807         }
 808
 809         if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
 810                 return -1;
 811
 812         callsite = find_callsite(evsel, sample);
 813
 814         /*
 815          * This is to find the current page (with correct gfp flags and
 816          * migrate type) at free event.
 817          */
 818         this.page = page;
 819         pstat = page_stat__findnew_page(&this);
 820         if (pstat == NULL)
 821                 return -ENOMEM;
 822
 823         pstat->nr_alloc++;
 824         pstat->alloc_bytes += bytes;
 825         pstat->callsite = callsite;
 826
 827         if (!live_page) {
 828                 pstat = page_stat__findnew_alloc(&this);
 829                 if (pstat == NULL)
 830                         return -ENOMEM;
 831
 832                 pstat->nr_alloc++;
 833                 pstat->alloc_bytes += bytes;
 834                 pstat->callsite = callsite;
 835         }
 836
 837         this.callsite = callsite;
 838         pstat = page_stat__findnew_caller(&this);
 839         if (pstat == NULL)
 840                 return -ENOMEM;
 841
 842         pstat->nr_alloc++;
 843         pstat->alloc_bytes += bytes;
 844
 845         order_stats[order][migrate_type]++;
 846
 847         return 0;
 848 }
 849
 850 static int perf_evsel__process_page_free_event(struct perf_evsel *evsel,
 851                                                 struct perf_sample *sample)
 852 {
 853         u64 page;
 854         unsigned int order = perf_evsel__intval(evsel, sample, "order");
 855         u64 bytes = kmem_page_size << order;
 856         struct page_stat *pstat;
 857         struct page_stat this = {
 858                 .order = order,
 859         };
 860
 861         if (use_pfn)
 862                 page = perf_evsel__intval(evsel, sample, "pfn");
 863         else
 864                 page = perf_evsel__intval(evsel, sample, "page");
 865
 866         nr_page_frees++;
 867         total_page_free_bytes += bytes;
 868
 869         this.page = page;
 870         pstat = page_stat__find_page(&this);
 871         if (pstat == NULL) {
 872                 pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
 873                           page, order);
 874
 875                 nr_page_nomatch++;
 876                 total_page_nomatch_bytes += bytes;
 877
 878                 return 0;
 879         }
 880
 881         this.gfp_flags = pstat->gfp_flags;
 882         this.migrate_type = pstat->migrate_type;
 883         this.callsite = pstat->callsite;
 884
 885         rb_erase(&pstat->node, &page_live_tree);
 886         free(pstat);
 887
 888         if (live_page) {
 889                 order_stats[this.order][this.migrate_type]--;
 890         } else {
 891                 pstat = page_stat__find_alloc(&this);
 892                 if (pstat == NULL)
 893                         return -ENOMEM;
 894
 895                 pstat->nr_free++;
 896                 pstat->free_bytes += bytes;
 897         }
 898
 899         pstat = page_stat__find_caller(&this);
 900         if (pstat == NULL)
 901                 return -ENOENT;
 902
 903         pstat->nr_free++;
 904         pstat->free_bytes += bytes;
 905
 906         if (live_page) {
 907                 pstat->nr_alloc--;
 908                 pstat->alloc_bytes -= bytes;
 909
 910                 if (pstat->nr_alloc == 0) {
 911                         rb_erase(&pstat->node, &page_caller_tree);
 912                         free(pstat);
 913                 }
 914         }
 915
 916         return 0;
 917 }
 918
 919 static bool perf_kmem__skip_sample(struct perf_sample *sample)
 920 {
 921         /* skip sample based on time? */
 922         if (perf_time__skip_sample(&ptime, sample->time))
 923                 return true;
 924
 925         return false;
 926 }
 927
 928 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
 929                                   struct perf_sample *sample);
 930
 931 static int process_sample_event(struct perf_tool *tool __maybe_unused,
 932                                 union perf_event *event,
 933                                 struct perf_sample *sample,
 934                                 struct perf_evsel *evsel,
 935                                 struct machine *machine)
 936 {
 937         int err = 0;
 938         struct thread *thread = machine__findnew_thread(machine, sample->pid,
 939                                                         sample->tid);
 940
 941         if (thread == NULL) {
 942                 pr_debug("problem processing %d event, skipping it.\n",
 943                          event->header.type);
 944                 return -1;
 945         }
 946
 947         if (perf_kmem__skip_sample(sample))
 948                 return 0;
 949
 950         dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
 951
 952         if (evsel->handler != NULL) {
 953                 tracepoint_handler f = evsel->handler;
 954                 err = f(evsel, sample);
 955         }
 956
 957         thread__put(thread);
 958
 959         return err;
 960 }
 961
 962 static struct perf_tool perf_kmem = {
 963         .sample          = process_sample_event,
 964         .comm            = perf_event__process_comm,
 965         .mmap            = perf_event__process_mmap,
 966         .mmap2           = perf_event__process_mmap2,
 967         .ordered_events  = true,
 968 };
 969
 970 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
 971 {
 972         if (n_alloc == 0)
 973                 return 0.0;
 974         else
 975                 return 100.0 - (100.0 * n_req / n_alloc);
 976 }
 977
 978 static void __print_slab_result(struct rb_root *root,
 979                                 struct perf_session *session,
 980                                 int n_lines, int is_caller)
 981 {
 982         struct rb_node *next;
 983         struct machine *machine = &session->machines.host;
 984
 985         printf("%.105s\n", graph_dotted_line);
 986         printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
 987         printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
 988         printf("%.105s\n", graph_dotted_line);
 989
 990         next = rb_first(root);
 991
 992         while (next && n_lines--) {
 993                 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
 994                                                    node);
 995                 struct symbol *sym = NULL;
 996                 struct map *map;
 997                 char buf[BUFSIZ];
 998                 u64 addr;
 999
1000                 if (is_caller) {
1001                         addr = data->call_site;
1002                         if (!raw_ip)
1003                                 sym = machine__find_kernel_function(machine, addr, &map);
1004                 } else
1005                         addr = data->ptr;
1006
1007                 if (sym != NULL)
1008                         snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1009                                  addr - map->unmap_ip(map, sym->start));
1010                 else
1011                         snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1012                 printf(" %-34s |", buf);
1013
1014                 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1015                        (unsigned long long)data->bytes_alloc,
1016                        (unsigned long)data->bytes_alloc / data->hit,
1017                        (unsigned long long)data->bytes_req,
1018                        (unsigned long)data->bytes_req / data->hit,
1019                        (unsigned long)data->hit,
1020                        (unsigned long)data->pingpong,
1021                        fragmentation(data->bytes_req, data->bytes_alloc));
1022
1023                 next = rb_next(next);
1024         }
1025
1026         if (n_lines == -1)
1027                 printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
1028
1029         printf("%.105s\n", graph_dotted_line);
1030 }
1031
1032 static const char * const migrate_type_str[] = {
1033         "UNMOVABL",
1034         "RECLAIM",
1035         "MOVABLE",
1036         "RESERVED",
1037         "CMA/ISLT",
1038         "UNKNOWN",
1039 };
1040
1041 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1042 {
1043         struct rb_node *next = rb_first(&page_alloc_sorted);
1044         struct machine *machine = &session->machines.host;
1045         const char *format;
1046         int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1047
1048         printf("\n%.105s\n", graph_dotted_line);
1049         printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1050                use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1051                gfp_len, "GFP flags");
1052         printf("%.105s\n", graph_dotted_line);
1053
1054         if (use_pfn)
1055                 format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1056         else
1057                 format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1058
1059         while (next && n_lines--) {
1060                 struct page_stat *data;
1061                 struct symbol *sym;
1062                 struct map *map;
1063                 char buf[32];
1064                 char *caller = buf;
1065
1066                 data = rb_entry(next, struct page_stat, node);
1067                 sym = machine__find_kernel_function(machine, data->callsite, &map);
1068                 if (sym && sym->name)
1069                         caller = sym->name;
1070                 else
1071                         scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1072
1073                 printf(format, (unsigned long long)data->page,
1074                        (unsigned long long)data->alloc_bytes / 1024,
1075                        data->nr_alloc, data->order,
1076                        migrate_type_str[data->migrate_type],
1077                        gfp_len, compact_gfp_string(data->gfp_flags), caller);
1078
1079                 next = rb_next(next);
1080         }
1081
1082         if (n_lines == -1) {
1083                 printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1084                        gfp_len, "...");
1085         }
1086
1087         printf("%.105s\n", graph_dotted_line);
1088 }
1089
1090 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1091 {
1092         struct rb_node *next = rb_first(&page_caller_sorted);
1093         struct machine *machine = &session->machines.host;
1094         int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1095
1096         printf("\n%.105s\n", graph_dotted_line);
1097         printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1098                live_page ? "Live" : "Total", gfp_len, "GFP flags");
1099         printf("%.105s\n", graph_dotted_line);
1100
1101         while (next && n_lines--) {
1102                 struct page_stat *data;
1103                 struct symbol *sym;
1104                 struct map *map;
1105                 char buf[32];
1106                 char *caller = buf;
1107
1108                 data = rb_entry(next, struct page_stat, node);
1109                 sym = machine__find_kernel_function(machine, data->callsite, &map);
1110                 if (sym && sym->name)
1111                         caller = sym->name;
1112                 else
1113                         scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1114
1115                 printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1116                        (unsigned long long)data->alloc_bytes / 1024,
1117                        data->nr_alloc, data->order,
1118                        migrate_type_str[data->migrate_type],
1119                        gfp_len, compact_gfp_string(data->gfp_flags), caller);
1120
1121                 next = rb_next(next);
1122         }
1123
1124         if (n_lines == -1) {
1125                 printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1126                        gfp_len, "...");
1127         }
1128
1129         printf("%.105s\n", graph_dotted_line);
1130 }
1131
1132 static void print_gfp_flags(void)
1133 {
1134         int i;
1135
1136         printf("#\n");
1137         printf("# GFP flags\n");
1138         printf("# ---------\n");
1139         for (i = 0; i < nr_gfps; i++) {
1140                 printf("# %08x: %*s: %s\n", gfps[i].flags,
1141                        (int) max_gfp_len, gfps[i].compact_str,
1142                        gfps[i].human_readable);
1143         }
1144 }
1145
1146 static void print_slab_summary(void)
1147 {
1148         printf("\nSUMMARY (SLAB allocator)");
1149         printf("\n========================\n");
1150         printf("Total bytes requested: %'lu\n", total_requested);
1151         printf("Total bytes allocated: %'lu\n", total_allocated);
1152         printf("Total bytes freed:     %'lu\n", total_freed);
1153         if (total_allocated > total_freed) {
1154                 printf("Net total bytes allocated: %'lu\n",
1155                 total_allocated - total_freed);
1156         }
1157         printf("Total bytes wasted on internal fragmentation: %'lu\n",
1158                total_allocated - total_requested);
1159         printf("Internal fragmentation: %f%%\n",
1160                fragmentation(total_requested, total_allocated));
1161         printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1162 }
1163
1164 static void print_page_summary(void)
1165 {
1166         int o, m;
1167         u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1168         u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1169
1170         printf("\nSUMMARY (page allocator)");
1171         printf("\n========================\n");
1172         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1173                nr_page_allocs, total_page_alloc_bytes / 1024);
1174         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1175                nr_page_frees, total_page_free_bytes / 1024);
1176         printf("\n");
1177
1178         printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1179                nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1180         printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1181                nr_page_allocs - nr_alloc_freed,
1182                (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1183         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1184                nr_page_nomatch, total_page_nomatch_bytes / 1024);
1185         printf("\n");
1186
1187         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1188                nr_page_fails, total_page_fail_bytes / 1024);
1189         printf("\n");
1190
1191         printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1192                "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1193         printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1194                graph_dotted_line, graph_dotted_line, graph_dotted_line,
1195                graph_dotted_line, graph_dotted_line);
1196
1197         for (o = 0; o < MAX_PAGE_ORDER; o++) {
1198                 printf("%5d", o);
1199                 for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1200                         if (order_stats[o][m])
1201                                 printf("  %'12d", order_stats[o][m]);
1202                         else
1203                                 printf("  %12c", '.');
1204                 }
1205                 printf("\n");
1206         }
1207 }
1208
1209 static void print_slab_result(struct perf_session *session)
1210 {
1211         if (caller_flag)
1212                 __print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1213         if (alloc_flag)
1214                 __print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1215         print_slab_summary();
1216 }
1217
1218 static void print_page_result(struct perf_session *session)
1219 {
1220         if (caller_flag || alloc_flag)
1221                 print_gfp_flags();
1222         if (caller_flag)
1223                 __print_page_caller_result(session, caller_lines);
1224         if (alloc_flag)
1225                 __print_page_alloc_result(session, alloc_lines);
1226         print_page_summary();
1227 }
1228
1229 static void print_result(struct perf_session *session)
1230 {
1231         if (kmem_slab)
1232                 print_slab_result(session);
1233         if (kmem_page)
1234                 print_page_result(session);
1235 }
1236
1237 static LIST_HEAD(slab_caller_sort);
1238 static LIST_HEAD(slab_alloc_sort);
1239 static LIST_HEAD(page_caller_sort);
1240 static LIST_HEAD(page_alloc_sort);
1241
1242 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1243                              struct list_head *sort_list)
1244 {
1245         struct rb_node **new = &(root->rb_node);
1246         struct rb_node *parent = NULL;
1247         struct sort_dimension *sort;
1248
1249         while (*new) {
1250                 struct alloc_stat *this;
1251                 int cmp = 0;
1252
1253                 this = rb_entry(*new, struct alloc_stat, node);
1254                 parent = *new;
1255
1256                 list_for_each_entry(sort, sort_list, list) {
1257                         cmp = sort->cmp(data, this);
1258                         if (cmp)
1259                                 break;
1260                 }
1261
1262                 if (cmp > 0)
1263                         new = &((*new)->rb_left);
1264                 else
1265                         new = &((*new)->rb_right);
1266         }
1267
1268         rb_link_node(&data->node, parent, new);
1269         rb_insert_color(&data->node, root);
1270 }
1271
1272 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1273                                struct list_head *sort_list)
1274 {
1275         struct rb_node *node;
1276         struct alloc_stat *data;
1277
1278         for (;;) {
1279                 node = rb_first(root);
1280                 if (!node)
1281                         break;
1282
1283                 rb_erase(node, root);
1284                 data = rb_entry(node, struct alloc_stat, node);
1285                 sort_slab_insert(root_sorted, data, sort_list);
1286         }
1287 }
1288
1289 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1290                              struct list_head *sort_list)
1291 {
1292         struct rb_node **new = &root->rb_node;
1293         struct rb_node *parent = NULL;
1294         struct sort_dimension *sort;
1295
1296         while (*new) {
1297                 struct page_stat *this;
1298                 int cmp = 0;
1299
1300                 this = rb_entry(*new, struct page_stat, node);
1301                 parent = *new;
1302
1303                 list_for_each_entry(sort, sort_list, list) {
1304                         cmp = sort->cmp(data, this);
1305                         if (cmp)
1306                                 break;
1307                 }
1308
1309                 if (cmp > 0)
1310                         new = &parent->rb_left;
1311                 else
1312                         new = &parent->rb_right;
1313         }
1314
1315         rb_link_node(&data->node, parent, new);
1316         rb_insert_color(&data->node, root);
1317 }
1318
1319 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1320                                struct list_head *sort_list)
1321 {
1322         struct rb_node *node;
1323         struct page_stat *data;
1324
1325         for (;;) {
1326                 node = rb_first(root);
1327                 if (!node)
1328                         break;
1329
1330                 rb_erase(node, root);
1331                 data = rb_entry(node, struct page_stat, node);
1332                 sort_page_insert(root_sorted, data, sort_list);
1333         }
1334 }
1335
1336 static void sort_result(void)
1337 {
1338         if (kmem_slab) {
1339                 __sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1340                                    &slab_alloc_sort);
1341                 __sort_slab_result(&root_caller_stat, &root_caller_sorted,
1342                                    &slab_caller_sort);
1343         }
1344         if (kmem_page) {
1345                 if (live_page)
1346                         __sort_page_result(&page_live_tree, &page_alloc_sorted,
1347                                            &page_alloc_sort);
1348                 else
1349                         __sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1350                                            &page_alloc_sort);
1351
1352                 __sort_page_result(&page_caller_tree, &page_caller_sorted,
1353                                    &page_caller_sort);
1354         }
1355 }
1356
1357 static int __cmd_kmem(struct perf_session *session)
1358 {
1359         int err = -EINVAL;
1360         struct perf_evsel *evsel;
1361         const struct perf_evsel_str_handler kmem_tracepoints[] = {
1362                 /* slab allocator */
1363                 { "kmem:kmalloc",               perf_evsel__process_alloc_event, },
1364                 { "kmem:kmem_cache_alloc",      perf_evsel__process_alloc_event, },
1365                 { "kmem:kmalloc_node",          perf_evsel__process_alloc_node_event, },
1366                 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
1367                 { "kmem:kfree",                 perf_evsel__process_free_event, },
1368                 { "kmem:kmem_cache_free",       perf_evsel__process_free_event, },
1369                 /* page allocator */
1370                 { "kmem:mm_page_alloc",         perf_evsel__process_page_alloc_event, },
1371                 { "kmem:mm_page_free",          perf_evsel__process_page_free_event, },
1372         };
1373
1374         if (!perf_session__has_traces(session, "kmem record"))
1375                 goto out;
1376
1377         if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1378                 pr_err("Initializing perf session tracepoint handlers failed\n");
1379                 goto out;
1380         }
1381
1382         evlist__for_each_entry(session->evlist, evsel) {
1383                 if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") &&
1384                     perf_evsel__field(evsel, "pfn")) {
1385                         use_pfn = true;
1386                         break;
1387                 }
1388         }
1389
1390         setup_pager();
1391         err = perf_session__process_events(session);
1392         if (err != 0) {
1393                 pr_err("error during process events: %d\n", err);
1394                 goto out;
1395         }
1396         sort_result();
1397         print_result(session);
1398 out:
1399         return err;
1400 }
1401
1402 /* slab sort keys */
1403 static int ptr_cmp(void *a, void *b)
1404 {
1405         struct alloc_stat *l = a;
1406         struct alloc_stat *r = b;
1407
1408         if (l->ptr < r->ptr)
1409                 return -1;
1410         else if (l->ptr > r->ptr)
1411                 return 1;
1412         return 0;
1413 }
1414
1415 static struct sort_dimension ptr_sort_dimension = {
1416         .name   = "ptr",
1417         .cmp    = ptr_cmp,
1418 };
1419
1420 static int slab_callsite_cmp(void *a, void *b)
1421 {
1422         struct alloc_stat *l = a;
1423         struct alloc_stat *r = b;
1424
1425         if (l->call_site < r->call_site)
1426                 return -1;
1427         else if (l->call_site > r->call_site)
1428                 return 1;
1429         return 0;
1430 }
1431
1432 static struct sort_dimension callsite_sort_dimension = {
1433         .name   = "callsite",
1434         .cmp    = slab_callsite_cmp,
1435 };
1436
1437 static int hit_cmp(void *a, void *b)
1438 {
1439         struct alloc_stat *l = a;
1440         struct alloc_stat *r = b;
1441
1442         if (l->hit < r->hit)
1443                 return -1;
1444         else if (l->hit > r->hit)
1445                 return 1;
1446         return 0;
1447 }
1448
1449 static struct sort_dimension hit_sort_dimension = {
1450         .name   = "hit",
1451         .cmp    = hit_cmp,
1452 };
1453
1454 static int bytes_cmp(void *a, void *b)
1455 {
1456         struct alloc_stat *l = a;
1457         struct alloc_stat *r = b;
1458
1459         if (l->bytes_alloc < r->bytes_alloc)
1460                 return -1;
1461         else if (l->bytes_alloc > r->bytes_alloc)
1462                 return 1;
1463         return 0;
1464 }
1465
1466 static struct sort_dimension bytes_sort_dimension = {
1467         .name   = "bytes",
1468         .cmp    = bytes_cmp,
1469 };
1470
1471 static int frag_cmp(void *a, void *b)
1472 {
1473         double x, y;
1474         struct alloc_stat *l = a;
1475         struct alloc_stat *r = b;
1476
1477         x = fragmentation(l->bytes_req, l->bytes_alloc);
1478         y = fragmentation(r->bytes_req, r->bytes_alloc);
1479
1480         if (x < y)
1481                 return -1;
1482         else if (x > y)
1483                 return 1;
1484         return 0;
1485 }
1486
1487 static struct sort_dimension frag_sort_dimension = {
1488         .name   = "frag",
1489         .cmp    = frag_cmp,
1490 };
1491
1492 static int pingpong_cmp(void *a, void *b)
1493 {
1494         struct alloc_stat *l = a;
1495         struct alloc_stat *r = b;
1496
1497         if (l->pingpong < r->pingpong)
1498                 return -1;
1499         else if (l->pingpong > r->pingpong)
1500                 return 1;
1501         return 0;
1502 }
1503
1504 static struct sort_dimension pingpong_sort_dimension = {
1505         .name   = "pingpong",
1506         .cmp    = pingpong_cmp,
1507 };
1508
1509 /* page sort keys */
1510 static int page_cmp(void *a, void *b)
1511 {
1512         struct page_stat *l = a;
1513         struct page_stat *r = b;
1514
1515         if (l->page < r->page)
1516                 return -1;
1517         else if (l->page > r->page)
1518                 return 1;
1519         return 0;
1520 }
1521
1522 static struct sort_dimension page_sort_dimension = {
1523         .name   = "page",
1524         .cmp    = page_cmp,
1525 };
1526
1527 static int page_callsite_cmp(void *a, void *b)
1528 {
1529         struct page_stat *l = a;
1530         struct page_stat *r = b;
1531
1532         if (l->callsite < r->callsite)
1533                 return -1;
1534         else if (l->callsite > r->callsite)
1535                 return 1;
1536         return 0;
1537 }
1538
1539 static struct sort_dimension page_callsite_sort_dimension = {
1540         .name   = "callsite",
1541         .cmp    = page_callsite_cmp,
1542 };
1543
1544 static int page_hit_cmp(void *a, void *b)
1545 {
1546         struct page_stat *l = a;
1547         struct page_stat *r = b;
1548
1549         if (l->nr_alloc < r->nr_alloc)
1550                 return -1;
1551         else if (l->nr_alloc > r->nr_alloc)
1552                 return 1;
1553         return 0;
1554 }
1555
1556 static struct sort_dimension page_hit_sort_dimension = {
1557         .name   = "hit",
1558         .cmp    = page_hit_cmp,
1559 };
1560
1561 static int page_bytes_cmp(void *a, void *b)
1562 {
1563         struct page_stat *l = a;
1564         struct page_stat *r = b;
1565
1566         if (l->alloc_bytes < r->alloc_bytes)
1567                 return -1;
1568         else if (l->alloc_bytes > r->alloc_bytes)
1569                 return 1;
1570         return 0;
1571 }
1572
1573 static struct sort_dimension page_bytes_sort_dimension = {
1574         .name   = "bytes",
1575         .cmp    = page_bytes_cmp,
1576 };
1577
1578 static int page_order_cmp(void *a, void *b)
1579 {
1580         struct page_stat *l = a;
1581         struct page_stat *r = b;
1582
1583         if (l->order < r->order)
1584                 return -1;
1585         else if (l->order > r->order)
1586                 return 1;
1587         return 0;
1588 }
1589
1590 static struct sort_dimension page_order_sort_dimension = {
1591         .name   = "order",
1592         .cmp    = page_order_cmp,
1593 };
1594
1595 static int migrate_type_cmp(void *a, void *b)
1596 {
1597         struct page_stat *l = a;
1598         struct page_stat *r = b;
1599
1600         /* for internal use to find free'd page */
1601         if (l->migrate_type == -1U)
1602                 return 0;
1603
1604         if (l->migrate_type < r->migrate_type)
1605                 return -1;
1606         else if (l->migrate_type > r->migrate_type)
1607                 return 1;
1608         return 0;
1609 }
1610
1611 static struct sort_dimension migrate_type_sort_dimension = {
1612         .name   = "migtype",
1613         .cmp    = migrate_type_cmp,
1614 };
1615
1616 static int gfp_flags_cmp(void *a, void *b)
1617 {
1618         struct page_stat *l = a;
1619         struct page_stat *r = b;
1620
1621         /* for internal use to find free'd page */
1622         if (l->gfp_flags == -1U)
1623                 return 0;
1624
1625         if (l->gfp_flags < r->gfp_flags)
1626                 return -1;
1627         else if (l->gfp_flags > r->gfp_flags)
1628                 return 1;
1629         return 0;
1630 }
1631
1632 static struct sort_dimension gfp_flags_sort_dimension = {
1633         .name   = "gfp",
1634         .cmp    = gfp_flags_cmp,
1635 };
1636
1637 static struct sort_dimension *slab_sorts[] = {
1638         &ptr_sort_dimension,
1639         &callsite_sort_dimension,
1640         &hit_sort_dimension,
1641         &bytes_sort_dimension,
1642         &frag_sort_dimension,
1643         &pingpong_sort_dimension,
1644 };
1645
1646 static struct sort_dimension *page_sorts[] = {
1647         &page_sort_dimension,
1648         &page_callsite_sort_dimension,
1649         &page_hit_sort_dimension,
1650         &page_bytes_sort_dimension,
1651         &page_order_sort_dimension,
1652         &migrate_type_sort_dimension,
1653         &gfp_flags_sort_dimension,
1654 };
1655
1656 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1657 {
1658         struct sort_dimension *sort;
1659         int i;
1660
1661         for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1662                 if (!strcmp(slab_sorts[i]->name, tok)) {
1663                         sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1664                         if (!sort) {
1665                                 pr_err("%s: memdup failed\n", __func__);
1666                                 return -1;
1667                         }
1668                         list_add_tail(&sort->list, list);
1669                         return 0;
1670                 }
1671         }
1672
1673         return -1;
1674 }
1675
1676 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1677 {
1678         struct sort_dimension *sort;
1679         int i;
1680
1681         for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1682                 if (!strcmp(page_sorts[i]->name, tok)) {
1683                         sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1684                         if (!sort) {
1685                                 pr_err("%s: memdup failed\n", __func__);
1686                                 return -1;
1687                         }
1688                         list_add_tail(&sort->list, list);
1689                         return 0;
1690                 }
1691         }
1692
1693         return -1;
1694 }
1695
1696 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1697 {
1698         char *tok;
1699         char *str = strdup(arg);
1700         char *pos = str;
1701
1702         if (!str) {
1703                 pr_err("%s: strdup failed\n", __func__);
1704                 return -1;
1705         }
1706
1707         while (true) {
1708                 tok = strsep(&pos, ",");
1709                 if (!tok)
1710                         break;
1711                 if (slab_sort_dimension__add(tok, sort_list) < 0) {
1712                         error("Unknown slab --sort key: '%s'", tok);
1713                         free(str);
1714                         return -1;
1715                 }
1716         }
1717
1718         free(str);
1719         return 0;
1720 }
1721
1722 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1723 {
1724         char *tok;
1725         char *str = strdup(arg);
1726         char *pos = str;
1727
1728         if (!str) {
1729                 pr_err("%s: strdup failed\n", __func__);
1730                 return -1;
1731         }
1732
1733         while (true) {
1734                 tok = strsep(&pos, ",");
1735                 if (!tok)
1736                         break;
1737                 if (page_sort_dimension__add(tok, sort_list) < 0) {
1738                         error("Unknown page --sort key: '%s'", tok);
1739                         free(str);
1740                         return -1;
1741                 }
1742         }
1743
1744         free(str);
1745         return 0;
1746 }
1747
1748 static int parse_sort_opt(const struct option *opt __maybe_unused,
1749                           const char *arg, int unset __maybe_unused)
1750 {
1751         if (!arg)
1752                 return -1;
1753
1754         if (kmem_page > kmem_slab ||
1755             (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1756                 if (caller_flag > alloc_flag)
1757                         return setup_page_sorting(&page_caller_sort, arg);
1758                 else
1759                         return setup_page_sorting(&page_alloc_sort, arg);
1760         } else {
1761                 if (caller_flag > alloc_flag)
1762                         return setup_slab_sorting(&slab_caller_sort, arg);
1763                 else
1764                         return setup_slab_sorting(&slab_alloc_sort, arg);
1765         }
1766
1767         return 0;
1768 }
1769
1770 static int parse_caller_opt(const struct option *opt __maybe_unused,
1771                             const char *arg __maybe_unused,
1772                             int unset __maybe_unused)
1773 {
1774         caller_flag = (alloc_flag + 1);
1775         return 0;
1776 }
1777
1778 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1779                            const char *arg __maybe_unused,
1780                            int unset __maybe_unused)
1781 {
1782         alloc_flag = (caller_flag + 1);
1783         return 0;
1784 }
1785
1786 static int parse_slab_opt(const struct option *opt __maybe_unused,
1787                           const char *arg __maybe_unused,
1788                           int unset __maybe_unused)
1789 {
1790         kmem_slab = (kmem_page + 1);
1791         return 0;
1792 }
1793
1794 static int parse_page_opt(const struct option *opt __maybe_unused,
1795                           const char *arg __maybe_unused,
1796                           int unset __maybe_unused)
1797 {
1798         kmem_page = (kmem_slab + 1);
1799         return 0;
1800 }
1801
1802 static int parse_line_opt(const struct option *opt __maybe_unused,
1803                           const char *arg, int unset __maybe_unused)
1804 {
1805         int lines;
1806
1807         if (!arg)
1808                 return -1;
1809
1810         lines = strtoul(arg, NULL, 10);
1811
1812         if (caller_flag > alloc_flag)
1813                 caller_lines = lines;
1814         else
1815                 alloc_lines = lines;
1816
1817         return 0;
1818 }
1819
1820 static int __cmd_record(int argc, const char **argv)
1821 {
1822         const char * const record_args[] = {
1823         "record", "-a", "-R", "-c", "1",
1824         };
1825         const char * const slab_events[] = {
1826         "-e", "kmem:kmalloc",
1827         "-e", "kmem:kmalloc_node",
1828         "-e", "kmem:kfree",
1829         "-e", "kmem:kmem_cache_alloc",
1830         "-e", "kmem:kmem_cache_alloc_node",
1831         "-e", "kmem:kmem_cache_free",
1832         };
1833         const char * const page_events[] = {
1834         "-e", "kmem:mm_page_alloc",
1835         "-e", "kmem:mm_page_free",
1836         };
1837         unsigned int rec_argc, i, j;
1838         const char **rec_argv;
1839
1840         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1841         if (kmem_slab)
1842                 rec_argc += ARRAY_SIZE(slab_events);
1843         if (kmem_page)
1844                 rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1845
1846         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1847
1848         if (rec_argv == NULL)
1849                 return -ENOMEM;
1850
1851         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1852                 rec_argv[i] = strdup(record_args[i]);
1853
1854         if (kmem_slab) {
1855                 for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1856                         rec_argv[i] = strdup(slab_events[j]);
1857         }
1858         if (kmem_page) {
1859                 rec_argv[i++] = strdup("-g");
1860
1861                 for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1862                         rec_argv[i] = strdup(page_events[j]);
1863         }
1864
1865         for (j = 1; j < (unsigned int)argc; j++, i++)
1866                 rec_argv[i] = argv[j];
1867
1868         return cmd_record(i, rec_argv, NULL);
1869 }
1870
1871 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1872 {
1873         if (!strcmp(var, "kmem.default")) {
1874                 if (!strcmp(value, "slab"))
1875                         kmem_default = KMEM_SLAB;
1876                 else if (!strcmp(value, "page"))
1877                         kmem_default = KMEM_PAGE;
1878                 else
1879                         pr_err("invalid default value ('slab' or 'page' required): %s\n",
1880                                value);
1881                 return 0;
1882         }
1883
1884         return 0;
1885 }
1886
1887 int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused)
1888 {
1889         const char * const default_slab_sort = "frag,hit,bytes";
1890         const char * const default_page_sort = "bytes,hit";
1891         struct perf_data_file file = {
1892                 .mode = PERF_DATA_MODE_READ,
1893         };
1894         const struct option kmem_options[] = {
1895         OPT_STRING('i', "input", &input_name, "file", "input file name"),
1896         OPT_INCR('v', "verbose", &verbose,
1897                     "be more verbose (show symbol address, etc)"),
1898         OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1899                            "show per-callsite statistics", parse_caller_opt),
1900         OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1901                            "show per-allocation statistics", parse_alloc_opt),
1902         OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1903                      "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1904                      "page, order, migtype, gfp", parse_sort_opt),
1905         OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1906         OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1907         OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"),
1908         OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1909                            parse_slab_opt),
1910         OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1911                            parse_page_opt),
1912         OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1913         OPT_STRING(0, "time", &time_str, "str",
1914                    "Time span of interest (start,stop)"),
1915         OPT_END()
1916         };
1917         const char *const kmem_subcommands[] = { "record", "stat", NULL };
1918         const char *kmem_usage[] = {
1919                 NULL,
1920                 NULL
1921         };
1922         struct perf_session *session;
1923         int ret = -1;
1924         const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1925
1926         perf_config(kmem_config, NULL);
1927         argc = parse_options_subcommand(argc, argv, kmem_options,
1928                                         kmem_subcommands, kmem_usage, 0);
1929
1930         if (!argc)
1931                 usage_with_options(kmem_usage, kmem_options);
1932
1933         if (kmem_slab == 0 && kmem_page == 0) {
1934                 if (kmem_default == KMEM_SLAB)
1935                         kmem_slab = 1;
1936                 else
1937                         kmem_page = 1;
1938         }
1939
1940         if (!strncmp(argv[0], "rec", 3)) {
1941                 symbol__init(NULL);
1942                 return __cmd_record(argc, argv);
1943         }
1944
1945         file.path = input_name;
1946
1947         kmem_session = session = perf_session__new(&file, false, &perf_kmem);
1948         if (session == NULL)
1949                 return -1;
1950
1951         if (kmem_slab) {
1952                 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
1953                                                           "kmem:kmalloc")) {
1954                         pr_err(errmsg, "slab", "slab");
1955                         goto out_delete;
1956                 }
1957         }
1958
1959         if (kmem_page) {
1960                 struct perf_evsel *evsel;
1961
1962                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
1963                                                              "kmem:mm_page_alloc");
1964                 if (evsel == NULL) {
1965                         pr_err(errmsg, "page", "page");
1966                         goto out_delete;
1967                 }
1968
1969                 kmem_page_size = pevent_get_page_size(evsel->tp_format->pevent);
1970                 symbol_conf.use_callchain = true;
1971         }
1972
1973         symbol__init(&session->header.env);
1974
1975         if (perf_time__parse_str(&ptime, time_str) != 0) {
1976                 pr_err("Invalid time string\n");
1977                 return -EINVAL;
1978         }
1979
1980         if (!strcmp(argv[0], "stat")) {
1981                 setlocale(LC_ALL, "");
1982
1983                 if (cpu__setup_cpunode_map())
1984                         goto out_delete;
1985
1986                 if (list_empty(&slab_caller_sort))
1987                         setup_slab_sorting(&slab_caller_sort, default_slab_sort);
1988                 if (list_empty(&slab_alloc_sort))
1989                         setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
1990                 if (list_empty(&page_caller_sort))
1991                         setup_page_sorting(&page_caller_sort, default_page_sort);
1992                 if (list_empty(&page_alloc_sort))
1993                         setup_page_sorting(&page_alloc_sort, default_page_sort);
1994
1995                 if (kmem_page) {
1996                         setup_page_sorting(&page_alloc_sort_input,
1997                                            "page,order,migtype,gfp");
1998                         setup_page_sorting(&page_caller_sort_input,
1999                                            "callsite,order,migtype,gfp");
2000                 }
2001                 ret = __cmd_kmem(session);
2002         } else
2003                 usage_with_options(kmem_usage, kmem_options);
2004
2005 out_delete:
2006         perf_session__delete(session);
2007
2008         return ret;
2009 }
2010