tools/perf/builtin-kmem.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "builtin.h"
   3 #include "perf.h"
   4
   5 #include "util/dso.h"
   6 #include "util/evlist.h"
   7 #include "util/evsel.h"
   8 #include "util/config.h"
   9 #include "util/map.h"
  10 #include "util/symbol.h"
  11 #include "util/thread.h"
  12 #include "util/header.h"
  13 #include "util/session.h"
  14 #include "util/tool.h"
  15 #include "util/callchain.h"
  16 #include "util/time-utils.h"
  17
  18 #include <subcmd/pager.h>
  19 #include <subcmd/parse-options.h>
  20 #include "util/trace-event.h"
  21 #include "util/data.h"
  22 #include "util/cpumap.h"
  23
  24 #include "util/debug.h"
  25 #include "util/string2.h"
  26
  27 #include <linux/kernel.h>
  28 #include <linux/rbtree.h>
  29 #include <linux/string.h>
  30 #include <linux/zalloc.h>
  31 #include <errno.h>
  32 #include <inttypes.h>
  33 #include <locale.h>
  34 #include <regex.h>
  35
  36 #include <linux/ctype.h>
  37
  38 static int      kmem_slab;
  39 static int      kmem_page;
  40
  41 static long     kmem_page_size;
  42 static enum {
  43         KMEM_SLAB,
  44         KMEM_PAGE,
  45 } kmem_default = KMEM_SLAB;  /* for backward compatibility */
  46
  47 struct alloc_stat;
  48 typedef int (*sort_fn_t)(void *, void *);
  49
  50 static int                      alloc_flag;
  51 static int                      caller_flag;
  52
  53 static int                      alloc_lines = -1;
  54 static int                      caller_lines = -1;
  55
  56 static bool                     raw_ip;
  57
  58 struct alloc_stat {
  59         u64     call_site;
  60         u64     ptr;
  61         u64     bytes_req;
  62         u64     bytes_alloc;
  63         u64     last_alloc;
  64         u32     hit;
  65         u32     pingpong;
  66
  67         short   alloc_cpu;
  68
  69         struct rb_node node;
  70 };
  71
  72 static struct rb_root root_alloc_stat;
  73 static struct rb_root root_alloc_sorted;
  74 static struct rb_root root_caller_stat;
  75 static struct rb_root root_caller_sorted;
  76
  77 static unsigned long total_requested, total_allocated, total_freed;
  78 static unsigned long nr_allocs, nr_cross_allocs;
  79
  80 /* filters for controlling start and stop of time of analysis */
  81 static struct perf_time_interval ptime;
  82 const char *time_str;
  83
  84 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
  85                              int bytes_req, int bytes_alloc, int cpu)
  86 {
  87         struct rb_node **node = &root_alloc_stat.rb_node;
  88         struct rb_node *parent = NULL;
  89         struct alloc_stat *data = NULL;
  90
  91         while (*node) {
  92                 parent = *node;
  93                 data = rb_entry(*node, struct alloc_stat, node);
  94
  95                 if (ptr > data->ptr)
  96                         node = &(*node)->rb_right;
  97                 else if (ptr < data->ptr)
  98                         node = &(*node)->rb_left;
  99                 else
 100                         break;
 101         }
 102
 103         if (data && data->ptr == ptr) {
 104                 data->hit++;
 105                 data->bytes_req += bytes_req;
 106                 data->bytes_alloc += bytes_alloc;
 107         } else {
 108                 data = malloc(sizeof(*data));
 109                 if (!data) {
 110                         pr_err("%s: malloc failed\n", __func__);
 111                         return -1;
 112                 }
 113                 data->ptr = ptr;
 114                 data->pingpong = 0;
 115                 data->hit = 1;
 116                 data->bytes_req = bytes_req;
 117                 data->bytes_alloc = bytes_alloc;
 118
 119                 rb_link_node(&data->node, parent, node);
 120                 rb_insert_color(&data->node, &root_alloc_stat);
 121         }
 122         data->call_site = call_site;
 123         data->alloc_cpu = cpu;
 124         data->last_alloc = bytes_alloc;
 125
 126         return 0;
 127 }
 128
 129 static int insert_caller_stat(unsigned long call_site,
 130                               int bytes_req, int bytes_alloc)
 131 {
 132         struct rb_node **node = &root_caller_stat.rb_node;
 133         struct rb_node *parent = NULL;
 134         struct alloc_stat *data = NULL;
 135
 136         while (*node) {
 137                 parent = *node;
 138                 data = rb_entry(*node, struct alloc_stat, node);
 139
 140                 if (call_site > data->call_site)
 141                         node = &(*node)->rb_right;
 142                 else if (call_site < data->call_site)
 143                         node = &(*node)->rb_left;
 144                 else
 145                         break;
 146         }
 147
 148         if (data && data->call_site == call_site) {
 149                 data->hit++;
 150                 data->bytes_req += bytes_req;
 151                 data->bytes_alloc += bytes_alloc;
 152         } else {
 153                 data = malloc(sizeof(*data));
 154                 if (!data) {
 155                         pr_err("%s: malloc failed\n", __func__);
 156                         return -1;
 157                 }
 158                 data->call_site = call_site;
 159                 data->pingpong = 0;
 160                 data->hit = 1;
 161                 data->bytes_req = bytes_req;
 162                 data->bytes_alloc = bytes_alloc;
 163
 164                 rb_link_node(&data->node, parent, node);
 165                 rb_insert_color(&data->node, &root_caller_stat);
 166         }
 167
 168         return 0;
 169 }
 170
 171 static int perf_evsel__process_alloc_event(struct evsel *evsel,
 172                                            struct perf_sample *sample)
 173 {
 174         unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
 175                       call_site = perf_evsel__intval(evsel, sample, "call_site");
 176         int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
 177             bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
 178
 179         if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
 180             insert_caller_stat(call_site, bytes_req, bytes_alloc))
 181                 return -1;
 182
 183         total_requested += bytes_req;
 184         total_allocated += bytes_alloc;
 185
 186         nr_allocs++;
 187         return 0;
 188 }
 189
 190 static int perf_evsel__process_alloc_node_event(struct evsel *evsel,
 191                                                 struct perf_sample *sample)
 192 {
 193         int ret = perf_evsel__process_alloc_event(evsel, sample);
 194
 195         if (!ret) {
 196                 int node1 = cpu__get_node(sample->cpu),
 197                     node2 = perf_evsel__intval(evsel, sample, "node");
 198
 199                 if (node1 != node2)
 200                         nr_cross_allocs++;
 201         }
 202
 203         return ret;
 204 }
 205
 206 static int ptr_cmp(void *, void *);
 207 static int slab_callsite_cmp(void *, void *);
 208
 209 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
 210                                             unsigned long call_site,
 211                                             struct rb_root *root,
 212                                             sort_fn_t sort_fn)
 213 {
 214         struct rb_node *node = root->rb_node;
 215         struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
 216
 217         while (node) {
 218                 struct alloc_stat *data;
 219                 int cmp;
 220
 221                 data = rb_entry(node, struct alloc_stat, node);
 222
 223                 cmp = sort_fn(&key, data);
 224                 if (cmp < 0)
 225                         node = node->rb_left;
 226                 else if (cmp > 0)
 227                         node = node->rb_right;
 228                 else
 229                         return data;
 230         }
 231         return NULL;
 232 }
 233
 234 static int perf_evsel__process_free_event(struct evsel *evsel,
 235                                           struct perf_sample *sample)
 236 {
 237         unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
 238         struct alloc_stat *s_alloc, *s_caller;
 239
 240         s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
 241         if (!s_alloc)
 242                 return 0;
 243
 244         total_freed += s_alloc->last_alloc;
 245
 246         if ((short)sample->cpu != s_alloc->alloc_cpu) {
 247                 s_alloc->pingpong++;
 248
 249                 s_caller = search_alloc_stat(0, s_alloc->call_site,
 250                                              &root_caller_stat,
 251                                              slab_callsite_cmp);
 252                 if (!s_caller)
 253                         return -1;
 254                 s_caller->pingpong++;
 255         }
 256         s_alloc->alloc_cpu = -1;
 257
 258         return 0;
 259 }
 260
 261 static u64 total_page_alloc_bytes;
 262 static u64 total_page_free_bytes;
 263 static u64 total_page_nomatch_bytes;
 264 static u64 total_page_fail_bytes;
 265 static unsigned long nr_page_allocs;
 266 static unsigned long nr_page_frees;
 267 static unsigned long nr_page_fails;
 268 static unsigned long nr_page_nomatch;
 269
 270 static bool use_pfn;
 271 static bool live_page;
 272 static struct perf_session *kmem_session;
 273
 274 #define MAX_MIGRATE_TYPES  6
 275 #define MAX_PAGE_ORDER     11
 276
 277 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
 278
 279 struct page_stat {
 280         struct rb_node  node;
 281         u64             page;
 282         u64             callsite;
 283         int             order;
 284         unsigned        gfp_flags;
 285         unsigned        migrate_type;
 286         u64             alloc_bytes;
 287         u64             free_bytes;
 288         int             nr_alloc;
 289         int             nr_free;
 290 };
 291
 292 static struct rb_root page_live_tree;
 293 static struct rb_root page_alloc_tree;
 294 static struct rb_root page_alloc_sorted;
 295 static struct rb_root page_caller_tree;
 296 static struct rb_root page_caller_sorted;
 297
 298 struct alloc_func {
 299         u64 start;
 300         u64 end;
 301         char *name;
 302 };
 303
 304 static int nr_alloc_funcs;
 305 static struct alloc_func *alloc_func_list;
 306
 307 static int funcmp(const void *a, const void *b)
 308 {
 309         const struct alloc_func *fa = a;
 310         const struct alloc_func *fb = b;
 311
 312         if (fa->start > fb->start)
 313                 return 1;
 314         else
 315                 return -1;
 316 }
 317
 318 static int callcmp(const void *a, const void *b)
 319 {
 320         const struct alloc_func *fa = a;
 321         const struct alloc_func *fb = b;
 322
 323         if (fb->start <= fa->start && fa->end < fb->end)
 324                 return 0;
 325
 326         if (fa->start > fb->start)
 327                 return 1;
 328         else
 329                 return -1;
 330 }
 331
 332 static int build_alloc_func_list(void)
 333 {
 334         int ret;
 335         struct map *kernel_map;
 336         struct symbol *sym;
 337         struct rb_node *node;
 338         struct alloc_func *func;
 339         struct machine *machine = &kmem_session->machines.host;
 340         regex_t alloc_func_regex;
 341         static const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
 342
 343         ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
 344         if (ret) {
 345                 char err[BUFSIZ];
 346
 347                 regerror(ret, &alloc_func_regex, err, sizeof(err));
 348                 pr_err("Invalid regex: %s\n%s", pattern, err);
 349                 return -EINVAL;
 350         }
 351
 352         kernel_map = machine__kernel_map(machine);
 353         if (map__load(kernel_map) < 0) {
 354                 pr_err("cannot load kernel map\n");
 355                 return -ENOENT;
 356         }
 357
 358         map__for_each_symbol(kernel_map, sym, node) {
 359                 if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
 360                         continue;
 361
 362                 func = realloc(alloc_func_list,
 363                                (nr_alloc_funcs + 1) * sizeof(*func));
 364                 if (func == NULL)
 365                         return -ENOMEM;
 366
 367                 pr_debug("alloc func: %s\n", sym->name);
 368                 func[nr_alloc_funcs].start = sym->start;
 369                 func[nr_alloc_funcs].end   = sym->end;
 370                 func[nr_alloc_funcs].name  = sym->name;
 371
 372                 alloc_func_list = func;
 373                 nr_alloc_funcs++;
 374         }
 375
 376         qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
 377
 378         regfree(&alloc_func_regex);
 379         return 0;
 380 }
 381
 382 /*
 383  * Find first non-memory allocation function from callchain.
 384  * The allocation functions are in the 'alloc_func_list'.
 385  */
 386 static u64 find_callsite(struct evsel *evsel, struct perf_sample *sample)
 387 {
 388         struct addr_location al;
 389         struct machine *machine = &kmem_session->machines.host;
 390         struct callchain_cursor_node *node;
 391
 392         if (alloc_func_list == NULL) {
 393                 if (build_alloc_func_list() < 0)
 394                         goto out;
 395         }
 396
 397         al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
 398         sample__resolve_callchain(sample, &callchain_cursor, NULL, evsel, &al, 16);
 399
 400         callchain_cursor_commit(&callchain_cursor);
 401         while (true) {
 402                 struct alloc_func key, *caller;
 403                 u64 addr;
 404
 405                 node = callchain_cursor_current(&callchain_cursor);
 406                 if (node == NULL)
 407                         break;
 408
 409                 key.start = key.end = node->ip;
 410                 caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
 411                                  sizeof(key), callcmp);
 412                 if (!caller) {
 413                         /* found */
 414                         if (node->map)
 415                                 addr = map__unmap_ip(node->map, node->ip);
 416                         else
 417                                 addr = node->ip;
 418
 419                         return addr;
 420                 } else
 421                         pr_debug3("skipping alloc function: %s\n", caller->name);
 422
 423                 callchain_cursor_advance(&callchain_cursor);
 424         }
 425
 426 out:
 427         pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
 428         return sample->ip;
 429 }
 430
 431 struct sort_dimension {
 432         const char              name[20];
 433         sort_fn_t               cmp;
 434         struct list_head        list;
 435 };
 436
 437 static LIST_HEAD(page_alloc_sort_input);
 438 static LIST_HEAD(page_caller_sort_input);
 439
 440 static struct page_stat *
 441 __page_stat__findnew_page(struct page_stat *pstat, bool create)
 442 {
 443         struct rb_node **node = &page_live_tree.rb_node;
 444         struct rb_node *parent = NULL;
 445         struct page_stat *data;
 446
 447         while (*node) {
 448                 s64 cmp;
 449
 450                 parent = *node;
 451                 data = rb_entry(*node, struct page_stat, node);
 452
 453                 cmp = data->page - pstat->page;
 454                 if (cmp < 0)
 455                         node = &parent->rb_left;
 456                 else if (cmp > 0)
 457                         node = &parent->rb_right;
 458                 else
 459                         return data;
 460         }
 461
 462         if (!create)
 463                 return NULL;
 464
 465         data = zalloc(sizeof(*data));
 466         if (data != NULL) {
 467                 data->page = pstat->page;
 468                 data->order = pstat->order;
 469                 data->gfp_flags = pstat->gfp_flags;
 470                 data->migrate_type = pstat->migrate_type;
 471
 472                 rb_link_node(&data->node, parent, node);
 473                 rb_insert_color(&data->node, &page_live_tree);
 474         }
 475
 476         return data;
 477 }
 478
 479 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
 480 {
 481         return __page_stat__findnew_page(pstat, false);
 482 }
 483
 484 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
 485 {
 486         return __page_stat__findnew_page(pstat, true);
 487 }
 488
 489 static struct page_stat *
 490 __page_stat__findnew_alloc(struct page_stat *pstat, bool create)
 491 {
 492         struct rb_node **node = &page_alloc_tree.rb_node;
 493         struct rb_node *parent = NULL;
 494         struct page_stat *data;
 495         struct sort_dimension *sort;
 496
 497         while (*node) {
 498                 int cmp = 0;
 499
 500                 parent = *node;
 501                 data = rb_entry(*node, struct page_stat, node);
 502
 503                 list_for_each_entry(sort, &page_alloc_sort_input, list) {
 504                         cmp = sort->cmp(pstat, data);
 505                         if (cmp)
 506                                 break;
 507                 }
 508
 509                 if (cmp < 0)
 510                         node = &parent->rb_left;
 511                 else if (cmp > 0)
 512                         node = &parent->rb_right;
 513                 else
 514                         return data;
 515         }
 516
 517         if (!create)
 518                 return NULL;
 519
 520         data = zalloc(sizeof(*data));
 521         if (data != NULL) {
 522                 data->page = pstat->page;
 523                 data->order = pstat->order;
 524                 data->gfp_flags = pstat->gfp_flags;
 525                 data->migrate_type = pstat->migrate_type;
 526
 527                 rb_link_node(&data->node, parent, node);
 528                 rb_insert_color(&data->node, &page_alloc_tree);
 529         }
 530
 531         return data;
 532 }
 533
 534 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
 535 {
 536         return __page_stat__findnew_alloc(pstat, false);
 537 }
 538
 539 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
 540 {
 541         return __page_stat__findnew_alloc(pstat, true);
 542 }
 543
 544 static struct page_stat *
 545 __page_stat__findnew_caller(struct page_stat *pstat, bool create)
 546 {
 547         struct rb_node **node = &page_caller_tree.rb_node;
 548         struct rb_node *parent = NULL;
 549         struct page_stat *data;
 550         struct sort_dimension *sort;
 551
 552         while (*node) {
 553                 int cmp = 0;
 554
 555                 parent = *node;
 556                 data = rb_entry(*node, struct page_stat, node);
 557
 558                 list_for_each_entry(sort, &page_caller_sort_input, list) {
 559                         cmp = sort->cmp(pstat, data);
 560                         if (cmp)
 561                                 break;
 562                 }
 563
 564                 if (cmp < 0)
 565                         node = &parent->rb_left;
 566                 else if (cmp > 0)
 567                         node = &parent->rb_right;
 568                 else
 569                         return data;
 570         }
 571
 572         if (!create)
 573                 return NULL;
 574
 575         data = zalloc(sizeof(*data));
 576         if (data != NULL) {
 577                 data->callsite = pstat->callsite;
 578                 data->order = pstat->order;
 579                 data->gfp_flags = pstat->gfp_flags;
 580                 data->migrate_type = pstat->migrate_type;
 581
 582                 rb_link_node(&data->node, parent, node);
 583                 rb_insert_color(&data->node, &page_caller_tree);
 584         }
 585
 586         return data;
 587 }
 588
 589 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
 590 {
 591         return __page_stat__findnew_caller(pstat, false);
 592 }
 593
 594 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
 595 {
 596         return __page_stat__findnew_caller(pstat, true);
 597 }
 598
 599 static bool valid_page(u64 pfn_or_page)
 600 {
 601         if (use_pfn && pfn_or_page == -1UL)
 602                 return false;
 603         if (!use_pfn && pfn_or_page == 0)
 604                 return false;
 605         return true;
 606 }
 607
 608 struct gfp_flag {
 609         unsigned int flags;
 610         char *compact_str;
 611         char *human_readable;
 612 };
 613
 614 static struct gfp_flag *gfps;
 615 static int nr_gfps;
 616
 617 static int gfpcmp(const void *a, const void *b)
 618 {
 619         const struct gfp_flag *fa = a;
 620         const struct gfp_flag *fb = b;
 621
 622         return fa->flags - fb->flags;
 623 }
 624
 625 /* see include/trace/events/mmflags.h */
 626 static const struct {
 627         const char *original;
 628         const char *compact;
 629 } gfp_compact_table[] = {
 630         { "GFP_TRANSHUGE",              "THP" },
 631         { "GFP_TRANSHUGE_LIGHT",        "THL" },
 632         { "GFP_HIGHUSER_MOVABLE",       "HUM" },
 633         { "GFP_HIGHUSER",               "HU" },
 634         { "GFP_USER",                   "U" },
 635         { "GFP_KERNEL_ACCOUNT",         "KAC" },
 636         { "GFP_KERNEL",                 "K" },
 637         { "GFP_NOFS",                   "NF" },
 638         { "GFP_ATOMIC",                 "A" },
 639         { "GFP_NOIO",                   "NI" },
 640         { "GFP_NOWAIT",                 "NW" },
 641         { "GFP_DMA",                    "D" },
 642         { "__GFP_HIGHMEM",              "HM" },
 643         { "GFP_DMA32",                  "D32" },
 644         { "__GFP_HIGH",                 "H" },
 645         { "__GFP_ATOMIC",               "_A" },
 646         { "__GFP_IO",                   "I" },
 647         { "__GFP_FS",                   "F" },
 648         { "__GFP_NOWARN",               "NWR" },
 649         { "__GFP_RETRY_MAYFAIL",        "R" },
 650         { "__GFP_NOFAIL",               "NF" },
 651         { "__GFP_NORETRY",              "NR" },
 652         { "__GFP_COMP",                 "C" },
 653         { "__GFP_ZERO",                 "Z" },
 654         { "__GFP_NOMEMALLOC",           "NMA" },
 655         { "__GFP_MEMALLOC",             "MA" },
 656         { "__GFP_HARDWALL",             "HW" },
 657         { "__GFP_THISNODE",             "TN" },
 658         { "__GFP_RECLAIMABLE",          "RC" },
 659         { "__GFP_MOVABLE",              "M" },
 660         { "__GFP_ACCOUNT",              "AC" },
 661         { "__GFP_WRITE",                "WR" },
 662         { "__GFP_RECLAIM",              "R" },
 663         { "__GFP_DIRECT_RECLAIM",       "DR" },
 664         { "__GFP_KSWAPD_RECLAIM",       "KR" },
 665 };
 666
 667 static size_t max_gfp_len;
 668
 669 static char *compact_gfp_flags(char *gfp_flags)
 670 {
 671         char *orig_flags = strdup(gfp_flags);
 672         char *new_flags = NULL;
 673         char *str, *pos = NULL;
 674         size_t len = 0;
 675
 676         if (orig_flags == NULL)
 677                 return NULL;
 678
 679         str = strtok_r(orig_flags, "|", &pos);
 680         while (str) {
 681                 size_t i;
 682                 char *new;
 683                 const char *cpt;
 684
 685                 for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
 686                         if (strcmp(gfp_compact_table[i].original, str))
 687                                 continue;
 688
 689                         cpt = gfp_compact_table[i].compact;
 690                         new = realloc(new_flags, len + strlen(cpt) + 2);
 691                         if (new == NULL) {
 692                                 free(new_flags);
 693                                 return NULL;
 694                         }
 695
 696                         new_flags = new;
 697
 698                         if (!len) {
 699                                 strcpy(new_flags, cpt);
 700                         } else {
 701                                 strcat(new_flags, "|");
 702                                 strcat(new_flags, cpt);
 703                                 len++;
 704                         }
 705
 706                         len += strlen(cpt);
 707                 }
 708
 709                 str = strtok_r(NULL, "|", &pos);
 710         }
 711
 712         if (max_gfp_len < len)
 713                 max_gfp_len = len;
 714
 715         free(orig_flags);
 716         return new_flags;
 717 }
 718
 719 static char *compact_gfp_string(unsigned long gfp_flags)
 720 {
 721         struct gfp_flag key = {
 722                 .flags = gfp_flags,
 723         };
 724         struct gfp_flag *gfp;
 725
 726         gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
 727         if (gfp)
 728                 return gfp->compact_str;
 729
 730         return NULL;
 731 }
 732
 733 static int parse_gfp_flags(struct evsel *evsel, struct perf_sample *sample,
 734                            unsigned int gfp_flags)
 735 {
 736         struct tep_record record = {
 737                 .cpu = sample->cpu,
 738                 .data = sample->raw_data,
 739                 .size = sample->raw_size,
 740         };
 741         struct trace_seq seq;
 742         char *str, *pos = NULL;
 743
 744         if (nr_gfps) {
 745                 struct gfp_flag key = {
 746                         .flags = gfp_flags,
 747                 };
 748
 749                 if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
 750                         return 0;
 751         }
 752
 753         trace_seq_init(&seq);
 754         tep_print_event(evsel->tp_format->tep,
 755                         &seq, &record, "%s", TEP_PRINT_INFO);
 756
 757         str = strtok_r(seq.buffer, " ", &pos);
 758         while (str) {
 759                 if (!strncmp(str, "gfp_flags=", 10)) {
 760                         struct gfp_flag *new;
 761
 762                         new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
 763                         if (new == NULL)
 764                                 return -ENOMEM;
 765
 766                         gfps = new;
 767                         new += nr_gfps++;
 768
 769                         new->flags = gfp_flags;
 770                         new->human_readable = strdup(str + 10);
 771                         new->compact_str = compact_gfp_flags(str + 10);
 772                         if (!new->human_readable || !new->compact_str)
 773                                 return -ENOMEM;
 774
 775                         qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
 776                 }
 777
 778                 str = strtok_r(NULL, " ", &pos);
 779         }
 780
 781         trace_seq_destroy(&seq);
 782         return 0;
 783 }
 784
 785 static int perf_evsel__process_page_alloc_event(struct evsel *evsel,
 786                                                 struct perf_sample *sample)
 787 {
 788         u64 page;
 789         unsigned int order = perf_evsel__intval(evsel, sample, "order");
 790         unsigned int gfp_flags = perf_evsel__intval(evsel, sample, "gfp_flags");
 791         unsigned int migrate_type = perf_evsel__intval(evsel, sample,
 792                                                        "migratetype");
 793         u64 bytes = kmem_page_size << order;
 794         u64 callsite;
 795         struct page_stat *pstat;
 796         struct page_stat this = {
 797                 .order = order,
 798                 .gfp_flags = gfp_flags,
 799                 .migrate_type = migrate_type,
 800         };
 801
 802         if (use_pfn)
 803                 page = perf_evsel__intval(evsel, sample, "pfn");
 804         else
 805                 page = perf_evsel__intval(evsel, sample, "page");
 806
 807         nr_page_allocs++;
 808         total_page_alloc_bytes += bytes;
 809
 810         if (!valid_page(page)) {
 811                 nr_page_fails++;
 812                 total_page_fail_bytes += bytes;
 813
 814                 return 0;
 815         }
 816
 817         if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
 818                 return -1;
 819
 820         callsite = find_callsite(evsel, sample);
 821
 822         /*
 823          * This is to find the current page (with correct gfp flags and
 824          * migrate type) at free event.
 825          */
 826         this.page = page;
 827         pstat = page_stat__findnew_page(&this);
 828         if (pstat == NULL)
 829                 return -ENOMEM;
 830
 831         pstat->nr_alloc++;
 832         pstat->alloc_bytes += bytes;
 833         pstat->callsite = callsite;
 834
 835         if (!live_page) {
 836                 pstat = page_stat__findnew_alloc(&this);
 837                 if (pstat == NULL)
 838                         return -ENOMEM;
 839
 840                 pstat->nr_alloc++;
 841                 pstat->alloc_bytes += bytes;
 842                 pstat->callsite = callsite;
 843         }
 844
 845         this.callsite = callsite;
 846         pstat = page_stat__findnew_caller(&this);
 847         if (pstat == NULL)
 848                 return -ENOMEM;
 849
 850         pstat->nr_alloc++;
 851         pstat->alloc_bytes += bytes;
 852
 853         order_stats[order][migrate_type]++;
 854
 855         return 0;
 856 }
 857
 858 static int perf_evsel__process_page_free_event(struct evsel *evsel,
 859                                                 struct perf_sample *sample)
 860 {
 861         u64 page;
 862         unsigned int order = perf_evsel__intval(evsel, sample, "order");
 863         u64 bytes = kmem_page_size << order;
 864         struct page_stat *pstat;
 865         struct page_stat this = {
 866                 .order = order,
 867         };
 868
 869         if (use_pfn)
 870                 page = perf_evsel__intval(evsel, sample, "pfn");
 871         else
 872                 page = perf_evsel__intval(evsel, sample, "page");
 873
 874         nr_page_frees++;
 875         total_page_free_bytes += bytes;
 876
 877         this.page = page;
 878         pstat = page_stat__find_page(&this);
 879         if (pstat == NULL) {
 880                 pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
 881                           page, order);
 882
 883                 nr_page_nomatch++;
 884                 total_page_nomatch_bytes += bytes;
 885
 886                 return 0;
 887         }
 888
 889         this.gfp_flags = pstat->gfp_flags;
 890         this.migrate_type = pstat->migrate_type;
 891         this.callsite = pstat->callsite;
 892
 893         rb_erase(&pstat->node, &page_live_tree);
 894         free(pstat);
 895
 896         if (live_page) {
 897                 order_stats[this.order][this.migrate_type]--;
 898         } else {
 899                 pstat = page_stat__find_alloc(&this);
 900                 if (pstat == NULL)
 901                         return -ENOMEM;
 902
 903                 pstat->nr_free++;
 904                 pstat->free_bytes += bytes;
 905         }
 906
 907         pstat = page_stat__find_caller(&this);
 908         if (pstat == NULL)
 909                 return -ENOENT;
 910
 911         pstat->nr_free++;
 912         pstat->free_bytes += bytes;
 913
 914         if (live_page) {
 915                 pstat->nr_alloc--;
 916                 pstat->alloc_bytes -= bytes;
 917
 918                 if (pstat->nr_alloc == 0) {
 919                         rb_erase(&pstat->node, &page_caller_tree);
 920                         free(pstat);
 921                 }
 922         }
 923
 924         return 0;
 925 }
 926
 927 static bool perf_kmem__skip_sample(struct perf_sample *sample)
 928 {
 929         /* skip sample based on time? */
 930         if (perf_time__skip_sample(&ptime, sample->time))
 931                 return true;
 932
 933         return false;
 934 }
 935
 936 typedef int (*tracepoint_handler)(struct evsel *evsel,
 937                                   struct perf_sample *sample);
 938
 939 static int process_sample_event(struct perf_tool *tool __maybe_unused,
 940                                 union perf_event *event,
 941                                 struct perf_sample *sample,
 942                                 struct evsel *evsel,
 943                                 struct machine *machine)
 944 {
 945         int err = 0;
 946         struct thread *thread = machine__findnew_thread(machine, sample->pid,
 947                                                         sample->tid);
 948
 949         if (thread == NULL) {
 950                 pr_debug("problem processing %d event, skipping it.\n",
 951                          event->header.type);
 952                 return -1;
 953         }
 954
 955         if (perf_kmem__skip_sample(sample))
 956                 return 0;
 957
 958         dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
 959
 960         if (evsel->handler != NULL) {
 961                 tracepoint_handler f = evsel->handler;
 962                 err = f(evsel, sample);
 963         }
 964
 965         thread__put(thread);
 966
 967         return err;
 968 }
 969
 970 static struct perf_tool perf_kmem = {
 971         .sample          = process_sample_event,
 972         .comm            = perf_event__process_comm,
 973         .mmap            = perf_event__process_mmap,
 974         .mmap2           = perf_event__process_mmap2,
 975         .namespaces      = perf_event__process_namespaces,
 976         .ordered_events  = true,
 977 };
 978
 979 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
 980 {
 981         if (n_alloc == 0)
 982                 return 0.0;
 983         else
 984                 return 100.0 - (100.0 * n_req / n_alloc);
 985 }
 986
 987 static void __print_slab_result(struct rb_root *root,
 988                                 struct perf_session *session,
 989                                 int n_lines, int is_caller)
 990 {
 991         struct rb_node *next;
 992         struct machine *machine = &session->machines.host;
 993
 994         printf("%.105s\n", graph_dotted_line);
 995         printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
 996         printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
 997         printf("%.105s\n", graph_dotted_line);
 998
 999         next = rb_first(root);
1000
1001         while (next && n_lines--) {
1002                 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
1003                                                    node);
1004                 struct symbol *sym = NULL;
1005                 struct map *map;
1006                 char buf[BUFSIZ];
1007                 u64 addr;
1008
1009                 if (is_caller) {
1010                         addr = data->call_site;
1011                         if (!raw_ip)
1012                                 sym = machine__find_kernel_symbol(machine, addr, &map);
1013                 } else
1014                         addr = data->ptr;
1015
1016                 if (sym != NULL)
1017                         snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1018                                  addr - map->unmap_ip(map, sym->start));
1019                 else
1020                         snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1021                 printf(" %-34s |", buf);
1022
1023                 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1024                        (unsigned long long)data->bytes_alloc,
1025                        (unsigned long)data->bytes_alloc / data->hit,
1026                        (unsigned long long)data->bytes_req,
1027                        (unsigned long)data->bytes_req / data->hit,
1028                        (unsigned long)data->hit,
1029                        (unsigned long)data->pingpong,
1030                        fragmentation(data->bytes_req, data->bytes_alloc));
1031
1032                 next = rb_next(next);
1033         }
1034
1035         if (n_lines == -1)
1036                 printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
1037
1038         printf("%.105s\n", graph_dotted_line);
1039 }
1040
1041 static const char * const migrate_type_str[] = {
1042         "UNMOVABL",
1043         "RECLAIM",
1044         "MOVABLE",
1045         "RESERVED",
1046         "CMA/ISLT",
1047         "UNKNOWN",
1048 };
1049
1050 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1051 {
1052         struct rb_node *next = rb_first(&page_alloc_sorted);
1053         struct machine *machine = &session->machines.host;
1054         const char *format;
1055         int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1056
1057         printf("\n%.105s\n", graph_dotted_line);
1058         printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1059                use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1060                gfp_len, "GFP flags");
1061         printf("%.105s\n", graph_dotted_line);
1062
1063         if (use_pfn)
1064                 format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1065         else
1066                 format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1067
1068         while (next && n_lines--) {
1069                 struct page_stat *data;
1070                 struct symbol *sym;
1071                 struct map *map;
1072                 char buf[32];
1073                 char *caller = buf;
1074
1075                 data = rb_entry(next, struct page_stat, node);
1076                 sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1077                 if (sym)
1078                         caller = sym->name;
1079                 else
1080                         scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1081
1082                 printf(format, (unsigned long long)data->page,
1083                        (unsigned long long)data->alloc_bytes / 1024,
1084                        data->nr_alloc, data->order,
1085                        migrate_type_str[data->migrate_type],
1086                        gfp_len, compact_gfp_string(data->gfp_flags), caller);
1087
1088                 next = rb_next(next);
1089         }
1090
1091         if (n_lines == -1) {
1092                 printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1093                        gfp_len, "...");
1094         }
1095
1096         printf("%.105s\n", graph_dotted_line);
1097 }
1098
1099 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1100 {
1101         struct rb_node *next = rb_first(&page_caller_sorted);
1102         struct machine *machine = &session->machines.host;
1103         int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1104
1105         printf("\n%.105s\n", graph_dotted_line);
1106         printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1107                live_page ? "Live" : "Total", gfp_len, "GFP flags");
1108         printf("%.105s\n", graph_dotted_line);
1109
1110         while (next && n_lines--) {
1111                 struct page_stat *data;
1112                 struct symbol *sym;
1113                 struct map *map;
1114                 char buf[32];
1115                 char *caller = buf;
1116
1117                 data = rb_entry(next, struct page_stat, node);
1118                 sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1119                 if (sym)
1120                         caller = sym->name;
1121                 else
1122                         scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1123
1124                 printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1125                        (unsigned long long)data->alloc_bytes / 1024,
1126                        data->nr_alloc, data->order,
1127                        migrate_type_str[data->migrate_type],
1128                        gfp_len, compact_gfp_string(data->gfp_flags), caller);
1129
1130                 next = rb_next(next);
1131         }
1132
1133         if (n_lines == -1) {
1134                 printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1135                        gfp_len, "...");
1136         }
1137
1138         printf("%.105s\n", graph_dotted_line);
1139 }
1140
1141 static void print_gfp_flags(void)
1142 {
1143         int i;
1144
1145         printf("#\n");
1146         printf("# GFP flags\n");
1147         printf("# ---------\n");
1148         for (i = 0; i < nr_gfps; i++) {
1149                 printf("# %08x: %*s: %s\n", gfps[i].flags,
1150                        (int) max_gfp_len, gfps[i].compact_str,
1151                        gfps[i].human_readable);
1152         }
1153 }
1154
1155 static void print_slab_summary(void)
1156 {
1157         printf("\nSUMMARY (SLAB allocator)");
1158         printf("\n========================\n");
1159         printf("Total bytes requested: %'lu\n", total_requested);
1160         printf("Total bytes allocated: %'lu\n", total_allocated);
1161         printf("Total bytes freed:     %'lu\n", total_freed);
1162         if (total_allocated > total_freed) {
1163                 printf("Net total bytes allocated: %'lu\n",
1164                 total_allocated - total_freed);
1165         }
1166         printf("Total bytes wasted on internal fragmentation: %'lu\n",
1167                total_allocated - total_requested);
1168         printf("Internal fragmentation: %f%%\n",
1169                fragmentation(total_requested, total_allocated));
1170         printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1171 }
1172
1173 static void print_page_summary(void)
1174 {
1175         int o, m;
1176         u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1177         u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1178
1179         printf("\nSUMMARY (page allocator)");
1180         printf("\n========================\n");
1181         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1182                nr_page_allocs, total_page_alloc_bytes / 1024);
1183         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1184                nr_page_frees, total_page_free_bytes / 1024);
1185         printf("\n");
1186
1187         printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1188                nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1189         printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1190                nr_page_allocs - nr_alloc_freed,
1191                (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1192         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1193                nr_page_nomatch, total_page_nomatch_bytes / 1024);
1194         printf("\n");
1195
1196         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1197                nr_page_fails, total_page_fail_bytes / 1024);
1198         printf("\n");
1199
1200         printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1201                "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1202         printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1203                graph_dotted_line, graph_dotted_line, graph_dotted_line,
1204                graph_dotted_line, graph_dotted_line);
1205
1206         for (o = 0; o < MAX_PAGE_ORDER; o++) {
1207                 printf("%5d", o);
1208                 for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1209                         if (order_stats[o][m])
1210                                 printf("  %'12d", order_stats[o][m]);
1211                         else
1212                                 printf("  %12c", '.');
1213                 }
1214                 printf("\n");
1215         }
1216 }
1217
1218 static void print_slab_result(struct perf_session *session)
1219 {
1220         if (caller_flag)
1221                 __print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1222         if (alloc_flag)
1223                 __print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1224         print_slab_summary();
1225 }
1226
1227 static void print_page_result(struct perf_session *session)
1228 {
1229         if (caller_flag || alloc_flag)
1230                 print_gfp_flags();
1231         if (caller_flag)
1232                 __print_page_caller_result(session, caller_lines);
1233         if (alloc_flag)
1234                 __print_page_alloc_result(session, alloc_lines);
1235         print_page_summary();
1236 }
1237
1238 static void print_result(struct perf_session *session)
1239 {
1240         if (kmem_slab)
1241                 print_slab_result(session);
1242         if (kmem_page)
1243                 print_page_result(session);
1244 }
1245
1246 static LIST_HEAD(slab_caller_sort);
1247 static LIST_HEAD(slab_alloc_sort);
1248 static LIST_HEAD(page_caller_sort);
1249 static LIST_HEAD(page_alloc_sort);
1250
1251 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1252                              struct list_head *sort_list)
1253 {
1254         struct rb_node **new = &(root->rb_node);
1255         struct rb_node *parent = NULL;
1256         struct sort_dimension *sort;
1257
1258         while (*new) {
1259                 struct alloc_stat *this;
1260                 int cmp = 0;
1261
1262                 this = rb_entry(*new, struct alloc_stat, node);
1263                 parent = *new;
1264
1265                 list_for_each_entry(sort, sort_list, list) {
1266                         cmp = sort->cmp(data, this);
1267                         if (cmp)
1268                                 break;
1269                 }
1270
1271                 if (cmp > 0)
1272                         new = &((*new)->rb_left);
1273                 else
1274                         new = &((*new)->rb_right);
1275         }
1276
1277         rb_link_node(&data->node, parent, new);
1278         rb_insert_color(&data->node, root);
1279 }
1280
1281 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1282                                struct list_head *sort_list)
1283 {
1284         struct rb_node *node;
1285         struct alloc_stat *data;
1286
1287         for (;;) {
1288                 node = rb_first(root);
1289                 if (!node)
1290                         break;
1291
1292                 rb_erase(node, root);
1293                 data = rb_entry(node, struct alloc_stat, node);
1294                 sort_slab_insert(root_sorted, data, sort_list);
1295         }
1296 }
1297
1298 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1299                              struct list_head *sort_list)
1300 {
1301         struct rb_node **new = &root->rb_node;
1302         struct rb_node *parent = NULL;
1303         struct sort_dimension *sort;
1304
1305         while (*new) {
1306                 struct page_stat *this;
1307                 int cmp = 0;
1308
1309                 this = rb_entry(*new, struct page_stat, node);
1310                 parent = *new;
1311
1312                 list_for_each_entry(sort, sort_list, list) {
1313                         cmp = sort->cmp(data, this);
1314                         if (cmp)
1315                                 break;
1316                 }
1317
1318                 if (cmp > 0)
1319                         new = &parent->rb_left;
1320                 else
1321                         new = &parent->rb_right;
1322         }
1323
1324         rb_link_node(&data->node, parent, new);
1325         rb_insert_color(&data->node, root);
1326 }
1327
1328 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1329                                struct list_head *sort_list)
1330 {
1331         struct rb_node *node;
1332         struct page_stat *data;
1333
1334         for (;;) {
1335                 node = rb_first(root);
1336                 if (!node)
1337                         break;
1338
1339                 rb_erase(node, root);
1340                 data = rb_entry(node, struct page_stat, node);
1341                 sort_page_insert(root_sorted, data, sort_list);
1342         }
1343 }
1344
1345 static void sort_result(void)
1346 {
1347         if (kmem_slab) {
1348                 __sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1349                                    &slab_alloc_sort);
1350                 __sort_slab_result(&root_caller_stat, &root_caller_sorted,
1351                                    &slab_caller_sort);
1352         }
1353         if (kmem_page) {
1354                 if (live_page)
1355                         __sort_page_result(&page_live_tree, &page_alloc_sorted,
1356                                            &page_alloc_sort);
1357                 else
1358                         __sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1359                                            &page_alloc_sort);
1360
1361                 __sort_page_result(&page_caller_tree, &page_caller_sorted,
1362                                    &page_caller_sort);
1363         }
1364 }
1365
1366 static int __cmd_kmem(struct perf_session *session)
1367 {
1368         int err = -EINVAL;
1369         struct evsel *evsel;
1370         const struct evsel_str_handler kmem_tracepoints[] = {
1371                 /* slab allocator */
1372                 { "kmem:kmalloc",               perf_evsel__process_alloc_event, },
1373                 { "kmem:kmem_cache_alloc",      perf_evsel__process_alloc_event, },
1374                 { "kmem:kmalloc_node",          perf_evsel__process_alloc_node_event, },
1375                 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
1376                 { "kmem:kfree",                 perf_evsel__process_free_event, },
1377                 { "kmem:kmem_cache_free",       perf_evsel__process_free_event, },
1378                 /* page allocator */
1379                 { "kmem:mm_page_alloc",         perf_evsel__process_page_alloc_event, },
1380                 { "kmem:mm_page_free",          perf_evsel__process_page_free_event, },
1381         };
1382
1383         if (!perf_session__has_traces(session, "kmem record"))
1384                 goto out;
1385
1386         if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1387                 pr_err("Initializing perf session tracepoint handlers failed\n");
1388                 goto out;
1389         }
1390
1391         evlist__for_each_entry(session->evlist, evsel) {
1392                 if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") &&
1393                     perf_evsel__field(evsel, "pfn")) {
1394                         use_pfn = true;
1395                         break;
1396                 }
1397         }
1398
1399         setup_pager();
1400         err = perf_session__process_events(session);
1401         if (err != 0) {
1402                 pr_err("error during process events: %d\n", err);
1403                 goto out;
1404         }
1405         sort_result();
1406         print_result(session);
1407 out:
1408         return err;
1409 }
1410
1411 /* slab sort keys */
1412 static int ptr_cmp(void *a, void *b)
1413 {
1414         struct alloc_stat *l = a;
1415         struct alloc_stat *r = b;
1416
1417         if (l->ptr < r->ptr)
1418                 return -1;
1419         else if (l->ptr > r->ptr)
1420                 return 1;
1421         return 0;
1422 }
1423
1424 static struct sort_dimension ptr_sort_dimension = {
1425         .name   = "ptr",
1426         .cmp    = ptr_cmp,
1427 };
1428
1429 static int slab_callsite_cmp(void *a, void *b)
1430 {
1431         struct alloc_stat *l = a;
1432         struct alloc_stat *r = b;
1433
1434         if (l->call_site < r->call_site)
1435                 return -1;
1436         else if (l->call_site > r->call_site)
1437                 return 1;
1438         return 0;
1439 }
1440
1441 static struct sort_dimension callsite_sort_dimension = {
1442         .name   = "callsite",
1443         .cmp    = slab_callsite_cmp,
1444 };
1445
1446 static int hit_cmp(void *a, void *b)
1447 {
1448         struct alloc_stat *l = a;
1449         struct alloc_stat *r = b;
1450
1451         if (l->hit < r->hit)
1452                 return -1;
1453         else if (l->hit > r->hit)
1454                 return 1;
1455         return 0;
1456 }
1457
1458 static struct sort_dimension hit_sort_dimension = {
1459         .name   = "hit",
1460         .cmp    = hit_cmp,
1461 };
1462
1463 static int bytes_cmp(void *a, void *b)
1464 {
1465         struct alloc_stat *l = a;
1466         struct alloc_stat *r = b;
1467
1468         if (l->bytes_alloc < r->bytes_alloc)
1469                 return -1;
1470         else if (l->bytes_alloc > r->bytes_alloc)
1471                 return 1;
1472         return 0;
1473 }
1474
1475 static struct sort_dimension bytes_sort_dimension = {
1476         .name   = "bytes",
1477         .cmp    = bytes_cmp,
1478 };
1479
1480 static int frag_cmp(void *a, void *b)
1481 {
1482         double x, y;
1483         struct alloc_stat *l = a;
1484         struct alloc_stat *r = b;
1485
1486         x = fragmentation(l->bytes_req, l->bytes_alloc);
1487         y = fragmentation(r->bytes_req, r->bytes_alloc);
1488
1489         if (x < y)
1490                 return -1;
1491         else if (x > y)
1492                 return 1;
1493         return 0;
1494 }
1495
1496 static struct sort_dimension frag_sort_dimension = {
1497         .name   = "frag",
1498         .cmp    = frag_cmp,
1499 };
1500
1501 static int pingpong_cmp(void *a, void *b)
1502 {
1503         struct alloc_stat *l = a;
1504         struct alloc_stat *r = b;
1505
1506         if (l->pingpong < r->pingpong)
1507                 return -1;
1508         else if (l->pingpong > r->pingpong)
1509                 return 1;
1510         return 0;
1511 }
1512
1513 static struct sort_dimension pingpong_sort_dimension = {
1514         .name   = "pingpong",
1515         .cmp    = pingpong_cmp,
1516 };
1517
1518 /* page sort keys */
1519 static int page_cmp(void *a, void *b)
1520 {
1521         struct page_stat *l = a;
1522         struct page_stat *r = b;
1523
1524         if (l->page < r->page)
1525                 return -1;
1526         else if (l->page > r->page)
1527                 return 1;
1528         return 0;
1529 }
1530
1531 static struct sort_dimension page_sort_dimension = {
1532         .name   = "page",
1533         .cmp    = page_cmp,
1534 };
1535
1536 static int page_callsite_cmp(void *a, void *b)
1537 {
1538         struct page_stat *l = a;
1539         struct page_stat *r = b;
1540
1541         if (l->callsite < r->callsite)
1542                 return -1;
1543         else if (l->callsite > r->callsite)
1544                 return 1;
1545         return 0;
1546 }
1547
1548 static struct sort_dimension page_callsite_sort_dimension = {
1549         .name   = "callsite",
1550         .cmp    = page_callsite_cmp,
1551 };
1552
1553 static int page_hit_cmp(void *a, void *b)
1554 {
1555         struct page_stat *l = a;
1556         struct page_stat *r = b;
1557
1558         if (l->nr_alloc < r->nr_alloc)
1559                 return -1;
1560         else if (l->nr_alloc > r->nr_alloc)
1561                 return 1;
1562         return 0;
1563 }
1564
1565 static struct sort_dimension page_hit_sort_dimension = {
1566         .name   = "hit",
1567         .cmp    = page_hit_cmp,
1568 };
1569
1570 static int page_bytes_cmp(void *a, void *b)
1571 {
1572         struct page_stat *l = a;
1573         struct page_stat *r = b;
1574
1575         if (l->alloc_bytes < r->alloc_bytes)
1576                 return -1;
1577         else if (l->alloc_bytes > r->alloc_bytes)
1578                 return 1;
1579         return 0;
1580 }
1581
1582 static struct sort_dimension page_bytes_sort_dimension = {
1583         .name   = "bytes",
1584         .cmp    = page_bytes_cmp,
1585 };
1586
1587 static int page_order_cmp(void *a, void *b)
1588 {
1589         struct page_stat *l = a;
1590         struct page_stat *r = b;
1591
1592         if (l->order < r->order)
1593                 return -1;
1594         else if (l->order > r->order)
1595                 return 1;
1596         return 0;
1597 }
1598
1599 static struct sort_dimension page_order_sort_dimension = {
1600         .name   = "order",
1601         .cmp    = page_order_cmp,
1602 };
1603
1604 static int migrate_type_cmp(void *a, void *b)
1605 {
1606         struct page_stat *l = a;
1607         struct page_stat *r = b;
1608
1609         /* for internal use to find free'd page */
1610         if (l->migrate_type == -1U)
1611                 return 0;
1612
1613         if (l->migrate_type < r->migrate_type)
1614                 return -1;
1615         else if (l->migrate_type > r->migrate_type)
1616                 return 1;
1617         return 0;
1618 }
1619
1620 static struct sort_dimension migrate_type_sort_dimension = {
1621         .name   = "migtype",
1622         .cmp    = migrate_type_cmp,
1623 };
1624
1625 static int gfp_flags_cmp(void *a, void *b)
1626 {
1627         struct page_stat *l = a;
1628         struct page_stat *r = b;
1629
1630         /* for internal use to find free'd page */
1631         if (l->gfp_flags == -1U)
1632                 return 0;
1633
1634         if (l->gfp_flags < r->gfp_flags)
1635                 return -1;
1636         else if (l->gfp_flags > r->gfp_flags)
1637                 return 1;
1638         return 0;
1639 }
1640
1641 static struct sort_dimension gfp_flags_sort_dimension = {
1642         .name   = "gfp",
1643         .cmp    = gfp_flags_cmp,
1644 };
1645
1646 static struct sort_dimension *slab_sorts[] = {
1647         &ptr_sort_dimension,
1648         &callsite_sort_dimension,
1649         &hit_sort_dimension,
1650         &bytes_sort_dimension,
1651         &frag_sort_dimension,
1652         &pingpong_sort_dimension,
1653 };
1654
1655 static struct sort_dimension *page_sorts[] = {
1656         &page_sort_dimension,
1657         &page_callsite_sort_dimension,
1658         &page_hit_sort_dimension,
1659         &page_bytes_sort_dimension,
1660         &page_order_sort_dimension,
1661         &migrate_type_sort_dimension,
1662         &gfp_flags_sort_dimension,
1663 };
1664
1665 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1666 {
1667         struct sort_dimension *sort;
1668         int i;
1669
1670         for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1671                 if (!strcmp(slab_sorts[i]->name, tok)) {
1672                         sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1673                         if (!sort) {
1674                                 pr_err("%s: memdup failed\n", __func__);
1675                                 return -1;
1676                         }
1677                         list_add_tail(&sort->list, list);
1678                         return 0;
1679                 }
1680         }
1681
1682         return -1;
1683 }
1684
1685 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1686 {
1687         struct sort_dimension *sort;
1688         int i;
1689
1690         for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1691                 if (!strcmp(page_sorts[i]->name, tok)) {
1692                         sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1693                         if (!sort) {
1694                                 pr_err("%s: memdup failed\n", __func__);
1695                                 return -1;
1696                         }
1697                         list_add_tail(&sort->list, list);
1698                         return 0;
1699                 }
1700         }
1701
1702         return -1;
1703 }
1704
1705 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1706 {
1707         char *tok;
1708         char *str = strdup(arg);
1709         char *pos = str;
1710
1711         if (!str) {
1712                 pr_err("%s: strdup failed\n", __func__);
1713                 return -1;
1714         }
1715
1716         while (true) {
1717                 tok = strsep(&pos, ",");
1718                 if (!tok)
1719                         break;
1720                 if (slab_sort_dimension__add(tok, sort_list) < 0) {
1721                         pr_err("Unknown slab --sort key: '%s'", tok);
1722                         free(str);
1723                         return -1;
1724                 }
1725         }
1726
1727         free(str);
1728         return 0;
1729 }
1730
1731 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1732 {
1733         char *tok;
1734         char *str = strdup(arg);
1735         char *pos = str;
1736
1737         if (!str) {
1738                 pr_err("%s: strdup failed\n", __func__);
1739                 return -1;
1740         }
1741
1742         while (true) {
1743                 tok = strsep(&pos, ",");
1744                 if (!tok)
1745                         break;
1746                 if (page_sort_dimension__add(tok, sort_list) < 0) {
1747                         pr_err("Unknown page --sort key: '%s'", tok);
1748                         free(str);
1749                         return -1;
1750                 }
1751         }
1752
1753         free(str);
1754         return 0;
1755 }
1756
1757 static int parse_sort_opt(const struct option *opt __maybe_unused,
1758                           const char *arg, int unset __maybe_unused)
1759 {
1760         if (!arg)
1761                 return -1;
1762
1763         if (kmem_page > kmem_slab ||
1764             (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1765                 if (caller_flag > alloc_flag)
1766                         return setup_page_sorting(&page_caller_sort, arg);
1767                 else
1768                         return setup_page_sorting(&page_alloc_sort, arg);
1769         } else {
1770                 if (caller_flag > alloc_flag)
1771                         return setup_slab_sorting(&slab_caller_sort, arg);
1772                 else
1773                         return setup_slab_sorting(&slab_alloc_sort, arg);
1774         }
1775
1776         return 0;
1777 }
1778
1779 static int parse_caller_opt(const struct option *opt __maybe_unused,
1780                             const char *arg __maybe_unused,
1781                             int unset __maybe_unused)
1782 {
1783         caller_flag = (alloc_flag + 1);
1784         return 0;
1785 }
1786
1787 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1788                            const char *arg __maybe_unused,
1789                            int unset __maybe_unused)
1790 {
1791         alloc_flag = (caller_flag + 1);
1792         return 0;
1793 }
1794
1795 static int parse_slab_opt(const struct option *opt __maybe_unused,
1796                           const char *arg __maybe_unused,
1797                           int unset __maybe_unused)
1798 {
1799         kmem_slab = (kmem_page + 1);
1800         return 0;
1801 }
1802
1803 static int parse_page_opt(const struct option *opt __maybe_unused,
1804                           const char *arg __maybe_unused,
1805                           int unset __maybe_unused)
1806 {
1807         kmem_page = (kmem_slab + 1);
1808         return 0;
1809 }
1810
1811 static int parse_line_opt(const struct option *opt __maybe_unused,
1812                           const char *arg, int unset __maybe_unused)
1813 {
1814         int lines;
1815
1816         if (!arg)
1817                 return -1;
1818
1819         lines = strtoul(arg, NULL, 10);
1820
1821         if (caller_flag > alloc_flag)
1822                 caller_lines = lines;
1823         else
1824                 alloc_lines = lines;
1825
1826         return 0;
1827 }
1828
1829 static int __cmd_record(int argc, const char **argv)
1830 {
1831         const char * const record_args[] = {
1832         "record", "-a", "-R", "-c", "1",
1833         };
1834         const char * const slab_events[] = {
1835         "-e", "kmem:kmalloc",
1836         "-e", "kmem:kmalloc_node",
1837         "-e", "kmem:kfree",
1838         "-e", "kmem:kmem_cache_alloc",
1839         "-e", "kmem:kmem_cache_alloc_node",
1840         "-e", "kmem:kmem_cache_free",
1841         };
1842         const char * const page_events[] = {
1843         "-e", "kmem:mm_page_alloc",
1844         "-e", "kmem:mm_page_free",
1845         };
1846         unsigned int rec_argc, i, j;
1847         const char **rec_argv;
1848
1849         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1850         if (kmem_slab)
1851                 rec_argc += ARRAY_SIZE(slab_events);
1852         if (kmem_page)
1853                 rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1854
1855         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1856
1857         if (rec_argv == NULL)
1858                 return -ENOMEM;
1859
1860         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1861                 rec_argv[i] = strdup(record_args[i]);
1862
1863         if (kmem_slab) {
1864                 for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1865                         rec_argv[i] = strdup(slab_events[j]);
1866         }
1867         if (kmem_page) {
1868                 rec_argv[i++] = strdup("-g");
1869
1870                 for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1871                         rec_argv[i] = strdup(page_events[j]);
1872         }
1873
1874         for (j = 1; j < (unsigned int)argc; j++, i++)
1875                 rec_argv[i] = argv[j];
1876
1877         return cmd_record(i, rec_argv);
1878 }
1879
1880 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1881 {
1882         if (!strcmp(var, "kmem.default")) {
1883                 if (!strcmp(value, "slab"))
1884                         kmem_default = KMEM_SLAB;
1885                 else if (!strcmp(value, "page"))
1886                         kmem_default = KMEM_PAGE;
1887                 else
1888                         pr_err("invalid default value ('slab' or 'page' required): %s\n",
1889                                value);
1890                 return 0;
1891         }
1892
1893         return 0;
1894 }
1895
1896 int cmd_kmem(int argc, const char **argv)
1897 {
1898         const char * const default_slab_sort = "frag,hit,bytes";
1899         const char * const default_page_sort = "bytes,hit";
1900         struct perf_data data = {
1901                 .mode = PERF_DATA_MODE_READ,
1902         };
1903         const struct option kmem_options[] = {
1904         OPT_STRING('i', "input", &input_name, "file", "input file name"),
1905         OPT_INCR('v', "verbose", &verbose,
1906                     "be more verbose (show symbol address, etc)"),
1907         OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1908                            "show per-callsite statistics", parse_caller_opt),
1909         OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1910                            "show per-allocation statistics", parse_alloc_opt),
1911         OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1912                      "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1913                      "page, order, migtype, gfp", parse_sort_opt),
1914         OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1915         OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1916         OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
1917         OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1918                            parse_slab_opt),
1919         OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1920                            parse_page_opt),
1921         OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1922         OPT_STRING(0, "time", &time_str, "str",
1923                    "Time span of interest (start,stop)"),
1924         OPT_END()
1925         };
1926         const char *const kmem_subcommands[] = { "record", "stat", NULL };
1927         const char *kmem_usage[] = {
1928                 NULL,
1929                 NULL
1930         };
1931         struct perf_session *session;
1932         static const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1933         int ret = perf_config(kmem_config, NULL);
1934
1935         if (ret)
1936                 return ret;
1937
1938         argc = parse_options_subcommand(argc, argv, kmem_options,
1939                                         kmem_subcommands, kmem_usage, 0);
1940
1941         if (!argc)
1942                 usage_with_options(kmem_usage, kmem_options);
1943
1944         if (kmem_slab == 0 && kmem_page == 0) {
1945                 if (kmem_default == KMEM_SLAB)
1946                         kmem_slab = 1;
1947                 else
1948                         kmem_page = 1;
1949         }
1950
1951         if (!strncmp(argv[0], "rec", 3)) {
1952                 symbol__init(NULL);
1953                 return __cmd_record(argc, argv);
1954         }
1955
1956         data.path = input_name;
1957
1958         kmem_session = session = perf_session__new(&data, false, &perf_kmem);
1959         if (session == NULL)
1960                 return -1;
1961
1962         ret = -1;
1963
1964         if (kmem_slab) {
1965                 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
1966                                                           "kmem:kmalloc")) {
1967                         pr_err(errmsg, "slab", "slab");
1968                         goto out_delete;
1969                 }
1970         }
1971
1972         if (kmem_page) {
1973                 struct evsel *evsel;
1974
1975                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
1976                                                              "kmem:mm_page_alloc");
1977                 if (evsel == NULL) {
1978                         pr_err(errmsg, "page", "page");
1979                         goto out_delete;
1980                 }
1981
1982                 kmem_page_size = tep_get_page_size(evsel->tp_format->tep);
1983                 symbol_conf.use_callchain = true;
1984         }
1985
1986         symbol__init(&session->header.env);
1987
1988         if (perf_time__parse_str(&ptime, time_str) != 0) {
1989                 pr_err("Invalid time string\n");
1990                 ret = -EINVAL;
1991                 goto out_delete;
1992         }
1993
1994         if (!strcmp(argv[0], "stat")) {
1995                 setlocale(LC_ALL, "");
1996
1997                 if (cpu__setup_cpunode_map())
1998                         goto out_delete;
1999
2000                 if (list_empty(&slab_caller_sort))
2001                         setup_slab_sorting(&slab_caller_sort, default_slab_sort);
2002                 if (list_empty(&slab_alloc_sort))
2003                         setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
2004                 if (list_empty(&page_caller_sort))
2005                         setup_page_sorting(&page_caller_sort, default_page_sort);
2006                 if (list_empty(&page_alloc_sort))
2007                         setup_page_sorting(&page_alloc_sort, default_page_sort);
2008
2009                 if (kmem_page) {
2010                         setup_page_sorting(&page_alloc_sort_input,
2011                                            "page,order,migtype,gfp");
2012                         setup_page_sorting(&page_caller_sort_input,
2013                                            "callsite,order,migtype,gfp");
2014                 }
2015                 ret = __cmd_kmem(session);
2016         } else
2017                 usage_with_options(kmem_usage, kmem_options);
2018
2019 out_delete:
2020         perf_session__delete(session);
2021
2022         return ret;
2023 }
2024