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