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 perf_evsel__process_alloc_event(struct evsel *evsel,
 173                                            struct perf_sample *sample)
 174 {
 175         unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
 176                       call_site = perf_evsel__intval(evsel, sample, "call_site");
 177         int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
 178             bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
 179
 180         if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
 181             insert_caller_stat(call_site, bytes_req, bytes_alloc))
 182                 return -1;
 183
 184         total_requested += bytes_req;
 185         total_allocated += bytes_alloc;
 186
 187         nr_allocs++;
 188         return 0;
 189 }
 190
 191 static int perf_evsel__process_alloc_node_event(struct evsel *evsel,
 192                                                 struct perf_sample *sample)
 193 {
 194         int ret = perf_evsel__process_alloc_event(evsel, sample);
 195
 196         if (!ret) {
 197                 int node1 = cpu__get_node(sample->cpu),
 198                     node2 = perf_evsel__intval(evsel, sample, "node");
 199
 200                 if (node1 != node2)
 201                         nr_cross_allocs++;
 202         }
 203
 204         return ret;
 205 }
 206
 207 static int ptr_cmp(void *, void *);
 208 static int slab_callsite_cmp(void *, void *);
 209
 210 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
 211                                             unsigned long call_site,
 212                                             struct rb_root *root,
 213                                             sort_fn_t sort_fn)
 214 {
 215         struct rb_node *node = root->rb_node;
 216         struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
 217
 218         while (node) {
 219                 struct alloc_stat *data;
 220                 int cmp;
 221
 222                 data = rb_entry(node, struct alloc_stat, node);
 223
 224                 cmp = sort_fn(&key, data);
 225                 if (cmp < 0)
 226                         node = node->rb_left;
 227                 else if (cmp > 0)
 228                         node = node->rb_right;
 229                 else
 230                         return data;
 231         }
 232         return NULL;
 233 }
 234
 235 static int perf_evsel__process_free_event(struct evsel *evsel,
 236                                           struct perf_sample *sample)
 237 {
 238         unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
 239         struct alloc_stat *s_alloc, *s_caller;
 240
 241         s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
 242         if (!s_alloc)
 243                 return 0;
 244
 245         total_freed += s_alloc->last_alloc;
 246
 247         if ((short)sample->cpu != s_alloc->alloc_cpu) {
 248                 s_alloc->pingpong++;
 249
 250                 s_caller = search_alloc_stat(0, s_alloc->call_site,
 251                                              &root_caller_stat,
 252                                              slab_callsite_cmp);
 253                 if (!s_caller)
 254                         return -1;
 255                 s_caller->pingpong++;
 256         }
 257         s_alloc->alloc_cpu = -1;
 258
 259         return 0;
 260 }
 261
 262 static u64 total_page_alloc_bytes;
 263 static u64 total_page_free_bytes;
 264 static u64 total_page_nomatch_bytes;
 265 static u64 total_page_fail_bytes;
 266 static unsigned long nr_page_allocs;
 267 static unsigned long nr_page_frees;
 268 static unsigned long nr_page_fails;
 269 static unsigned long nr_page_nomatch;
 270
 271 static bool use_pfn;
 272 static bool live_page;
 273 static struct perf_session *kmem_session;
 274
 275 #define MAX_MIGRATE_TYPES  6
 276 #define MAX_PAGE_ORDER     11
 277
 278 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
 279
 280 struct page_stat {
 281         struct rb_node  node;
 282         u64             page;
 283         u64             callsite;
 284         int             order;
 285         unsigned        gfp_flags;
 286         unsigned        migrate_type;
 287         u64             alloc_bytes;
 288         u64             free_bytes;
 289         int             nr_alloc;
 290         int             nr_free;
 291 };
 292
 293 static struct rb_root page_live_tree;
 294 static struct rb_root page_alloc_tree;
 295 static struct rb_root page_alloc_sorted;
 296 static struct rb_root page_caller_tree;
 297 static struct rb_root page_caller_sorted;
 298
 299 struct alloc_func {
 300         u64 start;
 301         u64 end;
 302         char *name;
 303 };
 304
 305 static int nr_alloc_funcs;
 306 static struct alloc_func *alloc_func_list;
 307
 308 static int funcmp(const void *a, const void *b)
 309 {
 310         const struct alloc_func *fa = a;
 311         const struct alloc_func *fb = b;
 312
 313         if (fa->start > fb->start)
 314                 return 1;
 315         else
 316                 return -1;
 317 }
 318
 319 static int callcmp(const void *a, const void *b)
 320 {
 321         const struct alloc_func *fa = a;
 322         const struct alloc_func *fb = b;
 323
 324         if (fb->start <= fa->start && fa->end < fb->end)
 325                 return 0;
 326
 327         if (fa->start > fb->start)
 328                 return 1;
 329         else
 330                 return -1;
 331 }
 332
 333 static int build_alloc_func_list(void)
 334 {
 335         int ret;
 336         struct map *kernel_map;
 337         struct symbol *sym;
 338         struct rb_node *node;
 339         struct alloc_func *func;
 340         struct machine *machine = &kmem_session->machines.host;
 341         regex_t alloc_func_regex;
 342         static const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
 343
 344         ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
 345         if (ret) {
 346                 char err[BUFSIZ];
 347
 348                 regerror(ret, &alloc_func_regex, err, sizeof(err));
 349                 pr_err("Invalid regex: %s\n%s", pattern, err);
 350                 return -EINVAL;
 351         }
 352
 353         kernel_map = machine__kernel_map(machine);
 354         if (map__load(kernel_map) < 0) {
 355                 pr_err("cannot load kernel map\n");
 356                 return -ENOENT;
 357         }
 358
 359         map__for_each_symbol(kernel_map, sym, node) {
 360                 if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
 361                         continue;
 362
 363                 func = realloc(alloc_func_list,
 364                                (nr_alloc_funcs + 1) * sizeof(*func));
 365                 if (func == NULL)
 366                         return -ENOMEM;
 367
 368                 pr_debug("alloc func: %s\n", sym->name);
 369                 func[nr_alloc_funcs].start = sym->start;
 370                 func[nr_alloc_funcs].end   = sym->end;
 371                 func[nr_alloc_funcs].name  = sym->name;
 372
 373                 alloc_func_list = func;
 374                 nr_alloc_funcs++;
 375         }
 376
 377         qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
 378
 379         regfree(&alloc_func_regex);
 380         return 0;
 381 }
 382
 383 /*
 384  * Find first non-memory allocation function from callchain.
 385  * The allocation functions are in the 'alloc_func_list'.
 386  */
 387 static u64 find_callsite(struct evsel *evsel, struct perf_sample *sample)
 388 {
 389         struct addr_location al;
 390         struct machine *machine = &kmem_session->machines.host;
 391         struct callchain_cursor_node *node;
 392
 393         if (alloc_func_list == NULL) {
 394                 if (build_alloc_func_list() < 0)
 395                         goto out;
 396         }
 397
 398         al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
 399         sample__resolve_callchain(sample, &callchain_cursor, NULL, evsel, &al, 16);
 400
 401         callchain_cursor_commit(&callchain_cursor);
 402         while (true) {
 403                 struct alloc_func key, *caller;
 404                 u64 addr;
 405
 406                 node = callchain_cursor_current(&callchain_cursor);
 407                 if (node == NULL)
 408                         break;
 409
 410                 key.start = key.end = node->ip;
 411                 caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
 412                                  sizeof(key), callcmp);
 413                 if (!caller) {
 414                         /* found */
 415                         if (node->ms.map)
 416                                 addr = map__unmap_ip(node->ms.map, node->ip);
 417                         else
 418                                 addr = node->ip;
 419
 420                         return addr;
 421                 } else
 422                         pr_debug3("skipping alloc function: %s\n", caller->name);
 423
 424                 callchain_cursor_advance(&callchain_cursor);
 425         }
 426
 427 out:
 428         pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
 429         return sample->ip;
 430 }
 431
 432 struct sort_dimension {
 433         const char              name[20];
 434         sort_fn_t               cmp;
 435         struct list_head        list;
 436 };
 437
 438 static LIST_HEAD(page_alloc_sort_input);
 439 static LIST_HEAD(page_caller_sort_input);
 440
 441 static struct page_stat *
 442 __page_stat__findnew_page(struct page_stat *pstat, bool create)
 443 {
 444         struct rb_node **node = &page_live_tree.rb_node;
 445         struct rb_node *parent = NULL;
 446         struct page_stat *data;
 447
 448         while (*node) {
 449                 s64 cmp;
 450
 451                 parent = *node;
 452                 data = rb_entry(*node, struct page_stat, node);
 453
 454                 cmp = data->page - pstat->page;
 455                 if (cmp < 0)
 456                         node = &parent->rb_left;
 457                 else if (cmp > 0)
 458                         node = &parent->rb_right;
 459                 else
 460                         return data;
 461         }
 462
 463         if (!create)
 464                 return NULL;
 465
 466         data = zalloc(sizeof(*data));
 467         if (data != NULL) {
 468                 data->page = pstat->page;
 469                 data->order = pstat->order;
 470                 data->gfp_flags = pstat->gfp_flags;
 471                 data->migrate_type = pstat->migrate_type;
 472
 473                 rb_link_node(&data->node, parent, node);
 474                 rb_insert_color(&data->node, &page_live_tree);
 475         }
 476
 477         return data;
 478 }
 479
 480 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
 481 {
 482         return __page_stat__findnew_page(pstat, false);
 483 }
 484
 485 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
 486 {
 487         return __page_stat__findnew_page(pstat, true);
 488 }
 489
 490 static struct page_stat *
 491 __page_stat__findnew_alloc(struct page_stat *pstat, bool create)
 492 {
 493         struct rb_node **node = &page_alloc_tree.rb_node;
 494         struct rb_node *parent = NULL;
 495         struct page_stat *data;
 496         struct sort_dimension *sort;
 497
 498         while (*node) {
 499                 int cmp = 0;
 500
 501                 parent = *node;
 502                 data = rb_entry(*node, struct page_stat, node);
 503
 504                 list_for_each_entry(sort, &page_alloc_sort_input, list) {
 505                         cmp = sort->cmp(pstat, data);
 506                         if (cmp)
 507                                 break;
 508                 }
 509
 510                 if (cmp < 0)
 511                         node = &parent->rb_left;
 512                 else if (cmp > 0)
 513                         node = &parent->rb_right;
 514                 else
 515                         return data;
 516         }
 517
 518         if (!create)
 519                 return NULL;
 520
 521         data = zalloc(sizeof(*data));
 522         if (data != NULL) {
 523                 data->page = pstat->page;
 524                 data->order = pstat->order;
 525                 data->gfp_flags = pstat->gfp_flags;
 526                 data->migrate_type = pstat->migrate_type;
 527
 528                 rb_link_node(&data->node, parent, node);
 529                 rb_insert_color(&data->node, &page_alloc_tree);
 530         }
 531
 532         return data;
 533 }
 534
 535 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
 536 {
 537         return __page_stat__findnew_alloc(pstat, false);
 538 }
 539
 540 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
 541 {
 542         return __page_stat__findnew_alloc(pstat, true);
 543 }
 544
 545 static struct page_stat *
 546 __page_stat__findnew_caller(struct page_stat *pstat, bool create)
 547 {
 548         struct rb_node **node = &page_caller_tree.rb_node;
 549         struct rb_node *parent = NULL;
 550         struct page_stat *data;
 551         struct sort_dimension *sort;
 552
 553         while (*node) {
 554                 int cmp = 0;
 555
 556                 parent = *node;
 557                 data = rb_entry(*node, struct page_stat, node);
 558
 559                 list_for_each_entry(sort, &page_caller_sort_input, list) {
 560                         cmp = sort->cmp(pstat, data);
 561                         if (cmp)
 562                                 break;
 563                 }
 564
 565                 if (cmp < 0)
 566                         node = &parent->rb_left;
 567                 else if (cmp > 0)
 568                         node = &parent->rb_right;
 569                 else
 570                         return data;
 571         }
 572
 573         if (!create)
 574                 return NULL;
 575
 576         data = zalloc(sizeof(*data));
 577         if (data != NULL) {
 578                 data->callsite = pstat->callsite;
 579                 data->order = pstat->order;
 580                 data->gfp_flags = pstat->gfp_flags;
 581                 data->migrate_type = pstat->migrate_type;
 582
 583                 rb_link_node(&data->node, parent, node);
 584                 rb_insert_color(&data->node, &page_caller_tree);
 585         }
 586
 587         return data;
 588 }
 589
 590 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
 591 {
 592         return __page_stat__findnew_caller(pstat, false);
 593 }
 594
 595 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
 596 {
 597         return __page_stat__findnew_caller(pstat, true);
 598 }
 599
 600 static bool valid_page(u64 pfn_or_page)
 601 {
 602         if (use_pfn && pfn_or_page == -1UL)
 603                 return false;
 604         if (!use_pfn && pfn_or_page == 0)
 605                 return false;
 606         return true;
 607 }
 608
 609 struct gfp_flag {
 610         unsigned int flags;
 611         char *compact_str;
 612         char *human_readable;
 613 };
 614
 615 static struct gfp_flag *gfps;
 616 static int nr_gfps;
 617
 618 static int gfpcmp(const void *a, const void *b)
 619 {
 620         const struct gfp_flag *fa = a;
 621         const struct gfp_flag *fb = b;
 622
 623         return fa->flags - fb->flags;
 624 }
 625
 626 /* see include/trace/events/mmflags.h */
 627 static const struct {
 628         const char *original;
 629         const char *compact;
 630 } gfp_compact_table[] = {
 631         { "GFP_TRANSHUGE",              "THP" },
 632         { "GFP_TRANSHUGE_LIGHT",        "THL" },
 633         { "GFP_HIGHUSER_MOVABLE",       "HUM" },
 634         { "GFP_HIGHUSER",               "HU" },
 635         { "GFP_USER",                   "U" },
 636         { "GFP_KERNEL_ACCOUNT",         "KAC" },
 637         { "GFP_KERNEL",                 "K" },
 638         { "GFP_NOFS",                   "NF" },
 639         { "GFP_ATOMIC",                 "A" },
 640         { "GFP_NOIO",                   "NI" },
 641         { "GFP_NOWAIT",                 "NW" },
 642         { "GFP_DMA",                    "D" },
 643         { "__GFP_HIGHMEM",              "HM" },
 644         { "GFP_DMA32",                  "D32" },
 645         { "__GFP_HIGH",                 "H" },
 646         { "__GFP_ATOMIC",               "_A" },
 647         { "__GFP_IO",                   "I" },
 648         { "__GFP_FS",                   "F" },
 649         { "__GFP_NOWARN",               "NWR" },
 650         { "__GFP_RETRY_MAYFAIL",        "R" },
 651         { "__GFP_NOFAIL",               "NF" },
 652         { "__GFP_NORETRY",              "NR" },
 653         { "__GFP_COMP",                 "C" },
 654         { "__GFP_ZERO",                 "Z" },
 655         { "__GFP_NOMEMALLOC",           "NMA" },
 656         { "__GFP_MEMALLOC",             "MA" },
 657         { "__GFP_HARDWALL",             "HW" },
 658         { "__GFP_THISNODE",             "TN" },
 659         { "__GFP_RECLAIMABLE",          "RC" },
 660         { "__GFP_MOVABLE",              "M" },
 661         { "__GFP_ACCOUNT",              "AC" },
 662         { "__GFP_WRITE",                "WR" },
 663         { "__GFP_RECLAIM",              "R" },
 664         { "__GFP_DIRECT_RECLAIM",       "DR" },
 665         { "__GFP_KSWAPD_RECLAIM",       "KR" },
 666 };
 667
 668 static size_t max_gfp_len;
 669
 670 static char *compact_gfp_flags(char *gfp_flags)
 671 {
 672         char *orig_flags = strdup(gfp_flags);
 673         char *new_flags = NULL;
 674         char *str, *pos = NULL;
 675         size_t len = 0;
 676
 677         if (orig_flags == NULL)
 678                 return NULL;
 679
 680         str = strtok_r(orig_flags, "|", &pos);
 681         while (str) {
 682                 size_t i;
 683                 char *new;
 684                 const char *cpt;
 685
 686                 for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
 687                         if (strcmp(gfp_compact_table[i].original, str))
 688                                 continue;
 689
 690                         cpt = gfp_compact_table[i].compact;
 691                         new = realloc(new_flags, len + strlen(cpt) + 2);
 692                         if (new == NULL) {
 693                                 free(new_flags);
 694                                 free(orig_flags);
 695                                 return NULL;
 696                         }
 697
 698                         new_flags = new;
 699
 700                         if (!len) {
 701                                 strcpy(new_flags, cpt);
 702                         } else {
 703                                 strcat(new_flags, "|");
 704                                 strcat(new_flags, cpt);
 705                                 len++;
 706                         }
 707
 708                         len += strlen(cpt);
 709                 }
 710
 711                 str = strtok_r(NULL, "|", &pos);
 712         }
 713
 714         if (max_gfp_len < len)
 715                 max_gfp_len = len;
 716
 717         free(orig_flags);
 718         return new_flags;
 719 }
 720
 721 static char *compact_gfp_string(unsigned long gfp_flags)
 722 {
 723         struct gfp_flag key = {
 724                 .flags = gfp_flags,
 725         };
 726         struct gfp_flag *gfp;
 727
 728         gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
 729         if (gfp)
 730                 return gfp->compact_str;
 731
 732         return NULL;
 733 }
 734
 735 static int parse_gfp_flags(struct evsel *evsel, struct perf_sample *sample,
 736                            unsigned int gfp_flags)
 737 {
 738         struct tep_record record = {
 739                 .cpu = sample->cpu,
 740                 .data = sample->raw_data,
 741                 .size = sample->raw_size,
 742         };
 743         struct trace_seq seq;
 744         char *str, *pos = NULL;
 745
 746         if (nr_gfps) {
 747                 struct gfp_flag key = {
 748                         .flags = gfp_flags,
 749                 };
 750
 751                 if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
 752                         return 0;
 753         }
 754
 755         trace_seq_init(&seq);
 756         tep_print_event(evsel->tp_format->tep,
 757                         &seq, &record, "%s", TEP_PRINT_INFO);
 758
 759         str = strtok_r(seq.buffer, " ", &pos);
 760         while (str) {
 761                 if (!strncmp(str, "gfp_flags=", 10)) {
 762                         struct gfp_flag *new;
 763
 764                         new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
 765                         if (new == NULL)
 766                                 return -ENOMEM;
 767
 768                         gfps = new;
 769                         new += nr_gfps++;
 770
 771                         new->flags = gfp_flags;
 772                         new->human_readable = strdup(str + 10);
 773                         new->compact_str = compact_gfp_flags(str + 10);
 774                         if (!new->human_readable || !new->compact_str)
 775                                 return -ENOMEM;
 776
 777                         qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
 778                 }
 779
 780                 str = strtok_r(NULL, " ", &pos);
 781         }
 782
 783         trace_seq_destroy(&seq);
 784         return 0;
 785 }
 786
 787 static int perf_evsel__process_page_alloc_event(struct evsel *evsel,
 788                                                 struct perf_sample *sample)
 789 {
 790         u64 page;
 791         unsigned int order = perf_evsel__intval(evsel, sample, "order");
 792         unsigned int gfp_flags = perf_evsel__intval(evsel, sample, "gfp_flags");
 793         unsigned int migrate_type = perf_evsel__intval(evsel, sample,
 794                                                        "migratetype");
 795         u64 bytes = kmem_page_size << order;
 796         u64 callsite;
 797         struct page_stat *pstat;
 798         struct page_stat this = {
 799                 .order = order,
 800                 .gfp_flags = gfp_flags,
 801                 .migrate_type = migrate_type,
 802         };
 803
 804         if (use_pfn)
 805                 page = perf_evsel__intval(evsel, sample, "pfn");
 806         else
 807                 page = perf_evsel__intval(evsel, sample, "page");
 808
 809         nr_page_allocs++;
 810         total_page_alloc_bytes += bytes;
 811
 812         if (!valid_page(page)) {
 813                 nr_page_fails++;
 814                 total_page_fail_bytes += bytes;
 815
 816                 return 0;
 817         }
 818
 819         if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
 820                 return -1;
 821
 822         callsite = find_callsite(evsel, sample);
 823
 824         /*
 825          * This is to find the current page (with correct gfp flags and
 826          * migrate type) at free event.
 827          */
 828         this.page = page;
 829         pstat = page_stat__findnew_page(&this);
 830         if (pstat == NULL)
 831                 return -ENOMEM;
 832
 833         pstat->nr_alloc++;
 834         pstat->alloc_bytes += bytes;
 835         pstat->callsite = callsite;
 836
 837         if (!live_page) {
 838                 pstat = page_stat__findnew_alloc(&this);
 839                 if (pstat == NULL)
 840                         return -ENOMEM;
 841
 842                 pstat->nr_alloc++;
 843                 pstat->alloc_bytes += bytes;
 844                 pstat->callsite = callsite;
 845         }
 846
 847         this.callsite = callsite;
 848         pstat = page_stat__findnew_caller(&this);
 849         if (pstat == NULL)
 850                 return -ENOMEM;
 851
 852         pstat->nr_alloc++;
 853         pstat->alloc_bytes += bytes;
 854
 855         order_stats[order][migrate_type]++;
 856
 857         return 0;
 858 }
 859
 860 static int perf_evsel__process_page_free_event(struct evsel *evsel,
 861                                                 struct perf_sample *sample)
 862 {
 863         u64 page;
 864         unsigned int order = perf_evsel__intval(evsel, sample, "order");
 865         u64 bytes = kmem_page_size << order;
 866         struct page_stat *pstat;
 867         struct page_stat this = {
 868                 .order = order,
 869         };
 870
 871         if (use_pfn)
 872                 page = perf_evsel__intval(evsel, sample, "pfn");
 873         else
 874                 page = perf_evsel__intval(evsel, sample, "page");
 875
 876         nr_page_frees++;
 877         total_page_free_bytes += bytes;
 878
 879         this.page = page;
 880         pstat = page_stat__find_page(&this);
 881         if (pstat == NULL) {
 882                 pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
 883                           page, order);
 884
 885                 nr_page_nomatch++;
 886                 total_page_nomatch_bytes += bytes;
 887
 888                 return 0;
 889         }
 890
 891         this.gfp_flags = pstat->gfp_flags;
 892         this.migrate_type = pstat->migrate_type;
 893         this.callsite = pstat->callsite;
 894
 895         rb_erase(&pstat->node, &page_live_tree);
 896         free(pstat);
 897
 898         if (live_page) {
 899                 order_stats[this.order][this.migrate_type]--;
 900         } else {
 901                 pstat = page_stat__find_alloc(&this);
 902                 if (pstat == NULL)
 903                         return -ENOMEM;
 904
 905                 pstat->nr_free++;
 906                 pstat->free_bytes += bytes;
 907         }
 908
 909         pstat = page_stat__find_caller(&this);
 910         if (pstat == NULL)
 911                 return -ENOENT;
 912
 913         pstat->nr_free++;
 914         pstat->free_bytes += bytes;
 915
 916         if (live_page) {
 917                 pstat->nr_alloc--;
 918                 pstat->alloc_bytes -= bytes;
 919
 920                 if (pstat->nr_alloc == 0) {
 921                         rb_erase(&pstat->node, &page_caller_tree);
 922                         free(pstat);
 923                 }
 924         }
 925
 926         return 0;
 927 }
 928
 929 static bool perf_kmem__skip_sample(struct perf_sample *sample)
 930 {
 931         /* skip sample based on time? */
 932         if (perf_time__skip_sample(&ptime, sample->time))
 933                 return true;
 934
 935         return false;
 936 }
 937
 938 typedef int (*tracepoint_handler)(struct evsel *evsel,
 939                                   struct perf_sample *sample);
 940
 941 static int process_sample_event(struct perf_tool *tool __maybe_unused,
 942                                 union perf_event *event,
 943                                 struct perf_sample *sample,
 944                                 struct evsel *evsel,
 945                                 struct machine *machine)
 946 {
 947         int err = 0;
 948         struct thread *thread = machine__findnew_thread(machine, sample->pid,
 949                                                         sample->tid);
 950
 951         if (thread == NULL) {
 952                 pr_debug("problem processing %d event, skipping it.\n",
 953                          event->header.type);
 954                 return -1;
 955         }
 956
 957         if (perf_kmem__skip_sample(sample))
 958                 return 0;
 959
 960         dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
 961
 962         if (evsel->handler != NULL) {
 963                 tracepoint_handler f = evsel->handler;
 964                 err = f(evsel, sample);
 965         }
 966
 967         thread__put(thread);
 968
 969         return err;
 970 }
 971
 972 static struct perf_tool perf_kmem = {
 973         .sample          = process_sample_event,
 974         .comm            = perf_event__process_comm,
 975         .mmap            = perf_event__process_mmap,
 976         .mmap2           = perf_event__process_mmap2,
 977         .namespaces      = perf_event__process_namespaces,
 978         .ordered_events  = true,
 979 };
 980
 981 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
 982 {
 983         if (n_alloc == 0)
 984                 return 0.0;
 985         else
 986                 return 100.0 - (100.0 * n_req / n_alloc);
 987 }
 988
 989 static void __print_slab_result(struct rb_root *root,
 990                                 struct perf_session *session,
 991                                 int n_lines, int is_caller)
 992 {
 993         struct rb_node *next;
 994         struct machine *machine = &session->machines.host;
 995
 996         printf("%.105s\n", graph_dotted_line);
 997         printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
 998         printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
 999         printf("%.105s\n", graph_dotted_line);
1000
1001         next = rb_first(root);
1002
1003         while (next && n_lines--) {
1004                 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
1005                                                    node);
1006                 struct symbol *sym = NULL;
1007                 struct map *map;
1008                 char buf[BUFSIZ];
1009                 u64 addr;
1010
1011                 if (is_caller) {
1012                         addr = data->call_site;
1013                         if (!raw_ip)
1014                                 sym = machine__find_kernel_symbol(machine, addr, &map);
1015                 } else
1016                         addr = data->ptr;
1017
1018                 if (sym != NULL)
1019                         snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1020                                  addr - map->unmap_ip(map, sym->start));
1021                 else
1022                         snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1023                 printf(" %-34s |", buf);
1024
1025                 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1026                        (unsigned long long)data->bytes_alloc,
1027                        (unsigned long)data->bytes_alloc / data->hit,
1028                        (unsigned long long)data->bytes_req,
1029                        (unsigned long)data->bytes_req / data->hit,
1030                        (unsigned long)data->hit,
1031                        (unsigned long)data->pingpong,
1032                        fragmentation(data->bytes_req, data->bytes_alloc));
1033
1034                 next = rb_next(next);
1035         }
1036
1037         if (n_lines == -1)
1038                 printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
1039
1040         printf("%.105s\n", graph_dotted_line);
1041 }
1042
1043 static const char * const migrate_type_str[] = {
1044         "UNMOVABL",
1045         "RECLAIM",
1046         "MOVABLE",
1047         "RESERVED",
1048         "CMA/ISLT",
1049         "UNKNOWN",
1050 };
1051
1052 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1053 {
1054         struct rb_node *next = rb_first(&page_alloc_sorted);
1055         struct machine *machine = &session->machines.host;
1056         const char *format;
1057         int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1058
1059         printf("\n%.105s\n", graph_dotted_line);
1060         printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1061                use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1062                gfp_len, "GFP flags");
1063         printf("%.105s\n", graph_dotted_line);
1064
1065         if (use_pfn)
1066                 format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1067         else
1068                 format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1069
1070         while (next && n_lines--) {
1071                 struct page_stat *data;
1072                 struct symbol *sym;
1073                 struct map *map;
1074                 char buf[32];
1075                 char *caller = buf;
1076
1077                 data = rb_entry(next, struct page_stat, node);
1078                 sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1079                 if (sym)
1080                         caller = sym->name;
1081                 else
1082                         scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1083
1084                 printf(format, (unsigned long long)data->page,
1085                        (unsigned long long)data->alloc_bytes / 1024,
1086                        data->nr_alloc, data->order,
1087                        migrate_type_str[data->migrate_type],
1088                        gfp_len, compact_gfp_string(data->gfp_flags), caller);
1089
1090                 next = rb_next(next);
1091         }
1092
1093         if (n_lines == -1) {
1094                 printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1095                        gfp_len, "...");
1096         }
1097
1098         printf("%.105s\n", graph_dotted_line);
1099 }
1100
1101 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1102 {
1103         struct rb_node *next = rb_first(&page_caller_sorted);
1104         struct machine *machine = &session->machines.host;
1105         int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1106
1107         printf("\n%.105s\n", graph_dotted_line);
1108         printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1109                live_page ? "Live" : "Total", gfp_len, "GFP flags");
1110         printf("%.105s\n", graph_dotted_line);
1111
1112         while (next && n_lines--) {
1113                 struct page_stat *data;
1114                 struct symbol *sym;
1115                 struct map *map;
1116                 char buf[32];
1117                 char *caller = buf;
1118
1119                 data = rb_entry(next, struct page_stat, node);
1120                 sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1121                 if (sym)
1122                         caller = sym->name;
1123                 else
1124                         scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1125
1126                 printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1127                        (unsigned long long)data->alloc_bytes / 1024,
1128                        data->nr_alloc, data->order,
1129                        migrate_type_str[data->migrate_type],
1130                        gfp_len, compact_gfp_string(data->gfp_flags), caller);
1131
1132                 next = rb_next(next);
1133         }
1134
1135         if (n_lines == -1) {
1136                 printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1137                        gfp_len, "...");
1138         }
1139
1140         printf("%.105s\n", graph_dotted_line);
1141 }
1142
1143 static void print_gfp_flags(void)
1144 {
1145         int i;
1146
1147         printf("#\n");
1148         printf("# GFP flags\n");
1149         printf("# ---------\n");
1150         for (i = 0; i < nr_gfps; i++) {
1151                 printf("# %08x: %*s: %s\n", gfps[i].flags,
1152                        (int) max_gfp_len, gfps[i].compact_str,
1153                        gfps[i].human_readable);
1154         }
1155 }
1156
1157 static void print_slab_summary(void)
1158 {
1159         printf("\nSUMMARY (SLAB allocator)");
1160         printf("\n========================\n");
1161         printf("Total bytes requested: %'lu\n", total_requested);
1162         printf("Total bytes allocated: %'lu\n", total_allocated);
1163         printf("Total bytes freed:     %'lu\n", total_freed);
1164         if (total_allocated > total_freed) {
1165                 printf("Net total bytes allocated: %'lu\n",
1166                 total_allocated - total_freed);
1167         }
1168         printf("Total bytes wasted on internal fragmentation: %'lu\n",
1169                total_allocated - total_requested);
1170         printf("Internal fragmentation: %f%%\n",
1171                fragmentation(total_requested, total_allocated));
1172         printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1173 }
1174
1175 static void print_page_summary(void)
1176 {
1177         int o, m;
1178         u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1179         u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1180
1181         printf("\nSUMMARY (page allocator)");
1182         printf("\n========================\n");
1183         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1184                nr_page_allocs, total_page_alloc_bytes / 1024);
1185         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1186                nr_page_frees, total_page_free_bytes / 1024);
1187         printf("\n");
1188
1189         printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1190                nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1191         printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1192                nr_page_allocs - nr_alloc_freed,
1193                (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1194         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1195                nr_page_nomatch, total_page_nomatch_bytes / 1024);
1196         printf("\n");
1197
1198         printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1199                nr_page_fails, total_page_fail_bytes / 1024);
1200         printf("\n");
1201
1202         printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1203                "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1204         printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1205                graph_dotted_line, graph_dotted_line, graph_dotted_line,
1206                graph_dotted_line, graph_dotted_line);
1207
1208         for (o = 0; o < MAX_PAGE_ORDER; o++) {
1209                 printf("%5d", o);
1210                 for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1211                         if (order_stats[o][m])
1212                                 printf("  %'12d", order_stats[o][m]);
1213                         else
1214                                 printf("  %12c", '.');
1215                 }
1216                 printf("\n");
1217         }
1218 }
1219
1220 static void print_slab_result(struct perf_session *session)
1221 {
1222         if (caller_flag)
1223                 __print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1224         if (alloc_flag)
1225                 __print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1226         print_slab_summary();
1227 }
1228
1229 static void print_page_result(struct perf_session *session)
1230 {
1231         if (caller_flag || alloc_flag)
1232                 print_gfp_flags();
1233         if (caller_flag)
1234                 __print_page_caller_result(session, caller_lines);
1235         if (alloc_flag)
1236                 __print_page_alloc_result(session, alloc_lines);
1237         print_page_summary();
1238 }
1239
1240 static void print_result(struct perf_session *session)
1241 {
1242         if (kmem_slab)
1243                 print_slab_result(session);
1244         if (kmem_page)
1245                 print_page_result(session);
1246 }
1247
1248 static LIST_HEAD(slab_caller_sort);
1249 static LIST_HEAD(slab_alloc_sort);
1250 static LIST_HEAD(page_caller_sort);
1251 static LIST_HEAD(page_alloc_sort);
1252
1253 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1254                              struct list_head *sort_list)
1255 {
1256         struct rb_node **new = &(root->rb_node);
1257         struct rb_node *parent = NULL;
1258         struct sort_dimension *sort;
1259
1260         while (*new) {
1261                 struct alloc_stat *this;
1262                 int cmp = 0;
1263
1264                 this = rb_entry(*new, struct alloc_stat, node);
1265                 parent = *new;
1266
1267                 list_for_each_entry(sort, sort_list, list) {
1268                         cmp = sort->cmp(data, this);
1269                         if (cmp)
1270                                 break;
1271                 }
1272
1273                 if (cmp > 0)
1274                         new = &((*new)->rb_left);
1275                 else
1276                         new = &((*new)->rb_right);
1277         }
1278
1279         rb_link_node(&data->node, parent, new);
1280         rb_insert_color(&data->node, root);
1281 }
1282
1283 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1284                                struct list_head *sort_list)
1285 {
1286         struct rb_node *node;
1287         struct alloc_stat *data;
1288
1289         for (;;) {
1290                 node = rb_first(root);
1291                 if (!node)
1292                         break;
1293
1294                 rb_erase(node, root);
1295                 data = rb_entry(node, struct alloc_stat, node);
1296                 sort_slab_insert(root_sorted, data, sort_list);
1297         }
1298 }
1299
1300 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1301                              struct list_head *sort_list)
1302 {
1303         struct rb_node **new = &root->rb_node;
1304         struct rb_node *parent = NULL;
1305         struct sort_dimension *sort;
1306
1307         while (*new) {
1308                 struct page_stat *this;
1309                 int cmp = 0;
1310
1311                 this = rb_entry(*new, struct page_stat, node);
1312                 parent = *new;
1313
1314                 list_for_each_entry(sort, sort_list, list) {
1315                         cmp = sort->cmp(data, this);
1316                         if (cmp)
1317                                 break;
1318                 }
1319
1320                 if (cmp > 0)
1321                         new = &parent->rb_left;
1322                 else
1323                         new = &parent->rb_right;
1324         }
1325
1326         rb_link_node(&data->node, parent, new);
1327         rb_insert_color(&data->node, root);
1328 }
1329
1330 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1331                                struct list_head *sort_list)
1332 {
1333         struct rb_node *node;
1334         struct page_stat *data;
1335
1336         for (;;) {
1337                 node = rb_first(root);
1338                 if (!node)
1339                         break;
1340
1341                 rb_erase(node, root);
1342                 data = rb_entry(node, struct page_stat, node);
1343                 sort_page_insert(root_sorted, data, sort_list);
1344         }
1345 }
1346
1347 static void sort_result(void)
1348 {
1349         if (kmem_slab) {
1350                 __sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1351                                    &slab_alloc_sort);
1352                 __sort_slab_result(&root_caller_stat, &root_caller_sorted,
1353                                    &slab_caller_sort);
1354         }
1355         if (kmem_page) {
1356                 if (live_page)
1357                         __sort_page_result(&page_live_tree, &page_alloc_sorted,
1358                                            &page_alloc_sort);
1359                 else
1360                         __sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1361                                            &page_alloc_sort);
1362
1363                 __sort_page_result(&page_caller_tree, &page_caller_sorted,
1364                                    &page_caller_sort);
1365         }
1366 }
1367
1368 static int __cmd_kmem(struct perf_session *session)
1369 {
1370         int err = -EINVAL;
1371         struct evsel *evsel;
1372         const struct evsel_str_handler kmem_tracepoints[] = {
1373                 /* slab allocator */
1374                 { "kmem:kmalloc",               perf_evsel__process_alloc_event, },
1375                 { "kmem:kmem_cache_alloc",      perf_evsel__process_alloc_event, },
1376                 { "kmem:kmalloc_node",          perf_evsel__process_alloc_node_event, },
1377                 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
1378                 { "kmem:kfree",                 perf_evsel__process_free_event, },
1379                 { "kmem:kmem_cache_free",       perf_evsel__process_free_event, },
1380                 /* page allocator */
1381                 { "kmem:mm_page_alloc",         perf_evsel__process_page_alloc_event, },
1382                 { "kmem:mm_page_free",          perf_evsel__process_page_free_event, },
1383         };
1384
1385         if (!perf_session__has_traces(session, "kmem record"))
1386                 goto out;
1387
1388         if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1389                 pr_err("Initializing perf session tracepoint handlers failed\n");
1390                 goto out;
1391         }
1392
1393         evlist__for_each_entry(session->evlist, evsel) {
1394                 if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") &&
1395                     perf_evsel__field(evsel, "pfn")) {
1396                         use_pfn = true;
1397                         break;
1398                 }
1399         }
1400
1401         setup_pager();
1402         err = perf_session__process_events(session);
1403         if (err != 0) {
1404                 pr_err("error during process events: %d\n", err);
1405                 goto out;
1406         }
1407         sort_result();
1408         print_result(session);
1409 out:
1410         return err;
1411 }
1412
1413 /* slab sort keys */
1414 static int ptr_cmp(void *a, void *b)
1415 {
1416         struct alloc_stat *l = a;
1417         struct alloc_stat *r = b;
1418
1419         if (l->ptr < r->ptr)
1420                 return -1;
1421         else if (l->ptr > r->ptr)
1422                 return 1;
1423         return 0;
1424 }
1425
1426 static struct sort_dimension ptr_sort_dimension = {
1427         .name   = "ptr",
1428         .cmp    = ptr_cmp,
1429 };
1430
1431 static int slab_callsite_cmp(void *a, void *b)
1432 {
1433         struct alloc_stat *l = a;
1434         struct alloc_stat *r = b;
1435
1436         if (l->call_site < r->call_site)
1437                 return -1;
1438         else if (l->call_site > r->call_site)
1439                 return 1;
1440         return 0;
1441 }
1442
1443 static struct sort_dimension callsite_sort_dimension = {
1444         .name   = "callsite",
1445         .cmp    = slab_callsite_cmp,
1446 };
1447
1448 static int hit_cmp(void *a, void *b)
1449 {
1450         struct alloc_stat *l = a;
1451         struct alloc_stat *r = b;
1452
1453         if (l->hit < r->hit)
1454                 return -1;
1455         else if (l->hit > r->hit)
1456                 return 1;
1457         return 0;
1458 }
1459
1460 static struct sort_dimension hit_sort_dimension = {
1461         .name   = "hit",
1462         .cmp    = hit_cmp,
1463 };
1464
1465 static int bytes_cmp(void *a, void *b)
1466 {
1467         struct alloc_stat *l = a;
1468         struct alloc_stat *r = b;
1469
1470         if (l->bytes_alloc < r->bytes_alloc)
1471                 return -1;
1472         else if (l->bytes_alloc > r->bytes_alloc)
1473                 return 1;
1474         return 0;
1475 }
1476
1477 static struct sort_dimension bytes_sort_dimension = {
1478         .name   = "bytes",
1479         .cmp    = bytes_cmp,
1480 };
1481
1482 static int frag_cmp(void *a, void *b)
1483 {
1484         double x, y;
1485         struct alloc_stat *l = a;
1486         struct alloc_stat *r = b;
1487
1488         x = fragmentation(l->bytes_req, l->bytes_alloc);
1489         y = fragmentation(r->bytes_req, r->bytes_alloc);
1490
1491         if (x < y)
1492                 return -1;
1493         else if (x > y)
1494                 return 1;
1495         return 0;
1496 }
1497
1498 static struct sort_dimension frag_sort_dimension = {
1499         .name   = "frag",
1500         .cmp    = frag_cmp,
1501 };
1502
1503 static int pingpong_cmp(void *a, void *b)
1504 {
1505         struct alloc_stat *l = a;
1506         struct alloc_stat *r = b;
1507
1508         if (l->pingpong < r->pingpong)
1509                 return -1;
1510         else if (l->pingpong > r->pingpong)
1511                 return 1;
1512         return 0;
1513 }
1514
1515 static struct sort_dimension pingpong_sort_dimension = {
1516         .name   = "pingpong",
1517         .cmp    = pingpong_cmp,
1518 };
1519
1520 /* page sort keys */
1521 static int page_cmp(void *a, void *b)
1522 {
1523         struct page_stat *l = a;
1524         struct page_stat *r = b;
1525
1526         if (l->page < r->page)
1527                 return -1;
1528         else if (l->page > r->page)
1529                 return 1;
1530         return 0;
1531 }
1532
1533 static struct sort_dimension page_sort_dimension = {
1534         .name   = "page",
1535         .cmp    = page_cmp,
1536 };
1537
1538 static int page_callsite_cmp(void *a, void *b)
1539 {
1540         struct page_stat *l = a;
1541         struct page_stat *r = b;
1542
1543         if (l->callsite < r->callsite)
1544                 return -1;
1545         else if (l->callsite > r->callsite)
1546                 return 1;
1547         return 0;
1548 }
1549
1550 static struct sort_dimension page_callsite_sort_dimension = {
1551         .name   = "callsite",
1552         .cmp    = page_callsite_cmp,
1553 };
1554
1555 static int page_hit_cmp(void *a, void *b)
1556 {
1557         struct page_stat *l = a;
1558         struct page_stat *r = b;
1559
1560         if (l->nr_alloc < r->nr_alloc)
1561                 return -1;
1562         else if (l->nr_alloc > r->nr_alloc)
1563                 return 1;
1564         return 0;
1565 }
1566
1567 static struct sort_dimension page_hit_sort_dimension = {
1568         .name   = "hit",
1569         .cmp    = page_hit_cmp,
1570 };
1571
1572 static int page_bytes_cmp(void *a, void *b)
1573 {
1574         struct page_stat *l = a;
1575         struct page_stat *r = b;
1576
1577         if (l->alloc_bytes < r->alloc_bytes)
1578                 return -1;
1579         else if (l->alloc_bytes > r->alloc_bytes)
1580                 return 1;
1581         return 0;
1582 }
1583
1584 static struct sort_dimension page_bytes_sort_dimension = {
1585         .name   = "bytes",
1586         .cmp    = page_bytes_cmp,
1587 };
1588
1589 static int page_order_cmp(void *a, void *b)
1590 {
1591         struct page_stat *l = a;
1592         struct page_stat *r = b;
1593
1594         if (l->order < r->order)
1595                 return -1;
1596         else if (l->order > r->order)
1597                 return 1;
1598         return 0;
1599 }
1600
1601 static struct sort_dimension page_order_sort_dimension = {
1602         .name   = "order",
1603         .cmp    = page_order_cmp,
1604 };
1605
1606 static int migrate_type_cmp(void *a, void *b)
1607 {
1608         struct page_stat *l = a;
1609         struct page_stat *r = b;
1610
1611         /* for internal use to find free'd page */
1612         if (l->migrate_type == -1U)
1613                 return 0;
1614
1615         if (l->migrate_type < r->migrate_type)
1616                 return -1;
1617         else if (l->migrate_type > r->migrate_type)
1618                 return 1;
1619         return 0;
1620 }
1621
1622 static struct sort_dimension migrate_type_sort_dimension = {
1623         .name   = "migtype",
1624         .cmp    = migrate_type_cmp,
1625 };
1626
1627 static int gfp_flags_cmp(void *a, void *b)
1628 {
1629         struct page_stat *l = a;
1630         struct page_stat *r = b;
1631
1632         /* for internal use to find free'd page */
1633         if (l->gfp_flags == -1U)
1634                 return 0;
1635
1636         if (l->gfp_flags < r->gfp_flags)
1637                 return -1;
1638         else if (l->gfp_flags > r->gfp_flags)
1639                 return 1;
1640         return 0;
1641 }
1642
1643 static struct sort_dimension gfp_flags_sort_dimension = {
1644         .name   = "gfp",
1645         .cmp    = gfp_flags_cmp,
1646 };
1647
1648 static struct sort_dimension *slab_sorts[] = {
1649         &ptr_sort_dimension,
1650         &callsite_sort_dimension,
1651         &hit_sort_dimension,
1652         &bytes_sort_dimension,
1653         &frag_sort_dimension,
1654         &pingpong_sort_dimension,
1655 };
1656
1657 static struct sort_dimension *page_sorts[] = {
1658         &page_sort_dimension,
1659         &page_callsite_sort_dimension,
1660         &page_hit_sort_dimension,
1661         &page_bytes_sort_dimension,
1662         &page_order_sort_dimension,
1663         &migrate_type_sort_dimension,
1664         &gfp_flags_sort_dimension,
1665 };
1666
1667 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1668 {
1669         struct sort_dimension *sort;
1670         int i;
1671
1672         for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1673                 if (!strcmp(slab_sorts[i]->name, tok)) {
1674                         sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1675                         if (!sort) {
1676                                 pr_err("%s: memdup failed\n", __func__);
1677                                 return -1;
1678                         }
1679                         list_add_tail(&sort->list, list);
1680                         return 0;
1681                 }
1682         }
1683
1684         return -1;
1685 }
1686
1687 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1688 {
1689         struct sort_dimension *sort;
1690         int i;
1691
1692         for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1693                 if (!strcmp(page_sorts[i]->name, tok)) {
1694                         sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1695                         if (!sort) {
1696                                 pr_err("%s: memdup failed\n", __func__);
1697                                 return -1;
1698                         }
1699                         list_add_tail(&sort->list, list);
1700                         return 0;
1701                 }
1702         }
1703
1704         return -1;
1705 }
1706
1707 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1708 {
1709         char *tok;
1710         char *str = strdup(arg);
1711         char *pos = str;
1712
1713         if (!str) {
1714                 pr_err("%s: strdup failed\n", __func__);
1715                 return -1;
1716         }
1717
1718         while (true) {
1719                 tok = strsep(&pos, ",");
1720                 if (!tok)
1721                         break;
1722                 if (slab_sort_dimension__add(tok, sort_list) < 0) {
1723                         pr_err("Unknown slab --sort key: '%s'", tok);
1724                         free(str);
1725                         return -1;
1726                 }
1727         }
1728
1729         free(str);
1730         return 0;
1731 }
1732
1733 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1734 {
1735         char *tok;
1736         char *str = strdup(arg);
1737         char *pos = str;
1738
1739         if (!str) {
1740                 pr_err("%s: strdup failed\n", __func__);
1741                 return -1;
1742         }
1743
1744         while (true) {
1745                 tok = strsep(&pos, ",");
1746                 if (!tok)
1747                         break;
1748                 if (page_sort_dimension__add(tok, sort_list) < 0) {
1749                         pr_err("Unknown page --sort key: '%s'", tok);
1750                         free(str);
1751                         return -1;
1752                 }
1753         }
1754
1755         free(str);
1756         return 0;
1757 }
1758
1759 static int parse_sort_opt(const struct option *opt __maybe_unused,
1760                           const char *arg, int unset __maybe_unused)
1761 {
1762         if (!arg)
1763                 return -1;
1764
1765         if (kmem_page > kmem_slab ||
1766             (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1767                 if (caller_flag > alloc_flag)
1768                         return setup_page_sorting(&page_caller_sort, arg);
1769                 else
1770                         return setup_page_sorting(&page_alloc_sort, arg);
1771         } else {
1772                 if (caller_flag > alloc_flag)
1773                         return setup_slab_sorting(&slab_caller_sort, arg);
1774                 else
1775                         return setup_slab_sorting(&slab_alloc_sort, arg);
1776         }
1777
1778         return 0;
1779 }
1780
1781 static int parse_caller_opt(const struct option *opt __maybe_unused,
1782                             const char *arg __maybe_unused,
1783                             int unset __maybe_unused)
1784 {
1785         caller_flag = (alloc_flag + 1);
1786         return 0;
1787 }
1788
1789 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1790                            const char *arg __maybe_unused,
1791                            int unset __maybe_unused)
1792 {
1793         alloc_flag = (caller_flag + 1);
1794         return 0;
1795 }
1796
1797 static int parse_slab_opt(const struct option *opt __maybe_unused,
1798                           const char *arg __maybe_unused,
1799                           int unset __maybe_unused)
1800 {
1801         kmem_slab = (kmem_page + 1);
1802         return 0;
1803 }
1804
1805 static int parse_page_opt(const struct option *opt __maybe_unused,
1806                           const char *arg __maybe_unused,
1807                           int unset __maybe_unused)
1808 {
1809         kmem_page = (kmem_slab + 1);
1810         return 0;
1811 }
1812
1813 static int parse_line_opt(const struct option *opt __maybe_unused,
1814                           const char *arg, int unset __maybe_unused)
1815 {
1816         int lines;
1817
1818         if (!arg)
1819                 return -1;
1820
1821         lines = strtoul(arg, NULL, 10);
1822
1823         if (caller_flag > alloc_flag)
1824                 caller_lines = lines;
1825         else
1826                 alloc_lines = lines;
1827
1828         return 0;
1829 }
1830
1831 static int __cmd_record(int argc, const char **argv)
1832 {
1833         const char * const record_args[] = {
1834         "record", "-a", "-R", "-c", "1",
1835         };
1836         const char * const slab_events[] = {
1837         "-e", "kmem:kmalloc",
1838         "-e", "kmem:kmalloc_node",
1839         "-e", "kmem:kfree",
1840         "-e", "kmem:kmem_cache_alloc",
1841         "-e", "kmem:kmem_cache_alloc_node",
1842         "-e", "kmem:kmem_cache_free",
1843         };
1844         const char * const page_events[] = {
1845         "-e", "kmem:mm_page_alloc",
1846         "-e", "kmem:mm_page_free",
1847         };
1848         unsigned int rec_argc, i, j;
1849         const char **rec_argv;
1850
1851         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1852         if (kmem_slab)
1853                 rec_argc += ARRAY_SIZE(slab_events);
1854         if (kmem_page)
1855                 rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1856
1857         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1858
1859         if (rec_argv == NULL)
1860                 return -ENOMEM;
1861
1862         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1863                 rec_argv[i] = strdup(record_args[i]);
1864
1865         if (kmem_slab) {
1866                 for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1867                         rec_argv[i] = strdup(slab_events[j]);
1868         }
1869         if (kmem_page) {
1870                 rec_argv[i++] = strdup("-g");
1871
1872                 for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1873                         rec_argv[i] = strdup(page_events[j]);
1874         }
1875
1876         for (j = 1; j < (unsigned int)argc; j++, i++)
1877                 rec_argv[i] = argv[j];
1878
1879         return cmd_record(i, rec_argv);
1880 }
1881
1882 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1883 {
1884         if (!strcmp(var, "kmem.default")) {
1885                 if (!strcmp(value, "slab"))
1886                         kmem_default = KMEM_SLAB;
1887                 else if (!strcmp(value, "page"))
1888                         kmem_default = KMEM_PAGE;
1889                 else
1890                         pr_err("invalid default value ('slab' or 'page' required): %s\n",
1891                                value);
1892                 return 0;
1893         }
1894
1895         return 0;
1896 }
1897
1898 int cmd_kmem(int argc, const char **argv)
1899 {
1900         const char * const default_slab_sort = "frag,hit,bytes";
1901         const char * const default_page_sort = "bytes,hit";
1902         struct perf_data data = {
1903                 .mode = PERF_DATA_MODE_READ,
1904         };
1905         const struct option kmem_options[] = {
1906         OPT_STRING('i', "input", &input_name, "file", "input file name"),
1907         OPT_INCR('v', "verbose", &verbose,
1908                     "be more verbose (show symbol address, etc)"),
1909         OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1910                            "show per-callsite statistics", parse_caller_opt),
1911         OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1912                            "show per-allocation statistics", parse_alloc_opt),
1913         OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1914                      "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1915                      "page, order, migtype, gfp", parse_sort_opt),
1916         OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1917         OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1918         OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
1919         OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1920                            parse_slab_opt),
1921         OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1922                            parse_page_opt),
1923         OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1924         OPT_STRING(0, "time", &time_str, "str",
1925                    "Time span of interest (start,stop)"),
1926         OPT_END()
1927         };
1928         const char *const kmem_subcommands[] = { "record", "stat", NULL };
1929         const char *kmem_usage[] = {
1930                 NULL,
1931                 NULL
1932         };
1933         struct perf_session *session;
1934         static const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1935         int ret = perf_config(kmem_config, NULL);
1936
1937         if (ret)
1938                 return ret;
1939
1940         argc = parse_options_subcommand(argc, argv, kmem_options,
1941                                         kmem_subcommands, kmem_usage, 0);
1942
1943         if (!argc)
1944                 usage_with_options(kmem_usage, kmem_options);
1945
1946         if (kmem_slab == 0 && kmem_page == 0) {
1947                 if (kmem_default == KMEM_SLAB)
1948                         kmem_slab = 1;
1949                 else
1950                         kmem_page = 1;
1951         }
1952
1953         if (!strncmp(argv[0], "rec", 3)) {
1954                 symbol__init(NULL);
1955                 return __cmd_record(argc, argv);
1956         }
1957
1958         data.path = input_name;
1959
1960         kmem_session = session = perf_session__new(&data, false, &perf_kmem);
1961         if (IS_ERR(session))
1962                 return PTR_ERR(session);
1963
1964         ret = -1;
1965
1966         if (kmem_slab) {
1967                 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
1968                                                           "kmem:kmalloc")) {
1969                         pr_err(errmsg, "slab", "slab");
1970                         goto out_delete;
1971                 }
1972         }
1973
1974         if (kmem_page) {
1975                 struct evsel *evsel;
1976
1977                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
1978                                                              "kmem:mm_page_alloc");
1979                 if (evsel == NULL) {
1980                         pr_err(errmsg, "page", "page");
1981                         goto out_delete;
1982                 }
1983
1984                 kmem_page_size = tep_get_page_size(evsel->tp_format->tep);
1985                 symbol_conf.use_callchain = true;
1986         }
1987
1988         symbol__init(&session->header.env);
1989
1990         if (perf_time__parse_str(&ptime, time_str) != 0) {
1991                 pr_err("Invalid time string\n");
1992                 ret = -EINVAL;
1993                 goto out_delete;
1994         }
1995
1996         if (!strcmp(argv[0], "stat")) {
1997                 setlocale(LC_ALL, "");
1998
1999                 if (cpu__setup_cpunode_map())
2000                         goto out_delete;
2001
2002                 if (list_empty(&slab_caller_sort))
2003                         setup_slab_sorting(&slab_caller_sort, default_slab_sort);
2004                 if (list_empty(&slab_alloc_sort))
2005                         setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
2006                 if (list_empty(&page_caller_sort))
2007                         setup_page_sorting(&page_caller_sort, default_page_sort);
2008                 if (list_empty(&page_alloc_sort))
2009                         setup_page_sorting(&page_alloc_sort, default_page_sort);
2010
2011                 if (kmem_page) {
2012                         setup_page_sorting(&page_alloc_sort_input,
2013                                            "page,order,migtype,gfp");
2014                         setup_page_sorting(&page_caller_sort_input,
2015                                            "callsite,order,migtype,gfp");
2016                 }
2017                 ret = __cmd_kmem(session);
2018         } else
2019                 usage_with_options(kmem_usage, kmem_options);
2020
2021 out_delete:
2022         perf_session__delete(session);
2023
2024         return ret;
2025 }
2026