builtin-kmem.c

   1// SPDX-License-Identifier: GPL-2.0
   2#include "builtin.h"
   3
   4#include "util/dso.h"
   5#include "util/evlist.h"
   6#include "util/evsel.h"
   7#include "util/config.h"
   8#include "util/map.h"
   9#include "util/symbol.h"
  10#include "util/thread.h"
  11#include "util/header.h"
  12#include "util/session.h"
  13#include "util/tool.h"
  14#include "util/callchain.h"
  15#include "util/time-utils.h"
  16#include <linux/err.h>
  17
  18#include <subcmd/pager.h>
  19#include <subcmd/parse-options.h>
  20#include "util/trace-event.h"
  21#include "util/data.h"
  22#include "util/cpumap.h"
  23
  24#include "util/debug.h"
  25#include "util/string2.h"
  26#include "util/util.h"
  27
  28#include <linux/kernel.h>
  29#include <linux/numa.h>
  30#include <linux/rbtree.h>
  31#include <linux/string.h>
  32#include <linux/zalloc.h>
  33#include <errno.h>
  34#include <inttypes.h>
  35#include <locale.h>
  36#include <regex.h>
  37
  38#include <linux/ctype.h>
  39#include <event-parse.h>
  40
  41static int	kmem_slab;
  42static int	kmem_page;
  43
  44static long	kmem_page_size;
  45static enum {
  46	KMEM_SLAB,
  47	KMEM_PAGE,
  48} kmem_default = KMEM_SLAB;  /* for backward compatibility */
  49
  50struct alloc_stat;
  51typedef int (*sort_fn_t)(void *, void *);
  52
  53static int			alloc_flag;
  54static int			caller_flag;
  55
  56static int			alloc_lines = -1;
  57static int			caller_lines = -1;
  58
  59static bool			raw_ip;
  60
  61struct alloc_stat {
  62	u64	call_site;
  63	u64	ptr;
  64	u64	bytes_req;
  65	u64	bytes_alloc;
  66	u64	last_alloc;
  67	u32	hit;
  68	u32	pingpong;
  69
  70	short	alloc_cpu;
  71
  72	struct rb_node node;
  73};
  74
  75static struct rb_root root_alloc_stat;
  76static struct rb_root root_alloc_sorted;
  77static struct rb_root root_caller_stat;
  78static struct rb_root root_caller_sorted;
  79
  80static unsigned long total_requested, total_allocated, total_freed;
  81static unsigned long nr_allocs, nr_cross_allocs;
  82
  83/* filters for controlling start and stop of time of analysis */
  84static struct perf_time_interval ptime;
  85const char *time_str;
  86
  87static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
  88			     int bytes_req, int bytes_alloc, int cpu)
  89{
  90	struct rb_node **node = &root_alloc_stat.rb_node;
  91	struct rb_node *parent = NULL;
  92	struct alloc_stat *data = NULL;
  93
  94	while (*node) {
  95		parent = *node;
  96		data = rb_entry(*node, struct alloc_stat, node);
  97
  98		if (ptr > data->ptr)
  99			node = &(*node)->rb_right;
 100		else if (ptr < data->ptr)
 101			node = &(*node)->rb_left;
 102		else
 103			break;
 104	}
 105
 106	if (data && data->ptr == ptr) {
 107		data->hit++;
 108		data->bytes_req += bytes_req;
 109		data->bytes_alloc += bytes_alloc;
 110	} else {
 111		data = malloc(sizeof(*data));
 112		if (!data) {
 113			pr_err("%s: malloc failed\n", __func__);
 114			return -1;
 115		}
 116		data->ptr = ptr;
 117		data->pingpong = 0;
 118		data->hit = 1;
 119		data->bytes_req = bytes_req;
 120		data->bytes_alloc = bytes_alloc;
 121
 122		rb_link_node(&data->node, parent, node);
 123		rb_insert_color(&data->node, &root_alloc_stat);
 124	}
 125	data->call_site = call_site;
 126	data->alloc_cpu = cpu;
 127	data->last_alloc = bytes_alloc;
 128
 129	return 0;
 130}
 131
 132static int insert_caller_stat(unsigned long call_site,
 133			      int bytes_req, int bytes_alloc)
 134{
 135	struct rb_node **node = &root_caller_stat.rb_node;
 136	struct rb_node *parent = NULL;
 137	struct alloc_stat *data = NULL;
 138
 139	while (*node) {
 140		parent = *node;
 141		data = rb_entry(*node, struct alloc_stat, node);
 142
 143		if (call_site > data->call_site)
 144			node = &(*node)->rb_right;
 145		else if (call_site < data->call_site)
 146			node = &(*node)->rb_left;
 147		else
 148			break;
 149	}
 150
 151	if (data && data->call_site == call_site) {
 152		data->hit++;
 153		data->bytes_req += bytes_req;
 154		data->bytes_alloc += bytes_alloc;
 155	} else {
 156		data = malloc(sizeof(*data));
 157		if (!data) {
 158			pr_err("%s: malloc failed\n", __func__);
 159			return -1;
 160		}
 161		data->call_site = call_site;
 162		data->pingpong = 0;
 163		data->hit = 1;
 164		data->bytes_req = bytes_req;
 165		data->bytes_alloc = bytes_alloc;
 166
 167		rb_link_node(&data->node, parent, node);
 168		rb_insert_color(&data->node, &root_caller_stat);
 169	}
 170
 171	return 0;
 172}
 173
 174static int evsel__process_alloc_event(struct evsel *evsel, struct perf_sample *sample)
 175{
 176	unsigned long ptr = evsel__intval(evsel, sample, "ptr"),
 177		      call_site = evsel__intval(evsel, sample, "call_site");
 178	int bytes_req = evsel__intval(evsel, sample, "bytes_req"),
 179	    bytes_alloc = evsel__intval(evsel, sample, "bytes_alloc");
 180
 181	if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
 182	    insert_caller_stat(call_site, bytes_req, bytes_alloc))
 183		return -1;
 184
 185	total_requested += bytes_req;
 186	total_allocated += bytes_alloc;
 187
 188	nr_allocs++;
 189
 190	/*
 191	 * Commit 11e9734bcb6a ("mm/slab_common: unify NUMA and UMA
 192	 * version of tracepoints") adds the field "node" into the
 193	 * tracepoints 'kmalloc' and 'kmem_cache_alloc'.
 194	 *
 195	 * The legacy tracepoints 'kmalloc_node' and 'kmem_cache_alloc_node'
 196	 * also contain the field "node".
 197	 *
 198	 * If the tracepoint contains the field "node" the tool stats the
 199	 * cross allocation.
 200	 */
 201	if (evsel__field(evsel, "node")) {
 202		int node1, node2;
 203
 204		node1 = cpu__get_node((struct perf_cpu){.cpu = sample->cpu});
 205		node2 = evsel__intval(evsel, sample, "node");
 206
 207		/*
 208		 * If the field "node" is NUMA_NO_NODE (-1), we don't take it
 209		 * as a cross allocation.
 210		 */
 211		if ((node2 != NUMA_NO_NODE) && (node1 != node2))
 212			nr_cross_allocs++;
 213	}
 214
 215	return 0;
 216}
 217
 218static int ptr_cmp(void *, void *);
 219static int slab_callsite_cmp(void *, void *);
 220
 221static struct alloc_stat *search_alloc_stat(unsigned long ptr,
 222					    unsigned long call_site,
 223					    struct rb_root *root,
 224					    sort_fn_t sort_fn)
 225{
 226	struct rb_node *node = root->rb_node;
 227	struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
 228
 229	while (node) {
 230		struct alloc_stat *data;
 231		int cmp;
 232
 233		data = rb_entry(node, struct alloc_stat, node);
 234
 235		cmp = sort_fn(&key, data);
 236		if (cmp < 0)
 237			node = node->rb_left;
 238		else if (cmp > 0)
 239			node = node->rb_right;
 240		else
 241			return data;
 242	}
 243	return NULL;
 244}
 245
 246static int evsel__process_free_event(struct evsel *evsel, struct perf_sample *sample)
 247{
 248	unsigned long ptr = evsel__intval(evsel, sample, "ptr");
 249	struct alloc_stat *s_alloc, *s_caller;
 250
 251	s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
 252	if (!s_alloc)
 253		return 0;
 254
 255	total_freed += s_alloc->last_alloc;
 256
 257	if ((short)sample->cpu != s_alloc->alloc_cpu) {
 258		s_alloc->pingpong++;
 259
 260		s_caller = search_alloc_stat(0, s_alloc->call_site,
 261					     &root_caller_stat,
 262					     slab_callsite_cmp);
 263		if (!s_caller)
 264			return -1;
 265		s_caller->pingpong++;
 266	}
 267	s_alloc->alloc_cpu = -1;
 268
 269	return 0;
 270}
 271
 272static u64 total_page_alloc_bytes;
 273static u64 total_page_free_bytes;
 274static u64 total_page_nomatch_bytes;
 275static u64 total_page_fail_bytes;
 276static unsigned long nr_page_allocs;
 277static unsigned long nr_page_frees;
 278static unsigned long nr_page_fails;
 279static unsigned long nr_page_nomatch;
 280
 281static bool use_pfn;
 282static bool live_page;
 283static struct perf_session *kmem_session;
 284
 285#define MAX_MIGRATE_TYPES  6
 286#define MAX_PAGE_ORDER     11
 287
 288static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
 289
 290struct page_stat {
 291	struct rb_node 	node;
 292	u64 		page;
 293	u64 		callsite;
 294	int 		order;
 295	unsigned 	gfp_flags;
 296	unsigned 	migrate_type;
 297	u64		alloc_bytes;
 298	u64 		free_bytes;
 299	int 		nr_alloc;
 300	int 		nr_free;
 301};
 302
 303static struct rb_root page_live_tree;
 304static struct rb_root page_alloc_tree;
 305static struct rb_root page_alloc_sorted;
 306static struct rb_root page_caller_tree;
 307static struct rb_root page_caller_sorted;
 308
 309struct alloc_func {
 310	u64 start;
 311	u64 end;
 312	char *name;
 313};
 314
 315static int nr_alloc_funcs;
 316static struct alloc_func *alloc_func_list;
 317
 318static int funcmp(const void *a, const void *b)
 319{
 320	const struct alloc_func *fa = a;
 321	const struct alloc_func *fb = b;
 322
 323	if (fa->start > fb->start)
 324		return 1;
 325	else
 326		return -1;
 327}
 328
 329static int callcmp(const void *a, const void *b)
 330{
 331	const struct alloc_func *fa = a;
 332	const struct alloc_func *fb = b;
 333
 334	if (fb->start <= fa->start && fa->end < fb->end)
 335		return 0;
 336
 337	if (fa->start > fb->start)
 338		return 1;
 339	else
 340		return -1;
 341}
 342
 343static int build_alloc_func_list(void)
 344{
 345	int ret;
 346	struct map *kernel_map;
 347	struct symbol *sym;
 348	struct rb_node *node;
 349	struct alloc_func *func;
 350	struct machine *machine = &kmem_session->machines.host;
 351	regex_t alloc_func_regex;
 352	static const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
 353
 354	ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
 355	if (ret) {
 356		char err[BUFSIZ];
 357
 358		regerror(ret, &alloc_func_regex, err, sizeof(err));
 359		pr_err("Invalid regex: %s\n%s", pattern, err);
 360		return -EINVAL;
 361	}
 362
 363	kernel_map = machine__kernel_map(machine);
 364	if (map__load(kernel_map) < 0) {
 365		pr_err("cannot load kernel map\n");
 366		return -ENOENT;
 367	}
 368
 369	map__for_each_symbol(kernel_map, sym, node) {
 370		if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
 371			continue;
 372
 373		func = realloc(alloc_func_list,
 374			       (nr_alloc_funcs + 1) * sizeof(*func));
 375		if (func == NULL)
 376			return -ENOMEM;
 377
 378		pr_debug("alloc func: %s\n", sym->name);
 379		func[nr_alloc_funcs].start = sym->start;
 380		func[nr_alloc_funcs].end   = sym->end;
 381		func[nr_alloc_funcs].name  = sym->name;
 382
 383		alloc_func_list = func;
 384		nr_alloc_funcs++;
 385	}
 386
 387	qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
 388
 389	regfree(&alloc_func_regex);
 390	return 0;
 391}
 392
 393/*
 394 * Find first non-memory allocation function from callchain.
 395 * The allocation functions are in the 'alloc_func_list'.
 396 */
 397static u64 find_callsite(struct evsel *evsel, struct perf_sample *sample)
 398{
 399	struct addr_location al;
 400	struct machine *machine = &kmem_session->machines.host;
 401	struct callchain_cursor_node *node;
 402	struct callchain_cursor *cursor;
 403	u64 result = sample->ip;
 404
 405	addr_location__init(&al);
 406	if (alloc_func_list == NULL) {
 407		if (build_alloc_func_list() < 0)
 408			goto out;
 409	}
 410
 411	al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
 412
 413	cursor = get_tls_callchain_cursor();
 414	if (cursor == NULL)
 415		goto out;
 416
 417	sample__resolve_callchain(sample, cursor, NULL, evsel, &al, 16);
 418
 419	callchain_cursor_commit(cursor);
 420	while (true) {
 421		struct alloc_func key, *caller;
 422		u64 addr;
 423
 424		node = callchain_cursor_current(cursor);
 425		if (node == NULL)
 426			break;
 427
 428		key.start = key.end = node->ip;
 429		caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
 430				 sizeof(key), callcmp);
 431		if (!caller) {
 432			/* found */
 433			if (node->ms.map)
 434				addr = map__dso_unmap_ip(node->ms.map, node->ip);
 435			else
 436				addr = node->ip;
 437
 438			result = addr;
 439			goto out;
 440		} else
 441			pr_debug3("skipping alloc function: %s\n", caller->name);
 442
 443		callchain_cursor_advance(cursor);
 444	}
 445
 446	pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
 447out:
 448	addr_location__exit(&al);
 449	return result;
 450}
 451
 452struct sort_dimension {
 453	const char		name[20];
 454	sort_fn_t		cmp;
 455	struct list_head	list;
 456};
 457
 458static LIST_HEAD(page_alloc_sort_input);
 459static LIST_HEAD(page_caller_sort_input);
 460
 461static struct page_stat *
 462__page_stat__findnew_page(struct page_stat *pstat, bool create)
 463{
 464	struct rb_node **node = &page_live_tree.rb_node;
 465	struct rb_node *parent = NULL;
 466	struct page_stat *data;
 467
 468	while (*node) {
 469		s64 cmp;
 470
 471		parent = *node;
 472		data = rb_entry(*node, struct page_stat, node);
 473
 474		cmp = data->page - pstat->page;
 475		if (cmp < 0)
 476			node = &parent->rb_left;
 477		else if (cmp > 0)
 478			node = &parent->rb_right;
 479		else
 480			return data;
 481	}
 482
 483	if (!create)
 484		return NULL;
 485
 486	data = zalloc(sizeof(*data));
 487	if (data != NULL) {
 488		data->page = pstat->page;
 489		data->order = pstat->order;
 490		data->gfp_flags = pstat->gfp_flags;
 491		data->migrate_type = pstat->migrate_type;
 492
 493		rb_link_node(&data->node, parent, node);
 494		rb_insert_color(&data->node, &page_live_tree);
 495	}
 496
 497	return data;
 498}
 499
 500static struct page_stat *page_stat__find_page(struct page_stat *pstat)
 501{
 502	return __page_stat__findnew_page(pstat, false);
 503}
 504
 505static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
 506{
 507	return __page_stat__findnew_page(pstat, true);
 508}
 509
 510static struct page_stat *
 511__page_stat__findnew_alloc(struct page_stat *pstat, bool create)
 512{
 513	struct rb_node **node = &page_alloc_tree.rb_node;
 514	struct rb_node *parent = NULL;
 515	struct page_stat *data;
 516	struct sort_dimension *sort;
 517
 518	while (*node) {
 519		int cmp = 0;
 520
 521		parent = *node;
 522		data = rb_entry(*node, struct page_stat, node);
 523
 524		list_for_each_entry(sort, &page_alloc_sort_input, list) {
 525			cmp = sort->cmp(pstat, data);
 526			if (cmp)
 527				break;
 528		}
 529
 530		if (cmp < 0)
 531			node = &parent->rb_left;
 532		else if (cmp > 0)
 533			node = &parent->rb_right;
 534		else
 535			return data;
 536	}
 537
 538	if (!create)
 539		return NULL;
 540
 541	data = zalloc(sizeof(*data));
 542	if (data != NULL) {
 543		data->page = pstat->page;
 544		data->order = pstat->order;
 545		data->gfp_flags = pstat->gfp_flags;
 546		data->migrate_type = pstat->migrate_type;
 547
 548		rb_link_node(&data->node, parent, node);
 549		rb_insert_color(&data->node, &page_alloc_tree);
 550	}
 551
 552	return data;
 553}
 554
 555static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
 556{
 557	return __page_stat__findnew_alloc(pstat, false);
 558}
 559
 560static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
 561{
 562	return __page_stat__findnew_alloc(pstat, true);
 563}
 564
 565static struct page_stat *
 566__page_stat__findnew_caller(struct page_stat *pstat, bool create)
 567{
 568	struct rb_node **node = &page_caller_tree.rb_node;
 569	struct rb_node *parent = NULL;
 570	struct page_stat *data;
 571	struct sort_dimension *sort;
 572
 573	while (*node) {
 574		int cmp = 0;
 575
 576		parent = *node;
 577		data = rb_entry(*node, struct page_stat, node);
 578
 579		list_for_each_entry(sort, &page_caller_sort_input, list) {
 580			cmp = sort->cmp(pstat, data);
 581			if (cmp)
 582				break;
 583		}
 584
 585		if (cmp < 0)
 586			node = &parent->rb_left;
 587		else if (cmp > 0)
 588			node = &parent->rb_right;
 589		else
 590			return data;
 591	}
 592
 593	if (!create)
 594		return NULL;
 595
 596	data = zalloc(sizeof(*data));
 597	if (data != NULL) {
 598		data->callsite = pstat->callsite;
 599		data->order = pstat->order;
 600		data->gfp_flags = pstat->gfp_flags;
 601		data->migrate_type = pstat->migrate_type;
 602
 603		rb_link_node(&data->node, parent, node);
 604		rb_insert_color(&data->node, &page_caller_tree);
 605	}
 606
 607	return data;
 608}
 609
 610static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
 611{
 612	return __page_stat__findnew_caller(pstat, false);
 613}
 614
 615static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
 616{
 617	return __page_stat__findnew_caller(pstat, true);
 618}
 619
 620static bool valid_page(u64 pfn_or_page)
 621{
 622	if (use_pfn && pfn_or_page == -1UL)
 623		return false;
 624	if (!use_pfn && pfn_or_page == 0)
 625		return false;
 626	return true;
 627}
 628
 629struct gfp_flag {
 630	unsigned int flags;
 631	char *compact_str;
 632	char *human_readable;
 633};
 634
 635static struct gfp_flag *gfps;
 636static int nr_gfps;
 637
 638static int gfpcmp(const void *a, const void *b)
 639{
 640	const struct gfp_flag *fa = a;
 641	const struct gfp_flag *fb = b;
 642
 643	return fa->flags - fb->flags;
 644}
 645
 646/* see include/trace/events/mmflags.h */
 647static const struct {
 648	const char *original;
 649	const char *compact;
 650} gfp_compact_table[] = {
 651	{ "GFP_TRANSHUGE",		"THP" },
 652	{ "GFP_TRANSHUGE_LIGHT",	"THL" },
 653	{ "GFP_HIGHUSER_MOVABLE",	"HUM" },
 654	{ "GFP_HIGHUSER",		"HU" },
 655	{ "GFP_USER",			"U" },
 656	{ "GFP_KERNEL_ACCOUNT",		"KAC" },
 657	{ "GFP_KERNEL",			"K" },
 658	{ "GFP_NOFS",			"NF" },
 659	{ "GFP_ATOMIC",			"A" },
 660	{ "GFP_NOIO",			"NI" },
 661	{ "GFP_NOWAIT",			"NW" },
 662	{ "GFP_DMA",			"D" },
 663	{ "__GFP_HIGHMEM",		"HM" },
 664	{ "GFP_DMA32",			"D32" },
 665	{ "__GFP_HIGH",			"H" },
 666	{ "__GFP_IO",			"I" },
 667	{ "__GFP_FS",			"F" },
 668	{ "__GFP_NOWARN",		"NWR" },
 669	{ "__GFP_RETRY_MAYFAIL",	"R" },
 670	{ "__GFP_NOFAIL",		"NF" },
 671	{ "__GFP_NORETRY",		"NR" },
 672	{ "__GFP_COMP",			"C" },
 673	{ "__GFP_ZERO",			"Z" },
 674	{ "__GFP_NOMEMALLOC",		"NMA" },
 675	{ "__GFP_MEMALLOC",		"MA" },
 676	{ "__GFP_HARDWALL",		"HW" },
 677	{ "__GFP_THISNODE",		"TN" },
 678	{ "__GFP_RECLAIMABLE",		"RC" },
 679	{ "__GFP_MOVABLE",		"M" },
 680	{ "__GFP_ACCOUNT",		"AC" },
 681	{ "__GFP_WRITE",		"WR" },
 682	{ "__GFP_RECLAIM",		"R" },
 683	{ "__GFP_DIRECT_RECLAIM",	"DR" },
 684	{ "__GFP_KSWAPD_RECLAIM",	"KR" },
 685};
 686
 687static size_t max_gfp_len;
 688
 689static char *compact_gfp_flags(char *gfp_flags)
 690{
 691	char *orig_flags = strdup(gfp_flags);
 692	char *new_flags = NULL;
 693	char *str, *pos = NULL;
 694	size_t len = 0;
 695
 696	if (orig_flags == NULL)
 697		return NULL;
 698
 699	str = strtok_r(orig_flags, "|", &pos);
 700	while (str) {
 701		size_t i;
 702		char *new;
 703		const char *cpt;
 704
 705		for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
 706			if (strcmp(gfp_compact_table[i].original, str))
 707				continue;
 708
 709			cpt = gfp_compact_table[i].compact;
 710			new = realloc(new_flags, len + strlen(cpt) + 2);
 711			if (new == NULL) {
 712				free(new_flags);
 713				free(orig_flags);
 714				return NULL;
 715			}
 716
 717			new_flags = new;
 718
 719			if (!len) {
 720				strcpy(new_flags, cpt);
 721			} else {
 722				strcat(new_flags, "|");
 723				strcat(new_flags, cpt);
 724				len++;
 725			}
 726
 727			len += strlen(cpt);
 728		}
 729
 730		str = strtok_r(NULL, "|", &pos);
 731	}
 732
 733	if (max_gfp_len < len)
 734		max_gfp_len = len;
 735
 736	free(orig_flags);
 737	return new_flags;
 738}
 739
 740static char *compact_gfp_string(unsigned long gfp_flags)
 741{
 742	struct gfp_flag key = {
 743		.flags = gfp_flags,
 744	};
 745	struct gfp_flag *gfp;
 746
 747	gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
 748	if (gfp)
 749		return gfp->compact_str;
 750
 751	return NULL;
 752}
 753
 754static int parse_gfp_flags(struct evsel *evsel, struct perf_sample *sample,
 755			   unsigned int gfp_flags)
 756{
 757	struct tep_record record = {
 758		.cpu = sample->cpu,
 759		.data = sample->raw_data,
 760		.size = sample->raw_size,
 761	};
 762	struct trace_seq seq;
 763	char *str, *pos = NULL;
 764	const struct tep_event *tp_format;
 765
 766	if (nr_gfps) {
 767		struct gfp_flag key = {
 768			.flags = gfp_flags,
 769		};
 770
 771		if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
 772			return 0;
 773	}
 774
 775	trace_seq_init(&seq);
 776	tp_format = evsel__tp_format(evsel);
 777	if (tp_format)
 778		tep_print_event(tp_format->tep, &seq, &record, "%s", TEP_PRINT_INFO);
 779
 780	str = strtok_r(seq.buffer, " ", &pos);
 781	while (str) {
 782		if (!strncmp(str, "gfp_flags=", 10)) {
 783			struct gfp_flag *new;
 784
 785			new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
 786			if (new == NULL)
 787				return -ENOMEM;
 788
 789			gfps = new;
 790			new += nr_gfps++;
 791
 792			new->flags = gfp_flags;
 793			new->human_readable = strdup(str + 10);
 794			new->compact_str = compact_gfp_flags(str + 10);
 795			if (!new->human_readable || !new->compact_str)
 796				return -ENOMEM;
 797
 798			qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
 799		}
 800
 801		str = strtok_r(NULL, " ", &pos);
 802	}
 803
 804	trace_seq_destroy(&seq);
 805	return 0;
 806}
 807
 808static int evsel__process_page_alloc_event(struct evsel *evsel, struct perf_sample *sample)
 809{
 810	u64 page;
 811	unsigned int order = evsel__intval(evsel, sample, "order");
 812	unsigned int gfp_flags = evsel__intval(evsel, sample, "gfp_flags");
 813	unsigned int migrate_type = evsel__intval(evsel, sample,
 814						       "migratetype");
 815	u64 bytes = kmem_page_size << order;
 816	u64 callsite;
 817	struct page_stat *pstat;
 818	struct page_stat this = {
 819		.order = order,
 820		.gfp_flags = gfp_flags,
 821		.migrate_type = migrate_type,
 822	};
 823
 824	if (use_pfn)
 825		page = evsel__intval(evsel, sample, "pfn");
 826	else
 827		page = evsel__intval(evsel, sample, "page");
 828
 829	nr_page_allocs++;
 830	total_page_alloc_bytes += bytes;
 831
 832	if (!valid_page(page)) {
 833		nr_page_fails++;
 834		total_page_fail_bytes += bytes;
 835
 836		return 0;
 837	}
 838
 839	if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
 840		return -1;
 841
 842	callsite = find_callsite(evsel, sample);
 843
 844	/*
 845	 * This is to find the current page (with correct gfp flags and
 846	 * migrate type) at free event.
 847	 */
 848	this.page = page;
 849	pstat = page_stat__findnew_page(&this);
 850	if (pstat == NULL)
 851		return -ENOMEM;
 852
 853	pstat->nr_alloc++;
 854	pstat->alloc_bytes += bytes;
 855	pstat->callsite = callsite;
 856
 857	if (!live_page) {
 858		pstat = page_stat__findnew_alloc(&this);
 859		if (pstat == NULL)
 860			return -ENOMEM;
 861
 862		pstat->nr_alloc++;
 863		pstat->alloc_bytes += bytes;
 864		pstat->callsite = callsite;
 865	}
 866
 867	this.callsite = callsite;
 868	pstat = page_stat__findnew_caller(&this);
 869	if (pstat == NULL)
 870		return -ENOMEM;
 871
 872	pstat->nr_alloc++;
 873	pstat->alloc_bytes += bytes;
 874
 875	order_stats[order][migrate_type]++;
 876
 877	return 0;
 878}
 879
 880static int evsel__process_page_free_event(struct evsel *evsel, struct perf_sample *sample)
 881{
 882	u64 page;
 883	unsigned int order = evsel__intval(evsel, sample, "order");
 884	u64 bytes = kmem_page_size << order;
 885	struct page_stat *pstat;
 886	struct page_stat this = {
 887		.order = order,
 888	};
 889
 890	if (use_pfn)
 891		page = evsel__intval(evsel, sample, "pfn");
 892	else
 893		page = evsel__intval(evsel, sample, "page");
 894
 895	nr_page_frees++;
 896	total_page_free_bytes += bytes;
 897
 898	this.page = page;
 899	pstat = page_stat__find_page(&this);
 900	if (pstat == NULL) {
 901		pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
 902			  page, order);
 903
 904		nr_page_nomatch++;
 905		total_page_nomatch_bytes += bytes;
 906
 907		return 0;
 908	}
 909
 910	this.gfp_flags = pstat->gfp_flags;
 911	this.migrate_type = pstat->migrate_type;
 912	this.callsite = pstat->callsite;
 913
 914	rb_erase(&pstat->node, &page_live_tree);
 915	free(pstat);
 916
 917	if (live_page) {
 918		order_stats[this.order][this.migrate_type]--;
 919	} else {
 920		pstat = page_stat__find_alloc(&this);
 921		if (pstat == NULL)
 922			return -ENOMEM;
 923
 924		pstat->nr_free++;
 925		pstat->free_bytes += bytes;
 926	}
 927
 928	pstat = page_stat__find_caller(&this);
 929	if (pstat == NULL)
 930		return -ENOENT;
 931
 932	pstat->nr_free++;
 933	pstat->free_bytes += bytes;
 934
 935	if (live_page) {
 936		pstat->nr_alloc--;
 937		pstat->alloc_bytes -= bytes;
 938
 939		if (pstat->nr_alloc == 0) {
 940			rb_erase(&pstat->node, &page_caller_tree);
 941			free(pstat);
 942		}
 943	}
 944
 945	return 0;
 946}
 947
 948static bool perf_kmem__skip_sample(struct perf_sample *sample)
 949{
 950	/* skip sample based on time? */
 951	if (perf_time__skip_sample(&ptime, sample->time))
 952		return true;
 953
 954	return false;
 955}
 956
 957typedef int (*tracepoint_handler)(struct evsel *evsel,
 958				  struct perf_sample *sample);
 959
 960static int process_sample_event(const struct perf_tool *tool __maybe_unused,
 961				union perf_event *event,
 962				struct perf_sample *sample,
 963				struct evsel *evsel,
 964				struct machine *machine)
 965{
 966	int err = 0;
 967	struct thread *thread = machine__findnew_thread(machine, sample->pid,
 968							sample->tid);
 969
 970	if (thread == NULL) {
 971		pr_debug("problem processing %d event, skipping it.\n",
 972			 event->header.type);
 973		return -1;
 974	}
 975
 976	if (perf_kmem__skip_sample(sample))
 977		return 0;
 978
 979	dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread__tid(thread));
 980
 981	if (evsel->handler != NULL) {
 982		tracepoint_handler f = evsel->handler;
 983		err = f(evsel, sample);
 984	}
 985
 986	thread__put(thread);
 987
 988	return err;
 989}
 990
 991static double fragmentation(unsigned long n_req, unsigned long n_alloc)
 992{
 993	if (n_alloc == 0)
 994		return 0.0;
 995	else
 996		return 100.0 - (100.0 * n_req / n_alloc);
 997}
 998
 999static void __print_slab_result(struct rb_root *root,
1000				struct perf_session *session,
1001				int n_lines, int is_caller)
1002{
1003	struct rb_node *next;
1004	struct machine *machine = &session->machines.host;
1005
1006	printf("%.105s\n", graph_dotted_line);
1007	printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
1008	printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
1009	printf("%.105s\n", graph_dotted_line);
1010
1011	next = rb_first(root);
1012
1013	while (next && n_lines--) {
1014		struct alloc_stat *data = rb_entry(next, struct alloc_stat,
1015						   node);
1016		struct symbol *sym = NULL;
1017		struct map *map;
1018		char buf[BUFSIZ];
1019		u64 addr;
1020
1021		if (is_caller) {
1022			addr = data->call_site;
1023			if (!raw_ip)
1024				sym = machine__find_kernel_symbol(machine, addr, &map);
1025		} else
1026			addr = data->ptr;
1027
1028		if (sym != NULL)
1029			snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1030				 addr - map__unmap_ip(map, sym->start));
1031		else
1032			snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1033		printf(" %-34s |", buf);
1034
1035		printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1036		       (unsigned long long)data->bytes_alloc,
1037		       (unsigned long)data->bytes_alloc / data->hit,
1038		       (unsigned long long)data->bytes_req,
1039		       (unsigned long)data->bytes_req / data->hit,
1040		       (unsigned long)data->hit,
1041		       (unsigned long)data->pingpong,
1042		       fragmentation(data->bytes_req, data->bytes_alloc));
1043
1044		next = rb_next(next);
1045	}
1046
1047	if (n_lines == -1)
1048		printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
1049
1050	printf("%.105s\n", graph_dotted_line);
1051}
1052
1053static const char * const migrate_type_str[] = {
1054	"UNMOVABL",
1055	"RECLAIM",
1056	"MOVABLE",
1057	"RESERVED",
1058	"CMA/ISLT",
1059	"UNKNOWN",
1060};
1061
1062static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1063{
1064	struct rb_node *next = rb_first(&page_alloc_sorted);
1065	struct machine *machine = &session->machines.host;
1066	const char *format;
1067	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1068
1069	printf("\n%.105s\n", graph_dotted_line);
1070	printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1071	       use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1072	       gfp_len, "GFP flags");
1073	printf("%.105s\n", graph_dotted_line);
1074
1075	if (use_pfn)
1076		format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1077	else
1078		format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1079
1080	while (next && n_lines--) {
1081		struct page_stat *data;
1082		struct symbol *sym;
1083		struct map *map;
1084		char buf[32];
1085		char *caller = buf;
1086
1087		data = rb_entry(next, struct page_stat, node);
1088		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1089		if (sym)
1090			caller = sym->name;
1091		else
1092			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1093
1094		printf(format, (unsigned long long)data->page,
1095		       (unsigned long long)data->alloc_bytes / 1024,
1096		       data->nr_alloc, data->order,
1097		       migrate_type_str[data->migrate_type],
1098		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1099
1100		next = rb_next(next);
1101	}
1102
1103	if (n_lines == -1) {
1104		printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1105		       gfp_len, "...");
1106	}
1107
1108	printf("%.105s\n", graph_dotted_line);
1109}
1110
1111static void __print_page_caller_result(struct perf_session *session, int n_lines)
1112{
1113	struct rb_node *next = rb_first(&page_caller_sorted);
1114	struct machine *machine = &session->machines.host;
1115	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1116
1117	printf("\n%.105s\n", graph_dotted_line);
1118	printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1119	       live_page ? "Live" : "Total", gfp_len, "GFP flags");
1120	printf("%.105s\n", graph_dotted_line);
1121
1122	while (next && n_lines--) {
1123		struct page_stat *data;
1124		struct symbol *sym;
1125		struct map *map;
1126		char buf[32];
1127		char *caller = buf;
1128
1129		data = rb_entry(next, struct page_stat, node);
1130		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1131		if (sym)
1132			caller = sym->name;
1133		else
1134			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1135
1136		printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1137		       (unsigned long long)data->alloc_bytes / 1024,
1138		       data->nr_alloc, data->order,
1139		       migrate_type_str[data->migrate_type],
1140		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1141
1142		next = rb_next(next);
1143	}
1144
1145	if (n_lines == -1) {
1146		printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1147		       gfp_len, "...");
1148	}
1149
1150	printf("%.105s\n", graph_dotted_line);
1151}
1152
1153static void print_gfp_flags(void)
1154{
1155	int i;
1156
1157	printf("#\n");
1158	printf("# GFP flags\n");
1159	printf("# ---------\n");
1160	for (i = 0; i < nr_gfps; i++) {
1161		printf("# %08x: %*s: %s\n", gfps[i].flags,
1162		       (int) max_gfp_len, gfps[i].compact_str,
1163		       gfps[i].human_readable);
1164	}
1165}
1166
1167static void print_slab_summary(void)
1168{
1169	printf("\nSUMMARY (SLAB allocator)");
1170	printf("\n========================\n");
1171	printf("Total bytes requested: %'lu\n", total_requested);
1172	printf("Total bytes allocated: %'lu\n", total_allocated);
1173	printf("Total bytes freed:     %'lu\n", total_freed);
1174	if (total_allocated > total_freed) {
1175		printf("Net total bytes allocated: %'lu\n",
1176		total_allocated - total_freed);
1177	}
1178	printf("Total bytes wasted on internal fragmentation: %'lu\n",
1179	       total_allocated - total_requested);
1180	printf("Internal fragmentation: %f%%\n",
1181	       fragmentation(total_requested, total_allocated));
1182	printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1183}
1184
1185static void print_page_summary(void)
1186{
1187	int o, m;
1188	u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1189	u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1190
1191	printf("\nSUMMARY (page allocator)");
1192	printf("\n========================\n");
1193	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1194	       nr_page_allocs, total_page_alloc_bytes / 1024);
1195	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1196	       nr_page_frees, total_page_free_bytes / 1024);
1197	printf("\n");
1198
1199	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1200	       nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1201	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1202	       nr_page_allocs - nr_alloc_freed,
1203	       (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1204	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1205	       nr_page_nomatch, total_page_nomatch_bytes / 1024);
1206	printf("\n");
1207
1208	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1209	       nr_page_fails, total_page_fail_bytes / 1024);
1210	printf("\n");
1211
1212	printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1213	       "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1214	printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1215	       graph_dotted_line, graph_dotted_line, graph_dotted_line,
1216	       graph_dotted_line, graph_dotted_line);
1217
1218	for (o = 0; o < MAX_PAGE_ORDER; o++) {
1219		printf("%5d", o);
1220		for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1221			if (order_stats[o][m])
1222				printf("  %'12d", order_stats[o][m]);
1223			else
1224				printf("  %12c", '.');
1225		}
1226		printf("\n");
1227	}
1228}
1229
1230static void print_slab_result(struct perf_session *session)
1231{
1232	if (caller_flag)
1233		__print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1234	if (alloc_flag)
1235		__print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1236	print_slab_summary();
1237}
1238
1239static void print_page_result(struct perf_session *session)
1240{
1241	if (caller_flag || alloc_flag)
1242		print_gfp_flags();
1243	if (caller_flag)
1244		__print_page_caller_result(session, caller_lines);
1245	if (alloc_flag)
1246		__print_page_alloc_result(session, alloc_lines);
1247	print_page_summary();
1248}
1249
1250static void print_result(struct perf_session *session)
1251{
1252	if (kmem_slab)
1253		print_slab_result(session);
1254	if (kmem_page)
1255		print_page_result(session);
1256}
1257
1258static LIST_HEAD(slab_caller_sort);
1259static LIST_HEAD(slab_alloc_sort);
1260static LIST_HEAD(page_caller_sort);
1261static LIST_HEAD(page_alloc_sort);
1262
1263static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1264			     struct list_head *sort_list)
1265{
1266	struct rb_node **new = &(root->rb_node);
1267	struct rb_node *parent = NULL;
1268	struct sort_dimension *sort;
1269
1270	while (*new) {
1271		struct alloc_stat *this;
1272		int cmp = 0;
1273
1274		this = rb_entry(*new, struct alloc_stat, node);
1275		parent = *new;
1276
1277		list_for_each_entry(sort, sort_list, list) {
1278			cmp = sort->cmp(data, this);
1279			if (cmp)
1280				break;
1281		}
1282
1283		if (cmp > 0)
1284			new = &((*new)->rb_left);
1285		else
1286			new = &((*new)->rb_right);
1287	}
1288
1289	rb_link_node(&data->node, parent, new);
1290	rb_insert_color(&data->node, root);
1291}
1292
1293static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1294			       struct list_head *sort_list)
1295{
1296	struct rb_node *node;
1297	struct alloc_stat *data;
1298
1299	for (;;) {
1300		node = rb_first(root);
1301		if (!node)
1302			break;
1303
1304		rb_erase(node, root);
1305		data = rb_entry(node, struct alloc_stat, node);
1306		sort_slab_insert(root_sorted, data, sort_list);
1307	}
1308}
1309
1310static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1311			     struct list_head *sort_list)
1312{
1313	struct rb_node **new = &root->rb_node;
1314	struct rb_node *parent = NULL;
1315	struct sort_dimension *sort;
1316
1317	while (*new) {
1318		struct page_stat *this;
1319		int cmp = 0;
1320
1321		this = rb_entry(*new, struct page_stat, node);
1322		parent = *new;
1323
1324		list_for_each_entry(sort, sort_list, list) {
1325			cmp = sort->cmp(data, this);
1326			if (cmp)
1327				break;
1328		}
1329
1330		if (cmp > 0)
1331			new = &parent->rb_left;
1332		else
1333			new = &parent->rb_right;
1334	}
1335
1336	rb_link_node(&data->node, parent, new);
1337	rb_insert_color(&data->node, root);
1338}
1339
1340static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1341			       struct list_head *sort_list)
1342{
1343	struct rb_node *node;
1344	struct page_stat *data;
1345
1346	for (;;) {
1347		node = rb_first(root);
1348		if (!node)
1349			break;
1350
1351		rb_erase(node, root);
1352		data = rb_entry(node, struct page_stat, node);
1353		sort_page_insert(root_sorted, data, sort_list);
1354	}
1355}
1356
1357static void sort_result(void)
1358{
1359	if (kmem_slab) {
1360		__sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1361				   &slab_alloc_sort);
1362		__sort_slab_result(&root_caller_stat, &root_caller_sorted,
1363				   &slab_caller_sort);
1364	}
1365	if (kmem_page) {
1366		if (live_page)
1367			__sort_page_result(&page_live_tree, &page_alloc_sorted,
1368					   &page_alloc_sort);
1369		else
1370			__sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1371					   &page_alloc_sort);
1372
1373		__sort_page_result(&page_caller_tree, &page_caller_sorted,
1374				   &page_caller_sort);
1375	}
1376}
1377
1378static int __cmd_kmem(struct perf_session *session)
1379{
1380	int err = -EINVAL;
1381	struct evsel *evsel;
1382	const struct evsel_str_handler kmem_tracepoints[] = {
1383		/* slab allocator */
1384		{ "kmem:kmalloc",		evsel__process_alloc_event, },
1385		{ "kmem:kmem_cache_alloc",	evsel__process_alloc_event, },
1386		{ "kmem:kmalloc_node",		evsel__process_alloc_event, },
1387		{ "kmem:kmem_cache_alloc_node", evsel__process_alloc_event, },
1388		{ "kmem:kfree",			evsel__process_free_event, },
1389		{ "kmem:kmem_cache_free",	evsel__process_free_event, },
1390		/* page allocator */
1391		{ "kmem:mm_page_alloc",		evsel__process_page_alloc_event, },
1392		{ "kmem:mm_page_free",		evsel__process_page_free_event, },
1393	};
1394
1395	if (!perf_session__has_traces(session, "kmem record"))
1396		goto out;
1397
1398	if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1399		pr_err("Initializing perf session tracepoint handlers failed\n");
1400		goto out;
1401	}
1402
1403	evlist__for_each_entry(session->evlist, evsel) {
1404		if (evsel__name_is(evsel, "kmem:mm_page_alloc") &&
1405		    evsel__field(evsel, "pfn")) {
1406			use_pfn = true;
1407			break;
1408		}
1409	}
1410
1411	setup_pager();
1412	err = perf_session__process_events(session);
1413	if (err != 0) {
1414		pr_err("error during process events: %d\n", err);
1415		goto out;
1416	}
1417	sort_result();
1418	print_result(session);
1419out:
1420	return err;
1421}
1422
1423/* slab sort keys */
1424static int ptr_cmp(void *a, void *b)
1425{
1426	struct alloc_stat *l = a;
1427	struct alloc_stat *r = b;
1428
1429	if (l->ptr < r->ptr)
1430		return -1;
1431	else if (l->ptr > r->ptr)
1432		return 1;
1433	return 0;
1434}
1435
1436static struct sort_dimension ptr_sort_dimension = {
1437	.name	= "ptr",
1438	.cmp	= ptr_cmp,
1439};
1440
1441static int slab_callsite_cmp(void *a, void *b)
1442{
1443	struct alloc_stat *l = a;
1444	struct alloc_stat *r = b;
1445
1446	if (l->call_site < r->call_site)
1447		return -1;
1448	else if (l->call_site > r->call_site)
1449		return 1;
1450	return 0;
1451}
1452
1453static struct sort_dimension callsite_sort_dimension = {
1454	.name	= "callsite",
1455	.cmp	= slab_callsite_cmp,
1456};
1457
1458static int hit_cmp(void *a, void *b)
1459{
1460	struct alloc_stat *l = a;
1461	struct alloc_stat *r = b;
1462
1463	if (l->hit < r->hit)
1464		return -1;
1465	else if (l->hit > r->hit)
1466		return 1;
1467	return 0;
1468}
1469
1470static struct sort_dimension hit_sort_dimension = {
1471	.name	= "hit",
1472	.cmp	= hit_cmp,
1473};
1474
1475static int bytes_cmp(void *a, void *b)
1476{
1477	struct alloc_stat *l = a;
1478	struct alloc_stat *r = b;
1479
1480	if (l->bytes_alloc < r->bytes_alloc)
1481		return -1;
1482	else if (l->bytes_alloc > r->bytes_alloc)
1483		return 1;
1484	return 0;
1485}
1486
1487static struct sort_dimension bytes_sort_dimension = {
1488	.name	= "bytes",
1489	.cmp	= bytes_cmp,
1490};
1491
1492static int frag_cmp(void *a, void *b)
1493{
1494	double x, y;
1495	struct alloc_stat *l = a;
1496	struct alloc_stat *r = b;
1497
1498	x = fragmentation(l->bytes_req, l->bytes_alloc);
1499	y = fragmentation(r->bytes_req, r->bytes_alloc);
1500
1501	if (x < y)
1502		return -1;
1503	else if (x > y)
1504		return 1;
1505	return 0;
1506}
1507
1508static struct sort_dimension frag_sort_dimension = {
1509	.name	= "frag",
1510	.cmp	= frag_cmp,
1511};
1512
1513static int pingpong_cmp(void *a, void *b)
1514{
1515	struct alloc_stat *l = a;
1516	struct alloc_stat *r = b;
1517
1518	if (l->pingpong < r->pingpong)
1519		return -1;
1520	else if (l->pingpong > r->pingpong)
1521		return 1;
1522	return 0;
1523}
1524
1525static struct sort_dimension pingpong_sort_dimension = {
1526	.name	= "pingpong",
1527	.cmp	= pingpong_cmp,
1528};
1529
1530/* page sort keys */
1531static int page_cmp(void *a, void *b)
1532{
1533	struct page_stat *l = a;
1534	struct page_stat *r = b;
1535
1536	if (l->page < r->page)
1537		return -1;
1538	else if (l->page > r->page)
1539		return 1;
1540	return 0;
1541}
1542
1543static struct sort_dimension page_sort_dimension = {
1544	.name	= "page",
1545	.cmp	= page_cmp,
1546};
1547
1548static int page_callsite_cmp(void *a, void *b)
1549{
1550	struct page_stat *l = a;
1551	struct page_stat *r = b;
1552
1553	if (l->callsite < r->callsite)
1554		return -1;
1555	else if (l->callsite > r->callsite)
1556		return 1;
1557	return 0;
1558}
1559
1560static struct sort_dimension page_callsite_sort_dimension = {
1561	.name	= "callsite",
1562	.cmp	= page_callsite_cmp,
1563};
1564
1565static int page_hit_cmp(void *a, void *b)
1566{
1567	struct page_stat *l = a;
1568	struct page_stat *r = b;
1569
1570	if (l->nr_alloc < r->nr_alloc)
1571		return -1;
1572	else if (l->nr_alloc > r->nr_alloc)
1573		return 1;
1574	return 0;
1575}
1576
1577static struct sort_dimension page_hit_sort_dimension = {
1578	.name	= "hit",
1579	.cmp	= page_hit_cmp,
1580};
1581
1582static int page_bytes_cmp(void *a, void *b)
1583{
1584	struct page_stat *l = a;
1585	struct page_stat *r = b;
1586
1587	if (l->alloc_bytes < r->alloc_bytes)
1588		return -1;
1589	else if (l->alloc_bytes > r->alloc_bytes)
1590		return 1;
1591	return 0;
1592}
1593
1594static struct sort_dimension page_bytes_sort_dimension = {
1595	.name	= "bytes",
1596	.cmp	= page_bytes_cmp,
1597};
1598
1599static int page_order_cmp(void *a, void *b)
1600{
1601	struct page_stat *l = a;
1602	struct page_stat *r = b;
1603
1604	if (l->order < r->order)
1605		return -1;
1606	else if (l->order > r->order)
1607		return 1;
1608	return 0;
1609}
1610
1611static struct sort_dimension page_order_sort_dimension = {
1612	.name	= "order",
1613	.cmp	= page_order_cmp,
1614};
1615
1616static int migrate_type_cmp(void *a, void *b)
1617{
1618	struct page_stat *l = a;
1619	struct page_stat *r = b;
1620
1621	/* for internal use to find free'd page */
1622	if (l->migrate_type == -1U)
1623		return 0;
1624
1625	if (l->migrate_type < r->migrate_type)
1626		return -1;
1627	else if (l->migrate_type > r->migrate_type)
1628		return 1;
1629	return 0;
1630}
1631
1632static struct sort_dimension migrate_type_sort_dimension = {
1633	.name	= "migtype",
1634	.cmp	= migrate_type_cmp,
1635};
1636
1637static int gfp_flags_cmp(void *a, void *b)
1638{
1639	struct page_stat *l = a;
1640	struct page_stat *r = b;
1641
1642	/* for internal use to find free'd page */
1643	if (l->gfp_flags == -1U)
1644		return 0;
1645
1646	if (l->gfp_flags < r->gfp_flags)
1647		return -1;
1648	else if (l->gfp_flags > r->gfp_flags)
1649		return 1;
1650	return 0;
1651}
1652
1653static struct sort_dimension gfp_flags_sort_dimension = {
1654	.name	= "gfp",
1655	.cmp	= gfp_flags_cmp,
1656};
1657
1658static struct sort_dimension *slab_sorts[] = {
1659	&ptr_sort_dimension,
1660	&callsite_sort_dimension,
1661	&hit_sort_dimension,
1662	&bytes_sort_dimension,
1663	&frag_sort_dimension,
1664	&pingpong_sort_dimension,
1665};
1666
1667static struct sort_dimension *page_sorts[] = {
1668	&page_sort_dimension,
1669	&page_callsite_sort_dimension,
1670	&page_hit_sort_dimension,
1671	&page_bytes_sort_dimension,
1672	&page_order_sort_dimension,
1673	&migrate_type_sort_dimension,
1674	&gfp_flags_sort_dimension,
1675};
1676
1677static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1678{
1679	struct sort_dimension *sort;
1680	int i;
1681
1682	for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1683		if (!strcmp(slab_sorts[i]->name, tok)) {
1684			sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1685			if (!sort) {
1686				pr_err("%s: memdup failed\n", __func__);
1687				return -1;
1688			}
1689			list_add_tail(&sort->list, list);
1690			return 0;
1691		}
1692	}
1693
1694	return -1;
1695}
1696
1697static int page_sort_dimension__add(const char *tok, struct list_head *list)
1698{
1699	struct sort_dimension *sort;
1700	int i;
1701
1702	for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1703		if (!strcmp(page_sorts[i]->name, tok)) {
1704			sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1705			if (!sort) {
1706				pr_err("%s: memdup failed\n", __func__);
1707				return -1;
1708			}
1709			list_add_tail(&sort->list, list);
1710			return 0;
1711		}
1712	}
1713
1714	return -1;
1715}
1716
1717static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1718{
1719	char *tok;
1720	char *str = strdup(arg);
1721	char *pos = str;
1722
1723	if (!str) {
1724		pr_err("%s: strdup failed\n", __func__);
1725		return -1;
1726	}
1727
1728	while (true) {
1729		tok = strsep(&pos, ",");
1730		if (!tok)
1731			break;
1732		if (slab_sort_dimension__add(tok, sort_list) < 0) {
1733			pr_err("Unknown slab --sort key: '%s'", tok);
1734			free(str);
1735			return -1;
1736		}
1737	}
1738
1739	free(str);
1740	return 0;
1741}
1742
1743static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1744{
1745	char *tok;
1746	char *str = strdup(arg);
1747	char *pos = str;
1748
1749	if (!str) {
1750		pr_err("%s: strdup failed\n", __func__);
1751		return -1;
1752	}
1753
1754	while (true) {
1755		tok = strsep(&pos, ",");
1756		if (!tok)
1757			break;
1758		if (page_sort_dimension__add(tok, sort_list) < 0) {
1759			pr_err("Unknown page --sort key: '%s'", tok);
1760			free(str);
1761			return -1;
1762		}
1763	}
1764
1765	free(str);
1766	return 0;
1767}
1768
1769static int parse_sort_opt(const struct option *opt __maybe_unused,
1770			  const char *arg, int unset __maybe_unused)
1771{
1772	if (!arg)
1773		return -1;
1774
1775	if (kmem_page > kmem_slab ||
1776	    (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1777		if (caller_flag > alloc_flag)
1778			return setup_page_sorting(&page_caller_sort, arg);
1779		else
1780			return setup_page_sorting(&page_alloc_sort, arg);
1781	} else {
1782		if (caller_flag > alloc_flag)
1783			return setup_slab_sorting(&slab_caller_sort, arg);
1784		else
1785			return setup_slab_sorting(&slab_alloc_sort, arg);
1786	}
1787
1788	return 0;
1789}
1790
1791static int parse_caller_opt(const struct option *opt __maybe_unused,
1792			    const char *arg __maybe_unused,
1793			    int unset __maybe_unused)
1794{
1795	caller_flag = (alloc_flag + 1);
1796	return 0;
1797}
1798
1799static int parse_alloc_opt(const struct option *opt __maybe_unused,
1800			   const char *arg __maybe_unused,
1801			   int unset __maybe_unused)
1802{
1803	alloc_flag = (caller_flag + 1);
1804	return 0;
1805}
1806
1807static int parse_slab_opt(const struct option *opt __maybe_unused,
1808			  const char *arg __maybe_unused,
1809			  int unset __maybe_unused)
1810{
1811	kmem_slab = (kmem_page + 1);
1812	return 0;
1813}
1814
1815static int parse_page_opt(const struct option *opt __maybe_unused,
1816			  const char *arg __maybe_unused,
1817			  int unset __maybe_unused)
1818{
1819	kmem_page = (kmem_slab + 1);
1820	return 0;
1821}
1822
1823static int parse_line_opt(const struct option *opt __maybe_unused,
1824			  const char *arg, int unset __maybe_unused)
1825{
1826	int lines;
1827
1828	if (!arg)
1829		return -1;
1830
1831	lines = strtoul(arg, NULL, 10);
1832
1833	if (caller_flag > alloc_flag)
1834		caller_lines = lines;
1835	else
1836		alloc_lines = lines;
1837
1838	return 0;
1839}
1840
1841static bool slab_legacy_tp_is_exposed(void)
1842{
1843	/*
1844	 * The tracepoints "kmem:kmalloc_node" and
1845	 * "kmem:kmem_cache_alloc_node" have been removed on the latest
1846	 * kernel, if the tracepoint "kmem:kmalloc_node" is existed it
1847	 * means the tool is running on an old kernel, we need to
1848	 * rollback to support these legacy tracepoints.
1849	 */
1850	return IS_ERR(trace_event__tp_format("kmem", "kmalloc_node")) ?
1851		false : true;
1852}
1853
1854static int __cmd_record(int argc, const char **argv)
1855{
1856	const char * const record_args[] = {
1857	"record", "-a", "-R", "-c", "1",
1858	};
1859	const char * const slab_events[] = {
1860	"-e", "kmem:kmalloc",
1861	"-e", "kmem:kfree",
1862	"-e", "kmem:kmem_cache_alloc",
1863	"-e", "kmem:kmem_cache_free",
1864	};
1865	const char * const slab_legacy_events[] = {
1866	"-e", "kmem:kmalloc_node",
1867	"-e", "kmem:kmem_cache_alloc_node",
1868	};
1869	const char * const page_events[] = {
1870	"-e", "kmem:mm_page_alloc",
1871	"-e", "kmem:mm_page_free",
1872	};
1873	unsigned int rec_argc, i, j;
1874	const char **rec_argv;
1875	unsigned int slab_legacy_tp_exposed = slab_legacy_tp_is_exposed();
1876
1877	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1878	if (kmem_slab) {
1879		rec_argc += ARRAY_SIZE(slab_events);
1880		if (slab_legacy_tp_exposed)
1881			rec_argc += ARRAY_SIZE(slab_legacy_events);
1882	}
1883	if (kmem_page)
1884		rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1885
1886	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1887
1888	if (rec_argv == NULL)
1889		return -ENOMEM;
1890
1891	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1892		rec_argv[i] = strdup(record_args[i]);
1893
1894	if (kmem_slab) {
1895		for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1896			rec_argv[i] = strdup(slab_events[j]);
1897		if (slab_legacy_tp_exposed) {
1898			for (j = 0; j < ARRAY_SIZE(slab_legacy_events); j++, i++)
1899				rec_argv[i] = strdup(slab_legacy_events[j]);
1900		}
1901	}
1902	if (kmem_page) {
1903		rec_argv[i++] = strdup("-g");
1904
1905		for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1906			rec_argv[i] = strdup(page_events[j]);
1907	}
1908
1909	for (j = 1; j < (unsigned int)argc; j++, i++)
1910		rec_argv[i] = argv[j];
1911
1912	return cmd_record(i, rec_argv);
1913}
1914
1915static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1916{
1917	if (!strcmp(var, "kmem.default")) {
1918		if (!strcmp(value, "slab"))
1919			kmem_default = KMEM_SLAB;
1920		else if (!strcmp(value, "page"))
1921			kmem_default = KMEM_PAGE;
1922		else
1923			pr_err("invalid default value ('slab' or 'page' required): %s\n",
1924			       value);
1925		return 0;
1926	}
1927
1928	return 0;
1929}
1930
1931int cmd_kmem(int argc, const char **argv)
1932{
1933	const char * const default_slab_sort = "frag,hit,bytes";
1934	const char * const default_page_sort = "bytes,hit";
1935	struct perf_data data = {
1936		.mode = PERF_DATA_MODE_READ,
1937	};
1938	const struct option kmem_options[] = {
1939	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1940	OPT_INCR('v', "verbose", &verbose,
1941		    "be more verbose (show symbol address, etc)"),
1942	OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1943			   "show per-callsite statistics", parse_caller_opt),
1944	OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1945			   "show per-allocation statistics", parse_alloc_opt),
1946	OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1947		     "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1948		     "page, order, migtype, gfp", parse_sort_opt),
1949	OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1950	OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1951	OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
1952	OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1953			   parse_slab_opt),
1954	OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1955			   parse_page_opt),
1956	OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1957	OPT_STRING(0, "time", &time_str, "str",
1958		   "Time span of interest (start,stop)"),
1959	OPT_END()
1960	};
1961	const char *const kmem_subcommands[] = { "record", "stat", NULL };
1962	const char *kmem_usage[] = {
1963		NULL,
1964		NULL
1965	};
1966	struct perf_session *session;
1967	struct perf_tool perf_kmem;
1968	static const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1969	int ret = perf_config(kmem_config, NULL);
1970
1971	if (ret)
1972		return ret;
1973
1974	argc = parse_options_subcommand(argc, argv, kmem_options,
1975					kmem_subcommands, kmem_usage,
1976					PARSE_OPT_STOP_AT_NON_OPTION);
1977
1978	if (!argc)
1979		usage_with_options(kmem_usage, kmem_options);
1980
1981	if (kmem_slab == 0 && kmem_page == 0) {
1982		if (kmem_default == KMEM_SLAB)
1983			kmem_slab = 1;
1984		else
1985			kmem_page = 1;
1986	}
1987
1988	if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
1989		symbol__init(NULL);
1990		return __cmd_record(argc, argv);
1991	}
1992
1993	data.path = input_name;
1994
1995	perf_tool__init(&perf_kmem, /*ordered_events=*/true);
1996	perf_kmem.sample	= process_sample_event;
1997	perf_kmem.comm		= perf_event__process_comm;
1998	perf_kmem.mmap		= perf_event__process_mmap;
1999	perf_kmem.mmap2		= perf_event__process_mmap2;
2000	perf_kmem.namespaces	= perf_event__process_namespaces;
2001
2002	kmem_session = session = perf_session__new(&data, &perf_kmem);
2003	if (IS_ERR(session))
2004		return PTR_ERR(session);
2005
2006	ret = -1;
2007
2008	if (kmem_slab) {
2009		if (!evlist__find_tracepoint_by_name(session->evlist, "kmem:kmalloc")) {
2010			pr_err(errmsg, "slab", "slab");
2011			goto out_delete;
2012		}
2013	}
2014
2015	if (kmem_page) {
2016		struct evsel *evsel = evlist__find_tracepoint_by_name(session->evlist, "kmem:mm_page_alloc");
2017		const struct tep_event *tp_format = evsel ? evsel__tp_format(evsel) : NULL;
2018
2019		if (tp_format == NULL) {
2020			pr_err(errmsg, "page", "page");
2021			goto out_delete;
2022		}
2023		kmem_page_size = tep_get_page_size(tp_format->tep);
2024		symbol_conf.use_callchain = true;
2025	}
2026
2027	symbol__init(perf_session__env(session));
2028
2029	if (perf_time__parse_str(&ptime, time_str) != 0) {
2030		pr_err("Invalid time string\n");
2031		ret = -EINVAL;
2032		goto out_delete;
2033	}
2034
2035	if (!strcmp(argv[0], "stat")) {
2036		setlocale(LC_ALL, "");
2037
2038		if (cpu__setup_cpunode_map())
2039			goto out_delete;
2040
2041		if (list_empty(&slab_caller_sort))
2042			setup_slab_sorting(&slab_caller_sort, default_slab_sort);
2043		if (list_empty(&slab_alloc_sort))
2044			setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
2045		if (list_empty(&page_caller_sort))
2046			setup_page_sorting(&page_caller_sort, default_page_sort);
2047		if (list_empty(&page_alloc_sort))
2048			setup_page_sorting(&page_alloc_sort, default_page_sort);
2049
2050		if (kmem_page) {
2051			setup_page_sorting(&page_alloc_sort_input,
2052					   "page,order,migtype,gfp");
2053			setup_page_sorting(&page_caller_sort_input,
2054					   "callsite,order,migtype,gfp");
2055		}
2056		ret = __cmd_kmem(session);
2057	} else
2058		usage_with_options(kmem_usage, kmem_options);
2059
2060out_delete:
2061	perf_session__delete(session);
2062	/* free usage string allocated by parse_options_subcommand */
2063	free((void *)kmem_usage[0]);
2064
2065	return ret;
2066}