xref: /linux/tools/perf/builtin-kmem.c (revision 06a130e42a5bfc84795464bff023bff4c16f58c5)
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 <traceevent/event-parse.h>
40 
41 static int	kmem_slab;
42 static int	kmem_page;
43 
44 static long	kmem_page_size;
45 static enum {
46 	KMEM_SLAB,
47 	KMEM_PAGE,
48 } kmem_default = KMEM_SLAB;  /* for backward compatibility */
49 
50 struct alloc_stat;
51 typedef int (*sort_fn_t)(void *, void *);
52 
53 static int			alloc_flag;
54 static int			caller_flag;
55 
56 static int			alloc_lines = -1;
57 static int			caller_lines = -1;
58 
59 static bool			raw_ip;
60 
61 struct 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 
75 static struct rb_root root_alloc_stat;
76 static struct rb_root root_alloc_sorted;
77 static struct rb_root root_caller_stat;
78 static struct rb_root root_caller_sorted;
79 
80 static unsigned long total_requested, total_allocated, total_freed;
81 static unsigned long nr_allocs, nr_cross_allocs;
82 
83 /* filters for controlling start and stop of time of analysis */
84 static struct perf_time_interval ptime;
85 const char *time_str;
86 
87 static 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 
132 static 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 
174 static 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 
218 static int ptr_cmp(void *, void *);
219 static int slab_callsite_cmp(void *, void *);
220 
221 static 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 
246 static 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 
272 static u64 total_page_alloc_bytes;
273 static u64 total_page_free_bytes;
274 static u64 total_page_nomatch_bytes;
275 static u64 total_page_fail_bytes;
276 static unsigned long nr_page_allocs;
277 static unsigned long nr_page_frees;
278 static unsigned long nr_page_fails;
279 static unsigned long nr_page_nomatch;
280 
281 static bool use_pfn;
282 static bool live_page;
283 static struct perf_session *kmem_session;
284 
285 #define MAX_MIGRATE_TYPES  6
286 #define MAX_PAGE_ORDER     11
287 
288 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
289 
290 struct 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 
303 static struct rb_root page_live_tree;
304 static struct rb_root page_alloc_tree;
305 static struct rb_root page_alloc_sorted;
306 static struct rb_root page_caller_tree;
307 static struct rb_root page_caller_sorted;
308 
309 struct alloc_func {
310 	u64 start;
311 	u64 end;
312 	char *name;
313 };
314 
315 static int nr_alloc_funcs;
316 static struct alloc_func *alloc_func_list;
317 
318 static 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 
329 static 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 
343 static 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  */
397 static 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);
447 out:
448 	addr_location__exit(&al);
449 	return result;
450 }
451 
452 struct sort_dimension {
453 	const char		name[20];
454 	sort_fn_t		cmp;
455 	struct list_head	list;
456 };
457 
458 static LIST_HEAD(page_alloc_sort_input);
459 static LIST_HEAD(page_caller_sort_input);
460 
461 static 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 
500 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
501 {
502 	return __page_stat__findnew_page(pstat, false);
503 }
504 
505 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
506 {
507 	return __page_stat__findnew_page(pstat, true);
508 }
509 
510 static 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 
555 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
556 {
557 	return __page_stat__findnew_alloc(pstat, false);
558 }
559 
560 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
561 {
562 	return __page_stat__findnew_alloc(pstat, true);
563 }
564 
565 static 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 
610 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
611 {
612 	return __page_stat__findnew_caller(pstat, false);
613 }
614 
615 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
616 {
617 	return __page_stat__findnew_caller(pstat, true);
618 }
619 
620 static 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 
629 struct gfp_flag {
630 	unsigned int flags;
631 	char *compact_str;
632 	char *human_readable;
633 };
634 
635 static struct gfp_flag *gfps;
636 static int nr_gfps;
637 
638 static 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 */
647 static 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 
687 static size_t max_gfp_len;
688 
689 static 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 
740 static 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 
754 static 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 
765 	if (nr_gfps) {
766 		struct gfp_flag key = {
767 			.flags = gfp_flags,
768 		};
769 
770 		if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
771 			return 0;
772 	}
773 
774 	trace_seq_init(&seq);
775 	tep_print_event(evsel->tp_format->tep,
776 			&seq, &record, "%s", TEP_PRINT_INFO);
777 
778 	str = strtok_r(seq.buffer, " ", &pos);
779 	while (str) {
780 		if (!strncmp(str, "gfp_flags=", 10)) {
781 			struct gfp_flag *new;
782 
783 			new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
784 			if (new == NULL)
785 				return -ENOMEM;
786 
787 			gfps = new;
788 			new += nr_gfps++;
789 
790 			new->flags = gfp_flags;
791 			new->human_readable = strdup(str + 10);
792 			new->compact_str = compact_gfp_flags(str + 10);
793 			if (!new->human_readable || !new->compact_str)
794 				return -ENOMEM;
795 
796 			qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
797 		}
798 
799 		str = strtok_r(NULL, " ", &pos);
800 	}
801 
802 	trace_seq_destroy(&seq);
803 	return 0;
804 }
805 
806 static int evsel__process_page_alloc_event(struct evsel *evsel, struct perf_sample *sample)
807 {
808 	u64 page;
809 	unsigned int order = evsel__intval(evsel, sample, "order");
810 	unsigned int gfp_flags = evsel__intval(evsel, sample, "gfp_flags");
811 	unsigned int migrate_type = evsel__intval(evsel, sample,
812 						       "migratetype");
813 	u64 bytes = kmem_page_size << order;
814 	u64 callsite;
815 	struct page_stat *pstat;
816 	struct page_stat this = {
817 		.order = order,
818 		.gfp_flags = gfp_flags,
819 		.migrate_type = migrate_type,
820 	};
821 
822 	if (use_pfn)
823 		page = evsel__intval(evsel, sample, "pfn");
824 	else
825 		page = evsel__intval(evsel, sample, "page");
826 
827 	nr_page_allocs++;
828 	total_page_alloc_bytes += bytes;
829 
830 	if (!valid_page(page)) {
831 		nr_page_fails++;
832 		total_page_fail_bytes += bytes;
833 
834 		return 0;
835 	}
836 
837 	if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
838 		return -1;
839 
840 	callsite = find_callsite(evsel, sample);
841 
842 	/*
843 	 * This is to find the current page (with correct gfp flags and
844 	 * migrate type) at free event.
845 	 */
846 	this.page = page;
847 	pstat = page_stat__findnew_page(&this);
848 	if (pstat == NULL)
849 		return -ENOMEM;
850 
851 	pstat->nr_alloc++;
852 	pstat->alloc_bytes += bytes;
853 	pstat->callsite = callsite;
854 
855 	if (!live_page) {
856 		pstat = page_stat__findnew_alloc(&this);
857 		if (pstat == NULL)
858 			return -ENOMEM;
859 
860 		pstat->nr_alloc++;
861 		pstat->alloc_bytes += bytes;
862 		pstat->callsite = callsite;
863 	}
864 
865 	this.callsite = callsite;
866 	pstat = page_stat__findnew_caller(&this);
867 	if (pstat == NULL)
868 		return -ENOMEM;
869 
870 	pstat->nr_alloc++;
871 	pstat->alloc_bytes += bytes;
872 
873 	order_stats[order][migrate_type]++;
874 
875 	return 0;
876 }
877 
878 static int evsel__process_page_free_event(struct evsel *evsel, struct perf_sample *sample)
879 {
880 	u64 page;
881 	unsigned int order = evsel__intval(evsel, sample, "order");
882 	u64 bytes = kmem_page_size << order;
883 	struct page_stat *pstat;
884 	struct page_stat this = {
885 		.order = order,
886 	};
887 
888 	if (use_pfn)
889 		page = evsel__intval(evsel, sample, "pfn");
890 	else
891 		page = evsel__intval(evsel, sample, "page");
892 
893 	nr_page_frees++;
894 	total_page_free_bytes += bytes;
895 
896 	this.page = page;
897 	pstat = page_stat__find_page(&this);
898 	if (pstat == NULL) {
899 		pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
900 			  page, order);
901 
902 		nr_page_nomatch++;
903 		total_page_nomatch_bytes += bytes;
904 
905 		return 0;
906 	}
907 
908 	this.gfp_flags = pstat->gfp_flags;
909 	this.migrate_type = pstat->migrate_type;
910 	this.callsite = pstat->callsite;
911 
912 	rb_erase(&pstat->node, &page_live_tree);
913 	free(pstat);
914 
915 	if (live_page) {
916 		order_stats[this.order][this.migrate_type]--;
917 	} else {
918 		pstat = page_stat__find_alloc(&this);
919 		if (pstat == NULL)
920 			return -ENOMEM;
921 
922 		pstat->nr_free++;
923 		pstat->free_bytes += bytes;
924 	}
925 
926 	pstat = page_stat__find_caller(&this);
927 	if (pstat == NULL)
928 		return -ENOENT;
929 
930 	pstat->nr_free++;
931 	pstat->free_bytes += bytes;
932 
933 	if (live_page) {
934 		pstat->nr_alloc--;
935 		pstat->alloc_bytes -= bytes;
936 
937 		if (pstat->nr_alloc == 0) {
938 			rb_erase(&pstat->node, &page_caller_tree);
939 			free(pstat);
940 		}
941 	}
942 
943 	return 0;
944 }
945 
946 static bool perf_kmem__skip_sample(struct perf_sample *sample)
947 {
948 	/* skip sample based on time? */
949 	if (perf_time__skip_sample(&ptime, sample->time))
950 		return true;
951 
952 	return false;
953 }
954 
955 typedef int (*tracepoint_handler)(struct evsel *evsel,
956 				  struct perf_sample *sample);
957 
958 static int process_sample_event(const struct perf_tool *tool __maybe_unused,
959 				union perf_event *event,
960 				struct perf_sample *sample,
961 				struct evsel *evsel,
962 				struct machine *machine)
963 {
964 	int err = 0;
965 	struct thread *thread = machine__findnew_thread(machine, sample->pid,
966 							sample->tid);
967 
968 	if (thread == NULL) {
969 		pr_debug("problem processing %d event, skipping it.\n",
970 			 event->header.type);
971 		return -1;
972 	}
973 
974 	if (perf_kmem__skip_sample(sample))
975 		return 0;
976 
977 	dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread__tid(thread));
978 
979 	if (evsel->handler != NULL) {
980 		tracepoint_handler f = evsel->handler;
981 		err = f(evsel, sample);
982 	}
983 
984 	thread__put(thread);
985 
986 	return err;
987 }
988 
989 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
990 {
991 	if (n_alloc == 0)
992 		return 0.0;
993 	else
994 		return 100.0 - (100.0 * n_req / n_alloc);
995 }
996 
997 static void __print_slab_result(struct rb_root *root,
998 				struct perf_session *session,
999 				int n_lines, int is_caller)
1000 {
1001 	struct rb_node *next;
1002 	struct machine *machine = &session->machines.host;
1003 
1004 	printf("%.105s\n", graph_dotted_line);
1005 	printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
1006 	printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
1007 	printf("%.105s\n", graph_dotted_line);
1008 
1009 	next = rb_first(root);
1010 
1011 	while (next && n_lines--) {
1012 		struct alloc_stat *data = rb_entry(next, struct alloc_stat,
1013 						   node);
1014 		struct symbol *sym = NULL;
1015 		struct map *map;
1016 		char buf[BUFSIZ];
1017 		u64 addr;
1018 
1019 		if (is_caller) {
1020 			addr = data->call_site;
1021 			if (!raw_ip)
1022 				sym = machine__find_kernel_symbol(machine, addr, &map);
1023 		} else
1024 			addr = data->ptr;
1025 
1026 		if (sym != NULL)
1027 			snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1028 				 addr - map__unmap_ip(map, sym->start));
1029 		else
1030 			snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1031 		printf(" %-34s |", buf);
1032 
1033 		printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1034 		       (unsigned long long)data->bytes_alloc,
1035 		       (unsigned long)data->bytes_alloc / data->hit,
1036 		       (unsigned long long)data->bytes_req,
1037 		       (unsigned long)data->bytes_req / data->hit,
1038 		       (unsigned long)data->hit,
1039 		       (unsigned long)data->pingpong,
1040 		       fragmentation(data->bytes_req, data->bytes_alloc));
1041 
1042 		next = rb_next(next);
1043 	}
1044 
1045 	if (n_lines == -1)
1046 		printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
1047 
1048 	printf("%.105s\n", graph_dotted_line);
1049 }
1050 
1051 static const char * const migrate_type_str[] = {
1052 	"UNMOVABL",
1053 	"RECLAIM",
1054 	"MOVABLE",
1055 	"RESERVED",
1056 	"CMA/ISLT",
1057 	"UNKNOWN",
1058 };
1059 
1060 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1061 {
1062 	struct rb_node *next = rb_first(&page_alloc_sorted);
1063 	struct machine *machine = &session->machines.host;
1064 	const char *format;
1065 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1066 
1067 	printf("\n%.105s\n", graph_dotted_line);
1068 	printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1069 	       use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1070 	       gfp_len, "GFP flags");
1071 	printf("%.105s\n", graph_dotted_line);
1072 
1073 	if (use_pfn)
1074 		format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1075 	else
1076 		format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1077 
1078 	while (next && n_lines--) {
1079 		struct page_stat *data;
1080 		struct symbol *sym;
1081 		struct map *map;
1082 		char buf[32];
1083 		char *caller = buf;
1084 
1085 		data = rb_entry(next, struct page_stat, node);
1086 		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1087 		if (sym)
1088 			caller = sym->name;
1089 		else
1090 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1091 
1092 		printf(format, (unsigned long long)data->page,
1093 		       (unsigned long long)data->alloc_bytes / 1024,
1094 		       data->nr_alloc, data->order,
1095 		       migrate_type_str[data->migrate_type],
1096 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1097 
1098 		next = rb_next(next);
1099 	}
1100 
1101 	if (n_lines == -1) {
1102 		printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1103 		       gfp_len, "...");
1104 	}
1105 
1106 	printf("%.105s\n", graph_dotted_line);
1107 }
1108 
1109 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1110 {
1111 	struct rb_node *next = rb_first(&page_caller_sorted);
1112 	struct machine *machine = &session->machines.host;
1113 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1114 
1115 	printf("\n%.105s\n", graph_dotted_line);
1116 	printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1117 	       live_page ? "Live" : "Total", gfp_len, "GFP flags");
1118 	printf("%.105s\n", graph_dotted_line);
1119 
1120 	while (next && n_lines--) {
1121 		struct page_stat *data;
1122 		struct symbol *sym;
1123 		struct map *map;
1124 		char buf[32];
1125 		char *caller = buf;
1126 
1127 		data = rb_entry(next, struct page_stat, node);
1128 		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1129 		if (sym)
1130 			caller = sym->name;
1131 		else
1132 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1133 
1134 		printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1135 		       (unsigned long long)data->alloc_bytes / 1024,
1136 		       data->nr_alloc, data->order,
1137 		       migrate_type_str[data->migrate_type],
1138 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1139 
1140 		next = rb_next(next);
1141 	}
1142 
1143 	if (n_lines == -1) {
1144 		printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1145 		       gfp_len, "...");
1146 	}
1147 
1148 	printf("%.105s\n", graph_dotted_line);
1149 }
1150 
1151 static void print_gfp_flags(void)
1152 {
1153 	int i;
1154 
1155 	printf("#\n");
1156 	printf("# GFP flags\n");
1157 	printf("# ---------\n");
1158 	for (i = 0; i < nr_gfps; i++) {
1159 		printf("# %08x: %*s: %s\n", gfps[i].flags,
1160 		       (int) max_gfp_len, gfps[i].compact_str,
1161 		       gfps[i].human_readable);
1162 	}
1163 }
1164 
1165 static void print_slab_summary(void)
1166 {
1167 	printf("\nSUMMARY (SLAB allocator)");
1168 	printf("\n========================\n");
1169 	printf("Total bytes requested: %'lu\n", total_requested);
1170 	printf("Total bytes allocated: %'lu\n", total_allocated);
1171 	printf("Total bytes freed:     %'lu\n", total_freed);
1172 	if (total_allocated > total_freed) {
1173 		printf("Net total bytes allocated: %'lu\n",
1174 		total_allocated - total_freed);
1175 	}
1176 	printf("Total bytes wasted on internal fragmentation: %'lu\n",
1177 	       total_allocated - total_requested);
1178 	printf("Internal fragmentation: %f%%\n",
1179 	       fragmentation(total_requested, total_allocated));
1180 	printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1181 }
1182 
1183 static void print_page_summary(void)
1184 {
1185 	int o, m;
1186 	u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1187 	u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1188 
1189 	printf("\nSUMMARY (page allocator)");
1190 	printf("\n========================\n");
1191 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1192 	       nr_page_allocs, total_page_alloc_bytes / 1024);
1193 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1194 	       nr_page_frees, total_page_free_bytes / 1024);
1195 	printf("\n");
1196 
1197 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1198 	       nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1199 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1200 	       nr_page_allocs - nr_alloc_freed,
1201 	       (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1202 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1203 	       nr_page_nomatch, total_page_nomatch_bytes / 1024);
1204 	printf("\n");
1205 
1206 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1207 	       nr_page_fails, total_page_fail_bytes / 1024);
1208 	printf("\n");
1209 
1210 	printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1211 	       "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1212 	printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1213 	       graph_dotted_line, graph_dotted_line, graph_dotted_line,
1214 	       graph_dotted_line, graph_dotted_line);
1215 
1216 	for (o = 0; o < MAX_PAGE_ORDER; o++) {
1217 		printf("%5d", o);
1218 		for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1219 			if (order_stats[o][m])
1220 				printf("  %'12d", order_stats[o][m]);
1221 			else
1222 				printf("  %12c", '.');
1223 		}
1224 		printf("\n");
1225 	}
1226 }
1227 
1228 static void print_slab_result(struct perf_session *session)
1229 {
1230 	if (caller_flag)
1231 		__print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1232 	if (alloc_flag)
1233 		__print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1234 	print_slab_summary();
1235 }
1236 
1237 static void print_page_result(struct perf_session *session)
1238 {
1239 	if (caller_flag || alloc_flag)
1240 		print_gfp_flags();
1241 	if (caller_flag)
1242 		__print_page_caller_result(session, caller_lines);
1243 	if (alloc_flag)
1244 		__print_page_alloc_result(session, alloc_lines);
1245 	print_page_summary();
1246 }
1247 
1248 static void print_result(struct perf_session *session)
1249 {
1250 	if (kmem_slab)
1251 		print_slab_result(session);
1252 	if (kmem_page)
1253 		print_page_result(session);
1254 }
1255 
1256 static LIST_HEAD(slab_caller_sort);
1257 static LIST_HEAD(slab_alloc_sort);
1258 static LIST_HEAD(page_caller_sort);
1259 static LIST_HEAD(page_alloc_sort);
1260 
1261 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1262 			     struct list_head *sort_list)
1263 {
1264 	struct rb_node **new = &(root->rb_node);
1265 	struct rb_node *parent = NULL;
1266 	struct sort_dimension *sort;
1267 
1268 	while (*new) {
1269 		struct alloc_stat *this;
1270 		int cmp = 0;
1271 
1272 		this = rb_entry(*new, struct alloc_stat, node);
1273 		parent = *new;
1274 
1275 		list_for_each_entry(sort, sort_list, list) {
1276 			cmp = sort->cmp(data, this);
1277 			if (cmp)
1278 				break;
1279 		}
1280 
1281 		if (cmp > 0)
1282 			new = &((*new)->rb_left);
1283 		else
1284 			new = &((*new)->rb_right);
1285 	}
1286 
1287 	rb_link_node(&data->node, parent, new);
1288 	rb_insert_color(&data->node, root);
1289 }
1290 
1291 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1292 			       struct list_head *sort_list)
1293 {
1294 	struct rb_node *node;
1295 	struct alloc_stat *data;
1296 
1297 	for (;;) {
1298 		node = rb_first(root);
1299 		if (!node)
1300 			break;
1301 
1302 		rb_erase(node, root);
1303 		data = rb_entry(node, struct alloc_stat, node);
1304 		sort_slab_insert(root_sorted, data, sort_list);
1305 	}
1306 }
1307 
1308 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1309 			     struct list_head *sort_list)
1310 {
1311 	struct rb_node **new = &root->rb_node;
1312 	struct rb_node *parent = NULL;
1313 	struct sort_dimension *sort;
1314 
1315 	while (*new) {
1316 		struct page_stat *this;
1317 		int cmp = 0;
1318 
1319 		this = rb_entry(*new, struct page_stat, node);
1320 		parent = *new;
1321 
1322 		list_for_each_entry(sort, sort_list, list) {
1323 			cmp = sort->cmp(data, this);
1324 			if (cmp)
1325 				break;
1326 		}
1327 
1328 		if (cmp > 0)
1329 			new = &parent->rb_left;
1330 		else
1331 			new = &parent->rb_right;
1332 	}
1333 
1334 	rb_link_node(&data->node, parent, new);
1335 	rb_insert_color(&data->node, root);
1336 }
1337 
1338 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1339 			       struct list_head *sort_list)
1340 {
1341 	struct rb_node *node;
1342 	struct page_stat *data;
1343 
1344 	for (;;) {
1345 		node = rb_first(root);
1346 		if (!node)
1347 			break;
1348 
1349 		rb_erase(node, root);
1350 		data = rb_entry(node, struct page_stat, node);
1351 		sort_page_insert(root_sorted, data, sort_list);
1352 	}
1353 }
1354 
1355 static void sort_result(void)
1356 {
1357 	if (kmem_slab) {
1358 		__sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1359 				   &slab_alloc_sort);
1360 		__sort_slab_result(&root_caller_stat, &root_caller_sorted,
1361 				   &slab_caller_sort);
1362 	}
1363 	if (kmem_page) {
1364 		if (live_page)
1365 			__sort_page_result(&page_live_tree, &page_alloc_sorted,
1366 					   &page_alloc_sort);
1367 		else
1368 			__sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1369 					   &page_alloc_sort);
1370 
1371 		__sort_page_result(&page_caller_tree, &page_caller_sorted,
1372 				   &page_caller_sort);
1373 	}
1374 }
1375 
1376 static int __cmd_kmem(struct perf_session *session)
1377 {
1378 	int err = -EINVAL;
1379 	struct evsel *evsel;
1380 	const struct evsel_str_handler kmem_tracepoints[] = {
1381 		/* slab allocator */
1382 		{ "kmem:kmalloc",		evsel__process_alloc_event, },
1383 		{ "kmem:kmem_cache_alloc",	evsel__process_alloc_event, },
1384 		{ "kmem:kmalloc_node",		evsel__process_alloc_event, },
1385 		{ "kmem:kmem_cache_alloc_node", evsel__process_alloc_event, },
1386 		{ "kmem:kfree",			evsel__process_free_event, },
1387 		{ "kmem:kmem_cache_free",	evsel__process_free_event, },
1388 		/* page allocator */
1389 		{ "kmem:mm_page_alloc",		evsel__process_page_alloc_event, },
1390 		{ "kmem:mm_page_free",		evsel__process_page_free_event, },
1391 	};
1392 
1393 	if (!perf_session__has_traces(session, "kmem record"))
1394 		goto out;
1395 
1396 	if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1397 		pr_err("Initializing perf session tracepoint handlers failed\n");
1398 		goto out;
1399 	}
1400 
1401 	evlist__for_each_entry(session->evlist, evsel) {
1402 		if (evsel__name_is(evsel, "kmem:mm_page_alloc") &&
1403 		    evsel__field(evsel, "pfn")) {
1404 			use_pfn = true;
1405 			break;
1406 		}
1407 	}
1408 
1409 	setup_pager();
1410 	err = perf_session__process_events(session);
1411 	if (err != 0) {
1412 		pr_err("error during process events: %d\n", err);
1413 		goto out;
1414 	}
1415 	sort_result();
1416 	print_result(session);
1417 out:
1418 	return err;
1419 }
1420 
1421 /* slab sort keys */
1422 static int ptr_cmp(void *a, void *b)
1423 {
1424 	struct alloc_stat *l = a;
1425 	struct alloc_stat *r = b;
1426 
1427 	if (l->ptr < r->ptr)
1428 		return -1;
1429 	else if (l->ptr > r->ptr)
1430 		return 1;
1431 	return 0;
1432 }
1433 
1434 static struct sort_dimension ptr_sort_dimension = {
1435 	.name	= "ptr",
1436 	.cmp	= ptr_cmp,
1437 };
1438 
1439 static int slab_callsite_cmp(void *a, void *b)
1440 {
1441 	struct alloc_stat *l = a;
1442 	struct alloc_stat *r = b;
1443 
1444 	if (l->call_site < r->call_site)
1445 		return -1;
1446 	else if (l->call_site > r->call_site)
1447 		return 1;
1448 	return 0;
1449 }
1450 
1451 static struct sort_dimension callsite_sort_dimension = {
1452 	.name	= "callsite",
1453 	.cmp	= slab_callsite_cmp,
1454 };
1455 
1456 static int hit_cmp(void *a, void *b)
1457 {
1458 	struct alloc_stat *l = a;
1459 	struct alloc_stat *r = b;
1460 
1461 	if (l->hit < r->hit)
1462 		return -1;
1463 	else if (l->hit > r->hit)
1464 		return 1;
1465 	return 0;
1466 }
1467 
1468 static struct sort_dimension hit_sort_dimension = {
1469 	.name	= "hit",
1470 	.cmp	= hit_cmp,
1471 };
1472 
1473 static int bytes_cmp(void *a, void *b)
1474 {
1475 	struct alloc_stat *l = a;
1476 	struct alloc_stat *r = b;
1477 
1478 	if (l->bytes_alloc < r->bytes_alloc)
1479 		return -1;
1480 	else if (l->bytes_alloc > r->bytes_alloc)
1481 		return 1;
1482 	return 0;
1483 }
1484 
1485 static struct sort_dimension bytes_sort_dimension = {
1486 	.name	= "bytes",
1487 	.cmp	= bytes_cmp,
1488 };
1489 
1490 static int frag_cmp(void *a, void *b)
1491 {
1492 	double x, y;
1493 	struct alloc_stat *l = a;
1494 	struct alloc_stat *r = b;
1495 
1496 	x = fragmentation(l->bytes_req, l->bytes_alloc);
1497 	y = fragmentation(r->bytes_req, r->bytes_alloc);
1498 
1499 	if (x < y)
1500 		return -1;
1501 	else if (x > y)
1502 		return 1;
1503 	return 0;
1504 }
1505 
1506 static struct sort_dimension frag_sort_dimension = {
1507 	.name	= "frag",
1508 	.cmp	= frag_cmp,
1509 };
1510 
1511 static int pingpong_cmp(void *a, void *b)
1512 {
1513 	struct alloc_stat *l = a;
1514 	struct alloc_stat *r = b;
1515 
1516 	if (l->pingpong < r->pingpong)
1517 		return -1;
1518 	else if (l->pingpong > r->pingpong)
1519 		return 1;
1520 	return 0;
1521 }
1522 
1523 static struct sort_dimension pingpong_sort_dimension = {
1524 	.name	= "pingpong",
1525 	.cmp	= pingpong_cmp,
1526 };
1527 
1528 /* page sort keys */
1529 static int page_cmp(void *a, void *b)
1530 {
1531 	struct page_stat *l = a;
1532 	struct page_stat *r = b;
1533 
1534 	if (l->page < r->page)
1535 		return -1;
1536 	else if (l->page > r->page)
1537 		return 1;
1538 	return 0;
1539 }
1540 
1541 static struct sort_dimension page_sort_dimension = {
1542 	.name	= "page",
1543 	.cmp	= page_cmp,
1544 };
1545 
1546 static int page_callsite_cmp(void *a, void *b)
1547 {
1548 	struct page_stat *l = a;
1549 	struct page_stat *r = b;
1550 
1551 	if (l->callsite < r->callsite)
1552 		return -1;
1553 	else if (l->callsite > r->callsite)
1554 		return 1;
1555 	return 0;
1556 }
1557 
1558 static struct sort_dimension page_callsite_sort_dimension = {
1559 	.name	= "callsite",
1560 	.cmp	= page_callsite_cmp,
1561 };
1562 
1563 static int page_hit_cmp(void *a, void *b)
1564 {
1565 	struct page_stat *l = a;
1566 	struct page_stat *r = b;
1567 
1568 	if (l->nr_alloc < r->nr_alloc)
1569 		return -1;
1570 	else if (l->nr_alloc > r->nr_alloc)
1571 		return 1;
1572 	return 0;
1573 }
1574 
1575 static struct sort_dimension page_hit_sort_dimension = {
1576 	.name	= "hit",
1577 	.cmp	= page_hit_cmp,
1578 };
1579 
1580 static int page_bytes_cmp(void *a, void *b)
1581 {
1582 	struct page_stat *l = a;
1583 	struct page_stat *r = b;
1584 
1585 	if (l->alloc_bytes < r->alloc_bytes)
1586 		return -1;
1587 	else if (l->alloc_bytes > r->alloc_bytes)
1588 		return 1;
1589 	return 0;
1590 }
1591 
1592 static struct sort_dimension page_bytes_sort_dimension = {
1593 	.name	= "bytes",
1594 	.cmp	= page_bytes_cmp,
1595 };
1596 
1597 static int page_order_cmp(void *a, void *b)
1598 {
1599 	struct page_stat *l = a;
1600 	struct page_stat *r = b;
1601 
1602 	if (l->order < r->order)
1603 		return -1;
1604 	else if (l->order > r->order)
1605 		return 1;
1606 	return 0;
1607 }
1608 
1609 static struct sort_dimension page_order_sort_dimension = {
1610 	.name	= "order",
1611 	.cmp	= page_order_cmp,
1612 };
1613 
1614 static int migrate_type_cmp(void *a, void *b)
1615 {
1616 	struct page_stat *l = a;
1617 	struct page_stat *r = b;
1618 
1619 	/* for internal use to find free'd page */
1620 	if (l->migrate_type == -1U)
1621 		return 0;
1622 
1623 	if (l->migrate_type < r->migrate_type)
1624 		return -1;
1625 	else if (l->migrate_type > r->migrate_type)
1626 		return 1;
1627 	return 0;
1628 }
1629 
1630 static struct sort_dimension migrate_type_sort_dimension = {
1631 	.name	= "migtype",
1632 	.cmp	= migrate_type_cmp,
1633 };
1634 
1635 static int gfp_flags_cmp(void *a, void *b)
1636 {
1637 	struct page_stat *l = a;
1638 	struct page_stat *r = b;
1639 
1640 	/* for internal use to find free'd page */
1641 	if (l->gfp_flags == -1U)
1642 		return 0;
1643 
1644 	if (l->gfp_flags < r->gfp_flags)
1645 		return -1;
1646 	else if (l->gfp_flags > r->gfp_flags)
1647 		return 1;
1648 	return 0;
1649 }
1650 
1651 static struct sort_dimension gfp_flags_sort_dimension = {
1652 	.name	= "gfp",
1653 	.cmp	= gfp_flags_cmp,
1654 };
1655 
1656 static struct sort_dimension *slab_sorts[] = {
1657 	&ptr_sort_dimension,
1658 	&callsite_sort_dimension,
1659 	&hit_sort_dimension,
1660 	&bytes_sort_dimension,
1661 	&frag_sort_dimension,
1662 	&pingpong_sort_dimension,
1663 };
1664 
1665 static struct sort_dimension *page_sorts[] = {
1666 	&page_sort_dimension,
1667 	&page_callsite_sort_dimension,
1668 	&page_hit_sort_dimension,
1669 	&page_bytes_sort_dimension,
1670 	&page_order_sort_dimension,
1671 	&migrate_type_sort_dimension,
1672 	&gfp_flags_sort_dimension,
1673 };
1674 
1675 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1676 {
1677 	struct sort_dimension *sort;
1678 	int i;
1679 
1680 	for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1681 		if (!strcmp(slab_sorts[i]->name, tok)) {
1682 			sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1683 			if (!sort) {
1684 				pr_err("%s: memdup failed\n", __func__);
1685 				return -1;
1686 			}
1687 			list_add_tail(&sort->list, list);
1688 			return 0;
1689 		}
1690 	}
1691 
1692 	return -1;
1693 }
1694 
1695 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1696 {
1697 	struct sort_dimension *sort;
1698 	int i;
1699 
1700 	for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1701 		if (!strcmp(page_sorts[i]->name, tok)) {
1702 			sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1703 			if (!sort) {
1704 				pr_err("%s: memdup failed\n", __func__);
1705 				return -1;
1706 			}
1707 			list_add_tail(&sort->list, list);
1708 			return 0;
1709 		}
1710 	}
1711 
1712 	return -1;
1713 }
1714 
1715 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1716 {
1717 	char *tok;
1718 	char *str = strdup(arg);
1719 	char *pos = str;
1720 
1721 	if (!str) {
1722 		pr_err("%s: strdup failed\n", __func__);
1723 		return -1;
1724 	}
1725 
1726 	while (true) {
1727 		tok = strsep(&pos, ",");
1728 		if (!tok)
1729 			break;
1730 		if (slab_sort_dimension__add(tok, sort_list) < 0) {
1731 			pr_err("Unknown slab --sort key: '%s'", tok);
1732 			free(str);
1733 			return -1;
1734 		}
1735 	}
1736 
1737 	free(str);
1738 	return 0;
1739 }
1740 
1741 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1742 {
1743 	char *tok;
1744 	char *str = strdup(arg);
1745 	char *pos = str;
1746 
1747 	if (!str) {
1748 		pr_err("%s: strdup failed\n", __func__);
1749 		return -1;
1750 	}
1751 
1752 	while (true) {
1753 		tok = strsep(&pos, ",");
1754 		if (!tok)
1755 			break;
1756 		if (page_sort_dimension__add(tok, sort_list) < 0) {
1757 			pr_err("Unknown page --sort key: '%s'", tok);
1758 			free(str);
1759 			return -1;
1760 		}
1761 	}
1762 
1763 	free(str);
1764 	return 0;
1765 }
1766 
1767 static int parse_sort_opt(const struct option *opt __maybe_unused,
1768 			  const char *arg, int unset __maybe_unused)
1769 {
1770 	if (!arg)
1771 		return -1;
1772 
1773 	if (kmem_page > kmem_slab ||
1774 	    (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1775 		if (caller_flag > alloc_flag)
1776 			return setup_page_sorting(&page_caller_sort, arg);
1777 		else
1778 			return setup_page_sorting(&page_alloc_sort, arg);
1779 	} else {
1780 		if (caller_flag > alloc_flag)
1781 			return setup_slab_sorting(&slab_caller_sort, arg);
1782 		else
1783 			return setup_slab_sorting(&slab_alloc_sort, arg);
1784 	}
1785 
1786 	return 0;
1787 }
1788 
1789 static int parse_caller_opt(const struct option *opt __maybe_unused,
1790 			    const char *arg __maybe_unused,
1791 			    int unset __maybe_unused)
1792 {
1793 	caller_flag = (alloc_flag + 1);
1794 	return 0;
1795 }
1796 
1797 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1798 			   const char *arg __maybe_unused,
1799 			   int unset __maybe_unused)
1800 {
1801 	alloc_flag = (caller_flag + 1);
1802 	return 0;
1803 }
1804 
1805 static int parse_slab_opt(const struct option *opt __maybe_unused,
1806 			  const char *arg __maybe_unused,
1807 			  int unset __maybe_unused)
1808 {
1809 	kmem_slab = (kmem_page + 1);
1810 	return 0;
1811 }
1812 
1813 static int parse_page_opt(const struct option *opt __maybe_unused,
1814 			  const char *arg __maybe_unused,
1815 			  int unset __maybe_unused)
1816 {
1817 	kmem_page = (kmem_slab + 1);
1818 	return 0;
1819 }
1820 
1821 static int parse_line_opt(const struct option *opt __maybe_unused,
1822 			  const char *arg, int unset __maybe_unused)
1823 {
1824 	int lines;
1825 
1826 	if (!arg)
1827 		return -1;
1828 
1829 	lines = strtoul(arg, NULL, 10);
1830 
1831 	if (caller_flag > alloc_flag)
1832 		caller_lines = lines;
1833 	else
1834 		alloc_lines = lines;
1835 
1836 	return 0;
1837 }
1838 
1839 static bool slab_legacy_tp_is_exposed(void)
1840 {
1841 	/*
1842 	 * The tracepoints "kmem:kmalloc_node" and
1843 	 * "kmem:kmem_cache_alloc_node" have been removed on the latest
1844 	 * kernel, if the tracepoint "kmem:kmalloc_node" is existed it
1845 	 * means the tool is running on an old kernel, we need to
1846 	 * rollback to support these legacy tracepoints.
1847 	 */
1848 	return IS_ERR(trace_event__tp_format("kmem", "kmalloc_node")) ?
1849 		false : true;
1850 }
1851 
1852 static int __cmd_record(int argc, const char **argv)
1853 {
1854 	const char * const record_args[] = {
1855 	"record", "-a", "-R", "-c", "1",
1856 	};
1857 	const char * const slab_events[] = {
1858 	"-e", "kmem:kmalloc",
1859 	"-e", "kmem:kfree",
1860 	"-e", "kmem:kmem_cache_alloc",
1861 	"-e", "kmem:kmem_cache_free",
1862 	};
1863 	const char * const slab_legacy_events[] = {
1864 	"-e", "kmem:kmalloc_node",
1865 	"-e", "kmem:kmem_cache_alloc_node",
1866 	};
1867 	const char * const page_events[] = {
1868 	"-e", "kmem:mm_page_alloc",
1869 	"-e", "kmem:mm_page_free",
1870 	};
1871 	unsigned int rec_argc, i, j;
1872 	const char **rec_argv;
1873 	unsigned int slab_legacy_tp_exposed = slab_legacy_tp_is_exposed();
1874 
1875 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1876 	if (kmem_slab) {
1877 		rec_argc += ARRAY_SIZE(slab_events);
1878 		if (slab_legacy_tp_exposed)
1879 			rec_argc += ARRAY_SIZE(slab_legacy_events);
1880 	}
1881 	if (kmem_page)
1882 		rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1883 
1884 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1885 
1886 	if (rec_argv == NULL)
1887 		return -ENOMEM;
1888 
1889 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1890 		rec_argv[i] = strdup(record_args[i]);
1891 
1892 	if (kmem_slab) {
1893 		for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1894 			rec_argv[i] = strdup(slab_events[j]);
1895 		if (slab_legacy_tp_exposed) {
1896 			for (j = 0; j < ARRAY_SIZE(slab_legacy_events); j++, i++)
1897 				rec_argv[i] = strdup(slab_legacy_events[j]);
1898 		}
1899 	}
1900 	if (kmem_page) {
1901 		rec_argv[i++] = strdup("-g");
1902 
1903 		for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1904 			rec_argv[i] = strdup(page_events[j]);
1905 	}
1906 
1907 	for (j = 1; j < (unsigned int)argc; j++, i++)
1908 		rec_argv[i] = argv[j];
1909 
1910 	return cmd_record(i, rec_argv);
1911 }
1912 
1913 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1914 {
1915 	if (!strcmp(var, "kmem.default")) {
1916 		if (!strcmp(value, "slab"))
1917 			kmem_default = KMEM_SLAB;
1918 		else if (!strcmp(value, "page"))
1919 			kmem_default = KMEM_PAGE;
1920 		else
1921 			pr_err("invalid default value ('slab' or 'page' required): %s\n",
1922 			       value);
1923 		return 0;
1924 	}
1925 
1926 	return 0;
1927 }
1928 
1929 int cmd_kmem(int argc, const char **argv)
1930 {
1931 	const char * const default_slab_sort = "frag,hit,bytes";
1932 	const char * const default_page_sort = "bytes,hit";
1933 	struct perf_data data = {
1934 		.mode = PERF_DATA_MODE_READ,
1935 	};
1936 	const struct option kmem_options[] = {
1937 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1938 	OPT_INCR('v', "verbose", &verbose,
1939 		    "be more verbose (show symbol address, etc)"),
1940 	OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1941 			   "show per-callsite statistics", parse_caller_opt),
1942 	OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1943 			   "show per-allocation statistics", parse_alloc_opt),
1944 	OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1945 		     "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1946 		     "page, order, migtype, gfp", parse_sort_opt),
1947 	OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1948 	OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1949 	OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
1950 	OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1951 			   parse_slab_opt),
1952 	OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1953 			   parse_page_opt),
1954 	OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1955 	OPT_STRING(0, "time", &time_str, "str",
1956 		   "Time span of interest (start,stop)"),
1957 	OPT_END()
1958 	};
1959 	const char *const kmem_subcommands[] = { "record", "stat", NULL };
1960 	const char *kmem_usage[] = {
1961 		NULL,
1962 		NULL
1963 	};
1964 	struct perf_session *session;
1965 	struct perf_tool perf_kmem;
1966 	static const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1967 	int ret = perf_config(kmem_config, NULL);
1968 
1969 	if (ret)
1970 		return ret;
1971 
1972 	argc = parse_options_subcommand(argc, argv, kmem_options,
1973 					kmem_subcommands, kmem_usage,
1974 					PARSE_OPT_STOP_AT_NON_OPTION);
1975 
1976 	if (!argc)
1977 		usage_with_options(kmem_usage, kmem_options);
1978 
1979 	if (kmem_slab == 0 && kmem_page == 0) {
1980 		if (kmem_default == KMEM_SLAB)
1981 			kmem_slab = 1;
1982 		else
1983 			kmem_page = 1;
1984 	}
1985 
1986 	if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
1987 		symbol__init(NULL);
1988 		return __cmd_record(argc, argv);
1989 	}
1990 
1991 	data.path = input_name;
1992 
1993 	perf_tool__init(&perf_kmem, /*ordered_events=*/true);
1994 	perf_kmem.sample	= process_sample_event;
1995 	perf_kmem.comm		= perf_event__process_comm;
1996 	perf_kmem.mmap		= perf_event__process_mmap;
1997 	perf_kmem.mmap2		= perf_event__process_mmap2;
1998 	perf_kmem.namespaces	= perf_event__process_namespaces;
1999 
2000 	kmem_session = session = perf_session__new(&data, &perf_kmem);
2001 	if (IS_ERR(session))
2002 		return PTR_ERR(session);
2003 
2004 	ret = -1;
2005 
2006 	if (kmem_slab) {
2007 		if (!evlist__find_tracepoint_by_name(session->evlist, "kmem:kmalloc")) {
2008 			pr_err(errmsg, "slab", "slab");
2009 			goto out_delete;
2010 		}
2011 	}
2012 
2013 	if (kmem_page) {
2014 		struct evsel *evsel = evlist__find_tracepoint_by_name(session->evlist, "kmem:mm_page_alloc");
2015 
2016 		if (evsel == NULL) {
2017 			pr_err(errmsg, "page", "page");
2018 			goto out_delete;
2019 		}
2020 
2021 		kmem_page_size = tep_get_page_size(evsel->tp_format->tep);
2022 		symbol_conf.use_callchain = true;
2023 	}
2024 
2025 	symbol__init(&session->header.env);
2026 
2027 	if (perf_time__parse_str(&ptime, time_str) != 0) {
2028 		pr_err("Invalid time string\n");
2029 		ret = -EINVAL;
2030 		goto out_delete;
2031 	}
2032 
2033 	if (!strcmp(argv[0], "stat")) {
2034 		setlocale(LC_ALL, "");
2035 
2036 		if (cpu__setup_cpunode_map())
2037 			goto out_delete;
2038 
2039 		if (list_empty(&slab_caller_sort))
2040 			setup_slab_sorting(&slab_caller_sort, default_slab_sort);
2041 		if (list_empty(&slab_alloc_sort))
2042 			setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
2043 		if (list_empty(&page_caller_sort))
2044 			setup_page_sorting(&page_caller_sort, default_page_sort);
2045 		if (list_empty(&page_alloc_sort))
2046 			setup_page_sorting(&page_alloc_sort, default_page_sort);
2047 
2048 		if (kmem_page) {
2049 			setup_page_sorting(&page_alloc_sort_input,
2050 					   "page,order,migtype,gfp");
2051 			setup_page_sorting(&page_caller_sort_input,
2052 					   "callsite,order,migtype,gfp");
2053 		}
2054 		ret = __cmd_kmem(session);
2055 	} else
2056 		usage_with_options(kmem_usage, kmem_options);
2057 
2058 out_delete:
2059 	perf_session__delete(session);
2060 	/* free usage string allocated by parse_options_subcommand */
2061 	free((void *)kmem_usage[0]);
2062 
2063 	return ret;
2064 }
2065