xref: /linux/tools/perf/builtin-kmem.c (revision 8c749ce93ee69e789e46b3be98de9e0cbfcf8ed8)
1 #include "builtin.h"
2 #include "perf.h"
3 
4 #include "util/evlist.h"
5 #include "util/evsel.h"
6 #include "util/util.h"
7 #include "util/cache.h"
8 #include "util/symbol.h"
9 #include "util/thread.h"
10 #include "util/header.h"
11 #include "util/session.h"
12 #include "util/tool.h"
13 #include "util/callchain.h"
14 
15 #include <subcmd/parse-options.h>
16 #include "util/trace-event.h"
17 #include "util/data.h"
18 #include "util/cpumap.h"
19 
20 #include "util/debug.h"
21 
22 #include <linux/rbtree.h>
23 #include <linux/string.h>
24 #include <locale.h>
25 #include <regex.h>
26 
27 static int	kmem_slab;
28 static int	kmem_page;
29 
30 static long	kmem_page_size;
31 static enum {
32 	KMEM_SLAB,
33 	KMEM_PAGE,
34 } kmem_default = KMEM_SLAB;  /* for backward compatibility */
35 
36 struct alloc_stat;
37 typedef int (*sort_fn_t)(void *, void *);
38 
39 static int			alloc_flag;
40 static int			caller_flag;
41 
42 static int			alloc_lines = -1;
43 static int			caller_lines = -1;
44 
45 static bool			raw_ip;
46 
47 struct alloc_stat {
48 	u64	call_site;
49 	u64	ptr;
50 	u64	bytes_req;
51 	u64	bytes_alloc;
52 	u32	hit;
53 	u32	pingpong;
54 
55 	short	alloc_cpu;
56 
57 	struct rb_node node;
58 };
59 
60 static struct rb_root root_alloc_stat;
61 static struct rb_root root_alloc_sorted;
62 static struct rb_root root_caller_stat;
63 static struct rb_root root_caller_sorted;
64 
65 static unsigned long total_requested, total_allocated;
66 static unsigned long nr_allocs, nr_cross_allocs;
67 
68 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
69 			     int bytes_req, int bytes_alloc, int cpu)
70 {
71 	struct rb_node **node = &root_alloc_stat.rb_node;
72 	struct rb_node *parent = NULL;
73 	struct alloc_stat *data = NULL;
74 
75 	while (*node) {
76 		parent = *node;
77 		data = rb_entry(*node, struct alloc_stat, node);
78 
79 		if (ptr > data->ptr)
80 			node = &(*node)->rb_right;
81 		else if (ptr < data->ptr)
82 			node = &(*node)->rb_left;
83 		else
84 			break;
85 	}
86 
87 	if (data && data->ptr == ptr) {
88 		data->hit++;
89 		data->bytes_req += bytes_req;
90 		data->bytes_alloc += bytes_alloc;
91 	} else {
92 		data = malloc(sizeof(*data));
93 		if (!data) {
94 			pr_err("%s: malloc failed\n", __func__);
95 			return -1;
96 		}
97 		data->ptr = ptr;
98 		data->pingpong = 0;
99 		data->hit = 1;
100 		data->bytes_req = bytes_req;
101 		data->bytes_alloc = bytes_alloc;
102 
103 		rb_link_node(&data->node, parent, node);
104 		rb_insert_color(&data->node, &root_alloc_stat);
105 	}
106 	data->call_site = call_site;
107 	data->alloc_cpu = cpu;
108 	return 0;
109 }
110 
111 static int insert_caller_stat(unsigned long call_site,
112 			      int bytes_req, int bytes_alloc)
113 {
114 	struct rb_node **node = &root_caller_stat.rb_node;
115 	struct rb_node *parent = NULL;
116 	struct alloc_stat *data = NULL;
117 
118 	while (*node) {
119 		parent = *node;
120 		data = rb_entry(*node, struct alloc_stat, node);
121 
122 		if (call_site > data->call_site)
123 			node = &(*node)->rb_right;
124 		else if (call_site < data->call_site)
125 			node = &(*node)->rb_left;
126 		else
127 			break;
128 	}
129 
130 	if (data && data->call_site == call_site) {
131 		data->hit++;
132 		data->bytes_req += bytes_req;
133 		data->bytes_alloc += bytes_alloc;
134 	} else {
135 		data = malloc(sizeof(*data));
136 		if (!data) {
137 			pr_err("%s: malloc failed\n", __func__);
138 			return -1;
139 		}
140 		data->call_site = call_site;
141 		data->pingpong = 0;
142 		data->hit = 1;
143 		data->bytes_req = bytes_req;
144 		data->bytes_alloc = bytes_alloc;
145 
146 		rb_link_node(&data->node, parent, node);
147 		rb_insert_color(&data->node, &root_caller_stat);
148 	}
149 
150 	return 0;
151 }
152 
153 static int perf_evsel__process_alloc_event(struct perf_evsel *evsel,
154 					   struct perf_sample *sample)
155 {
156 	unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
157 		      call_site = perf_evsel__intval(evsel, sample, "call_site");
158 	int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
159 	    bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
160 
161 	if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
162 	    insert_caller_stat(call_site, bytes_req, bytes_alloc))
163 		return -1;
164 
165 	total_requested += bytes_req;
166 	total_allocated += bytes_alloc;
167 
168 	nr_allocs++;
169 	return 0;
170 }
171 
172 static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,
173 						struct perf_sample *sample)
174 {
175 	int ret = perf_evsel__process_alloc_event(evsel, sample);
176 
177 	if (!ret) {
178 		int node1 = cpu__get_node(sample->cpu),
179 		    node2 = perf_evsel__intval(evsel, sample, "node");
180 
181 		if (node1 != node2)
182 			nr_cross_allocs++;
183 	}
184 
185 	return ret;
186 }
187 
188 static int ptr_cmp(void *, void *);
189 static int slab_callsite_cmp(void *, void *);
190 
191 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
192 					    unsigned long call_site,
193 					    struct rb_root *root,
194 					    sort_fn_t sort_fn)
195 {
196 	struct rb_node *node = root->rb_node;
197 	struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
198 
199 	while (node) {
200 		struct alloc_stat *data;
201 		int cmp;
202 
203 		data = rb_entry(node, struct alloc_stat, node);
204 
205 		cmp = sort_fn(&key, data);
206 		if (cmp < 0)
207 			node = node->rb_left;
208 		else if (cmp > 0)
209 			node = node->rb_right;
210 		else
211 			return data;
212 	}
213 	return NULL;
214 }
215 
216 static int perf_evsel__process_free_event(struct perf_evsel *evsel,
217 					  struct perf_sample *sample)
218 {
219 	unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
220 	struct alloc_stat *s_alloc, *s_caller;
221 
222 	s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
223 	if (!s_alloc)
224 		return 0;
225 
226 	if ((short)sample->cpu != s_alloc->alloc_cpu) {
227 		s_alloc->pingpong++;
228 
229 		s_caller = search_alloc_stat(0, s_alloc->call_site,
230 					     &root_caller_stat,
231 					     slab_callsite_cmp);
232 		if (!s_caller)
233 			return -1;
234 		s_caller->pingpong++;
235 	}
236 	s_alloc->alloc_cpu = -1;
237 
238 	return 0;
239 }
240 
241 static u64 total_page_alloc_bytes;
242 static u64 total_page_free_bytes;
243 static u64 total_page_nomatch_bytes;
244 static u64 total_page_fail_bytes;
245 static unsigned long nr_page_allocs;
246 static unsigned long nr_page_frees;
247 static unsigned long nr_page_fails;
248 static unsigned long nr_page_nomatch;
249 
250 static bool use_pfn;
251 static bool live_page;
252 static struct perf_session *kmem_session;
253 
254 #define MAX_MIGRATE_TYPES  6
255 #define MAX_PAGE_ORDER     11
256 
257 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
258 
259 struct page_stat {
260 	struct rb_node 	node;
261 	u64 		page;
262 	u64 		callsite;
263 	int 		order;
264 	unsigned 	gfp_flags;
265 	unsigned 	migrate_type;
266 	u64		alloc_bytes;
267 	u64 		free_bytes;
268 	int 		nr_alloc;
269 	int 		nr_free;
270 };
271 
272 static struct rb_root page_live_tree;
273 static struct rb_root page_alloc_tree;
274 static struct rb_root page_alloc_sorted;
275 static struct rb_root page_caller_tree;
276 static struct rb_root page_caller_sorted;
277 
278 struct alloc_func {
279 	u64 start;
280 	u64 end;
281 	char *name;
282 };
283 
284 static int nr_alloc_funcs;
285 static struct alloc_func *alloc_func_list;
286 
287 static int funcmp(const void *a, const void *b)
288 {
289 	const struct alloc_func *fa = a;
290 	const struct alloc_func *fb = b;
291 
292 	if (fa->start > fb->start)
293 		return 1;
294 	else
295 		return -1;
296 }
297 
298 static int callcmp(const void *a, const void *b)
299 {
300 	const struct alloc_func *fa = a;
301 	const struct alloc_func *fb = b;
302 
303 	if (fb->start <= fa->start && fa->end < fb->end)
304 		return 0;
305 
306 	if (fa->start > fb->start)
307 		return 1;
308 	else
309 		return -1;
310 }
311 
312 static int build_alloc_func_list(void)
313 {
314 	int ret;
315 	struct map *kernel_map;
316 	struct symbol *sym;
317 	struct rb_node *node;
318 	struct alloc_func *func;
319 	struct machine *machine = &kmem_session->machines.host;
320 	regex_t alloc_func_regex;
321 	const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
322 
323 	ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
324 	if (ret) {
325 		char err[BUFSIZ];
326 
327 		regerror(ret, &alloc_func_regex, err, sizeof(err));
328 		pr_err("Invalid regex: %s\n%s", pattern, err);
329 		return -EINVAL;
330 	}
331 
332 	kernel_map = machine__kernel_map(machine);
333 	if (map__load(kernel_map, NULL) < 0) {
334 		pr_err("cannot load kernel map\n");
335 		return -ENOENT;
336 	}
337 
338 	map__for_each_symbol(kernel_map, sym, node) {
339 		if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
340 			continue;
341 
342 		func = realloc(alloc_func_list,
343 			       (nr_alloc_funcs + 1) * sizeof(*func));
344 		if (func == NULL)
345 			return -ENOMEM;
346 
347 		pr_debug("alloc func: %s\n", sym->name);
348 		func[nr_alloc_funcs].start = sym->start;
349 		func[nr_alloc_funcs].end   = sym->end;
350 		func[nr_alloc_funcs].name  = sym->name;
351 
352 		alloc_func_list = func;
353 		nr_alloc_funcs++;
354 	}
355 
356 	qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
357 
358 	regfree(&alloc_func_regex);
359 	return 0;
360 }
361 
362 /*
363  * Find first non-memory allocation function from callchain.
364  * The allocation functions are in the 'alloc_func_list'.
365  */
366 static u64 find_callsite(struct perf_evsel *evsel, struct perf_sample *sample)
367 {
368 	struct addr_location al;
369 	struct machine *machine = &kmem_session->machines.host;
370 	struct callchain_cursor_node *node;
371 
372 	if (alloc_func_list == NULL) {
373 		if (build_alloc_func_list() < 0)
374 			goto out;
375 	}
376 
377 	al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
378 	sample__resolve_callchain(sample, NULL, evsel, &al, 16);
379 
380 	callchain_cursor_commit(&callchain_cursor);
381 	while (true) {
382 		struct alloc_func key, *caller;
383 		u64 addr;
384 
385 		node = callchain_cursor_current(&callchain_cursor);
386 		if (node == NULL)
387 			break;
388 
389 		key.start = key.end = node->ip;
390 		caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
391 				 sizeof(key), callcmp);
392 		if (!caller) {
393 			/* found */
394 			if (node->map)
395 				addr = map__unmap_ip(node->map, node->ip);
396 			else
397 				addr = node->ip;
398 
399 			return addr;
400 		} else
401 			pr_debug3("skipping alloc function: %s\n", caller->name);
402 
403 		callchain_cursor_advance(&callchain_cursor);
404 	}
405 
406 out:
407 	pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
408 	return sample->ip;
409 }
410 
411 struct sort_dimension {
412 	const char		name[20];
413 	sort_fn_t		cmp;
414 	struct list_head	list;
415 };
416 
417 static LIST_HEAD(page_alloc_sort_input);
418 static LIST_HEAD(page_caller_sort_input);
419 
420 static struct page_stat *
421 __page_stat__findnew_page(struct page_stat *pstat, bool create)
422 {
423 	struct rb_node **node = &page_live_tree.rb_node;
424 	struct rb_node *parent = NULL;
425 	struct page_stat *data;
426 
427 	while (*node) {
428 		s64 cmp;
429 
430 		parent = *node;
431 		data = rb_entry(*node, struct page_stat, node);
432 
433 		cmp = data->page - pstat->page;
434 		if (cmp < 0)
435 			node = &parent->rb_left;
436 		else if (cmp > 0)
437 			node = &parent->rb_right;
438 		else
439 			return data;
440 	}
441 
442 	if (!create)
443 		return NULL;
444 
445 	data = zalloc(sizeof(*data));
446 	if (data != NULL) {
447 		data->page = pstat->page;
448 		data->order = pstat->order;
449 		data->gfp_flags = pstat->gfp_flags;
450 		data->migrate_type = pstat->migrate_type;
451 
452 		rb_link_node(&data->node, parent, node);
453 		rb_insert_color(&data->node, &page_live_tree);
454 	}
455 
456 	return data;
457 }
458 
459 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
460 {
461 	return __page_stat__findnew_page(pstat, false);
462 }
463 
464 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
465 {
466 	return __page_stat__findnew_page(pstat, true);
467 }
468 
469 static struct page_stat *
470 __page_stat__findnew_alloc(struct page_stat *pstat, bool create)
471 {
472 	struct rb_node **node = &page_alloc_tree.rb_node;
473 	struct rb_node *parent = NULL;
474 	struct page_stat *data;
475 	struct sort_dimension *sort;
476 
477 	while (*node) {
478 		int cmp = 0;
479 
480 		parent = *node;
481 		data = rb_entry(*node, struct page_stat, node);
482 
483 		list_for_each_entry(sort, &page_alloc_sort_input, list) {
484 			cmp = sort->cmp(pstat, data);
485 			if (cmp)
486 				break;
487 		}
488 
489 		if (cmp < 0)
490 			node = &parent->rb_left;
491 		else if (cmp > 0)
492 			node = &parent->rb_right;
493 		else
494 			return data;
495 	}
496 
497 	if (!create)
498 		return NULL;
499 
500 	data = zalloc(sizeof(*data));
501 	if (data != NULL) {
502 		data->page = pstat->page;
503 		data->order = pstat->order;
504 		data->gfp_flags = pstat->gfp_flags;
505 		data->migrate_type = pstat->migrate_type;
506 
507 		rb_link_node(&data->node, parent, node);
508 		rb_insert_color(&data->node, &page_alloc_tree);
509 	}
510 
511 	return data;
512 }
513 
514 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
515 {
516 	return __page_stat__findnew_alloc(pstat, false);
517 }
518 
519 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
520 {
521 	return __page_stat__findnew_alloc(pstat, true);
522 }
523 
524 static struct page_stat *
525 __page_stat__findnew_caller(struct page_stat *pstat, bool create)
526 {
527 	struct rb_node **node = &page_caller_tree.rb_node;
528 	struct rb_node *parent = NULL;
529 	struct page_stat *data;
530 	struct sort_dimension *sort;
531 
532 	while (*node) {
533 		int cmp = 0;
534 
535 		parent = *node;
536 		data = rb_entry(*node, struct page_stat, node);
537 
538 		list_for_each_entry(sort, &page_caller_sort_input, list) {
539 			cmp = sort->cmp(pstat, data);
540 			if (cmp)
541 				break;
542 		}
543 
544 		if (cmp < 0)
545 			node = &parent->rb_left;
546 		else if (cmp > 0)
547 			node = &parent->rb_right;
548 		else
549 			return data;
550 	}
551 
552 	if (!create)
553 		return NULL;
554 
555 	data = zalloc(sizeof(*data));
556 	if (data != NULL) {
557 		data->callsite = pstat->callsite;
558 		data->order = pstat->order;
559 		data->gfp_flags = pstat->gfp_flags;
560 		data->migrate_type = pstat->migrate_type;
561 
562 		rb_link_node(&data->node, parent, node);
563 		rb_insert_color(&data->node, &page_caller_tree);
564 	}
565 
566 	return data;
567 }
568 
569 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
570 {
571 	return __page_stat__findnew_caller(pstat, false);
572 }
573 
574 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
575 {
576 	return __page_stat__findnew_caller(pstat, true);
577 }
578 
579 static bool valid_page(u64 pfn_or_page)
580 {
581 	if (use_pfn && pfn_or_page == -1UL)
582 		return false;
583 	if (!use_pfn && pfn_or_page == 0)
584 		return false;
585 	return true;
586 }
587 
588 struct gfp_flag {
589 	unsigned int flags;
590 	char *compact_str;
591 	char *human_readable;
592 };
593 
594 static struct gfp_flag *gfps;
595 static int nr_gfps;
596 
597 static int gfpcmp(const void *a, const void *b)
598 {
599 	const struct gfp_flag *fa = a;
600 	const struct gfp_flag *fb = b;
601 
602 	return fa->flags - fb->flags;
603 }
604 
605 /* see include/trace/events/gfpflags.h */
606 static const struct {
607 	const char *original;
608 	const char *compact;
609 } gfp_compact_table[] = {
610 	{ "GFP_TRANSHUGE",		"THP" },
611 	{ "GFP_HIGHUSER_MOVABLE",	"HUM" },
612 	{ "GFP_HIGHUSER",		"HU" },
613 	{ "GFP_USER",			"U" },
614 	{ "GFP_TEMPORARY",		"TMP" },
615 	{ "GFP_KERNEL",			"K" },
616 	{ "GFP_NOFS",			"NF" },
617 	{ "GFP_ATOMIC",			"A" },
618 	{ "GFP_NOIO",			"NI" },
619 	{ "GFP_HIGH",			"H" },
620 	{ "GFP_WAIT",			"W" },
621 	{ "GFP_IO",			"I" },
622 	{ "GFP_COLD",			"CO" },
623 	{ "GFP_NOWARN",			"NWR" },
624 	{ "GFP_REPEAT",			"R" },
625 	{ "GFP_NOFAIL",			"NF" },
626 	{ "GFP_NORETRY",		"NR" },
627 	{ "GFP_COMP",			"C" },
628 	{ "GFP_ZERO",			"Z" },
629 	{ "GFP_NOMEMALLOC",		"NMA" },
630 	{ "GFP_MEMALLOC",		"MA" },
631 	{ "GFP_HARDWALL",		"HW" },
632 	{ "GFP_THISNODE",		"TN" },
633 	{ "GFP_RECLAIMABLE",		"RC" },
634 	{ "GFP_MOVABLE",		"M" },
635 	{ "GFP_NOTRACK",		"NT" },
636 	{ "GFP_NO_KSWAPD",		"NK" },
637 	{ "GFP_OTHER_NODE",		"ON" },
638 	{ "GFP_NOWAIT",			"NW" },
639 };
640 
641 static size_t max_gfp_len;
642 
643 static char *compact_gfp_flags(char *gfp_flags)
644 {
645 	char *orig_flags = strdup(gfp_flags);
646 	char *new_flags = NULL;
647 	char *str, *pos = NULL;
648 	size_t len = 0;
649 
650 	if (orig_flags == NULL)
651 		return NULL;
652 
653 	str = strtok_r(orig_flags, "|", &pos);
654 	while (str) {
655 		size_t i;
656 		char *new;
657 		const char *cpt;
658 
659 		for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
660 			if (strcmp(gfp_compact_table[i].original, str))
661 				continue;
662 
663 			cpt = gfp_compact_table[i].compact;
664 			new = realloc(new_flags, len + strlen(cpt) + 2);
665 			if (new == NULL) {
666 				free(new_flags);
667 				return NULL;
668 			}
669 
670 			new_flags = new;
671 
672 			if (!len) {
673 				strcpy(new_flags, cpt);
674 			} else {
675 				strcat(new_flags, "|");
676 				strcat(new_flags, cpt);
677 				len++;
678 			}
679 
680 			len += strlen(cpt);
681 		}
682 
683 		str = strtok_r(NULL, "|", &pos);
684 	}
685 
686 	if (max_gfp_len < len)
687 		max_gfp_len = len;
688 
689 	free(orig_flags);
690 	return new_flags;
691 }
692 
693 static char *compact_gfp_string(unsigned long gfp_flags)
694 {
695 	struct gfp_flag key = {
696 		.flags = gfp_flags,
697 	};
698 	struct gfp_flag *gfp;
699 
700 	gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
701 	if (gfp)
702 		return gfp->compact_str;
703 
704 	return NULL;
705 }
706 
707 static int parse_gfp_flags(struct perf_evsel *evsel, struct perf_sample *sample,
708 			   unsigned int gfp_flags)
709 {
710 	struct pevent_record record = {
711 		.cpu = sample->cpu,
712 		.data = sample->raw_data,
713 		.size = sample->raw_size,
714 	};
715 	struct trace_seq seq;
716 	char *str, *pos = NULL;
717 
718 	if (nr_gfps) {
719 		struct gfp_flag key = {
720 			.flags = gfp_flags,
721 		};
722 
723 		if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
724 			return 0;
725 	}
726 
727 	trace_seq_init(&seq);
728 	pevent_event_info(&seq, evsel->tp_format, &record);
729 
730 	str = strtok_r(seq.buffer, " ", &pos);
731 	while (str) {
732 		if (!strncmp(str, "gfp_flags=", 10)) {
733 			struct gfp_flag *new;
734 
735 			new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
736 			if (new == NULL)
737 				return -ENOMEM;
738 
739 			gfps = new;
740 			new += nr_gfps++;
741 
742 			new->flags = gfp_flags;
743 			new->human_readable = strdup(str + 10);
744 			new->compact_str = compact_gfp_flags(str + 10);
745 			if (!new->human_readable || !new->compact_str)
746 				return -ENOMEM;
747 
748 			qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
749 		}
750 
751 		str = strtok_r(NULL, " ", &pos);
752 	}
753 
754 	trace_seq_destroy(&seq);
755 	return 0;
756 }
757 
758 static int perf_evsel__process_page_alloc_event(struct perf_evsel *evsel,
759 						struct perf_sample *sample)
760 {
761 	u64 page;
762 	unsigned int order = perf_evsel__intval(evsel, sample, "order");
763 	unsigned int gfp_flags = perf_evsel__intval(evsel, sample, "gfp_flags");
764 	unsigned int migrate_type = perf_evsel__intval(evsel, sample,
765 						       "migratetype");
766 	u64 bytes = kmem_page_size << order;
767 	u64 callsite;
768 	struct page_stat *pstat;
769 	struct page_stat this = {
770 		.order = order,
771 		.gfp_flags = gfp_flags,
772 		.migrate_type = migrate_type,
773 	};
774 
775 	if (use_pfn)
776 		page = perf_evsel__intval(evsel, sample, "pfn");
777 	else
778 		page = perf_evsel__intval(evsel, sample, "page");
779 
780 	nr_page_allocs++;
781 	total_page_alloc_bytes += bytes;
782 
783 	if (!valid_page(page)) {
784 		nr_page_fails++;
785 		total_page_fail_bytes += bytes;
786 
787 		return 0;
788 	}
789 
790 	if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
791 		return -1;
792 
793 	callsite = find_callsite(evsel, sample);
794 
795 	/*
796 	 * This is to find the current page (with correct gfp flags and
797 	 * migrate type) at free event.
798 	 */
799 	this.page = page;
800 	pstat = page_stat__findnew_page(&this);
801 	if (pstat == NULL)
802 		return -ENOMEM;
803 
804 	pstat->nr_alloc++;
805 	pstat->alloc_bytes += bytes;
806 	pstat->callsite = callsite;
807 
808 	if (!live_page) {
809 		pstat = page_stat__findnew_alloc(&this);
810 		if (pstat == NULL)
811 			return -ENOMEM;
812 
813 		pstat->nr_alloc++;
814 		pstat->alloc_bytes += bytes;
815 		pstat->callsite = callsite;
816 	}
817 
818 	this.callsite = callsite;
819 	pstat = page_stat__findnew_caller(&this);
820 	if (pstat == NULL)
821 		return -ENOMEM;
822 
823 	pstat->nr_alloc++;
824 	pstat->alloc_bytes += bytes;
825 
826 	order_stats[order][migrate_type]++;
827 
828 	return 0;
829 }
830 
831 static int perf_evsel__process_page_free_event(struct perf_evsel *evsel,
832 						struct perf_sample *sample)
833 {
834 	u64 page;
835 	unsigned int order = perf_evsel__intval(evsel, sample, "order");
836 	u64 bytes = kmem_page_size << order;
837 	struct page_stat *pstat;
838 	struct page_stat this = {
839 		.order = order,
840 	};
841 
842 	if (use_pfn)
843 		page = perf_evsel__intval(evsel, sample, "pfn");
844 	else
845 		page = perf_evsel__intval(evsel, sample, "page");
846 
847 	nr_page_frees++;
848 	total_page_free_bytes += bytes;
849 
850 	this.page = page;
851 	pstat = page_stat__find_page(&this);
852 	if (pstat == NULL) {
853 		pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
854 			  page, order);
855 
856 		nr_page_nomatch++;
857 		total_page_nomatch_bytes += bytes;
858 
859 		return 0;
860 	}
861 
862 	this.gfp_flags = pstat->gfp_flags;
863 	this.migrate_type = pstat->migrate_type;
864 	this.callsite = pstat->callsite;
865 
866 	rb_erase(&pstat->node, &page_live_tree);
867 	free(pstat);
868 
869 	if (live_page) {
870 		order_stats[this.order][this.migrate_type]--;
871 	} else {
872 		pstat = page_stat__find_alloc(&this);
873 		if (pstat == NULL)
874 			return -ENOMEM;
875 
876 		pstat->nr_free++;
877 		pstat->free_bytes += bytes;
878 	}
879 
880 	pstat = page_stat__find_caller(&this);
881 	if (pstat == NULL)
882 		return -ENOENT;
883 
884 	pstat->nr_free++;
885 	pstat->free_bytes += bytes;
886 
887 	if (live_page) {
888 		pstat->nr_alloc--;
889 		pstat->alloc_bytes -= bytes;
890 
891 		if (pstat->nr_alloc == 0) {
892 			rb_erase(&pstat->node, &page_caller_tree);
893 			free(pstat);
894 		}
895 	}
896 
897 	return 0;
898 }
899 
900 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
901 				  struct perf_sample *sample);
902 
903 static int process_sample_event(struct perf_tool *tool __maybe_unused,
904 				union perf_event *event,
905 				struct perf_sample *sample,
906 				struct perf_evsel *evsel,
907 				struct machine *machine)
908 {
909 	int err = 0;
910 	struct thread *thread = machine__findnew_thread(machine, sample->pid,
911 							sample->tid);
912 
913 	if (thread == NULL) {
914 		pr_debug("problem processing %d event, skipping it.\n",
915 			 event->header.type);
916 		return -1;
917 	}
918 
919 	dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
920 
921 	if (evsel->handler != NULL) {
922 		tracepoint_handler f = evsel->handler;
923 		err = f(evsel, sample);
924 	}
925 
926 	thread__put(thread);
927 
928 	return err;
929 }
930 
931 static struct perf_tool perf_kmem = {
932 	.sample		 = process_sample_event,
933 	.comm		 = perf_event__process_comm,
934 	.mmap		 = perf_event__process_mmap,
935 	.mmap2		 = perf_event__process_mmap2,
936 	.ordered_events	 = true,
937 };
938 
939 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
940 {
941 	if (n_alloc == 0)
942 		return 0.0;
943 	else
944 		return 100.0 - (100.0 * n_req / n_alloc);
945 }
946 
947 static void __print_slab_result(struct rb_root *root,
948 				struct perf_session *session,
949 				int n_lines, int is_caller)
950 {
951 	struct rb_node *next;
952 	struct machine *machine = &session->machines.host;
953 
954 	printf("%.105s\n", graph_dotted_line);
955 	printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
956 	printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
957 	printf("%.105s\n", graph_dotted_line);
958 
959 	next = rb_first(root);
960 
961 	while (next && n_lines--) {
962 		struct alloc_stat *data = rb_entry(next, struct alloc_stat,
963 						   node);
964 		struct symbol *sym = NULL;
965 		struct map *map;
966 		char buf[BUFSIZ];
967 		u64 addr;
968 
969 		if (is_caller) {
970 			addr = data->call_site;
971 			if (!raw_ip)
972 				sym = machine__find_kernel_function(machine, addr, &map, NULL);
973 		} else
974 			addr = data->ptr;
975 
976 		if (sym != NULL)
977 			snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
978 				 addr - map->unmap_ip(map, sym->start));
979 		else
980 			snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
981 		printf(" %-34s |", buf);
982 
983 		printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
984 		       (unsigned long long)data->bytes_alloc,
985 		       (unsigned long)data->bytes_alloc / data->hit,
986 		       (unsigned long long)data->bytes_req,
987 		       (unsigned long)data->bytes_req / data->hit,
988 		       (unsigned long)data->hit,
989 		       (unsigned long)data->pingpong,
990 		       fragmentation(data->bytes_req, data->bytes_alloc));
991 
992 		next = rb_next(next);
993 	}
994 
995 	if (n_lines == -1)
996 		printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
997 
998 	printf("%.105s\n", graph_dotted_line);
999 }
1000 
1001 static const char * const migrate_type_str[] = {
1002 	"UNMOVABL",
1003 	"RECLAIM",
1004 	"MOVABLE",
1005 	"RESERVED",
1006 	"CMA/ISLT",
1007 	"UNKNOWN",
1008 };
1009 
1010 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1011 {
1012 	struct rb_node *next = rb_first(&page_alloc_sorted);
1013 	struct machine *machine = &session->machines.host;
1014 	const char *format;
1015 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1016 
1017 	printf("\n%.105s\n", graph_dotted_line);
1018 	printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1019 	       use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1020 	       gfp_len, "GFP flags");
1021 	printf("%.105s\n", graph_dotted_line);
1022 
1023 	if (use_pfn)
1024 		format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1025 	else
1026 		format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1027 
1028 	while (next && n_lines--) {
1029 		struct page_stat *data;
1030 		struct symbol *sym;
1031 		struct map *map;
1032 		char buf[32];
1033 		char *caller = buf;
1034 
1035 		data = rb_entry(next, struct page_stat, node);
1036 		sym = machine__find_kernel_function(machine, data->callsite,
1037 						    &map, NULL);
1038 		if (sym && sym->name)
1039 			caller = sym->name;
1040 		else
1041 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1042 
1043 		printf(format, (unsigned long long)data->page,
1044 		       (unsigned long long)data->alloc_bytes / 1024,
1045 		       data->nr_alloc, data->order,
1046 		       migrate_type_str[data->migrate_type],
1047 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1048 
1049 		next = rb_next(next);
1050 	}
1051 
1052 	if (n_lines == -1) {
1053 		printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1054 		       gfp_len, "...");
1055 	}
1056 
1057 	printf("%.105s\n", graph_dotted_line);
1058 }
1059 
1060 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1061 {
1062 	struct rb_node *next = rb_first(&page_caller_sorted);
1063 	struct machine *machine = &session->machines.host;
1064 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1065 
1066 	printf("\n%.105s\n", graph_dotted_line);
1067 	printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1068 	       live_page ? "Live" : "Total", gfp_len, "GFP flags");
1069 	printf("%.105s\n", graph_dotted_line);
1070 
1071 	while (next && n_lines--) {
1072 		struct page_stat *data;
1073 		struct symbol *sym;
1074 		struct map *map;
1075 		char buf[32];
1076 		char *caller = buf;
1077 
1078 		data = rb_entry(next, struct page_stat, node);
1079 		sym = machine__find_kernel_function(machine, data->callsite,
1080 						    &map, NULL);
1081 		if (sym && sym->name)
1082 			caller = sym->name;
1083 		else
1084 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1085 
1086 		printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1087 		       (unsigned long long)data->alloc_bytes / 1024,
1088 		       data->nr_alloc, data->order,
1089 		       migrate_type_str[data->migrate_type],
1090 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1091 
1092 		next = rb_next(next);
1093 	}
1094 
1095 	if (n_lines == -1) {
1096 		printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1097 		       gfp_len, "...");
1098 	}
1099 
1100 	printf("%.105s\n", graph_dotted_line);
1101 }
1102 
1103 static void print_gfp_flags(void)
1104 {
1105 	int i;
1106 
1107 	printf("#\n");
1108 	printf("# GFP flags\n");
1109 	printf("# ---------\n");
1110 	for (i = 0; i < nr_gfps; i++) {
1111 		printf("# %08x: %*s: %s\n", gfps[i].flags,
1112 		       (int) max_gfp_len, gfps[i].compact_str,
1113 		       gfps[i].human_readable);
1114 	}
1115 }
1116 
1117 static void print_slab_summary(void)
1118 {
1119 	printf("\nSUMMARY (SLAB allocator)");
1120 	printf("\n========================\n");
1121 	printf("Total bytes requested: %'lu\n", total_requested);
1122 	printf("Total bytes allocated: %'lu\n", total_allocated);
1123 	printf("Total bytes wasted on internal fragmentation: %'lu\n",
1124 	       total_allocated - total_requested);
1125 	printf("Internal fragmentation: %f%%\n",
1126 	       fragmentation(total_requested, total_allocated));
1127 	printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1128 }
1129 
1130 static void print_page_summary(void)
1131 {
1132 	int o, m;
1133 	u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1134 	u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1135 
1136 	printf("\nSUMMARY (page allocator)");
1137 	printf("\n========================\n");
1138 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1139 	       nr_page_allocs, total_page_alloc_bytes / 1024);
1140 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1141 	       nr_page_frees, total_page_free_bytes / 1024);
1142 	printf("\n");
1143 
1144 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1145 	       nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1146 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1147 	       nr_page_allocs - nr_alloc_freed,
1148 	       (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1149 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1150 	       nr_page_nomatch, total_page_nomatch_bytes / 1024);
1151 	printf("\n");
1152 
1153 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1154 	       nr_page_fails, total_page_fail_bytes / 1024);
1155 	printf("\n");
1156 
1157 	printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1158 	       "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1159 	printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1160 	       graph_dotted_line, graph_dotted_line, graph_dotted_line,
1161 	       graph_dotted_line, graph_dotted_line);
1162 
1163 	for (o = 0; o < MAX_PAGE_ORDER; o++) {
1164 		printf("%5d", o);
1165 		for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1166 			if (order_stats[o][m])
1167 				printf("  %'12d", order_stats[o][m]);
1168 			else
1169 				printf("  %12c", '.');
1170 		}
1171 		printf("\n");
1172 	}
1173 }
1174 
1175 static void print_slab_result(struct perf_session *session)
1176 {
1177 	if (caller_flag)
1178 		__print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1179 	if (alloc_flag)
1180 		__print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1181 	print_slab_summary();
1182 }
1183 
1184 static void print_page_result(struct perf_session *session)
1185 {
1186 	if (caller_flag || alloc_flag)
1187 		print_gfp_flags();
1188 	if (caller_flag)
1189 		__print_page_caller_result(session, caller_lines);
1190 	if (alloc_flag)
1191 		__print_page_alloc_result(session, alloc_lines);
1192 	print_page_summary();
1193 }
1194 
1195 static void print_result(struct perf_session *session)
1196 {
1197 	if (kmem_slab)
1198 		print_slab_result(session);
1199 	if (kmem_page)
1200 		print_page_result(session);
1201 }
1202 
1203 static LIST_HEAD(slab_caller_sort);
1204 static LIST_HEAD(slab_alloc_sort);
1205 static LIST_HEAD(page_caller_sort);
1206 static LIST_HEAD(page_alloc_sort);
1207 
1208 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1209 			     struct list_head *sort_list)
1210 {
1211 	struct rb_node **new = &(root->rb_node);
1212 	struct rb_node *parent = NULL;
1213 	struct sort_dimension *sort;
1214 
1215 	while (*new) {
1216 		struct alloc_stat *this;
1217 		int cmp = 0;
1218 
1219 		this = rb_entry(*new, struct alloc_stat, node);
1220 		parent = *new;
1221 
1222 		list_for_each_entry(sort, sort_list, list) {
1223 			cmp = sort->cmp(data, this);
1224 			if (cmp)
1225 				break;
1226 		}
1227 
1228 		if (cmp > 0)
1229 			new = &((*new)->rb_left);
1230 		else
1231 			new = &((*new)->rb_right);
1232 	}
1233 
1234 	rb_link_node(&data->node, parent, new);
1235 	rb_insert_color(&data->node, root);
1236 }
1237 
1238 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1239 			       struct list_head *sort_list)
1240 {
1241 	struct rb_node *node;
1242 	struct alloc_stat *data;
1243 
1244 	for (;;) {
1245 		node = rb_first(root);
1246 		if (!node)
1247 			break;
1248 
1249 		rb_erase(node, root);
1250 		data = rb_entry(node, struct alloc_stat, node);
1251 		sort_slab_insert(root_sorted, data, sort_list);
1252 	}
1253 }
1254 
1255 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1256 			     struct list_head *sort_list)
1257 {
1258 	struct rb_node **new = &root->rb_node;
1259 	struct rb_node *parent = NULL;
1260 	struct sort_dimension *sort;
1261 
1262 	while (*new) {
1263 		struct page_stat *this;
1264 		int cmp = 0;
1265 
1266 		this = rb_entry(*new, struct page_stat, node);
1267 		parent = *new;
1268 
1269 		list_for_each_entry(sort, sort_list, list) {
1270 			cmp = sort->cmp(data, this);
1271 			if (cmp)
1272 				break;
1273 		}
1274 
1275 		if (cmp > 0)
1276 			new = &parent->rb_left;
1277 		else
1278 			new = &parent->rb_right;
1279 	}
1280 
1281 	rb_link_node(&data->node, parent, new);
1282 	rb_insert_color(&data->node, root);
1283 }
1284 
1285 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1286 			       struct list_head *sort_list)
1287 {
1288 	struct rb_node *node;
1289 	struct page_stat *data;
1290 
1291 	for (;;) {
1292 		node = rb_first(root);
1293 		if (!node)
1294 			break;
1295 
1296 		rb_erase(node, root);
1297 		data = rb_entry(node, struct page_stat, node);
1298 		sort_page_insert(root_sorted, data, sort_list);
1299 	}
1300 }
1301 
1302 static void sort_result(void)
1303 {
1304 	if (kmem_slab) {
1305 		__sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1306 				   &slab_alloc_sort);
1307 		__sort_slab_result(&root_caller_stat, &root_caller_sorted,
1308 				   &slab_caller_sort);
1309 	}
1310 	if (kmem_page) {
1311 		if (live_page)
1312 			__sort_page_result(&page_live_tree, &page_alloc_sorted,
1313 					   &page_alloc_sort);
1314 		else
1315 			__sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1316 					   &page_alloc_sort);
1317 
1318 		__sort_page_result(&page_caller_tree, &page_caller_sorted,
1319 				   &page_caller_sort);
1320 	}
1321 }
1322 
1323 static int __cmd_kmem(struct perf_session *session)
1324 {
1325 	int err = -EINVAL;
1326 	struct perf_evsel *evsel;
1327 	const struct perf_evsel_str_handler kmem_tracepoints[] = {
1328 		/* slab allocator */
1329 		{ "kmem:kmalloc",		perf_evsel__process_alloc_event, },
1330     		{ "kmem:kmem_cache_alloc",	perf_evsel__process_alloc_event, },
1331 		{ "kmem:kmalloc_node",		perf_evsel__process_alloc_node_event, },
1332     		{ "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
1333 		{ "kmem:kfree",			perf_evsel__process_free_event, },
1334     		{ "kmem:kmem_cache_free",	perf_evsel__process_free_event, },
1335 		/* page allocator */
1336 		{ "kmem:mm_page_alloc",		perf_evsel__process_page_alloc_event, },
1337 		{ "kmem:mm_page_free",		perf_evsel__process_page_free_event, },
1338 	};
1339 
1340 	if (!perf_session__has_traces(session, "kmem record"))
1341 		goto out;
1342 
1343 	if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1344 		pr_err("Initializing perf session tracepoint handlers failed\n");
1345 		goto out;
1346 	}
1347 
1348 	evlist__for_each(session->evlist, evsel) {
1349 		if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") &&
1350 		    perf_evsel__field(evsel, "pfn")) {
1351 			use_pfn = true;
1352 			break;
1353 		}
1354 	}
1355 
1356 	setup_pager();
1357 	err = perf_session__process_events(session);
1358 	if (err != 0) {
1359 		pr_err("error during process events: %d\n", err);
1360 		goto out;
1361 	}
1362 	sort_result();
1363 	print_result(session);
1364 out:
1365 	return err;
1366 }
1367 
1368 /* slab sort keys */
1369 static int ptr_cmp(void *a, void *b)
1370 {
1371 	struct alloc_stat *l = a;
1372 	struct alloc_stat *r = b;
1373 
1374 	if (l->ptr < r->ptr)
1375 		return -1;
1376 	else if (l->ptr > r->ptr)
1377 		return 1;
1378 	return 0;
1379 }
1380 
1381 static struct sort_dimension ptr_sort_dimension = {
1382 	.name	= "ptr",
1383 	.cmp	= ptr_cmp,
1384 };
1385 
1386 static int slab_callsite_cmp(void *a, void *b)
1387 {
1388 	struct alloc_stat *l = a;
1389 	struct alloc_stat *r = b;
1390 
1391 	if (l->call_site < r->call_site)
1392 		return -1;
1393 	else if (l->call_site > r->call_site)
1394 		return 1;
1395 	return 0;
1396 }
1397 
1398 static struct sort_dimension callsite_sort_dimension = {
1399 	.name	= "callsite",
1400 	.cmp	= slab_callsite_cmp,
1401 };
1402 
1403 static int hit_cmp(void *a, void *b)
1404 {
1405 	struct alloc_stat *l = a;
1406 	struct alloc_stat *r = b;
1407 
1408 	if (l->hit < r->hit)
1409 		return -1;
1410 	else if (l->hit > r->hit)
1411 		return 1;
1412 	return 0;
1413 }
1414 
1415 static struct sort_dimension hit_sort_dimension = {
1416 	.name	= "hit",
1417 	.cmp	= hit_cmp,
1418 };
1419 
1420 static int bytes_cmp(void *a, void *b)
1421 {
1422 	struct alloc_stat *l = a;
1423 	struct alloc_stat *r = b;
1424 
1425 	if (l->bytes_alloc < r->bytes_alloc)
1426 		return -1;
1427 	else if (l->bytes_alloc > r->bytes_alloc)
1428 		return 1;
1429 	return 0;
1430 }
1431 
1432 static struct sort_dimension bytes_sort_dimension = {
1433 	.name	= "bytes",
1434 	.cmp	= bytes_cmp,
1435 };
1436 
1437 static int frag_cmp(void *a, void *b)
1438 {
1439 	double x, y;
1440 	struct alloc_stat *l = a;
1441 	struct alloc_stat *r = b;
1442 
1443 	x = fragmentation(l->bytes_req, l->bytes_alloc);
1444 	y = fragmentation(r->bytes_req, r->bytes_alloc);
1445 
1446 	if (x < y)
1447 		return -1;
1448 	else if (x > y)
1449 		return 1;
1450 	return 0;
1451 }
1452 
1453 static struct sort_dimension frag_sort_dimension = {
1454 	.name	= "frag",
1455 	.cmp	= frag_cmp,
1456 };
1457 
1458 static int pingpong_cmp(void *a, void *b)
1459 {
1460 	struct alloc_stat *l = a;
1461 	struct alloc_stat *r = b;
1462 
1463 	if (l->pingpong < r->pingpong)
1464 		return -1;
1465 	else if (l->pingpong > r->pingpong)
1466 		return 1;
1467 	return 0;
1468 }
1469 
1470 static struct sort_dimension pingpong_sort_dimension = {
1471 	.name	= "pingpong",
1472 	.cmp	= pingpong_cmp,
1473 };
1474 
1475 /* page sort keys */
1476 static int page_cmp(void *a, void *b)
1477 {
1478 	struct page_stat *l = a;
1479 	struct page_stat *r = b;
1480 
1481 	if (l->page < r->page)
1482 		return -1;
1483 	else if (l->page > r->page)
1484 		return 1;
1485 	return 0;
1486 }
1487 
1488 static struct sort_dimension page_sort_dimension = {
1489 	.name	= "page",
1490 	.cmp	= page_cmp,
1491 };
1492 
1493 static int page_callsite_cmp(void *a, void *b)
1494 {
1495 	struct page_stat *l = a;
1496 	struct page_stat *r = b;
1497 
1498 	if (l->callsite < r->callsite)
1499 		return -1;
1500 	else if (l->callsite > r->callsite)
1501 		return 1;
1502 	return 0;
1503 }
1504 
1505 static struct sort_dimension page_callsite_sort_dimension = {
1506 	.name	= "callsite",
1507 	.cmp	= page_callsite_cmp,
1508 };
1509 
1510 static int page_hit_cmp(void *a, void *b)
1511 {
1512 	struct page_stat *l = a;
1513 	struct page_stat *r = b;
1514 
1515 	if (l->nr_alloc < r->nr_alloc)
1516 		return -1;
1517 	else if (l->nr_alloc > r->nr_alloc)
1518 		return 1;
1519 	return 0;
1520 }
1521 
1522 static struct sort_dimension page_hit_sort_dimension = {
1523 	.name	= "hit",
1524 	.cmp	= page_hit_cmp,
1525 };
1526 
1527 static int page_bytes_cmp(void *a, void *b)
1528 {
1529 	struct page_stat *l = a;
1530 	struct page_stat *r = b;
1531 
1532 	if (l->alloc_bytes < r->alloc_bytes)
1533 		return -1;
1534 	else if (l->alloc_bytes > r->alloc_bytes)
1535 		return 1;
1536 	return 0;
1537 }
1538 
1539 static struct sort_dimension page_bytes_sort_dimension = {
1540 	.name	= "bytes",
1541 	.cmp	= page_bytes_cmp,
1542 };
1543 
1544 static int page_order_cmp(void *a, void *b)
1545 {
1546 	struct page_stat *l = a;
1547 	struct page_stat *r = b;
1548 
1549 	if (l->order < r->order)
1550 		return -1;
1551 	else if (l->order > r->order)
1552 		return 1;
1553 	return 0;
1554 }
1555 
1556 static struct sort_dimension page_order_sort_dimension = {
1557 	.name	= "order",
1558 	.cmp	= page_order_cmp,
1559 };
1560 
1561 static int migrate_type_cmp(void *a, void *b)
1562 {
1563 	struct page_stat *l = a;
1564 	struct page_stat *r = b;
1565 
1566 	/* for internal use to find free'd page */
1567 	if (l->migrate_type == -1U)
1568 		return 0;
1569 
1570 	if (l->migrate_type < r->migrate_type)
1571 		return -1;
1572 	else if (l->migrate_type > r->migrate_type)
1573 		return 1;
1574 	return 0;
1575 }
1576 
1577 static struct sort_dimension migrate_type_sort_dimension = {
1578 	.name	= "migtype",
1579 	.cmp	= migrate_type_cmp,
1580 };
1581 
1582 static int gfp_flags_cmp(void *a, void *b)
1583 {
1584 	struct page_stat *l = a;
1585 	struct page_stat *r = b;
1586 
1587 	/* for internal use to find free'd page */
1588 	if (l->gfp_flags == -1U)
1589 		return 0;
1590 
1591 	if (l->gfp_flags < r->gfp_flags)
1592 		return -1;
1593 	else if (l->gfp_flags > r->gfp_flags)
1594 		return 1;
1595 	return 0;
1596 }
1597 
1598 static struct sort_dimension gfp_flags_sort_dimension = {
1599 	.name	= "gfp",
1600 	.cmp	= gfp_flags_cmp,
1601 };
1602 
1603 static struct sort_dimension *slab_sorts[] = {
1604 	&ptr_sort_dimension,
1605 	&callsite_sort_dimension,
1606 	&hit_sort_dimension,
1607 	&bytes_sort_dimension,
1608 	&frag_sort_dimension,
1609 	&pingpong_sort_dimension,
1610 };
1611 
1612 static struct sort_dimension *page_sorts[] = {
1613 	&page_sort_dimension,
1614 	&page_callsite_sort_dimension,
1615 	&page_hit_sort_dimension,
1616 	&page_bytes_sort_dimension,
1617 	&page_order_sort_dimension,
1618 	&migrate_type_sort_dimension,
1619 	&gfp_flags_sort_dimension,
1620 };
1621 
1622 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1623 {
1624 	struct sort_dimension *sort;
1625 	int i;
1626 
1627 	for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1628 		if (!strcmp(slab_sorts[i]->name, tok)) {
1629 			sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1630 			if (!sort) {
1631 				pr_err("%s: memdup failed\n", __func__);
1632 				return -1;
1633 			}
1634 			list_add_tail(&sort->list, list);
1635 			return 0;
1636 		}
1637 	}
1638 
1639 	return -1;
1640 }
1641 
1642 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1643 {
1644 	struct sort_dimension *sort;
1645 	int i;
1646 
1647 	for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1648 		if (!strcmp(page_sorts[i]->name, tok)) {
1649 			sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1650 			if (!sort) {
1651 				pr_err("%s: memdup failed\n", __func__);
1652 				return -1;
1653 			}
1654 			list_add_tail(&sort->list, list);
1655 			return 0;
1656 		}
1657 	}
1658 
1659 	return -1;
1660 }
1661 
1662 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1663 {
1664 	char *tok;
1665 	char *str = strdup(arg);
1666 	char *pos = str;
1667 
1668 	if (!str) {
1669 		pr_err("%s: strdup failed\n", __func__);
1670 		return -1;
1671 	}
1672 
1673 	while (true) {
1674 		tok = strsep(&pos, ",");
1675 		if (!tok)
1676 			break;
1677 		if (slab_sort_dimension__add(tok, sort_list) < 0) {
1678 			error("Unknown slab --sort key: '%s'", tok);
1679 			free(str);
1680 			return -1;
1681 		}
1682 	}
1683 
1684 	free(str);
1685 	return 0;
1686 }
1687 
1688 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1689 {
1690 	char *tok;
1691 	char *str = strdup(arg);
1692 	char *pos = str;
1693 
1694 	if (!str) {
1695 		pr_err("%s: strdup failed\n", __func__);
1696 		return -1;
1697 	}
1698 
1699 	while (true) {
1700 		tok = strsep(&pos, ",");
1701 		if (!tok)
1702 			break;
1703 		if (page_sort_dimension__add(tok, sort_list) < 0) {
1704 			error("Unknown page --sort key: '%s'", tok);
1705 			free(str);
1706 			return -1;
1707 		}
1708 	}
1709 
1710 	free(str);
1711 	return 0;
1712 }
1713 
1714 static int parse_sort_opt(const struct option *opt __maybe_unused,
1715 			  const char *arg, int unset __maybe_unused)
1716 {
1717 	if (!arg)
1718 		return -1;
1719 
1720 	if (kmem_page > kmem_slab ||
1721 	    (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1722 		if (caller_flag > alloc_flag)
1723 			return setup_page_sorting(&page_caller_sort, arg);
1724 		else
1725 			return setup_page_sorting(&page_alloc_sort, arg);
1726 	} else {
1727 		if (caller_flag > alloc_flag)
1728 			return setup_slab_sorting(&slab_caller_sort, arg);
1729 		else
1730 			return setup_slab_sorting(&slab_alloc_sort, arg);
1731 	}
1732 
1733 	return 0;
1734 }
1735 
1736 static int parse_caller_opt(const struct option *opt __maybe_unused,
1737 			    const char *arg __maybe_unused,
1738 			    int unset __maybe_unused)
1739 {
1740 	caller_flag = (alloc_flag + 1);
1741 	return 0;
1742 }
1743 
1744 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1745 			   const char *arg __maybe_unused,
1746 			   int unset __maybe_unused)
1747 {
1748 	alloc_flag = (caller_flag + 1);
1749 	return 0;
1750 }
1751 
1752 static int parse_slab_opt(const struct option *opt __maybe_unused,
1753 			  const char *arg __maybe_unused,
1754 			  int unset __maybe_unused)
1755 {
1756 	kmem_slab = (kmem_page + 1);
1757 	return 0;
1758 }
1759 
1760 static int parse_page_opt(const struct option *opt __maybe_unused,
1761 			  const char *arg __maybe_unused,
1762 			  int unset __maybe_unused)
1763 {
1764 	kmem_page = (kmem_slab + 1);
1765 	return 0;
1766 }
1767 
1768 static int parse_line_opt(const struct option *opt __maybe_unused,
1769 			  const char *arg, int unset __maybe_unused)
1770 {
1771 	int lines;
1772 
1773 	if (!arg)
1774 		return -1;
1775 
1776 	lines = strtoul(arg, NULL, 10);
1777 
1778 	if (caller_flag > alloc_flag)
1779 		caller_lines = lines;
1780 	else
1781 		alloc_lines = lines;
1782 
1783 	return 0;
1784 }
1785 
1786 static int __cmd_record(int argc, const char **argv)
1787 {
1788 	const char * const record_args[] = {
1789 	"record", "-a", "-R", "-c", "1",
1790 	};
1791 	const char * const slab_events[] = {
1792 	"-e", "kmem:kmalloc",
1793 	"-e", "kmem:kmalloc_node",
1794 	"-e", "kmem:kfree",
1795 	"-e", "kmem:kmem_cache_alloc",
1796 	"-e", "kmem:kmem_cache_alloc_node",
1797 	"-e", "kmem:kmem_cache_free",
1798 	};
1799 	const char * const page_events[] = {
1800 	"-e", "kmem:mm_page_alloc",
1801 	"-e", "kmem:mm_page_free",
1802 	};
1803 	unsigned int rec_argc, i, j;
1804 	const char **rec_argv;
1805 
1806 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1807 	if (kmem_slab)
1808 		rec_argc += ARRAY_SIZE(slab_events);
1809 	if (kmem_page)
1810 		rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1811 
1812 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1813 
1814 	if (rec_argv == NULL)
1815 		return -ENOMEM;
1816 
1817 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1818 		rec_argv[i] = strdup(record_args[i]);
1819 
1820 	if (kmem_slab) {
1821 		for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1822 			rec_argv[i] = strdup(slab_events[j]);
1823 	}
1824 	if (kmem_page) {
1825 		rec_argv[i++] = strdup("-g");
1826 
1827 		for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1828 			rec_argv[i] = strdup(page_events[j]);
1829 	}
1830 
1831 	for (j = 1; j < (unsigned int)argc; j++, i++)
1832 		rec_argv[i] = argv[j];
1833 
1834 	return cmd_record(i, rec_argv, NULL);
1835 }
1836 
1837 static int kmem_config(const char *var, const char *value, void *cb)
1838 {
1839 	if (!strcmp(var, "kmem.default")) {
1840 		if (!strcmp(value, "slab"))
1841 			kmem_default = KMEM_SLAB;
1842 		else if (!strcmp(value, "page"))
1843 			kmem_default = KMEM_PAGE;
1844 		else
1845 			pr_err("invalid default value ('slab' or 'page' required): %s\n",
1846 			       value);
1847 		return 0;
1848 	}
1849 
1850 	return perf_default_config(var, value, cb);
1851 }
1852 
1853 int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused)
1854 {
1855 	const char * const default_slab_sort = "frag,hit,bytes";
1856 	const char * const default_page_sort = "bytes,hit";
1857 	struct perf_data_file file = {
1858 		.mode = PERF_DATA_MODE_READ,
1859 	};
1860 	const struct option kmem_options[] = {
1861 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1862 	OPT_INCR('v', "verbose", &verbose,
1863 		    "be more verbose (show symbol address, etc)"),
1864 	OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1865 			   "show per-callsite statistics", parse_caller_opt),
1866 	OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1867 			   "show per-allocation statistics", parse_alloc_opt),
1868 	OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1869 		     "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1870 		     "page, order, migtype, gfp", parse_sort_opt),
1871 	OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1872 	OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1873 	OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"),
1874 	OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1875 			   parse_slab_opt),
1876 	OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1877 			   parse_page_opt),
1878 	OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1879 	OPT_END()
1880 	};
1881 	const char *const kmem_subcommands[] = { "record", "stat", NULL };
1882 	const char *kmem_usage[] = {
1883 		NULL,
1884 		NULL
1885 	};
1886 	struct perf_session *session;
1887 	int ret = -1;
1888 	const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1889 
1890 	perf_config(kmem_config, NULL);
1891 	argc = parse_options_subcommand(argc, argv, kmem_options,
1892 					kmem_subcommands, kmem_usage, 0);
1893 
1894 	if (!argc)
1895 		usage_with_options(kmem_usage, kmem_options);
1896 
1897 	if (kmem_slab == 0 && kmem_page == 0) {
1898 		if (kmem_default == KMEM_SLAB)
1899 			kmem_slab = 1;
1900 		else
1901 			kmem_page = 1;
1902 	}
1903 
1904 	if (!strncmp(argv[0], "rec", 3)) {
1905 		symbol__init(NULL);
1906 		return __cmd_record(argc, argv);
1907 	}
1908 
1909 	file.path = input_name;
1910 
1911 	kmem_session = session = perf_session__new(&file, false, &perf_kmem);
1912 	if (session == NULL)
1913 		return -1;
1914 
1915 	if (kmem_slab) {
1916 		if (!perf_evlist__find_tracepoint_by_name(session->evlist,
1917 							  "kmem:kmalloc")) {
1918 			pr_err(errmsg, "slab", "slab");
1919 			goto out_delete;
1920 		}
1921 	}
1922 
1923 	if (kmem_page) {
1924 		struct perf_evsel *evsel;
1925 
1926 		evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
1927 							     "kmem:mm_page_alloc");
1928 		if (evsel == NULL) {
1929 			pr_err(errmsg, "page", "page");
1930 			goto out_delete;
1931 		}
1932 
1933 		kmem_page_size = pevent_get_page_size(evsel->tp_format->pevent);
1934 		symbol_conf.use_callchain = true;
1935 	}
1936 
1937 	symbol__init(&session->header.env);
1938 
1939 	if (!strcmp(argv[0], "stat")) {
1940 		setlocale(LC_ALL, "");
1941 
1942 		if (cpu__setup_cpunode_map())
1943 			goto out_delete;
1944 
1945 		if (list_empty(&slab_caller_sort))
1946 			setup_slab_sorting(&slab_caller_sort, default_slab_sort);
1947 		if (list_empty(&slab_alloc_sort))
1948 			setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
1949 		if (list_empty(&page_caller_sort))
1950 			setup_page_sorting(&page_caller_sort, default_page_sort);
1951 		if (list_empty(&page_alloc_sort))
1952 			setup_page_sorting(&page_alloc_sort, default_page_sort);
1953 
1954 		if (kmem_page) {
1955 			setup_page_sorting(&page_alloc_sort_input,
1956 					   "page,order,migtype,gfp");
1957 			setup_page_sorting(&page_caller_sort_input,
1958 					   "callsite,order,migtype,gfp");
1959 		}
1960 		ret = __cmd_kmem(session);
1961 	} else
1962 		usage_with_options(kmem_usage, kmem_options);
1963 
1964 out_delete:
1965 	perf_session__delete(session);
1966 
1967 	return ret;
1968 }
1969 
1970