xref: /linux/tools/perf/builtin-kmem.c (revision 827634added7f38b7d724cab1dccdb2b004c13c3)
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 "util/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->vmlinux_maps[MAP__FUNCTION];
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;
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;
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 	struct thread *thread = machine__findnew_thread(machine, sample->pid,
910 							sample->tid);
911 
912 	if (thread == NULL) {
913 		pr_debug("problem processing %d event, skipping it.\n",
914 			 event->header.type);
915 		return -1;
916 	}
917 
918 	dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
919 
920 	if (evsel->handler != NULL) {
921 		tracepoint_handler f = evsel->handler;
922 		return f(evsel, sample);
923 	}
924 
925 	return 0;
926 }
927 
928 static struct perf_tool perf_kmem = {
929 	.sample		 = process_sample_event,
930 	.comm		 = perf_event__process_comm,
931 	.mmap		 = perf_event__process_mmap,
932 	.mmap2		 = perf_event__process_mmap2,
933 	.ordered_events	 = true,
934 };
935 
936 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
937 {
938 	if (n_alloc == 0)
939 		return 0.0;
940 	else
941 		return 100.0 - (100.0 * n_req / n_alloc);
942 }
943 
944 static void __print_slab_result(struct rb_root *root,
945 				struct perf_session *session,
946 				int n_lines, int is_caller)
947 {
948 	struct rb_node *next;
949 	struct machine *machine = &session->machines.host;
950 
951 	printf("%.105s\n", graph_dotted_line);
952 	printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
953 	printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
954 	printf("%.105s\n", graph_dotted_line);
955 
956 	next = rb_first(root);
957 
958 	while (next && n_lines--) {
959 		struct alloc_stat *data = rb_entry(next, struct alloc_stat,
960 						   node);
961 		struct symbol *sym = NULL;
962 		struct map *map;
963 		char buf[BUFSIZ];
964 		u64 addr;
965 
966 		if (is_caller) {
967 			addr = data->call_site;
968 			if (!raw_ip)
969 				sym = machine__find_kernel_function(machine, addr, &map, NULL);
970 		} else
971 			addr = data->ptr;
972 
973 		if (sym != NULL)
974 			snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
975 				 addr - map->unmap_ip(map, sym->start));
976 		else
977 			snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
978 		printf(" %-34s |", buf);
979 
980 		printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
981 		       (unsigned long long)data->bytes_alloc,
982 		       (unsigned long)data->bytes_alloc / data->hit,
983 		       (unsigned long long)data->bytes_req,
984 		       (unsigned long)data->bytes_req / data->hit,
985 		       (unsigned long)data->hit,
986 		       (unsigned long)data->pingpong,
987 		       fragmentation(data->bytes_req, data->bytes_alloc));
988 
989 		next = rb_next(next);
990 	}
991 
992 	if (n_lines == -1)
993 		printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
994 
995 	printf("%.105s\n", graph_dotted_line);
996 }
997 
998 static const char * const migrate_type_str[] = {
999 	"UNMOVABL",
1000 	"RECLAIM",
1001 	"MOVABLE",
1002 	"RESERVED",
1003 	"CMA/ISLT",
1004 	"UNKNOWN",
1005 };
1006 
1007 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1008 {
1009 	struct rb_node *next = rb_first(&page_alloc_sorted);
1010 	struct machine *machine = &session->machines.host;
1011 	const char *format;
1012 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1013 
1014 	printf("\n%.105s\n", graph_dotted_line);
1015 	printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1016 	       use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1017 	       gfp_len, "GFP flags");
1018 	printf("%.105s\n", graph_dotted_line);
1019 
1020 	if (use_pfn)
1021 		format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1022 	else
1023 		format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1024 
1025 	while (next && n_lines--) {
1026 		struct page_stat *data;
1027 		struct symbol *sym;
1028 		struct map *map;
1029 		char buf[32];
1030 		char *caller = buf;
1031 
1032 		data = rb_entry(next, struct page_stat, node);
1033 		sym = machine__find_kernel_function(machine, data->callsite,
1034 						    &map, NULL);
1035 		if (sym && sym->name)
1036 			caller = sym->name;
1037 		else
1038 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1039 
1040 		printf(format, (unsigned long long)data->page,
1041 		       (unsigned long long)data->alloc_bytes / 1024,
1042 		       data->nr_alloc, data->order,
1043 		       migrate_type_str[data->migrate_type],
1044 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1045 
1046 		next = rb_next(next);
1047 	}
1048 
1049 	if (n_lines == -1) {
1050 		printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1051 		       gfp_len, "...");
1052 	}
1053 
1054 	printf("%.105s\n", graph_dotted_line);
1055 }
1056 
1057 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1058 {
1059 	struct rb_node *next = rb_first(&page_caller_sorted);
1060 	struct machine *machine = &session->machines.host;
1061 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1062 
1063 	printf("\n%.105s\n", graph_dotted_line);
1064 	printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1065 	       live_page ? "Live" : "Total", gfp_len, "GFP flags");
1066 	printf("%.105s\n", graph_dotted_line);
1067 
1068 	while (next && n_lines--) {
1069 		struct page_stat *data;
1070 		struct symbol *sym;
1071 		struct map *map;
1072 		char buf[32];
1073 		char *caller = buf;
1074 
1075 		data = rb_entry(next, struct page_stat, node);
1076 		sym = machine__find_kernel_function(machine, data->callsite,
1077 						    &map, NULL);
1078 		if (sym && sym->name)
1079 			caller = sym->name;
1080 		else
1081 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1082 
1083 		printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1084 		       (unsigned long long)data->alloc_bytes / 1024,
1085 		       data->nr_alloc, data->order,
1086 		       migrate_type_str[data->migrate_type],
1087 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1088 
1089 		next = rb_next(next);
1090 	}
1091 
1092 	if (n_lines == -1) {
1093 		printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1094 		       gfp_len, "...");
1095 	}
1096 
1097 	printf("%.105s\n", graph_dotted_line);
1098 }
1099 
1100 static void print_gfp_flags(void)
1101 {
1102 	int i;
1103 
1104 	printf("#\n");
1105 	printf("# GFP flags\n");
1106 	printf("# ---------\n");
1107 	for (i = 0; i < nr_gfps; i++) {
1108 		printf("# %08x: %*s: %s\n", gfps[i].flags,
1109 		       (int) max_gfp_len, gfps[i].compact_str,
1110 		       gfps[i].human_readable);
1111 	}
1112 }
1113 
1114 static void print_slab_summary(void)
1115 {
1116 	printf("\nSUMMARY (SLAB allocator)");
1117 	printf("\n========================\n");
1118 	printf("Total bytes requested: %'lu\n", total_requested);
1119 	printf("Total bytes allocated: %'lu\n", total_allocated);
1120 	printf("Total bytes wasted on internal fragmentation: %'lu\n",
1121 	       total_allocated - total_requested);
1122 	printf("Internal fragmentation: %f%%\n",
1123 	       fragmentation(total_requested, total_allocated));
1124 	printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1125 }
1126 
1127 static void print_page_summary(void)
1128 {
1129 	int o, m;
1130 	u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1131 	u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1132 
1133 	printf("\nSUMMARY (page allocator)");
1134 	printf("\n========================\n");
1135 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1136 	       nr_page_allocs, total_page_alloc_bytes / 1024);
1137 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1138 	       nr_page_frees, total_page_free_bytes / 1024);
1139 	printf("\n");
1140 
1141 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1142 	       nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1143 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1144 	       nr_page_allocs - nr_alloc_freed,
1145 	       (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1146 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1147 	       nr_page_nomatch, total_page_nomatch_bytes / 1024);
1148 	printf("\n");
1149 
1150 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1151 	       nr_page_fails, total_page_fail_bytes / 1024);
1152 	printf("\n");
1153 
1154 	printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1155 	       "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1156 	printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1157 	       graph_dotted_line, graph_dotted_line, graph_dotted_line,
1158 	       graph_dotted_line, graph_dotted_line);
1159 
1160 	for (o = 0; o < MAX_PAGE_ORDER; o++) {
1161 		printf("%5d", o);
1162 		for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1163 			if (order_stats[o][m])
1164 				printf("  %'12d", order_stats[o][m]);
1165 			else
1166 				printf("  %12c", '.');
1167 		}
1168 		printf("\n");
1169 	}
1170 }
1171 
1172 static void print_slab_result(struct perf_session *session)
1173 {
1174 	if (caller_flag)
1175 		__print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1176 	if (alloc_flag)
1177 		__print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1178 	print_slab_summary();
1179 }
1180 
1181 static void print_page_result(struct perf_session *session)
1182 {
1183 	if (caller_flag || alloc_flag)
1184 		print_gfp_flags();
1185 	if (caller_flag)
1186 		__print_page_caller_result(session, caller_lines);
1187 	if (alloc_flag)
1188 		__print_page_alloc_result(session, alloc_lines);
1189 	print_page_summary();
1190 }
1191 
1192 static void print_result(struct perf_session *session)
1193 {
1194 	if (kmem_slab)
1195 		print_slab_result(session);
1196 	if (kmem_page)
1197 		print_page_result(session);
1198 }
1199 
1200 static LIST_HEAD(slab_caller_sort);
1201 static LIST_HEAD(slab_alloc_sort);
1202 static LIST_HEAD(page_caller_sort);
1203 static LIST_HEAD(page_alloc_sort);
1204 
1205 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1206 			     struct list_head *sort_list)
1207 {
1208 	struct rb_node **new = &(root->rb_node);
1209 	struct rb_node *parent = NULL;
1210 	struct sort_dimension *sort;
1211 
1212 	while (*new) {
1213 		struct alloc_stat *this;
1214 		int cmp = 0;
1215 
1216 		this = rb_entry(*new, struct alloc_stat, node);
1217 		parent = *new;
1218 
1219 		list_for_each_entry(sort, sort_list, list) {
1220 			cmp = sort->cmp(data, this);
1221 			if (cmp)
1222 				break;
1223 		}
1224 
1225 		if (cmp > 0)
1226 			new = &((*new)->rb_left);
1227 		else
1228 			new = &((*new)->rb_right);
1229 	}
1230 
1231 	rb_link_node(&data->node, parent, new);
1232 	rb_insert_color(&data->node, root);
1233 }
1234 
1235 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1236 			       struct list_head *sort_list)
1237 {
1238 	struct rb_node *node;
1239 	struct alloc_stat *data;
1240 
1241 	for (;;) {
1242 		node = rb_first(root);
1243 		if (!node)
1244 			break;
1245 
1246 		rb_erase(node, root);
1247 		data = rb_entry(node, struct alloc_stat, node);
1248 		sort_slab_insert(root_sorted, data, sort_list);
1249 	}
1250 }
1251 
1252 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1253 			     struct list_head *sort_list)
1254 {
1255 	struct rb_node **new = &root->rb_node;
1256 	struct rb_node *parent = NULL;
1257 	struct sort_dimension *sort;
1258 
1259 	while (*new) {
1260 		struct page_stat *this;
1261 		int cmp = 0;
1262 
1263 		this = rb_entry(*new, struct page_stat, node);
1264 		parent = *new;
1265 
1266 		list_for_each_entry(sort, sort_list, list) {
1267 			cmp = sort->cmp(data, this);
1268 			if (cmp)
1269 				break;
1270 		}
1271 
1272 		if (cmp > 0)
1273 			new = &parent->rb_left;
1274 		else
1275 			new = &parent->rb_right;
1276 	}
1277 
1278 	rb_link_node(&data->node, parent, new);
1279 	rb_insert_color(&data->node, root);
1280 }
1281 
1282 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1283 			       struct list_head *sort_list)
1284 {
1285 	struct rb_node *node;
1286 	struct page_stat *data;
1287 
1288 	for (;;) {
1289 		node = rb_first(root);
1290 		if (!node)
1291 			break;
1292 
1293 		rb_erase(node, root);
1294 		data = rb_entry(node, struct page_stat, node);
1295 		sort_page_insert(root_sorted, data, sort_list);
1296 	}
1297 }
1298 
1299 static void sort_result(void)
1300 {
1301 	if (kmem_slab) {
1302 		__sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1303 				   &slab_alloc_sort);
1304 		__sort_slab_result(&root_caller_stat, &root_caller_sorted,
1305 				   &slab_caller_sort);
1306 	}
1307 	if (kmem_page) {
1308 		if (live_page)
1309 			__sort_page_result(&page_live_tree, &page_alloc_sorted,
1310 					   &page_alloc_sort);
1311 		else
1312 			__sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1313 					   &page_alloc_sort);
1314 
1315 		__sort_page_result(&page_caller_tree, &page_caller_sorted,
1316 				   &page_caller_sort);
1317 	}
1318 }
1319 
1320 static int __cmd_kmem(struct perf_session *session)
1321 {
1322 	int err = -EINVAL;
1323 	struct perf_evsel *evsel;
1324 	const struct perf_evsel_str_handler kmem_tracepoints[] = {
1325 		/* slab allocator */
1326 		{ "kmem:kmalloc",		perf_evsel__process_alloc_event, },
1327     		{ "kmem:kmem_cache_alloc",	perf_evsel__process_alloc_event, },
1328 		{ "kmem:kmalloc_node",		perf_evsel__process_alloc_node_event, },
1329     		{ "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
1330 		{ "kmem:kfree",			perf_evsel__process_free_event, },
1331     		{ "kmem:kmem_cache_free",	perf_evsel__process_free_event, },
1332 		/* page allocator */
1333 		{ "kmem:mm_page_alloc",		perf_evsel__process_page_alloc_event, },
1334 		{ "kmem:mm_page_free",		perf_evsel__process_page_free_event, },
1335 	};
1336 
1337 	if (!perf_session__has_traces(session, "kmem record"))
1338 		goto out;
1339 
1340 	if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1341 		pr_err("Initializing perf session tracepoint handlers failed\n");
1342 		goto out;
1343 	}
1344 
1345 	evlist__for_each(session->evlist, evsel) {
1346 		if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") &&
1347 		    perf_evsel__field(evsel, "pfn")) {
1348 			use_pfn = true;
1349 			break;
1350 		}
1351 	}
1352 
1353 	setup_pager();
1354 	err = perf_session__process_events(session);
1355 	if (err != 0) {
1356 		pr_err("error during process events: %d\n", err);
1357 		goto out;
1358 	}
1359 	sort_result();
1360 	print_result(session);
1361 out:
1362 	return err;
1363 }
1364 
1365 /* slab sort keys */
1366 static int ptr_cmp(void *a, void *b)
1367 {
1368 	struct alloc_stat *l = a;
1369 	struct alloc_stat *r = b;
1370 
1371 	if (l->ptr < r->ptr)
1372 		return -1;
1373 	else if (l->ptr > r->ptr)
1374 		return 1;
1375 	return 0;
1376 }
1377 
1378 static struct sort_dimension ptr_sort_dimension = {
1379 	.name	= "ptr",
1380 	.cmp	= ptr_cmp,
1381 };
1382 
1383 static int slab_callsite_cmp(void *a, void *b)
1384 {
1385 	struct alloc_stat *l = a;
1386 	struct alloc_stat *r = b;
1387 
1388 	if (l->call_site < r->call_site)
1389 		return -1;
1390 	else if (l->call_site > r->call_site)
1391 		return 1;
1392 	return 0;
1393 }
1394 
1395 static struct sort_dimension callsite_sort_dimension = {
1396 	.name	= "callsite",
1397 	.cmp	= slab_callsite_cmp,
1398 };
1399 
1400 static int hit_cmp(void *a, void *b)
1401 {
1402 	struct alloc_stat *l = a;
1403 	struct alloc_stat *r = b;
1404 
1405 	if (l->hit < r->hit)
1406 		return -1;
1407 	else if (l->hit > r->hit)
1408 		return 1;
1409 	return 0;
1410 }
1411 
1412 static struct sort_dimension hit_sort_dimension = {
1413 	.name	= "hit",
1414 	.cmp	= hit_cmp,
1415 };
1416 
1417 static int bytes_cmp(void *a, void *b)
1418 {
1419 	struct alloc_stat *l = a;
1420 	struct alloc_stat *r = b;
1421 
1422 	if (l->bytes_alloc < r->bytes_alloc)
1423 		return -1;
1424 	else if (l->bytes_alloc > r->bytes_alloc)
1425 		return 1;
1426 	return 0;
1427 }
1428 
1429 static struct sort_dimension bytes_sort_dimension = {
1430 	.name	= "bytes",
1431 	.cmp	= bytes_cmp,
1432 };
1433 
1434 static int frag_cmp(void *a, void *b)
1435 {
1436 	double x, y;
1437 	struct alloc_stat *l = a;
1438 	struct alloc_stat *r = b;
1439 
1440 	x = fragmentation(l->bytes_req, l->bytes_alloc);
1441 	y = fragmentation(r->bytes_req, r->bytes_alloc);
1442 
1443 	if (x < y)
1444 		return -1;
1445 	else if (x > y)
1446 		return 1;
1447 	return 0;
1448 }
1449 
1450 static struct sort_dimension frag_sort_dimension = {
1451 	.name	= "frag",
1452 	.cmp	= frag_cmp,
1453 };
1454 
1455 static int pingpong_cmp(void *a, void *b)
1456 {
1457 	struct alloc_stat *l = a;
1458 	struct alloc_stat *r = b;
1459 
1460 	if (l->pingpong < r->pingpong)
1461 		return -1;
1462 	else if (l->pingpong > r->pingpong)
1463 		return 1;
1464 	return 0;
1465 }
1466 
1467 static struct sort_dimension pingpong_sort_dimension = {
1468 	.name	= "pingpong",
1469 	.cmp	= pingpong_cmp,
1470 };
1471 
1472 /* page sort keys */
1473 static int page_cmp(void *a, void *b)
1474 {
1475 	struct page_stat *l = a;
1476 	struct page_stat *r = b;
1477 
1478 	if (l->page < r->page)
1479 		return -1;
1480 	else if (l->page > r->page)
1481 		return 1;
1482 	return 0;
1483 }
1484 
1485 static struct sort_dimension page_sort_dimension = {
1486 	.name	= "page",
1487 	.cmp	= page_cmp,
1488 };
1489 
1490 static int page_callsite_cmp(void *a, void *b)
1491 {
1492 	struct page_stat *l = a;
1493 	struct page_stat *r = b;
1494 
1495 	if (l->callsite < r->callsite)
1496 		return -1;
1497 	else if (l->callsite > r->callsite)
1498 		return 1;
1499 	return 0;
1500 }
1501 
1502 static struct sort_dimension page_callsite_sort_dimension = {
1503 	.name	= "callsite",
1504 	.cmp	= page_callsite_cmp,
1505 };
1506 
1507 static int page_hit_cmp(void *a, void *b)
1508 {
1509 	struct page_stat *l = a;
1510 	struct page_stat *r = b;
1511 
1512 	if (l->nr_alloc < r->nr_alloc)
1513 		return -1;
1514 	else if (l->nr_alloc > r->nr_alloc)
1515 		return 1;
1516 	return 0;
1517 }
1518 
1519 static struct sort_dimension page_hit_sort_dimension = {
1520 	.name	= "hit",
1521 	.cmp	= page_hit_cmp,
1522 };
1523 
1524 static int page_bytes_cmp(void *a, void *b)
1525 {
1526 	struct page_stat *l = a;
1527 	struct page_stat *r = b;
1528 
1529 	if (l->alloc_bytes < r->alloc_bytes)
1530 		return -1;
1531 	else if (l->alloc_bytes > r->alloc_bytes)
1532 		return 1;
1533 	return 0;
1534 }
1535 
1536 static struct sort_dimension page_bytes_sort_dimension = {
1537 	.name	= "bytes",
1538 	.cmp	= page_bytes_cmp,
1539 };
1540 
1541 static int page_order_cmp(void *a, void *b)
1542 {
1543 	struct page_stat *l = a;
1544 	struct page_stat *r = b;
1545 
1546 	if (l->order < r->order)
1547 		return -1;
1548 	else if (l->order > r->order)
1549 		return 1;
1550 	return 0;
1551 }
1552 
1553 static struct sort_dimension page_order_sort_dimension = {
1554 	.name	= "order",
1555 	.cmp	= page_order_cmp,
1556 };
1557 
1558 static int migrate_type_cmp(void *a, void *b)
1559 {
1560 	struct page_stat *l = a;
1561 	struct page_stat *r = b;
1562 
1563 	/* for internal use to find free'd page */
1564 	if (l->migrate_type == -1U)
1565 		return 0;
1566 
1567 	if (l->migrate_type < r->migrate_type)
1568 		return -1;
1569 	else if (l->migrate_type > r->migrate_type)
1570 		return 1;
1571 	return 0;
1572 }
1573 
1574 static struct sort_dimension migrate_type_sort_dimension = {
1575 	.name	= "migtype",
1576 	.cmp	= migrate_type_cmp,
1577 };
1578 
1579 static int gfp_flags_cmp(void *a, void *b)
1580 {
1581 	struct page_stat *l = a;
1582 	struct page_stat *r = b;
1583 
1584 	/* for internal use to find free'd page */
1585 	if (l->gfp_flags == -1U)
1586 		return 0;
1587 
1588 	if (l->gfp_flags < r->gfp_flags)
1589 		return -1;
1590 	else if (l->gfp_flags > r->gfp_flags)
1591 		return 1;
1592 	return 0;
1593 }
1594 
1595 static struct sort_dimension gfp_flags_sort_dimension = {
1596 	.name	= "gfp",
1597 	.cmp	= gfp_flags_cmp,
1598 };
1599 
1600 static struct sort_dimension *slab_sorts[] = {
1601 	&ptr_sort_dimension,
1602 	&callsite_sort_dimension,
1603 	&hit_sort_dimension,
1604 	&bytes_sort_dimension,
1605 	&frag_sort_dimension,
1606 	&pingpong_sort_dimension,
1607 };
1608 
1609 static struct sort_dimension *page_sorts[] = {
1610 	&page_sort_dimension,
1611 	&page_callsite_sort_dimension,
1612 	&page_hit_sort_dimension,
1613 	&page_bytes_sort_dimension,
1614 	&page_order_sort_dimension,
1615 	&migrate_type_sort_dimension,
1616 	&gfp_flags_sort_dimension,
1617 };
1618 
1619 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1620 {
1621 	struct sort_dimension *sort;
1622 	int i;
1623 
1624 	for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1625 		if (!strcmp(slab_sorts[i]->name, tok)) {
1626 			sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1627 			if (!sort) {
1628 				pr_err("%s: memdup failed\n", __func__);
1629 				return -1;
1630 			}
1631 			list_add_tail(&sort->list, list);
1632 			return 0;
1633 		}
1634 	}
1635 
1636 	return -1;
1637 }
1638 
1639 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1640 {
1641 	struct sort_dimension *sort;
1642 	int i;
1643 
1644 	for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1645 		if (!strcmp(page_sorts[i]->name, tok)) {
1646 			sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1647 			if (!sort) {
1648 				pr_err("%s: memdup failed\n", __func__);
1649 				return -1;
1650 			}
1651 			list_add_tail(&sort->list, list);
1652 			return 0;
1653 		}
1654 	}
1655 
1656 	return -1;
1657 }
1658 
1659 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1660 {
1661 	char *tok;
1662 	char *str = strdup(arg);
1663 	char *pos = str;
1664 
1665 	if (!str) {
1666 		pr_err("%s: strdup failed\n", __func__);
1667 		return -1;
1668 	}
1669 
1670 	while (true) {
1671 		tok = strsep(&pos, ",");
1672 		if (!tok)
1673 			break;
1674 		if (slab_sort_dimension__add(tok, sort_list) < 0) {
1675 			error("Unknown slab --sort key: '%s'", tok);
1676 			free(str);
1677 			return -1;
1678 		}
1679 	}
1680 
1681 	free(str);
1682 	return 0;
1683 }
1684 
1685 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1686 {
1687 	char *tok;
1688 	char *str = strdup(arg);
1689 	char *pos = str;
1690 
1691 	if (!str) {
1692 		pr_err("%s: strdup failed\n", __func__);
1693 		return -1;
1694 	}
1695 
1696 	while (true) {
1697 		tok = strsep(&pos, ",");
1698 		if (!tok)
1699 			break;
1700 		if (page_sort_dimension__add(tok, sort_list) < 0) {
1701 			error("Unknown page --sort key: '%s'", tok);
1702 			free(str);
1703 			return -1;
1704 		}
1705 	}
1706 
1707 	free(str);
1708 	return 0;
1709 }
1710 
1711 static int parse_sort_opt(const struct option *opt __maybe_unused,
1712 			  const char *arg, int unset __maybe_unused)
1713 {
1714 	if (!arg)
1715 		return -1;
1716 
1717 	if (kmem_page > kmem_slab ||
1718 	    (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1719 		if (caller_flag > alloc_flag)
1720 			return setup_page_sorting(&page_caller_sort, arg);
1721 		else
1722 			return setup_page_sorting(&page_alloc_sort, arg);
1723 	} else {
1724 		if (caller_flag > alloc_flag)
1725 			return setup_slab_sorting(&slab_caller_sort, arg);
1726 		else
1727 			return setup_slab_sorting(&slab_alloc_sort, arg);
1728 	}
1729 
1730 	return 0;
1731 }
1732 
1733 static int parse_caller_opt(const struct option *opt __maybe_unused,
1734 			    const char *arg __maybe_unused,
1735 			    int unset __maybe_unused)
1736 {
1737 	caller_flag = (alloc_flag + 1);
1738 	return 0;
1739 }
1740 
1741 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1742 			   const char *arg __maybe_unused,
1743 			   int unset __maybe_unused)
1744 {
1745 	alloc_flag = (caller_flag + 1);
1746 	return 0;
1747 }
1748 
1749 static int parse_slab_opt(const struct option *opt __maybe_unused,
1750 			  const char *arg __maybe_unused,
1751 			  int unset __maybe_unused)
1752 {
1753 	kmem_slab = (kmem_page + 1);
1754 	return 0;
1755 }
1756 
1757 static int parse_page_opt(const struct option *opt __maybe_unused,
1758 			  const char *arg __maybe_unused,
1759 			  int unset __maybe_unused)
1760 {
1761 	kmem_page = (kmem_slab + 1);
1762 	return 0;
1763 }
1764 
1765 static int parse_line_opt(const struct option *opt __maybe_unused,
1766 			  const char *arg, int unset __maybe_unused)
1767 {
1768 	int lines;
1769 
1770 	if (!arg)
1771 		return -1;
1772 
1773 	lines = strtoul(arg, NULL, 10);
1774 
1775 	if (caller_flag > alloc_flag)
1776 		caller_lines = lines;
1777 	else
1778 		alloc_lines = lines;
1779 
1780 	return 0;
1781 }
1782 
1783 static int __cmd_record(int argc, const char **argv)
1784 {
1785 	const char * const record_args[] = {
1786 	"record", "-a", "-R", "-c", "1",
1787 	};
1788 	const char * const slab_events[] = {
1789 	"-e", "kmem:kmalloc",
1790 	"-e", "kmem:kmalloc_node",
1791 	"-e", "kmem:kfree",
1792 	"-e", "kmem:kmem_cache_alloc",
1793 	"-e", "kmem:kmem_cache_alloc_node",
1794 	"-e", "kmem:kmem_cache_free",
1795 	};
1796 	const char * const page_events[] = {
1797 	"-e", "kmem:mm_page_alloc",
1798 	"-e", "kmem:mm_page_free",
1799 	};
1800 	unsigned int rec_argc, i, j;
1801 	const char **rec_argv;
1802 
1803 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1804 	if (kmem_slab)
1805 		rec_argc += ARRAY_SIZE(slab_events);
1806 	if (kmem_page)
1807 		rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1808 
1809 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1810 
1811 	if (rec_argv == NULL)
1812 		return -ENOMEM;
1813 
1814 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1815 		rec_argv[i] = strdup(record_args[i]);
1816 
1817 	if (kmem_slab) {
1818 		for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1819 			rec_argv[i] = strdup(slab_events[j]);
1820 	}
1821 	if (kmem_page) {
1822 		rec_argv[i++] = strdup("-g");
1823 
1824 		for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1825 			rec_argv[i] = strdup(page_events[j]);
1826 	}
1827 
1828 	for (j = 1; j < (unsigned int)argc; j++, i++)
1829 		rec_argv[i] = argv[j];
1830 
1831 	return cmd_record(i, rec_argv, NULL);
1832 }
1833 
1834 static int kmem_config(const char *var, const char *value, void *cb)
1835 {
1836 	if (!strcmp(var, "kmem.default")) {
1837 		if (!strcmp(value, "slab"))
1838 			kmem_default = KMEM_SLAB;
1839 		else if (!strcmp(value, "page"))
1840 			kmem_default = KMEM_PAGE;
1841 		else
1842 			pr_err("invalid default value ('slab' or 'page' required): %s\n",
1843 			       value);
1844 		return 0;
1845 	}
1846 
1847 	return perf_default_config(var, value, cb);
1848 }
1849 
1850 int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused)
1851 {
1852 	const char * const default_slab_sort = "frag,hit,bytes";
1853 	const char * const default_page_sort = "bytes,hit";
1854 	struct perf_data_file file = {
1855 		.mode = PERF_DATA_MODE_READ,
1856 	};
1857 	const struct option kmem_options[] = {
1858 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1859 	OPT_INCR('v', "verbose", &verbose,
1860 		    "be more verbose (show symbol address, etc)"),
1861 	OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1862 			   "show per-callsite statistics", parse_caller_opt),
1863 	OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1864 			   "show per-allocation statistics", parse_alloc_opt),
1865 	OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1866 		     "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1867 		     "page, order, migtype, gfp", parse_sort_opt),
1868 	OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1869 	OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1870 	OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"),
1871 	OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1872 			   parse_slab_opt),
1873 	OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1874 			   parse_page_opt),
1875 	OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1876 	OPT_END()
1877 	};
1878 	const char *const kmem_subcommands[] = { "record", "stat", NULL };
1879 	const char *kmem_usage[] = {
1880 		NULL,
1881 		NULL
1882 	};
1883 	struct perf_session *session;
1884 	int ret = -1;
1885 	const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1886 
1887 	perf_config(kmem_config, NULL);
1888 	argc = parse_options_subcommand(argc, argv, kmem_options,
1889 					kmem_subcommands, kmem_usage, 0);
1890 
1891 	if (!argc)
1892 		usage_with_options(kmem_usage, kmem_options);
1893 
1894 	if (kmem_slab == 0 && kmem_page == 0) {
1895 		if (kmem_default == KMEM_SLAB)
1896 			kmem_slab = 1;
1897 		else
1898 			kmem_page = 1;
1899 	}
1900 
1901 	if (!strncmp(argv[0], "rec", 3)) {
1902 		symbol__init(NULL);
1903 		return __cmd_record(argc, argv);
1904 	}
1905 
1906 	file.path = input_name;
1907 
1908 	kmem_session = session = perf_session__new(&file, false, &perf_kmem);
1909 	if (session == NULL)
1910 		return -1;
1911 
1912 	if (kmem_slab) {
1913 		if (!perf_evlist__find_tracepoint_by_name(session->evlist,
1914 							  "kmem:kmalloc")) {
1915 			pr_err(errmsg, "slab", "slab");
1916 			return -1;
1917 		}
1918 	}
1919 
1920 	if (kmem_page) {
1921 		struct perf_evsel *evsel;
1922 
1923 		evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
1924 							     "kmem:mm_page_alloc");
1925 		if (evsel == NULL) {
1926 			pr_err(errmsg, "page", "page");
1927 			return -1;
1928 		}
1929 
1930 		kmem_page_size = pevent_get_page_size(evsel->tp_format->pevent);
1931 		symbol_conf.use_callchain = true;
1932 	}
1933 
1934 	symbol__init(&session->header.env);
1935 
1936 	if (!strcmp(argv[0], "stat")) {
1937 		setlocale(LC_ALL, "");
1938 
1939 		if (cpu__setup_cpunode_map())
1940 			goto out_delete;
1941 
1942 		if (list_empty(&slab_caller_sort))
1943 			setup_slab_sorting(&slab_caller_sort, default_slab_sort);
1944 		if (list_empty(&slab_alloc_sort))
1945 			setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
1946 		if (list_empty(&page_caller_sort))
1947 			setup_page_sorting(&page_caller_sort, default_page_sort);
1948 		if (list_empty(&page_alloc_sort))
1949 			setup_page_sorting(&page_alloc_sort, default_page_sort);
1950 
1951 		if (kmem_page) {
1952 			setup_page_sorting(&page_alloc_sort_input,
1953 					   "page,order,migtype,gfp");
1954 			setup_page_sorting(&page_caller_sort_input,
1955 					   "callsite,order,migtype,gfp");
1956 		}
1957 		ret = __cmd_kmem(session);
1958 	} else
1959 		usage_with_options(kmem_usage, kmem_options);
1960 
1961 out_delete:
1962 	perf_session__delete(session);
1963 
1964 	return ret;
1965 }
1966 
1967