xref: /linux/tools/perf/builtin-kmem.c (revision bfd5bb6f90af092aa345b15cd78143956a13c2a8)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 #include "perf.h"
4 
5 #include "util/evlist.h"
6 #include "util/evsel.h"
7 #include "util/util.h"
8 #include "util/config.h"
9 #include "util/symbol.h"
10 #include "util/thread.h"
11 #include "util/header.h"
12 #include "util/session.h"
13 #include "util/tool.h"
14 #include "util/callchain.h"
15 #include "util/time-utils.h"
16 
17 #include <subcmd/parse-options.h>
18 #include "util/trace-event.h"
19 #include "util/data.h"
20 #include "util/cpumap.h"
21 
22 #include "util/debug.h"
23 
24 #include <linux/kernel.h>
25 #include <linux/rbtree.h>
26 #include <linux/string.h>
27 #include <errno.h>
28 #include <inttypes.h>
29 #include <locale.h>
30 #include <regex.h>
31 
32 #include "sane_ctype.h"
33 
34 static int	kmem_slab;
35 static int	kmem_page;
36 
37 static long	kmem_page_size;
38 static enum {
39 	KMEM_SLAB,
40 	KMEM_PAGE,
41 } kmem_default = KMEM_SLAB;  /* for backward compatibility */
42 
43 struct alloc_stat;
44 typedef int (*sort_fn_t)(void *, void *);
45 
46 static int			alloc_flag;
47 static int			caller_flag;
48 
49 static int			alloc_lines = -1;
50 static int			caller_lines = -1;
51 
52 static bool			raw_ip;
53 
54 struct alloc_stat {
55 	u64	call_site;
56 	u64	ptr;
57 	u64	bytes_req;
58 	u64	bytes_alloc;
59 	u64	last_alloc;
60 	u32	hit;
61 	u32	pingpong;
62 
63 	short	alloc_cpu;
64 
65 	struct rb_node node;
66 };
67 
68 static struct rb_root root_alloc_stat;
69 static struct rb_root root_alloc_sorted;
70 static struct rb_root root_caller_stat;
71 static struct rb_root root_caller_sorted;
72 
73 static unsigned long total_requested, total_allocated, total_freed;
74 static unsigned long nr_allocs, nr_cross_allocs;
75 
76 /* filters for controlling start and stop of time of analysis */
77 static struct perf_time_interval ptime;
78 const char *time_str;
79 
80 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
81 			     int bytes_req, int bytes_alloc, int cpu)
82 {
83 	struct rb_node **node = &root_alloc_stat.rb_node;
84 	struct rb_node *parent = NULL;
85 	struct alloc_stat *data = NULL;
86 
87 	while (*node) {
88 		parent = *node;
89 		data = rb_entry(*node, struct alloc_stat, node);
90 
91 		if (ptr > data->ptr)
92 			node = &(*node)->rb_right;
93 		else if (ptr < data->ptr)
94 			node = &(*node)->rb_left;
95 		else
96 			break;
97 	}
98 
99 	if (data && data->ptr == ptr) {
100 		data->hit++;
101 		data->bytes_req += bytes_req;
102 		data->bytes_alloc += bytes_alloc;
103 	} else {
104 		data = malloc(sizeof(*data));
105 		if (!data) {
106 			pr_err("%s: malloc failed\n", __func__);
107 			return -1;
108 		}
109 		data->ptr = ptr;
110 		data->pingpong = 0;
111 		data->hit = 1;
112 		data->bytes_req = bytes_req;
113 		data->bytes_alloc = bytes_alloc;
114 
115 		rb_link_node(&data->node, parent, node);
116 		rb_insert_color(&data->node, &root_alloc_stat);
117 	}
118 	data->call_site = call_site;
119 	data->alloc_cpu = cpu;
120 	data->last_alloc = bytes_alloc;
121 
122 	return 0;
123 }
124 
125 static int insert_caller_stat(unsigned long call_site,
126 			      int bytes_req, int bytes_alloc)
127 {
128 	struct rb_node **node = &root_caller_stat.rb_node;
129 	struct rb_node *parent = NULL;
130 	struct alloc_stat *data = NULL;
131 
132 	while (*node) {
133 		parent = *node;
134 		data = rb_entry(*node, struct alloc_stat, node);
135 
136 		if (call_site > data->call_site)
137 			node = &(*node)->rb_right;
138 		else if (call_site < data->call_site)
139 			node = &(*node)->rb_left;
140 		else
141 			break;
142 	}
143 
144 	if (data && data->call_site == call_site) {
145 		data->hit++;
146 		data->bytes_req += bytes_req;
147 		data->bytes_alloc += bytes_alloc;
148 	} else {
149 		data = malloc(sizeof(*data));
150 		if (!data) {
151 			pr_err("%s: malloc failed\n", __func__);
152 			return -1;
153 		}
154 		data->call_site = call_site;
155 		data->pingpong = 0;
156 		data->hit = 1;
157 		data->bytes_req = bytes_req;
158 		data->bytes_alloc = bytes_alloc;
159 
160 		rb_link_node(&data->node, parent, node);
161 		rb_insert_color(&data->node, &root_caller_stat);
162 	}
163 
164 	return 0;
165 }
166 
167 static int perf_evsel__process_alloc_event(struct perf_evsel *evsel,
168 					   struct perf_sample *sample)
169 {
170 	unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
171 		      call_site = perf_evsel__intval(evsel, sample, "call_site");
172 	int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
173 	    bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
174 
175 	if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
176 	    insert_caller_stat(call_site, bytes_req, bytes_alloc))
177 		return -1;
178 
179 	total_requested += bytes_req;
180 	total_allocated += bytes_alloc;
181 
182 	nr_allocs++;
183 	return 0;
184 }
185 
186 static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,
187 						struct perf_sample *sample)
188 {
189 	int ret = perf_evsel__process_alloc_event(evsel, sample);
190 
191 	if (!ret) {
192 		int node1 = cpu__get_node(sample->cpu),
193 		    node2 = perf_evsel__intval(evsel, sample, "node");
194 
195 		if (node1 != node2)
196 			nr_cross_allocs++;
197 	}
198 
199 	return ret;
200 }
201 
202 static int ptr_cmp(void *, void *);
203 static int slab_callsite_cmp(void *, void *);
204 
205 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
206 					    unsigned long call_site,
207 					    struct rb_root *root,
208 					    sort_fn_t sort_fn)
209 {
210 	struct rb_node *node = root->rb_node;
211 	struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
212 
213 	while (node) {
214 		struct alloc_stat *data;
215 		int cmp;
216 
217 		data = rb_entry(node, struct alloc_stat, node);
218 
219 		cmp = sort_fn(&key, data);
220 		if (cmp < 0)
221 			node = node->rb_left;
222 		else if (cmp > 0)
223 			node = node->rb_right;
224 		else
225 			return data;
226 	}
227 	return NULL;
228 }
229 
230 static int perf_evsel__process_free_event(struct perf_evsel *evsel,
231 					  struct perf_sample *sample)
232 {
233 	unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
234 	struct alloc_stat *s_alloc, *s_caller;
235 
236 	s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
237 	if (!s_alloc)
238 		return 0;
239 
240 	total_freed += s_alloc->last_alloc;
241 
242 	if ((short)sample->cpu != s_alloc->alloc_cpu) {
243 		s_alloc->pingpong++;
244 
245 		s_caller = search_alloc_stat(0, s_alloc->call_site,
246 					     &root_caller_stat,
247 					     slab_callsite_cmp);
248 		if (!s_caller)
249 			return -1;
250 		s_caller->pingpong++;
251 	}
252 	s_alloc->alloc_cpu = -1;
253 
254 	return 0;
255 }
256 
257 static u64 total_page_alloc_bytes;
258 static u64 total_page_free_bytes;
259 static u64 total_page_nomatch_bytes;
260 static u64 total_page_fail_bytes;
261 static unsigned long nr_page_allocs;
262 static unsigned long nr_page_frees;
263 static unsigned long nr_page_fails;
264 static unsigned long nr_page_nomatch;
265 
266 static bool use_pfn;
267 static bool live_page;
268 static struct perf_session *kmem_session;
269 
270 #define MAX_MIGRATE_TYPES  6
271 #define MAX_PAGE_ORDER     11
272 
273 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
274 
275 struct page_stat {
276 	struct rb_node 	node;
277 	u64 		page;
278 	u64 		callsite;
279 	int 		order;
280 	unsigned 	gfp_flags;
281 	unsigned 	migrate_type;
282 	u64		alloc_bytes;
283 	u64 		free_bytes;
284 	int 		nr_alloc;
285 	int 		nr_free;
286 };
287 
288 static struct rb_root page_live_tree;
289 static struct rb_root page_alloc_tree;
290 static struct rb_root page_alloc_sorted;
291 static struct rb_root page_caller_tree;
292 static struct rb_root page_caller_sorted;
293 
294 struct alloc_func {
295 	u64 start;
296 	u64 end;
297 	char *name;
298 };
299 
300 static int nr_alloc_funcs;
301 static struct alloc_func *alloc_func_list;
302 
303 static int funcmp(const void *a, const void *b)
304 {
305 	const struct alloc_func *fa = a;
306 	const struct alloc_func *fb = b;
307 
308 	if (fa->start > fb->start)
309 		return 1;
310 	else
311 		return -1;
312 }
313 
314 static int callcmp(const void *a, const void *b)
315 {
316 	const struct alloc_func *fa = a;
317 	const struct alloc_func *fb = b;
318 
319 	if (fb->start <= fa->start && fa->end < fb->end)
320 		return 0;
321 
322 	if (fa->start > fb->start)
323 		return 1;
324 	else
325 		return -1;
326 }
327 
328 static int build_alloc_func_list(void)
329 {
330 	int ret;
331 	struct map *kernel_map;
332 	struct symbol *sym;
333 	struct rb_node *node;
334 	struct alloc_func *func;
335 	struct machine *machine = &kmem_session->machines.host;
336 	regex_t alloc_func_regex;
337 	const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
338 
339 	ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
340 	if (ret) {
341 		char err[BUFSIZ];
342 
343 		regerror(ret, &alloc_func_regex, err, sizeof(err));
344 		pr_err("Invalid regex: %s\n%s", pattern, err);
345 		return -EINVAL;
346 	}
347 
348 	kernel_map = machine__kernel_map(machine);
349 	if (map__load(kernel_map) < 0) {
350 		pr_err("cannot load kernel map\n");
351 		return -ENOENT;
352 	}
353 
354 	map__for_each_symbol(kernel_map, sym, node) {
355 		if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
356 			continue;
357 
358 		func = realloc(alloc_func_list,
359 			       (nr_alloc_funcs + 1) * sizeof(*func));
360 		if (func == NULL)
361 			return -ENOMEM;
362 
363 		pr_debug("alloc func: %s\n", sym->name);
364 		func[nr_alloc_funcs].start = sym->start;
365 		func[nr_alloc_funcs].end   = sym->end;
366 		func[nr_alloc_funcs].name  = sym->name;
367 
368 		alloc_func_list = func;
369 		nr_alloc_funcs++;
370 	}
371 
372 	qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
373 
374 	regfree(&alloc_func_regex);
375 	return 0;
376 }
377 
378 /*
379  * Find first non-memory allocation function from callchain.
380  * The allocation functions are in the 'alloc_func_list'.
381  */
382 static u64 find_callsite(struct perf_evsel *evsel, struct perf_sample *sample)
383 {
384 	struct addr_location al;
385 	struct machine *machine = &kmem_session->machines.host;
386 	struct callchain_cursor_node *node;
387 
388 	if (alloc_func_list == NULL) {
389 		if (build_alloc_func_list() < 0)
390 			goto out;
391 	}
392 
393 	al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
394 	sample__resolve_callchain(sample, &callchain_cursor, NULL, evsel, &al, 16);
395 
396 	callchain_cursor_commit(&callchain_cursor);
397 	while (true) {
398 		struct alloc_func key, *caller;
399 		u64 addr;
400 
401 		node = callchain_cursor_current(&callchain_cursor);
402 		if (node == NULL)
403 			break;
404 
405 		key.start = key.end = node->ip;
406 		caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
407 				 sizeof(key), callcmp);
408 		if (!caller) {
409 			/* found */
410 			if (node->map)
411 				addr = map__unmap_ip(node->map, node->ip);
412 			else
413 				addr = node->ip;
414 
415 			return addr;
416 		} else
417 			pr_debug3("skipping alloc function: %s\n", caller->name);
418 
419 		callchain_cursor_advance(&callchain_cursor);
420 	}
421 
422 out:
423 	pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
424 	return sample->ip;
425 }
426 
427 struct sort_dimension {
428 	const char		name[20];
429 	sort_fn_t		cmp;
430 	struct list_head	list;
431 };
432 
433 static LIST_HEAD(page_alloc_sort_input);
434 static LIST_HEAD(page_caller_sort_input);
435 
436 static struct page_stat *
437 __page_stat__findnew_page(struct page_stat *pstat, bool create)
438 {
439 	struct rb_node **node = &page_live_tree.rb_node;
440 	struct rb_node *parent = NULL;
441 	struct page_stat *data;
442 
443 	while (*node) {
444 		s64 cmp;
445 
446 		parent = *node;
447 		data = rb_entry(*node, struct page_stat, node);
448 
449 		cmp = data->page - pstat->page;
450 		if (cmp < 0)
451 			node = &parent->rb_left;
452 		else if (cmp > 0)
453 			node = &parent->rb_right;
454 		else
455 			return data;
456 	}
457 
458 	if (!create)
459 		return NULL;
460 
461 	data = zalloc(sizeof(*data));
462 	if (data != NULL) {
463 		data->page = pstat->page;
464 		data->order = pstat->order;
465 		data->gfp_flags = pstat->gfp_flags;
466 		data->migrate_type = pstat->migrate_type;
467 
468 		rb_link_node(&data->node, parent, node);
469 		rb_insert_color(&data->node, &page_live_tree);
470 	}
471 
472 	return data;
473 }
474 
475 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
476 {
477 	return __page_stat__findnew_page(pstat, false);
478 }
479 
480 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
481 {
482 	return __page_stat__findnew_page(pstat, true);
483 }
484 
485 static struct page_stat *
486 __page_stat__findnew_alloc(struct page_stat *pstat, bool create)
487 {
488 	struct rb_node **node = &page_alloc_tree.rb_node;
489 	struct rb_node *parent = NULL;
490 	struct page_stat *data;
491 	struct sort_dimension *sort;
492 
493 	while (*node) {
494 		int cmp = 0;
495 
496 		parent = *node;
497 		data = rb_entry(*node, struct page_stat, node);
498 
499 		list_for_each_entry(sort, &page_alloc_sort_input, list) {
500 			cmp = sort->cmp(pstat, data);
501 			if (cmp)
502 				break;
503 		}
504 
505 		if (cmp < 0)
506 			node = &parent->rb_left;
507 		else if (cmp > 0)
508 			node = &parent->rb_right;
509 		else
510 			return data;
511 	}
512 
513 	if (!create)
514 		return NULL;
515 
516 	data = zalloc(sizeof(*data));
517 	if (data != NULL) {
518 		data->page = pstat->page;
519 		data->order = pstat->order;
520 		data->gfp_flags = pstat->gfp_flags;
521 		data->migrate_type = pstat->migrate_type;
522 
523 		rb_link_node(&data->node, parent, node);
524 		rb_insert_color(&data->node, &page_alloc_tree);
525 	}
526 
527 	return data;
528 }
529 
530 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
531 {
532 	return __page_stat__findnew_alloc(pstat, false);
533 }
534 
535 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
536 {
537 	return __page_stat__findnew_alloc(pstat, true);
538 }
539 
540 static struct page_stat *
541 __page_stat__findnew_caller(struct page_stat *pstat, bool create)
542 {
543 	struct rb_node **node = &page_caller_tree.rb_node;
544 	struct rb_node *parent = NULL;
545 	struct page_stat *data;
546 	struct sort_dimension *sort;
547 
548 	while (*node) {
549 		int cmp = 0;
550 
551 		parent = *node;
552 		data = rb_entry(*node, struct page_stat, node);
553 
554 		list_for_each_entry(sort, &page_caller_sort_input, list) {
555 			cmp = sort->cmp(pstat, data);
556 			if (cmp)
557 				break;
558 		}
559 
560 		if (cmp < 0)
561 			node = &parent->rb_left;
562 		else if (cmp > 0)
563 			node = &parent->rb_right;
564 		else
565 			return data;
566 	}
567 
568 	if (!create)
569 		return NULL;
570 
571 	data = zalloc(sizeof(*data));
572 	if (data != NULL) {
573 		data->callsite = pstat->callsite;
574 		data->order = pstat->order;
575 		data->gfp_flags = pstat->gfp_flags;
576 		data->migrate_type = pstat->migrate_type;
577 
578 		rb_link_node(&data->node, parent, node);
579 		rb_insert_color(&data->node, &page_caller_tree);
580 	}
581 
582 	return data;
583 }
584 
585 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
586 {
587 	return __page_stat__findnew_caller(pstat, false);
588 }
589 
590 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
591 {
592 	return __page_stat__findnew_caller(pstat, true);
593 }
594 
595 static bool valid_page(u64 pfn_or_page)
596 {
597 	if (use_pfn && pfn_or_page == -1UL)
598 		return false;
599 	if (!use_pfn && pfn_or_page == 0)
600 		return false;
601 	return true;
602 }
603 
604 struct gfp_flag {
605 	unsigned int flags;
606 	char *compact_str;
607 	char *human_readable;
608 };
609 
610 static struct gfp_flag *gfps;
611 static int nr_gfps;
612 
613 static int gfpcmp(const void *a, const void *b)
614 {
615 	const struct gfp_flag *fa = a;
616 	const struct gfp_flag *fb = b;
617 
618 	return fa->flags - fb->flags;
619 }
620 
621 /* see include/trace/events/mmflags.h */
622 static const struct {
623 	const char *original;
624 	const char *compact;
625 } gfp_compact_table[] = {
626 	{ "GFP_TRANSHUGE",		"THP" },
627 	{ "GFP_TRANSHUGE_LIGHT",	"THL" },
628 	{ "GFP_HIGHUSER_MOVABLE",	"HUM" },
629 	{ "GFP_HIGHUSER",		"HU" },
630 	{ "GFP_USER",			"U" },
631 	{ "GFP_KERNEL_ACCOUNT",		"KAC" },
632 	{ "GFP_KERNEL",			"K" },
633 	{ "GFP_NOFS",			"NF" },
634 	{ "GFP_ATOMIC",			"A" },
635 	{ "GFP_NOIO",			"NI" },
636 	{ "GFP_NOWAIT",			"NW" },
637 	{ "GFP_DMA",			"D" },
638 	{ "__GFP_HIGHMEM",		"HM" },
639 	{ "GFP_DMA32",			"D32" },
640 	{ "__GFP_HIGH",			"H" },
641 	{ "__GFP_ATOMIC",		"_A" },
642 	{ "__GFP_IO",			"I" },
643 	{ "__GFP_FS",			"F" },
644 	{ "__GFP_NOWARN",		"NWR" },
645 	{ "__GFP_RETRY_MAYFAIL",	"R" },
646 	{ "__GFP_NOFAIL",		"NF" },
647 	{ "__GFP_NORETRY",		"NR" },
648 	{ "__GFP_COMP",			"C" },
649 	{ "__GFP_ZERO",			"Z" },
650 	{ "__GFP_NOMEMALLOC",		"NMA" },
651 	{ "__GFP_MEMALLOC",		"MA" },
652 	{ "__GFP_HARDWALL",		"HW" },
653 	{ "__GFP_THISNODE",		"TN" },
654 	{ "__GFP_RECLAIMABLE",		"RC" },
655 	{ "__GFP_MOVABLE",		"M" },
656 	{ "__GFP_ACCOUNT",		"AC" },
657 	{ "__GFP_WRITE",		"WR" },
658 	{ "__GFP_RECLAIM",		"R" },
659 	{ "__GFP_DIRECT_RECLAIM",	"DR" },
660 	{ "__GFP_KSWAPD_RECLAIM",	"KR" },
661 };
662 
663 static size_t max_gfp_len;
664 
665 static char *compact_gfp_flags(char *gfp_flags)
666 {
667 	char *orig_flags = strdup(gfp_flags);
668 	char *new_flags = NULL;
669 	char *str, *pos = NULL;
670 	size_t len = 0;
671 
672 	if (orig_flags == NULL)
673 		return NULL;
674 
675 	str = strtok_r(orig_flags, "|", &pos);
676 	while (str) {
677 		size_t i;
678 		char *new;
679 		const char *cpt;
680 
681 		for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
682 			if (strcmp(gfp_compact_table[i].original, str))
683 				continue;
684 
685 			cpt = gfp_compact_table[i].compact;
686 			new = realloc(new_flags, len + strlen(cpt) + 2);
687 			if (new == NULL) {
688 				free(new_flags);
689 				return NULL;
690 			}
691 
692 			new_flags = new;
693 
694 			if (!len) {
695 				strcpy(new_flags, cpt);
696 			} else {
697 				strcat(new_flags, "|");
698 				strcat(new_flags, cpt);
699 				len++;
700 			}
701 
702 			len += strlen(cpt);
703 		}
704 
705 		str = strtok_r(NULL, "|", &pos);
706 	}
707 
708 	if (max_gfp_len < len)
709 		max_gfp_len = len;
710 
711 	free(orig_flags);
712 	return new_flags;
713 }
714 
715 static char *compact_gfp_string(unsigned long gfp_flags)
716 {
717 	struct gfp_flag key = {
718 		.flags = gfp_flags,
719 	};
720 	struct gfp_flag *gfp;
721 
722 	gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
723 	if (gfp)
724 		return gfp->compact_str;
725 
726 	return NULL;
727 }
728 
729 static int parse_gfp_flags(struct perf_evsel *evsel, struct perf_sample *sample,
730 			   unsigned int gfp_flags)
731 {
732 	struct pevent_record record = {
733 		.cpu = sample->cpu,
734 		.data = sample->raw_data,
735 		.size = sample->raw_size,
736 	};
737 	struct trace_seq seq;
738 	char *str, *pos = NULL;
739 
740 	if (nr_gfps) {
741 		struct gfp_flag key = {
742 			.flags = gfp_flags,
743 		};
744 
745 		if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
746 			return 0;
747 	}
748 
749 	trace_seq_init(&seq);
750 	pevent_event_info(&seq, evsel->tp_format, &record);
751 
752 	str = strtok_r(seq.buffer, " ", &pos);
753 	while (str) {
754 		if (!strncmp(str, "gfp_flags=", 10)) {
755 			struct gfp_flag *new;
756 
757 			new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
758 			if (new == NULL)
759 				return -ENOMEM;
760 
761 			gfps = new;
762 			new += nr_gfps++;
763 
764 			new->flags = gfp_flags;
765 			new->human_readable = strdup(str + 10);
766 			new->compact_str = compact_gfp_flags(str + 10);
767 			if (!new->human_readable || !new->compact_str)
768 				return -ENOMEM;
769 
770 			qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
771 		}
772 
773 		str = strtok_r(NULL, " ", &pos);
774 	}
775 
776 	trace_seq_destroy(&seq);
777 	return 0;
778 }
779 
780 static int perf_evsel__process_page_alloc_event(struct perf_evsel *evsel,
781 						struct perf_sample *sample)
782 {
783 	u64 page;
784 	unsigned int order = perf_evsel__intval(evsel, sample, "order");
785 	unsigned int gfp_flags = perf_evsel__intval(evsel, sample, "gfp_flags");
786 	unsigned int migrate_type = perf_evsel__intval(evsel, sample,
787 						       "migratetype");
788 	u64 bytes = kmem_page_size << order;
789 	u64 callsite;
790 	struct page_stat *pstat;
791 	struct page_stat this = {
792 		.order = order,
793 		.gfp_flags = gfp_flags,
794 		.migrate_type = migrate_type,
795 	};
796 
797 	if (use_pfn)
798 		page = perf_evsel__intval(evsel, sample, "pfn");
799 	else
800 		page = perf_evsel__intval(evsel, sample, "page");
801 
802 	nr_page_allocs++;
803 	total_page_alloc_bytes += bytes;
804 
805 	if (!valid_page(page)) {
806 		nr_page_fails++;
807 		total_page_fail_bytes += bytes;
808 
809 		return 0;
810 	}
811 
812 	if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
813 		return -1;
814 
815 	callsite = find_callsite(evsel, sample);
816 
817 	/*
818 	 * This is to find the current page (with correct gfp flags and
819 	 * migrate type) at free event.
820 	 */
821 	this.page = page;
822 	pstat = page_stat__findnew_page(&this);
823 	if (pstat == NULL)
824 		return -ENOMEM;
825 
826 	pstat->nr_alloc++;
827 	pstat->alloc_bytes += bytes;
828 	pstat->callsite = callsite;
829 
830 	if (!live_page) {
831 		pstat = page_stat__findnew_alloc(&this);
832 		if (pstat == NULL)
833 			return -ENOMEM;
834 
835 		pstat->nr_alloc++;
836 		pstat->alloc_bytes += bytes;
837 		pstat->callsite = callsite;
838 	}
839 
840 	this.callsite = callsite;
841 	pstat = page_stat__findnew_caller(&this);
842 	if (pstat == NULL)
843 		return -ENOMEM;
844 
845 	pstat->nr_alloc++;
846 	pstat->alloc_bytes += bytes;
847 
848 	order_stats[order][migrate_type]++;
849 
850 	return 0;
851 }
852 
853 static int perf_evsel__process_page_free_event(struct perf_evsel *evsel,
854 						struct perf_sample *sample)
855 {
856 	u64 page;
857 	unsigned int order = perf_evsel__intval(evsel, sample, "order");
858 	u64 bytes = kmem_page_size << order;
859 	struct page_stat *pstat;
860 	struct page_stat this = {
861 		.order = order,
862 	};
863 
864 	if (use_pfn)
865 		page = perf_evsel__intval(evsel, sample, "pfn");
866 	else
867 		page = perf_evsel__intval(evsel, sample, "page");
868 
869 	nr_page_frees++;
870 	total_page_free_bytes += bytes;
871 
872 	this.page = page;
873 	pstat = page_stat__find_page(&this);
874 	if (pstat == NULL) {
875 		pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
876 			  page, order);
877 
878 		nr_page_nomatch++;
879 		total_page_nomatch_bytes += bytes;
880 
881 		return 0;
882 	}
883 
884 	this.gfp_flags = pstat->gfp_flags;
885 	this.migrate_type = pstat->migrate_type;
886 	this.callsite = pstat->callsite;
887 
888 	rb_erase(&pstat->node, &page_live_tree);
889 	free(pstat);
890 
891 	if (live_page) {
892 		order_stats[this.order][this.migrate_type]--;
893 	} else {
894 		pstat = page_stat__find_alloc(&this);
895 		if (pstat == NULL)
896 			return -ENOMEM;
897 
898 		pstat->nr_free++;
899 		pstat->free_bytes += bytes;
900 	}
901 
902 	pstat = page_stat__find_caller(&this);
903 	if (pstat == NULL)
904 		return -ENOENT;
905 
906 	pstat->nr_free++;
907 	pstat->free_bytes += bytes;
908 
909 	if (live_page) {
910 		pstat->nr_alloc--;
911 		pstat->alloc_bytes -= bytes;
912 
913 		if (pstat->nr_alloc == 0) {
914 			rb_erase(&pstat->node, &page_caller_tree);
915 			free(pstat);
916 		}
917 	}
918 
919 	return 0;
920 }
921 
922 static bool perf_kmem__skip_sample(struct perf_sample *sample)
923 {
924 	/* skip sample based on time? */
925 	if (perf_time__skip_sample(&ptime, sample->time))
926 		return true;
927 
928 	return false;
929 }
930 
931 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
932 				  struct perf_sample *sample);
933 
934 static int process_sample_event(struct perf_tool *tool __maybe_unused,
935 				union perf_event *event,
936 				struct perf_sample *sample,
937 				struct perf_evsel *evsel,
938 				struct machine *machine)
939 {
940 	int err = 0;
941 	struct thread *thread = machine__findnew_thread(machine, sample->pid,
942 							sample->tid);
943 
944 	if (thread == NULL) {
945 		pr_debug("problem processing %d event, skipping it.\n",
946 			 event->header.type);
947 		return -1;
948 	}
949 
950 	if (perf_kmem__skip_sample(sample))
951 		return 0;
952 
953 	dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
954 
955 	if (evsel->handler != NULL) {
956 		tracepoint_handler f = evsel->handler;
957 		err = f(evsel, sample);
958 	}
959 
960 	thread__put(thread);
961 
962 	return err;
963 }
964 
965 static struct perf_tool perf_kmem = {
966 	.sample		 = process_sample_event,
967 	.comm		 = perf_event__process_comm,
968 	.mmap		 = perf_event__process_mmap,
969 	.mmap2		 = perf_event__process_mmap2,
970 	.namespaces	 = perf_event__process_namespaces,
971 	.ordered_events	 = true,
972 };
973 
974 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
975 {
976 	if (n_alloc == 0)
977 		return 0.0;
978 	else
979 		return 100.0 - (100.0 * n_req / n_alloc);
980 }
981 
982 static void __print_slab_result(struct rb_root *root,
983 				struct perf_session *session,
984 				int n_lines, int is_caller)
985 {
986 	struct rb_node *next;
987 	struct machine *machine = &session->machines.host;
988 
989 	printf("%.105s\n", graph_dotted_line);
990 	printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
991 	printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
992 	printf("%.105s\n", graph_dotted_line);
993 
994 	next = rb_first(root);
995 
996 	while (next && n_lines--) {
997 		struct alloc_stat *data = rb_entry(next, struct alloc_stat,
998 						   node);
999 		struct symbol *sym = NULL;
1000 		struct map *map;
1001 		char buf[BUFSIZ];
1002 		u64 addr;
1003 
1004 		if (is_caller) {
1005 			addr = data->call_site;
1006 			if (!raw_ip)
1007 				sym = machine__find_kernel_symbol(machine, addr, &map);
1008 		} else
1009 			addr = data->ptr;
1010 
1011 		if (sym != NULL)
1012 			snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1013 				 addr - map->unmap_ip(map, sym->start));
1014 		else
1015 			snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1016 		printf(" %-34s |", buf);
1017 
1018 		printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1019 		       (unsigned long long)data->bytes_alloc,
1020 		       (unsigned long)data->bytes_alloc / data->hit,
1021 		       (unsigned long long)data->bytes_req,
1022 		       (unsigned long)data->bytes_req / data->hit,
1023 		       (unsigned long)data->hit,
1024 		       (unsigned long)data->pingpong,
1025 		       fragmentation(data->bytes_req, data->bytes_alloc));
1026 
1027 		next = rb_next(next);
1028 	}
1029 
1030 	if (n_lines == -1)
1031 		printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
1032 
1033 	printf("%.105s\n", graph_dotted_line);
1034 }
1035 
1036 static const char * const migrate_type_str[] = {
1037 	"UNMOVABL",
1038 	"RECLAIM",
1039 	"MOVABLE",
1040 	"RESERVED",
1041 	"CMA/ISLT",
1042 	"UNKNOWN",
1043 };
1044 
1045 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1046 {
1047 	struct rb_node *next = rb_first(&page_alloc_sorted);
1048 	struct machine *machine = &session->machines.host;
1049 	const char *format;
1050 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1051 
1052 	printf("\n%.105s\n", graph_dotted_line);
1053 	printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1054 	       use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1055 	       gfp_len, "GFP flags");
1056 	printf("%.105s\n", graph_dotted_line);
1057 
1058 	if (use_pfn)
1059 		format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1060 	else
1061 		format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1062 
1063 	while (next && n_lines--) {
1064 		struct page_stat *data;
1065 		struct symbol *sym;
1066 		struct map *map;
1067 		char buf[32];
1068 		char *caller = buf;
1069 
1070 		data = rb_entry(next, struct page_stat, node);
1071 		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1072 		if (sym)
1073 			caller = sym->name;
1074 		else
1075 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1076 
1077 		printf(format, (unsigned long long)data->page,
1078 		       (unsigned long long)data->alloc_bytes / 1024,
1079 		       data->nr_alloc, data->order,
1080 		       migrate_type_str[data->migrate_type],
1081 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1082 
1083 		next = rb_next(next);
1084 	}
1085 
1086 	if (n_lines == -1) {
1087 		printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1088 		       gfp_len, "...");
1089 	}
1090 
1091 	printf("%.105s\n", graph_dotted_line);
1092 }
1093 
1094 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1095 {
1096 	struct rb_node *next = rb_first(&page_caller_sorted);
1097 	struct machine *machine = &session->machines.host;
1098 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1099 
1100 	printf("\n%.105s\n", graph_dotted_line);
1101 	printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1102 	       live_page ? "Live" : "Total", gfp_len, "GFP flags");
1103 	printf("%.105s\n", graph_dotted_line);
1104 
1105 	while (next && n_lines--) {
1106 		struct page_stat *data;
1107 		struct symbol *sym;
1108 		struct map *map;
1109 		char buf[32];
1110 		char *caller = buf;
1111 
1112 		data = rb_entry(next, struct page_stat, node);
1113 		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1114 		if (sym)
1115 			caller = sym->name;
1116 		else
1117 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1118 
1119 		printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1120 		       (unsigned long long)data->alloc_bytes / 1024,
1121 		       data->nr_alloc, data->order,
1122 		       migrate_type_str[data->migrate_type],
1123 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1124 
1125 		next = rb_next(next);
1126 	}
1127 
1128 	if (n_lines == -1) {
1129 		printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1130 		       gfp_len, "...");
1131 	}
1132 
1133 	printf("%.105s\n", graph_dotted_line);
1134 }
1135 
1136 static void print_gfp_flags(void)
1137 {
1138 	int i;
1139 
1140 	printf("#\n");
1141 	printf("# GFP flags\n");
1142 	printf("# ---------\n");
1143 	for (i = 0; i < nr_gfps; i++) {
1144 		printf("# %08x: %*s: %s\n", gfps[i].flags,
1145 		       (int) max_gfp_len, gfps[i].compact_str,
1146 		       gfps[i].human_readable);
1147 	}
1148 }
1149 
1150 static void print_slab_summary(void)
1151 {
1152 	printf("\nSUMMARY (SLAB allocator)");
1153 	printf("\n========================\n");
1154 	printf("Total bytes requested: %'lu\n", total_requested);
1155 	printf("Total bytes allocated: %'lu\n", total_allocated);
1156 	printf("Total bytes freed:     %'lu\n", total_freed);
1157 	if (total_allocated > total_freed) {
1158 		printf("Net total bytes allocated: %'lu\n",
1159 		total_allocated - total_freed);
1160 	}
1161 	printf("Total bytes wasted on internal fragmentation: %'lu\n",
1162 	       total_allocated - total_requested);
1163 	printf("Internal fragmentation: %f%%\n",
1164 	       fragmentation(total_requested, total_allocated));
1165 	printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1166 }
1167 
1168 static void print_page_summary(void)
1169 {
1170 	int o, m;
1171 	u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1172 	u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1173 
1174 	printf("\nSUMMARY (page allocator)");
1175 	printf("\n========================\n");
1176 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1177 	       nr_page_allocs, total_page_alloc_bytes / 1024);
1178 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1179 	       nr_page_frees, total_page_free_bytes / 1024);
1180 	printf("\n");
1181 
1182 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1183 	       nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1184 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1185 	       nr_page_allocs - nr_alloc_freed,
1186 	       (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1187 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1188 	       nr_page_nomatch, total_page_nomatch_bytes / 1024);
1189 	printf("\n");
1190 
1191 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1192 	       nr_page_fails, total_page_fail_bytes / 1024);
1193 	printf("\n");
1194 
1195 	printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1196 	       "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1197 	printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1198 	       graph_dotted_line, graph_dotted_line, graph_dotted_line,
1199 	       graph_dotted_line, graph_dotted_line);
1200 
1201 	for (o = 0; o < MAX_PAGE_ORDER; o++) {
1202 		printf("%5d", o);
1203 		for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1204 			if (order_stats[o][m])
1205 				printf("  %'12d", order_stats[o][m]);
1206 			else
1207 				printf("  %12c", '.');
1208 		}
1209 		printf("\n");
1210 	}
1211 }
1212 
1213 static void print_slab_result(struct perf_session *session)
1214 {
1215 	if (caller_flag)
1216 		__print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1217 	if (alloc_flag)
1218 		__print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1219 	print_slab_summary();
1220 }
1221 
1222 static void print_page_result(struct perf_session *session)
1223 {
1224 	if (caller_flag || alloc_flag)
1225 		print_gfp_flags();
1226 	if (caller_flag)
1227 		__print_page_caller_result(session, caller_lines);
1228 	if (alloc_flag)
1229 		__print_page_alloc_result(session, alloc_lines);
1230 	print_page_summary();
1231 }
1232 
1233 static void print_result(struct perf_session *session)
1234 {
1235 	if (kmem_slab)
1236 		print_slab_result(session);
1237 	if (kmem_page)
1238 		print_page_result(session);
1239 }
1240 
1241 static LIST_HEAD(slab_caller_sort);
1242 static LIST_HEAD(slab_alloc_sort);
1243 static LIST_HEAD(page_caller_sort);
1244 static LIST_HEAD(page_alloc_sort);
1245 
1246 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1247 			     struct list_head *sort_list)
1248 {
1249 	struct rb_node **new = &(root->rb_node);
1250 	struct rb_node *parent = NULL;
1251 	struct sort_dimension *sort;
1252 
1253 	while (*new) {
1254 		struct alloc_stat *this;
1255 		int cmp = 0;
1256 
1257 		this = rb_entry(*new, struct alloc_stat, node);
1258 		parent = *new;
1259 
1260 		list_for_each_entry(sort, sort_list, list) {
1261 			cmp = sort->cmp(data, this);
1262 			if (cmp)
1263 				break;
1264 		}
1265 
1266 		if (cmp > 0)
1267 			new = &((*new)->rb_left);
1268 		else
1269 			new = &((*new)->rb_right);
1270 	}
1271 
1272 	rb_link_node(&data->node, parent, new);
1273 	rb_insert_color(&data->node, root);
1274 }
1275 
1276 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1277 			       struct list_head *sort_list)
1278 {
1279 	struct rb_node *node;
1280 	struct alloc_stat *data;
1281 
1282 	for (;;) {
1283 		node = rb_first(root);
1284 		if (!node)
1285 			break;
1286 
1287 		rb_erase(node, root);
1288 		data = rb_entry(node, struct alloc_stat, node);
1289 		sort_slab_insert(root_sorted, data, sort_list);
1290 	}
1291 }
1292 
1293 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1294 			     struct list_head *sort_list)
1295 {
1296 	struct rb_node **new = &root->rb_node;
1297 	struct rb_node *parent = NULL;
1298 	struct sort_dimension *sort;
1299 
1300 	while (*new) {
1301 		struct page_stat *this;
1302 		int cmp = 0;
1303 
1304 		this = rb_entry(*new, struct page_stat, node);
1305 		parent = *new;
1306 
1307 		list_for_each_entry(sort, sort_list, list) {
1308 			cmp = sort->cmp(data, this);
1309 			if (cmp)
1310 				break;
1311 		}
1312 
1313 		if (cmp > 0)
1314 			new = &parent->rb_left;
1315 		else
1316 			new = &parent->rb_right;
1317 	}
1318 
1319 	rb_link_node(&data->node, parent, new);
1320 	rb_insert_color(&data->node, root);
1321 }
1322 
1323 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1324 			       struct list_head *sort_list)
1325 {
1326 	struct rb_node *node;
1327 	struct page_stat *data;
1328 
1329 	for (;;) {
1330 		node = rb_first(root);
1331 		if (!node)
1332 			break;
1333 
1334 		rb_erase(node, root);
1335 		data = rb_entry(node, struct page_stat, node);
1336 		sort_page_insert(root_sorted, data, sort_list);
1337 	}
1338 }
1339 
1340 static void sort_result(void)
1341 {
1342 	if (kmem_slab) {
1343 		__sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1344 				   &slab_alloc_sort);
1345 		__sort_slab_result(&root_caller_stat, &root_caller_sorted,
1346 				   &slab_caller_sort);
1347 	}
1348 	if (kmem_page) {
1349 		if (live_page)
1350 			__sort_page_result(&page_live_tree, &page_alloc_sorted,
1351 					   &page_alloc_sort);
1352 		else
1353 			__sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1354 					   &page_alloc_sort);
1355 
1356 		__sort_page_result(&page_caller_tree, &page_caller_sorted,
1357 				   &page_caller_sort);
1358 	}
1359 }
1360 
1361 static int __cmd_kmem(struct perf_session *session)
1362 {
1363 	int err = -EINVAL;
1364 	struct perf_evsel *evsel;
1365 	const struct perf_evsel_str_handler kmem_tracepoints[] = {
1366 		/* slab allocator */
1367 		{ "kmem:kmalloc",		perf_evsel__process_alloc_event, },
1368     		{ "kmem:kmem_cache_alloc",	perf_evsel__process_alloc_event, },
1369 		{ "kmem:kmalloc_node",		perf_evsel__process_alloc_node_event, },
1370     		{ "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
1371 		{ "kmem:kfree",			perf_evsel__process_free_event, },
1372     		{ "kmem:kmem_cache_free",	perf_evsel__process_free_event, },
1373 		/* page allocator */
1374 		{ "kmem:mm_page_alloc",		perf_evsel__process_page_alloc_event, },
1375 		{ "kmem:mm_page_free",		perf_evsel__process_page_free_event, },
1376 	};
1377 
1378 	if (!perf_session__has_traces(session, "kmem record"))
1379 		goto out;
1380 
1381 	if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1382 		pr_err("Initializing perf session tracepoint handlers failed\n");
1383 		goto out;
1384 	}
1385 
1386 	evlist__for_each_entry(session->evlist, evsel) {
1387 		if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") &&
1388 		    perf_evsel__field(evsel, "pfn")) {
1389 			use_pfn = true;
1390 			break;
1391 		}
1392 	}
1393 
1394 	setup_pager();
1395 	err = perf_session__process_events(session);
1396 	if (err != 0) {
1397 		pr_err("error during process events: %d\n", err);
1398 		goto out;
1399 	}
1400 	sort_result();
1401 	print_result(session);
1402 out:
1403 	return err;
1404 }
1405 
1406 /* slab sort keys */
1407 static int ptr_cmp(void *a, void *b)
1408 {
1409 	struct alloc_stat *l = a;
1410 	struct alloc_stat *r = b;
1411 
1412 	if (l->ptr < r->ptr)
1413 		return -1;
1414 	else if (l->ptr > r->ptr)
1415 		return 1;
1416 	return 0;
1417 }
1418 
1419 static struct sort_dimension ptr_sort_dimension = {
1420 	.name	= "ptr",
1421 	.cmp	= ptr_cmp,
1422 };
1423 
1424 static int slab_callsite_cmp(void *a, void *b)
1425 {
1426 	struct alloc_stat *l = a;
1427 	struct alloc_stat *r = b;
1428 
1429 	if (l->call_site < r->call_site)
1430 		return -1;
1431 	else if (l->call_site > r->call_site)
1432 		return 1;
1433 	return 0;
1434 }
1435 
1436 static struct sort_dimension callsite_sort_dimension = {
1437 	.name	= "callsite",
1438 	.cmp	= slab_callsite_cmp,
1439 };
1440 
1441 static int hit_cmp(void *a, void *b)
1442 {
1443 	struct alloc_stat *l = a;
1444 	struct alloc_stat *r = b;
1445 
1446 	if (l->hit < r->hit)
1447 		return -1;
1448 	else if (l->hit > r->hit)
1449 		return 1;
1450 	return 0;
1451 }
1452 
1453 static struct sort_dimension hit_sort_dimension = {
1454 	.name	= "hit",
1455 	.cmp	= hit_cmp,
1456 };
1457 
1458 static int bytes_cmp(void *a, void *b)
1459 {
1460 	struct alloc_stat *l = a;
1461 	struct alloc_stat *r = b;
1462 
1463 	if (l->bytes_alloc < r->bytes_alloc)
1464 		return -1;
1465 	else if (l->bytes_alloc > r->bytes_alloc)
1466 		return 1;
1467 	return 0;
1468 }
1469 
1470 static struct sort_dimension bytes_sort_dimension = {
1471 	.name	= "bytes",
1472 	.cmp	= bytes_cmp,
1473 };
1474 
1475 static int frag_cmp(void *a, void *b)
1476 {
1477 	double x, y;
1478 	struct alloc_stat *l = a;
1479 	struct alloc_stat *r = b;
1480 
1481 	x = fragmentation(l->bytes_req, l->bytes_alloc);
1482 	y = fragmentation(r->bytes_req, r->bytes_alloc);
1483 
1484 	if (x < y)
1485 		return -1;
1486 	else if (x > y)
1487 		return 1;
1488 	return 0;
1489 }
1490 
1491 static struct sort_dimension frag_sort_dimension = {
1492 	.name	= "frag",
1493 	.cmp	= frag_cmp,
1494 };
1495 
1496 static int pingpong_cmp(void *a, void *b)
1497 {
1498 	struct alloc_stat *l = a;
1499 	struct alloc_stat *r = b;
1500 
1501 	if (l->pingpong < r->pingpong)
1502 		return -1;
1503 	else if (l->pingpong > r->pingpong)
1504 		return 1;
1505 	return 0;
1506 }
1507 
1508 static struct sort_dimension pingpong_sort_dimension = {
1509 	.name	= "pingpong",
1510 	.cmp	= pingpong_cmp,
1511 };
1512 
1513 /* page sort keys */
1514 static int page_cmp(void *a, void *b)
1515 {
1516 	struct page_stat *l = a;
1517 	struct page_stat *r = b;
1518 
1519 	if (l->page < r->page)
1520 		return -1;
1521 	else if (l->page > r->page)
1522 		return 1;
1523 	return 0;
1524 }
1525 
1526 static struct sort_dimension page_sort_dimension = {
1527 	.name	= "page",
1528 	.cmp	= page_cmp,
1529 };
1530 
1531 static int page_callsite_cmp(void *a, void *b)
1532 {
1533 	struct page_stat *l = a;
1534 	struct page_stat *r = b;
1535 
1536 	if (l->callsite < r->callsite)
1537 		return -1;
1538 	else if (l->callsite > r->callsite)
1539 		return 1;
1540 	return 0;
1541 }
1542 
1543 static struct sort_dimension page_callsite_sort_dimension = {
1544 	.name	= "callsite",
1545 	.cmp	= page_callsite_cmp,
1546 };
1547 
1548 static int page_hit_cmp(void *a, void *b)
1549 {
1550 	struct page_stat *l = a;
1551 	struct page_stat *r = b;
1552 
1553 	if (l->nr_alloc < r->nr_alloc)
1554 		return -1;
1555 	else if (l->nr_alloc > r->nr_alloc)
1556 		return 1;
1557 	return 0;
1558 }
1559 
1560 static struct sort_dimension page_hit_sort_dimension = {
1561 	.name	= "hit",
1562 	.cmp	= page_hit_cmp,
1563 };
1564 
1565 static int page_bytes_cmp(void *a, void *b)
1566 {
1567 	struct page_stat *l = a;
1568 	struct page_stat *r = b;
1569 
1570 	if (l->alloc_bytes < r->alloc_bytes)
1571 		return -1;
1572 	else if (l->alloc_bytes > r->alloc_bytes)
1573 		return 1;
1574 	return 0;
1575 }
1576 
1577 static struct sort_dimension page_bytes_sort_dimension = {
1578 	.name	= "bytes",
1579 	.cmp	= page_bytes_cmp,
1580 };
1581 
1582 static int page_order_cmp(void *a, void *b)
1583 {
1584 	struct page_stat *l = a;
1585 	struct page_stat *r = b;
1586 
1587 	if (l->order < r->order)
1588 		return -1;
1589 	else if (l->order > r->order)
1590 		return 1;
1591 	return 0;
1592 }
1593 
1594 static struct sort_dimension page_order_sort_dimension = {
1595 	.name	= "order",
1596 	.cmp	= page_order_cmp,
1597 };
1598 
1599 static int migrate_type_cmp(void *a, void *b)
1600 {
1601 	struct page_stat *l = a;
1602 	struct page_stat *r = b;
1603 
1604 	/* for internal use to find free'd page */
1605 	if (l->migrate_type == -1U)
1606 		return 0;
1607 
1608 	if (l->migrate_type < r->migrate_type)
1609 		return -1;
1610 	else if (l->migrate_type > r->migrate_type)
1611 		return 1;
1612 	return 0;
1613 }
1614 
1615 static struct sort_dimension migrate_type_sort_dimension = {
1616 	.name	= "migtype",
1617 	.cmp	= migrate_type_cmp,
1618 };
1619 
1620 static int gfp_flags_cmp(void *a, void *b)
1621 {
1622 	struct page_stat *l = a;
1623 	struct page_stat *r = b;
1624 
1625 	/* for internal use to find free'd page */
1626 	if (l->gfp_flags == -1U)
1627 		return 0;
1628 
1629 	if (l->gfp_flags < r->gfp_flags)
1630 		return -1;
1631 	else if (l->gfp_flags > r->gfp_flags)
1632 		return 1;
1633 	return 0;
1634 }
1635 
1636 static struct sort_dimension gfp_flags_sort_dimension = {
1637 	.name	= "gfp",
1638 	.cmp	= gfp_flags_cmp,
1639 };
1640 
1641 static struct sort_dimension *slab_sorts[] = {
1642 	&ptr_sort_dimension,
1643 	&callsite_sort_dimension,
1644 	&hit_sort_dimension,
1645 	&bytes_sort_dimension,
1646 	&frag_sort_dimension,
1647 	&pingpong_sort_dimension,
1648 };
1649 
1650 static struct sort_dimension *page_sorts[] = {
1651 	&page_sort_dimension,
1652 	&page_callsite_sort_dimension,
1653 	&page_hit_sort_dimension,
1654 	&page_bytes_sort_dimension,
1655 	&page_order_sort_dimension,
1656 	&migrate_type_sort_dimension,
1657 	&gfp_flags_sort_dimension,
1658 };
1659 
1660 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1661 {
1662 	struct sort_dimension *sort;
1663 	int i;
1664 
1665 	for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1666 		if (!strcmp(slab_sorts[i]->name, tok)) {
1667 			sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1668 			if (!sort) {
1669 				pr_err("%s: memdup failed\n", __func__);
1670 				return -1;
1671 			}
1672 			list_add_tail(&sort->list, list);
1673 			return 0;
1674 		}
1675 	}
1676 
1677 	return -1;
1678 }
1679 
1680 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1681 {
1682 	struct sort_dimension *sort;
1683 	int i;
1684 
1685 	for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1686 		if (!strcmp(page_sorts[i]->name, tok)) {
1687 			sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1688 			if (!sort) {
1689 				pr_err("%s: memdup failed\n", __func__);
1690 				return -1;
1691 			}
1692 			list_add_tail(&sort->list, list);
1693 			return 0;
1694 		}
1695 	}
1696 
1697 	return -1;
1698 }
1699 
1700 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1701 {
1702 	char *tok;
1703 	char *str = strdup(arg);
1704 	char *pos = str;
1705 
1706 	if (!str) {
1707 		pr_err("%s: strdup failed\n", __func__);
1708 		return -1;
1709 	}
1710 
1711 	while (true) {
1712 		tok = strsep(&pos, ",");
1713 		if (!tok)
1714 			break;
1715 		if (slab_sort_dimension__add(tok, sort_list) < 0) {
1716 			pr_err("Unknown slab --sort key: '%s'", tok);
1717 			free(str);
1718 			return -1;
1719 		}
1720 	}
1721 
1722 	free(str);
1723 	return 0;
1724 }
1725 
1726 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1727 {
1728 	char *tok;
1729 	char *str = strdup(arg);
1730 	char *pos = str;
1731 
1732 	if (!str) {
1733 		pr_err("%s: strdup failed\n", __func__);
1734 		return -1;
1735 	}
1736 
1737 	while (true) {
1738 		tok = strsep(&pos, ",");
1739 		if (!tok)
1740 			break;
1741 		if (page_sort_dimension__add(tok, sort_list) < 0) {
1742 			pr_err("Unknown page --sort key: '%s'", tok);
1743 			free(str);
1744 			return -1;
1745 		}
1746 	}
1747 
1748 	free(str);
1749 	return 0;
1750 }
1751 
1752 static int parse_sort_opt(const struct option *opt __maybe_unused,
1753 			  const char *arg, int unset __maybe_unused)
1754 {
1755 	if (!arg)
1756 		return -1;
1757 
1758 	if (kmem_page > kmem_slab ||
1759 	    (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1760 		if (caller_flag > alloc_flag)
1761 			return setup_page_sorting(&page_caller_sort, arg);
1762 		else
1763 			return setup_page_sorting(&page_alloc_sort, arg);
1764 	} else {
1765 		if (caller_flag > alloc_flag)
1766 			return setup_slab_sorting(&slab_caller_sort, arg);
1767 		else
1768 			return setup_slab_sorting(&slab_alloc_sort, arg);
1769 	}
1770 
1771 	return 0;
1772 }
1773 
1774 static int parse_caller_opt(const struct option *opt __maybe_unused,
1775 			    const char *arg __maybe_unused,
1776 			    int unset __maybe_unused)
1777 {
1778 	caller_flag = (alloc_flag + 1);
1779 	return 0;
1780 }
1781 
1782 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1783 			   const char *arg __maybe_unused,
1784 			   int unset __maybe_unused)
1785 {
1786 	alloc_flag = (caller_flag + 1);
1787 	return 0;
1788 }
1789 
1790 static int parse_slab_opt(const struct option *opt __maybe_unused,
1791 			  const char *arg __maybe_unused,
1792 			  int unset __maybe_unused)
1793 {
1794 	kmem_slab = (kmem_page + 1);
1795 	return 0;
1796 }
1797 
1798 static int parse_page_opt(const struct option *opt __maybe_unused,
1799 			  const char *arg __maybe_unused,
1800 			  int unset __maybe_unused)
1801 {
1802 	kmem_page = (kmem_slab + 1);
1803 	return 0;
1804 }
1805 
1806 static int parse_line_opt(const struct option *opt __maybe_unused,
1807 			  const char *arg, int unset __maybe_unused)
1808 {
1809 	int lines;
1810 
1811 	if (!arg)
1812 		return -1;
1813 
1814 	lines = strtoul(arg, NULL, 10);
1815 
1816 	if (caller_flag > alloc_flag)
1817 		caller_lines = lines;
1818 	else
1819 		alloc_lines = lines;
1820 
1821 	return 0;
1822 }
1823 
1824 static int __cmd_record(int argc, const char **argv)
1825 {
1826 	const char * const record_args[] = {
1827 	"record", "-a", "-R", "-c", "1",
1828 	};
1829 	const char * const slab_events[] = {
1830 	"-e", "kmem:kmalloc",
1831 	"-e", "kmem:kmalloc_node",
1832 	"-e", "kmem:kfree",
1833 	"-e", "kmem:kmem_cache_alloc",
1834 	"-e", "kmem:kmem_cache_alloc_node",
1835 	"-e", "kmem:kmem_cache_free",
1836 	};
1837 	const char * const page_events[] = {
1838 	"-e", "kmem:mm_page_alloc",
1839 	"-e", "kmem:mm_page_free",
1840 	};
1841 	unsigned int rec_argc, i, j;
1842 	const char **rec_argv;
1843 
1844 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1845 	if (kmem_slab)
1846 		rec_argc += ARRAY_SIZE(slab_events);
1847 	if (kmem_page)
1848 		rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1849 
1850 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1851 
1852 	if (rec_argv == NULL)
1853 		return -ENOMEM;
1854 
1855 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1856 		rec_argv[i] = strdup(record_args[i]);
1857 
1858 	if (kmem_slab) {
1859 		for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1860 			rec_argv[i] = strdup(slab_events[j]);
1861 	}
1862 	if (kmem_page) {
1863 		rec_argv[i++] = strdup("-g");
1864 
1865 		for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1866 			rec_argv[i] = strdup(page_events[j]);
1867 	}
1868 
1869 	for (j = 1; j < (unsigned int)argc; j++, i++)
1870 		rec_argv[i] = argv[j];
1871 
1872 	return cmd_record(i, rec_argv);
1873 }
1874 
1875 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1876 {
1877 	if (!strcmp(var, "kmem.default")) {
1878 		if (!strcmp(value, "slab"))
1879 			kmem_default = KMEM_SLAB;
1880 		else if (!strcmp(value, "page"))
1881 			kmem_default = KMEM_PAGE;
1882 		else
1883 			pr_err("invalid default value ('slab' or 'page' required): %s\n",
1884 			       value);
1885 		return 0;
1886 	}
1887 
1888 	return 0;
1889 }
1890 
1891 int cmd_kmem(int argc, const char **argv)
1892 {
1893 	const char * const default_slab_sort = "frag,hit,bytes";
1894 	const char * const default_page_sort = "bytes,hit";
1895 	struct perf_data data = {
1896 		.mode = PERF_DATA_MODE_READ,
1897 	};
1898 	const struct option kmem_options[] = {
1899 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1900 	OPT_INCR('v', "verbose", &verbose,
1901 		    "be more verbose (show symbol address, etc)"),
1902 	OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1903 			   "show per-callsite statistics", parse_caller_opt),
1904 	OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1905 			   "show per-allocation statistics", parse_alloc_opt),
1906 	OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1907 		     "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1908 		     "page, order, migtype, gfp", parse_sort_opt),
1909 	OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1910 	OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1911 	OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
1912 	OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1913 			   parse_slab_opt),
1914 	OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1915 			   parse_page_opt),
1916 	OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1917 	OPT_STRING(0, "time", &time_str, "str",
1918 		   "Time span of interest (start,stop)"),
1919 	OPT_END()
1920 	};
1921 	const char *const kmem_subcommands[] = { "record", "stat", NULL };
1922 	const char *kmem_usage[] = {
1923 		NULL,
1924 		NULL
1925 	};
1926 	struct perf_session *session;
1927 	const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1928 	int ret = perf_config(kmem_config, NULL);
1929 
1930 	if (ret)
1931 		return ret;
1932 
1933 	argc = parse_options_subcommand(argc, argv, kmem_options,
1934 					kmem_subcommands, kmem_usage, 0);
1935 
1936 	if (!argc)
1937 		usage_with_options(kmem_usage, kmem_options);
1938 
1939 	if (kmem_slab == 0 && kmem_page == 0) {
1940 		if (kmem_default == KMEM_SLAB)
1941 			kmem_slab = 1;
1942 		else
1943 			kmem_page = 1;
1944 	}
1945 
1946 	if (!strncmp(argv[0], "rec", 3)) {
1947 		symbol__init(NULL);
1948 		return __cmd_record(argc, argv);
1949 	}
1950 
1951 	data.file.path = input_name;
1952 
1953 	kmem_session = session = perf_session__new(&data, false, &perf_kmem);
1954 	if (session == NULL)
1955 		return -1;
1956 
1957 	ret = -1;
1958 
1959 	if (kmem_slab) {
1960 		if (!perf_evlist__find_tracepoint_by_name(session->evlist,
1961 							  "kmem:kmalloc")) {
1962 			pr_err(errmsg, "slab", "slab");
1963 			goto out_delete;
1964 		}
1965 	}
1966 
1967 	if (kmem_page) {
1968 		struct perf_evsel *evsel;
1969 
1970 		evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
1971 							     "kmem:mm_page_alloc");
1972 		if (evsel == NULL) {
1973 			pr_err(errmsg, "page", "page");
1974 			goto out_delete;
1975 		}
1976 
1977 		kmem_page_size = pevent_get_page_size(evsel->tp_format->pevent);
1978 		symbol_conf.use_callchain = true;
1979 	}
1980 
1981 	symbol__init(&session->header.env);
1982 
1983 	if (perf_time__parse_str(&ptime, time_str) != 0) {
1984 		pr_err("Invalid time string\n");
1985 		ret = -EINVAL;
1986 		goto out_delete;
1987 	}
1988 
1989 	if (!strcmp(argv[0], "stat")) {
1990 		setlocale(LC_ALL, "");
1991 
1992 		if (cpu__setup_cpunode_map())
1993 			goto out_delete;
1994 
1995 		if (list_empty(&slab_caller_sort))
1996 			setup_slab_sorting(&slab_caller_sort, default_slab_sort);
1997 		if (list_empty(&slab_alloc_sort))
1998 			setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
1999 		if (list_empty(&page_caller_sort))
2000 			setup_page_sorting(&page_caller_sort, default_page_sort);
2001 		if (list_empty(&page_alloc_sort))
2002 			setup_page_sorting(&page_alloc_sort, default_page_sort);
2003 
2004 		if (kmem_page) {
2005 			setup_page_sorting(&page_alloc_sort_input,
2006 					   "page,order,migtype,gfp");
2007 			setup_page_sorting(&page_caller_sort_input,
2008 					   "callsite,order,migtype,gfp");
2009 		}
2010 		ret = __cmd_kmem(session);
2011 	} else
2012 		usage_with_options(kmem_usage, kmem_options);
2013 
2014 out_delete:
2015 	perf_session__delete(session);
2016 
2017 	return ret;
2018 }
2019 
2020