xref: /openbmc/linux/tools/perf/builtin-kmem.c (revision 1a59d1b8)
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/map.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/tool.h"
15 #include "util/callchain.h"
16 #include "util/time-utils.h"
17 
18 #include <subcmd/parse-options.h>
19 #include "util/trace-event.h"
20 #include "util/data.h"
21 #include "util/cpumap.h"
22 
23 #include "util/debug.h"
24 
25 #include <linux/kernel.h>
26 #include <linux/rbtree.h>
27 #include <linux/string.h>
28 #include <errno.h>
29 #include <inttypes.h>
30 #include <locale.h>
31 #include <regex.h>
32 
33 #include "sane_ctype.h"
34 
35 static int	kmem_slab;
36 static int	kmem_page;
37 
38 static long	kmem_page_size;
39 static enum {
40 	KMEM_SLAB,
41 	KMEM_PAGE,
42 } kmem_default = KMEM_SLAB;  /* for backward compatibility */
43 
44 struct alloc_stat;
45 typedef int (*sort_fn_t)(void *, void *);
46 
47 static int			alloc_flag;
48 static int			caller_flag;
49 
50 static int			alloc_lines = -1;
51 static int			caller_lines = -1;
52 
53 static bool			raw_ip;
54 
55 struct alloc_stat {
56 	u64	call_site;
57 	u64	ptr;
58 	u64	bytes_req;
59 	u64	bytes_alloc;
60 	u64	last_alloc;
61 	u32	hit;
62 	u32	pingpong;
63 
64 	short	alloc_cpu;
65 
66 	struct rb_node node;
67 };
68 
69 static struct rb_root root_alloc_stat;
70 static struct rb_root root_alloc_sorted;
71 static struct rb_root root_caller_stat;
72 static struct rb_root root_caller_sorted;
73 
74 static unsigned long total_requested, total_allocated, total_freed;
75 static unsigned long nr_allocs, nr_cross_allocs;
76 
77 /* filters for controlling start and stop of time of analysis */
78 static struct perf_time_interval ptime;
79 const char *time_str;
80 
81 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
82 			     int bytes_req, int bytes_alloc, int cpu)
83 {
84 	struct rb_node **node = &root_alloc_stat.rb_node;
85 	struct rb_node *parent = NULL;
86 	struct alloc_stat *data = NULL;
87 
88 	while (*node) {
89 		parent = *node;
90 		data = rb_entry(*node, struct alloc_stat, node);
91 
92 		if (ptr > data->ptr)
93 			node = &(*node)->rb_right;
94 		else if (ptr < data->ptr)
95 			node = &(*node)->rb_left;
96 		else
97 			break;
98 	}
99 
100 	if (data && data->ptr == ptr) {
101 		data->hit++;
102 		data->bytes_req += bytes_req;
103 		data->bytes_alloc += bytes_alloc;
104 	} else {
105 		data = malloc(sizeof(*data));
106 		if (!data) {
107 			pr_err("%s: malloc failed\n", __func__);
108 			return -1;
109 		}
110 		data->ptr = ptr;
111 		data->pingpong = 0;
112 		data->hit = 1;
113 		data->bytes_req = bytes_req;
114 		data->bytes_alloc = bytes_alloc;
115 
116 		rb_link_node(&data->node, parent, node);
117 		rb_insert_color(&data->node, &root_alloc_stat);
118 	}
119 	data->call_site = call_site;
120 	data->alloc_cpu = cpu;
121 	data->last_alloc = bytes_alloc;
122 
123 	return 0;
124 }
125 
126 static int insert_caller_stat(unsigned long call_site,
127 			      int bytes_req, int bytes_alloc)
128 {
129 	struct rb_node **node = &root_caller_stat.rb_node;
130 	struct rb_node *parent = NULL;
131 	struct alloc_stat *data = NULL;
132 
133 	while (*node) {
134 		parent = *node;
135 		data = rb_entry(*node, struct alloc_stat, node);
136 
137 		if (call_site > data->call_site)
138 			node = &(*node)->rb_right;
139 		else if (call_site < data->call_site)
140 			node = &(*node)->rb_left;
141 		else
142 			break;
143 	}
144 
145 	if (data && data->call_site == call_site) {
146 		data->hit++;
147 		data->bytes_req += bytes_req;
148 		data->bytes_alloc += bytes_alloc;
149 	} else {
150 		data = malloc(sizeof(*data));
151 		if (!data) {
152 			pr_err("%s: malloc failed\n", __func__);
153 			return -1;
154 		}
155 		data->call_site = call_site;
156 		data->pingpong = 0;
157 		data->hit = 1;
158 		data->bytes_req = bytes_req;
159 		data->bytes_alloc = bytes_alloc;
160 
161 		rb_link_node(&data->node, parent, node);
162 		rb_insert_color(&data->node, &root_caller_stat);
163 	}
164 
165 	return 0;
166 }
167 
168 static int perf_evsel__process_alloc_event(struct perf_evsel *evsel,
169 					   struct perf_sample *sample)
170 {
171 	unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
172 		      call_site = perf_evsel__intval(evsel, sample, "call_site");
173 	int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
174 	    bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
175 
176 	if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
177 	    insert_caller_stat(call_site, bytes_req, bytes_alloc))
178 		return -1;
179 
180 	total_requested += bytes_req;
181 	total_allocated += bytes_alloc;
182 
183 	nr_allocs++;
184 	return 0;
185 }
186 
187 static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,
188 						struct perf_sample *sample)
189 {
190 	int ret = perf_evsel__process_alloc_event(evsel, sample);
191 
192 	if (!ret) {
193 		int node1 = cpu__get_node(sample->cpu),
194 		    node2 = perf_evsel__intval(evsel, sample, "node");
195 
196 		if (node1 != node2)
197 			nr_cross_allocs++;
198 	}
199 
200 	return ret;
201 }
202 
203 static int ptr_cmp(void *, void *);
204 static int slab_callsite_cmp(void *, void *);
205 
206 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
207 					    unsigned long call_site,
208 					    struct rb_root *root,
209 					    sort_fn_t sort_fn)
210 {
211 	struct rb_node *node = root->rb_node;
212 	struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
213 
214 	while (node) {
215 		struct alloc_stat *data;
216 		int cmp;
217 
218 		data = rb_entry(node, struct alloc_stat, node);
219 
220 		cmp = sort_fn(&key, data);
221 		if (cmp < 0)
222 			node = node->rb_left;
223 		else if (cmp > 0)
224 			node = node->rb_right;
225 		else
226 			return data;
227 	}
228 	return NULL;
229 }
230 
231 static int perf_evsel__process_free_event(struct perf_evsel *evsel,
232 					  struct perf_sample *sample)
233 {
234 	unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
235 	struct alloc_stat *s_alloc, *s_caller;
236 
237 	s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
238 	if (!s_alloc)
239 		return 0;
240 
241 	total_freed += s_alloc->last_alloc;
242 
243 	if ((short)sample->cpu != s_alloc->alloc_cpu) {
244 		s_alloc->pingpong++;
245 
246 		s_caller = search_alloc_stat(0, s_alloc->call_site,
247 					     &root_caller_stat,
248 					     slab_callsite_cmp);
249 		if (!s_caller)
250 			return -1;
251 		s_caller->pingpong++;
252 	}
253 	s_alloc->alloc_cpu = -1;
254 
255 	return 0;
256 }
257 
258 static u64 total_page_alloc_bytes;
259 static u64 total_page_free_bytes;
260 static u64 total_page_nomatch_bytes;
261 static u64 total_page_fail_bytes;
262 static unsigned long nr_page_allocs;
263 static unsigned long nr_page_frees;
264 static unsigned long nr_page_fails;
265 static unsigned long nr_page_nomatch;
266 
267 static bool use_pfn;
268 static bool live_page;
269 static struct perf_session *kmem_session;
270 
271 #define MAX_MIGRATE_TYPES  6
272 #define MAX_PAGE_ORDER     11
273 
274 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
275 
276 struct page_stat {
277 	struct rb_node 	node;
278 	u64 		page;
279 	u64 		callsite;
280 	int 		order;
281 	unsigned 	gfp_flags;
282 	unsigned 	migrate_type;
283 	u64		alloc_bytes;
284 	u64 		free_bytes;
285 	int 		nr_alloc;
286 	int 		nr_free;
287 };
288 
289 static struct rb_root page_live_tree;
290 static struct rb_root page_alloc_tree;
291 static struct rb_root page_alloc_sorted;
292 static struct rb_root page_caller_tree;
293 static struct rb_root page_caller_sorted;
294 
295 struct alloc_func {
296 	u64 start;
297 	u64 end;
298 	char *name;
299 };
300 
301 static int nr_alloc_funcs;
302 static struct alloc_func *alloc_func_list;
303 
304 static int funcmp(const void *a, const void *b)
305 {
306 	const struct alloc_func *fa = a;
307 	const struct alloc_func *fb = b;
308 
309 	if (fa->start > fb->start)
310 		return 1;
311 	else
312 		return -1;
313 }
314 
315 static int callcmp(const void *a, const void *b)
316 {
317 	const struct alloc_func *fa = a;
318 	const struct alloc_func *fb = b;
319 
320 	if (fb->start <= fa->start && fa->end < fb->end)
321 		return 0;
322 
323 	if (fa->start > fb->start)
324 		return 1;
325 	else
326 		return -1;
327 }
328 
329 static int build_alloc_func_list(void)
330 {
331 	int ret;
332 	struct map *kernel_map;
333 	struct symbol *sym;
334 	struct rb_node *node;
335 	struct alloc_func *func;
336 	struct machine *machine = &kmem_session->machines.host;
337 	regex_t alloc_func_regex;
338 	static const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
339 
340 	ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
341 	if (ret) {
342 		char err[BUFSIZ];
343 
344 		regerror(ret, &alloc_func_regex, err, sizeof(err));
345 		pr_err("Invalid regex: %s\n%s", pattern, err);
346 		return -EINVAL;
347 	}
348 
349 	kernel_map = machine__kernel_map(machine);
350 	if (map__load(kernel_map) < 0) {
351 		pr_err("cannot load kernel map\n");
352 		return -ENOENT;
353 	}
354 
355 	map__for_each_symbol(kernel_map, sym, node) {
356 		if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
357 			continue;
358 
359 		func = realloc(alloc_func_list,
360 			       (nr_alloc_funcs + 1) * sizeof(*func));
361 		if (func == NULL)
362 			return -ENOMEM;
363 
364 		pr_debug("alloc func: %s\n", sym->name);
365 		func[nr_alloc_funcs].start = sym->start;
366 		func[nr_alloc_funcs].end   = sym->end;
367 		func[nr_alloc_funcs].name  = sym->name;
368 
369 		alloc_func_list = func;
370 		nr_alloc_funcs++;
371 	}
372 
373 	qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
374 
375 	regfree(&alloc_func_regex);
376 	return 0;
377 }
378 
379 /*
380  * Find first non-memory allocation function from callchain.
381  * The allocation functions are in the 'alloc_func_list'.
382  */
383 static u64 find_callsite(struct perf_evsel *evsel, struct perf_sample *sample)
384 {
385 	struct addr_location al;
386 	struct machine *machine = &kmem_session->machines.host;
387 	struct callchain_cursor_node *node;
388 
389 	if (alloc_func_list == NULL) {
390 		if (build_alloc_func_list() < 0)
391 			goto out;
392 	}
393 
394 	al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
395 	sample__resolve_callchain(sample, &callchain_cursor, NULL, evsel, &al, 16);
396 
397 	callchain_cursor_commit(&callchain_cursor);
398 	while (true) {
399 		struct alloc_func key, *caller;
400 		u64 addr;
401 
402 		node = callchain_cursor_current(&callchain_cursor);
403 		if (node == NULL)
404 			break;
405 
406 		key.start = key.end = node->ip;
407 		caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
408 				 sizeof(key), callcmp);
409 		if (!caller) {
410 			/* found */
411 			if (node->map)
412 				addr = map__unmap_ip(node->map, node->ip);
413 			else
414 				addr = node->ip;
415 
416 			return addr;
417 		} else
418 			pr_debug3("skipping alloc function: %s\n", caller->name);
419 
420 		callchain_cursor_advance(&callchain_cursor);
421 	}
422 
423 out:
424 	pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
425 	return sample->ip;
426 }
427 
428 struct sort_dimension {
429 	const char		name[20];
430 	sort_fn_t		cmp;
431 	struct list_head	list;
432 };
433 
434 static LIST_HEAD(page_alloc_sort_input);
435 static LIST_HEAD(page_caller_sort_input);
436 
437 static struct page_stat *
438 __page_stat__findnew_page(struct page_stat *pstat, bool create)
439 {
440 	struct rb_node **node = &page_live_tree.rb_node;
441 	struct rb_node *parent = NULL;
442 	struct page_stat *data;
443 
444 	while (*node) {
445 		s64 cmp;
446 
447 		parent = *node;
448 		data = rb_entry(*node, struct page_stat, node);
449 
450 		cmp = data->page - pstat->page;
451 		if (cmp < 0)
452 			node = &parent->rb_left;
453 		else if (cmp > 0)
454 			node = &parent->rb_right;
455 		else
456 			return data;
457 	}
458 
459 	if (!create)
460 		return NULL;
461 
462 	data = zalloc(sizeof(*data));
463 	if (data != NULL) {
464 		data->page = pstat->page;
465 		data->order = pstat->order;
466 		data->gfp_flags = pstat->gfp_flags;
467 		data->migrate_type = pstat->migrate_type;
468 
469 		rb_link_node(&data->node, parent, node);
470 		rb_insert_color(&data->node, &page_live_tree);
471 	}
472 
473 	return data;
474 }
475 
476 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
477 {
478 	return __page_stat__findnew_page(pstat, false);
479 }
480 
481 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
482 {
483 	return __page_stat__findnew_page(pstat, true);
484 }
485 
486 static struct page_stat *
487 __page_stat__findnew_alloc(struct page_stat *pstat, bool create)
488 {
489 	struct rb_node **node = &page_alloc_tree.rb_node;
490 	struct rb_node *parent = NULL;
491 	struct page_stat *data;
492 	struct sort_dimension *sort;
493 
494 	while (*node) {
495 		int cmp = 0;
496 
497 		parent = *node;
498 		data = rb_entry(*node, struct page_stat, node);
499 
500 		list_for_each_entry(sort, &page_alloc_sort_input, list) {
501 			cmp = sort->cmp(pstat, data);
502 			if (cmp)
503 				break;
504 		}
505 
506 		if (cmp < 0)
507 			node = &parent->rb_left;
508 		else if (cmp > 0)
509 			node = &parent->rb_right;
510 		else
511 			return data;
512 	}
513 
514 	if (!create)
515 		return NULL;
516 
517 	data = zalloc(sizeof(*data));
518 	if (data != NULL) {
519 		data->page = pstat->page;
520 		data->order = pstat->order;
521 		data->gfp_flags = pstat->gfp_flags;
522 		data->migrate_type = pstat->migrate_type;
523 
524 		rb_link_node(&data->node, parent, node);
525 		rb_insert_color(&data->node, &page_alloc_tree);
526 	}
527 
528 	return data;
529 }
530 
531 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
532 {
533 	return __page_stat__findnew_alloc(pstat, false);
534 }
535 
536 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
537 {
538 	return __page_stat__findnew_alloc(pstat, true);
539 }
540 
541 static struct page_stat *
542 __page_stat__findnew_caller(struct page_stat *pstat, bool create)
543 {
544 	struct rb_node **node = &page_caller_tree.rb_node;
545 	struct rb_node *parent = NULL;
546 	struct page_stat *data;
547 	struct sort_dimension *sort;
548 
549 	while (*node) {
550 		int cmp = 0;
551 
552 		parent = *node;
553 		data = rb_entry(*node, struct page_stat, node);
554 
555 		list_for_each_entry(sort, &page_caller_sort_input, list) {
556 			cmp = sort->cmp(pstat, data);
557 			if (cmp)
558 				break;
559 		}
560 
561 		if (cmp < 0)
562 			node = &parent->rb_left;
563 		else if (cmp > 0)
564 			node = &parent->rb_right;
565 		else
566 			return data;
567 	}
568 
569 	if (!create)
570 		return NULL;
571 
572 	data = zalloc(sizeof(*data));
573 	if (data != NULL) {
574 		data->callsite = pstat->callsite;
575 		data->order = pstat->order;
576 		data->gfp_flags = pstat->gfp_flags;
577 		data->migrate_type = pstat->migrate_type;
578 
579 		rb_link_node(&data->node, parent, node);
580 		rb_insert_color(&data->node, &page_caller_tree);
581 	}
582 
583 	return data;
584 }
585 
586 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
587 {
588 	return __page_stat__findnew_caller(pstat, false);
589 }
590 
591 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
592 {
593 	return __page_stat__findnew_caller(pstat, true);
594 }
595 
596 static bool valid_page(u64 pfn_or_page)
597 {
598 	if (use_pfn && pfn_or_page == -1UL)
599 		return false;
600 	if (!use_pfn && pfn_or_page == 0)
601 		return false;
602 	return true;
603 }
604 
605 struct gfp_flag {
606 	unsigned int flags;
607 	char *compact_str;
608 	char *human_readable;
609 };
610 
611 static struct gfp_flag *gfps;
612 static int nr_gfps;
613 
614 static int gfpcmp(const void *a, const void *b)
615 {
616 	const struct gfp_flag *fa = a;
617 	const struct gfp_flag *fb = b;
618 
619 	return fa->flags - fb->flags;
620 }
621 
622 /* see include/trace/events/mmflags.h */
623 static const struct {
624 	const char *original;
625 	const char *compact;
626 } gfp_compact_table[] = {
627 	{ "GFP_TRANSHUGE",		"THP" },
628 	{ "GFP_TRANSHUGE_LIGHT",	"THL" },
629 	{ "GFP_HIGHUSER_MOVABLE",	"HUM" },
630 	{ "GFP_HIGHUSER",		"HU" },
631 	{ "GFP_USER",			"U" },
632 	{ "GFP_KERNEL_ACCOUNT",		"KAC" },
633 	{ "GFP_KERNEL",			"K" },
634 	{ "GFP_NOFS",			"NF" },
635 	{ "GFP_ATOMIC",			"A" },
636 	{ "GFP_NOIO",			"NI" },
637 	{ "GFP_NOWAIT",			"NW" },
638 	{ "GFP_DMA",			"D" },
639 	{ "__GFP_HIGHMEM",		"HM" },
640 	{ "GFP_DMA32",			"D32" },
641 	{ "__GFP_HIGH",			"H" },
642 	{ "__GFP_ATOMIC",		"_A" },
643 	{ "__GFP_IO",			"I" },
644 	{ "__GFP_FS",			"F" },
645 	{ "__GFP_NOWARN",		"NWR" },
646 	{ "__GFP_RETRY_MAYFAIL",	"R" },
647 	{ "__GFP_NOFAIL",		"NF" },
648 	{ "__GFP_NORETRY",		"NR" },
649 	{ "__GFP_COMP",			"C" },
650 	{ "__GFP_ZERO",			"Z" },
651 	{ "__GFP_NOMEMALLOC",		"NMA" },
652 	{ "__GFP_MEMALLOC",		"MA" },
653 	{ "__GFP_HARDWALL",		"HW" },
654 	{ "__GFP_THISNODE",		"TN" },
655 	{ "__GFP_RECLAIMABLE",		"RC" },
656 	{ "__GFP_MOVABLE",		"M" },
657 	{ "__GFP_ACCOUNT",		"AC" },
658 	{ "__GFP_WRITE",		"WR" },
659 	{ "__GFP_RECLAIM",		"R" },
660 	{ "__GFP_DIRECT_RECLAIM",	"DR" },
661 	{ "__GFP_KSWAPD_RECLAIM",	"KR" },
662 };
663 
664 static size_t max_gfp_len;
665 
666 static char *compact_gfp_flags(char *gfp_flags)
667 {
668 	char *orig_flags = strdup(gfp_flags);
669 	char *new_flags = NULL;
670 	char *str, *pos = NULL;
671 	size_t len = 0;
672 
673 	if (orig_flags == NULL)
674 		return NULL;
675 
676 	str = strtok_r(orig_flags, "|", &pos);
677 	while (str) {
678 		size_t i;
679 		char *new;
680 		const char *cpt;
681 
682 		for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
683 			if (strcmp(gfp_compact_table[i].original, str))
684 				continue;
685 
686 			cpt = gfp_compact_table[i].compact;
687 			new = realloc(new_flags, len + strlen(cpt) + 2);
688 			if (new == NULL) {
689 				free(new_flags);
690 				return NULL;
691 			}
692 
693 			new_flags = new;
694 
695 			if (!len) {
696 				strcpy(new_flags, cpt);
697 			} else {
698 				strcat(new_flags, "|");
699 				strcat(new_flags, cpt);
700 				len++;
701 			}
702 
703 			len += strlen(cpt);
704 		}
705 
706 		str = strtok_r(NULL, "|", &pos);
707 	}
708 
709 	if (max_gfp_len < len)
710 		max_gfp_len = len;
711 
712 	free(orig_flags);
713 	return new_flags;
714 }
715 
716 static char *compact_gfp_string(unsigned long gfp_flags)
717 {
718 	struct gfp_flag key = {
719 		.flags = gfp_flags,
720 	};
721 	struct gfp_flag *gfp;
722 
723 	gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
724 	if (gfp)
725 		return gfp->compact_str;
726 
727 	return NULL;
728 }
729 
730 static int parse_gfp_flags(struct perf_evsel *evsel, struct perf_sample *sample,
731 			   unsigned int gfp_flags)
732 {
733 	struct tep_record record = {
734 		.cpu = sample->cpu,
735 		.data = sample->raw_data,
736 		.size = sample->raw_size,
737 	};
738 	struct trace_seq seq;
739 	char *str, *pos = NULL;
740 
741 	if (nr_gfps) {
742 		struct gfp_flag key = {
743 			.flags = gfp_flags,
744 		};
745 
746 		if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
747 			return 0;
748 	}
749 
750 	trace_seq_init(&seq);
751 	tep_event_info(&seq, evsel->tp_format, &record);
752 
753 	str = strtok_r(seq.buffer, " ", &pos);
754 	while (str) {
755 		if (!strncmp(str, "gfp_flags=", 10)) {
756 			struct gfp_flag *new;
757 
758 			new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
759 			if (new == NULL)
760 				return -ENOMEM;
761 
762 			gfps = new;
763 			new += nr_gfps++;
764 
765 			new->flags = gfp_flags;
766 			new->human_readable = strdup(str + 10);
767 			new->compact_str = compact_gfp_flags(str + 10);
768 			if (!new->human_readable || !new->compact_str)
769 				return -ENOMEM;
770 
771 			qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
772 		}
773 
774 		str = strtok_r(NULL, " ", &pos);
775 	}
776 
777 	trace_seq_destroy(&seq);
778 	return 0;
779 }
780 
781 static int perf_evsel__process_page_alloc_event(struct perf_evsel *evsel,
782 						struct perf_sample *sample)
783 {
784 	u64 page;
785 	unsigned int order = perf_evsel__intval(evsel, sample, "order");
786 	unsigned int gfp_flags = perf_evsel__intval(evsel, sample, "gfp_flags");
787 	unsigned int migrate_type = perf_evsel__intval(evsel, sample,
788 						       "migratetype");
789 	u64 bytes = kmem_page_size << order;
790 	u64 callsite;
791 	struct page_stat *pstat;
792 	struct page_stat this = {
793 		.order = order,
794 		.gfp_flags = gfp_flags,
795 		.migrate_type = migrate_type,
796 	};
797 
798 	if (use_pfn)
799 		page = perf_evsel__intval(evsel, sample, "pfn");
800 	else
801 		page = perf_evsel__intval(evsel, sample, "page");
802 
803 	nr_page_allocs++;
804 	total_page_alloc_bytes += bytes;
805 
806 	if (!valid_page(page)) {
807 		nr_page_fails++;
808 		total_page_fail_bytes += bytes;
809 
810 		return 0;
811 	}
812 
813 	if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
814 		return -1;
815 
816 	callsite = find_callsite(evsel, sample);
817 
818 	/*
819 	 * This is to find the current page (with correct gfp flags and
820 	 * migrate type) at free event.
821 	 */
822 	this.page = page;
823 	pstat = page_stat__findnew_page(&this);
824 	if (pstat == NULL)
825 		return -ENOMEM;
826 
827 	pstat->nr_alloc++;
828 	pstat->alloc_bytes += bytes;
829 	pstat->callsite = callsite;
830 
831 	if (!live_page) {
832 		pstat = page_stat__findnew_alloc(&this);
833 		if (pstat == NULL)
834 			return -ENOMEM;
835 
836 		pstat->nr_alloc++;
837 		pstat->alloc_bytes += bytes;
838 		pstat->callsite = callsite;
839 	}
840 
841 	this.callsite = callsite;
842 	pstat = page_stat__findnew_caller(&this);
843 	if (pstat == NULL)
844 		return -ENOMEM;
845 
846 	pstat->nr_alloc++;
847 	pstat->alloc_bytes += bytes;
848 
849 	order_stats[order][migrate_type]++;
850 
851 	return 0;
852 }
853 
854 static int perf_evsel__process_page_free_event(struct perf_evsel *evsel,
855 						struct perf_sample *sample)
856 {
857 	u64 page;
858 	unsigned int order = perf_evsel__intval(evsel, sample, "order");
859 	u64 bytes = kmem_page_size << order;
860 	struct page_stat *pstat;
861 	struct page_stat this = {
862 		.order = order,
863 	};
864 
865 	if (use_pfn)
866 		page = perf_evsel__intval(evsel, sample, "pfn");
867 	else
868 		page = perf_evsel__intval(evsel, sample, "page");
869 
870 	nr_page_frees++;
871 	total_page_free_bytes += bytes;
872 
873 	this.page = page;
874 	pstat = page_stat__find_page(&this);
875 	if (pstat == NULL) {
876 		pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
877 			  page, order);
878 
879 		nr_page_nomatch++;
880 		total_page_nomatch_bytes += bytes;
881 
882 		return 0;
883 	}
884 
885 	this.gfp_flags = pstat->gfp_flags;
886 	this.migrate_type = pstat->migrate_type;
887 	this.callsite = pstat->callsite;
888 
889 	rb_erase(&pstat->node, &page_live_tree);
890 	free(pstat);
891 
892 	if (live_page) {
893 		order_stats[this.order][this.migrate_type]--;
894 	} else {
895 		pstat = page_stat__find_alloc(&this);
896 		if (pstat == NULL)
897 			return -ENOMEM;
898 
899 		pstat->nr_free++;
900 		pstat->free_bytes += bytes;
901 	}
902 
903 	pstat = page_stat__find_caller(&this);
904 	if (pstat == NULL)
905 		return -ENOENT;
906 
907 	pstat->nr_free++;
908 	pstat->free_bytes += bytes;
909 
910 	if (live_page) {
911 		pstat->nr_alloc--;
912 		pstat->alloc_bytes -= bytes;
913 
914 		if (pstat->nr_alloc == 0) {
915 			rb_erase(&pstat->node, &page_caller_tree);
916 			free(pstat);
917 		}
918 	}
919 
920 	return 0;
921 }
922 
923 static bool perf_kmem__skip_sample(struct perf_sample *sample)
924 {
925 	/* skip sample based on time? */
926 	if (perf_time__skip_sample(&ptime, sample->time))
927 		return true;
928 
929 	return false;
930 }
931 
932 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
933 				  struct perf_sample *sample);
934 
935 static int process_sample_event(struct perf_tool *tool __maybe_unused,
936 				union perf_event *event,
937 				struct perf_sample *sample,
938 				struct perf_evsel *evsel,
939 				struct machine *machine)
940 {
941 	int err = 0;
942 	struct thread *thread = machine__findnew_thread(machine, sample->pid,
943 							sample->tid);
944 
945 	if (thread == NULL) {
946 		pr_debug("problem processing %d event, skipping it.\n",
947 			 event->header.type);
948 		return -1;
949 	}
950 
951 	if (perf_kmem__skip_sample(sample))
952 		return 0;
953 
954 	dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
955 
956 	if (evsel->handler != NULL) {
957 		tracepoint_handler f = evsel->handler;
958 		err = f(evsel, sample);
959 	}
960 
961 	thread__put(thread);
962 
963 	return err;
964 }
965 
966 static struct perf_tool perf_kmem = {
967 	.sample		 = process_sample_event,
968 	.comm		 = perf_event__process_comm,
969 	.mmap		 = perf_event__process_mmap,
970 	.mmap2		 = perf_event__process_mmap2,
971 	.namespaces	 = perf_event__process_namespaces,
972 	.ordered_events	 = true,
973 };
974 
975 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
976 {
977 	if (n_alloc == 0)
978 		return 0.0;
979 	else
980 		return 100.0 - (100.0 * n_req / n_alloc);
981 }
982 
983 static void __print_slab_result(struct rb_root *root,
984 				struct perf_session *session,
985 				int n_lines, int is_caller)
986 {
987 	struct rb_node *next;
988 	struct machine *machine = &session->machines.host;
989 
990 	printf("%.105s\n", graph_dotted_line);
991 	printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
992 	printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
993 	printf("%.105s\n", graph_dotted_line);
994 
995 	next = rb_first(root);
996 
997 	while (next && n_lines--) {
998 		struct alloc_stat *data = rb_entry(next, struct alloc_stat,
999 						   node);
1000 		struct symbol *sym = NULL;
1001 		struct map *map;
1002 		char buf[BUFSIZ];
1003 		u64 addr;
1004 
1005 		if (is_caller) {
1006 			addr = data->call_site;
1007 			if (!raw_ip)
1008 				sym = machine__find_kernel_symbol(machine, addr, &map);
1009 		} else
1010 			addr = data->ptr;
1011 
1012 		if (sym != NULL)
1013 			snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1014 				 addr - map->unmap_ip(map, sym->start));
1015 		else
1016 			snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1017 		printf(" %-34s |", buf);
1018 
1019 		printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1020 		       (unsigned long long)data->bytes_alloc,
1021 		       (unsigned long)data->bytes_alloc / data->hit,
1022 		       (unsigned long long)data->bytes_req,
1023 		       (unsigned long)data->bytes_req / data->hit,
1024 		       (unsigned long)data->hit,
1025 		       (unsigned long)data->pingpong,
1026 		       fragmentation(data->bytes_req, data->bytes_alloc));
1027 
1028 		next = rb_next(next);
1029 	}
1030 
1031 	if (n_lines == -1)
1032 		printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
1033 
1034 	printf("%.105s\n", graph_dotted_line);
1035 }
1036 
1037 static const char * const migrate_type_str[] = {
1038 	"UNMOVABL",
1039 	"RECLAIM",
1040 	"MOVABLE",
1041 	"RESERVED",
1042 	"CMA/ISLT",
1043 	"UNKNOWN",
1044 };
1045 
1046 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1047 {
1048 	struct rb_node *next = rb_first(&page_alloc_sorted);
1049 	struct machine *machine = &session->machines.host;
1050 	const char *format;
1051 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1052 
1053 	printf("\n%.105s\n", graph_dotted_line);
1054 	printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1055 	       use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1056 	       gfp_len, "GFP flags");
1057 	printf("%.105s\n", graph_dotted_line);
1058 
1059 	if (use_pfn)
1060 		format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1061 	else
1062 		format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1063 
1064 	while (next && n_lines--) {
1065 		struct page_stat *data;
1066 		struct symbol *sym;
1067 		struct map *map;
1068 		char buf[32];
1069 		char *caller = buf;
1070 
1071 		data = rb_entry(next, struct page_stat, node);
1072 		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1073 		if (sym)
1074 			caller = sym->name;
1075 		else
1076 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1077 
1078 		printf(format, (unsigned long long)data->page,
1079 		       (unsigned long long)data->alloc_bytes / 1024,
1080 		       data->nr_alloc, data->order,
1081 		       migrate_type_str[data->migrate_type],
1082 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1083 
1084 		next = rb_next(next);
1085 	}
1086 
1087 	if (n_lines == -1) {
1088 		printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1089 		       gfp_len, "...");
1090 	}
1091 
1092 	printf("%.105s\n", graph_dotted_line);
1093 }
1094 
1095 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1096 {
1097 	struct rb_node *next = rb_first(&page_caller_sorted);
1098 	struct machine *machine = &session->machines.host;
1099 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1100 
1101 	printf("\n%.105s\n", graph_dotted_line);
1102 	printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1103 	       live_page ? "Live" : "Total", gfp_len, "GFP flags");
1104 	printf("%.105s\n", graph_dotted_line);
1105 
1106 	while (next && n_lines--) {
1107 		struct page_stat *data;
1108 		struct symbol *sym;
1109 		struct map *map;
1110 		char buf[32];
1111 		char *caller = buf;
1112 
1113 		data = rb_entry(next, struct page_stat, node);
1114 		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1115 		if (sym)
1116 			caller = sym->name;
1117 		else
1118 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1119 
1120 		printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1121 		       (unsigned long long)data->alloc_bytes / 1024,
1122 		       data->nr_alloc, data->order,
1123 		       migrate_type_str[data->migrate_type],
1124 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1125 
1126 		next = rb_next(next);
1127 	}
1128 
1129 	if (n_lines == -1) {
1130 		printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1131 		       gfp_len, "...");
1132 	}
1133 
1134 	printf("%.105s\n", graph_dotted_line);
1135 }
1136 
1137 static void print_gfp_flags(void)
1138 {
1139 	int i;
1140 
1141 	printf("#\n");
1142 	printf("# GFP flags\n");
1143 	printf("# ---------\n");
1144 	for (i = 0; i < nr_gfps; i++) {
1145 		printf("# %08x: %*s: %s\n", gfps[i].flags,
1146 		       (int) max_gfp_len, gfps[i].compact_str,
1147 		       gfps[i].human_readable);
1148 	}
1149 }
1150 
1151 static void print_slab_summary(void)
1152 {
1153 	printf("\nSUMMARY (SLAB allocator)");
1154 	printf("\n========================\n");
1155 	printf("Total bytes requested: %'lu\n", total_requested);
1156 	printf("Total bytes allocated: %'lu\n", total_allocated);
1157 	printf("Total bytes freed:     %'lu\n", total_freed);
1158 	if (total_allocated > total_freed) {
1159 		printf("Net total bytes allocated: %'lu\n",
1160 		total_allocated - total_freed);
1161 	}
1162 	printf("Total bytes wasted on internal fragmentation: %'lu\n",
1163 	       total_allocated - total_requested);
1164 	printf("Internal fragmentation: %f%%\n",
1165 	       fragmentation(total_requested, total_allocated));
1166 	printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1167 }
1168 
1169 static void print_page_summary(void)
1170 {
1171 	int o, m;
1172 	u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1173 	u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1174 
1175 	printf("\nSUMMARY (page allocator)");
1176 	printf("\n========================\n");
1177 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1178 	       nr_page_allocs, total_page_alloc_bytes / 1024);
1179 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1180 	       nr_page_frees, total_page_free_bytes / 1024);
1181 	printf("\n");
1182 
1183 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1184 	       nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1185 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1186 	       nr_page_allocs - nr_alloc_freed,
1187 	       (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1188 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1189 	       nr_page_nomatch, total_page_nomatch_bytes / 1024);
1190 	printf("\n");
1191 
1192 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1193 	       nr_page_fails, total_page_fail_bytes / 1024);
1194 	printf("\n");
1195 
1196 	printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1197 	       "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1198 	printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1199 	       graph_dotted_line, graph_dotted_line, graph_dotted_line,
1200 	       graph_dotted_line, graph_dotted_line);
1201 
1202 	for (o = 0; o < MAX_PAGE_ORDER; o++) {
1203 		printf("%5d", o);
1204 		for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1205 			if (order_stats[o][m])
1206 				printf("  %'12d", order_stats[o][m]);
1207 			else
1208 				printf("  %12c", '.');
1209 		}
1210 		printf("\n");
1211 	}
1212 }
1213 
1214 static void print_slab_result(struct perf_session *session)
1215 {
1216 	if (caller_flag)
1217 		__print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1218 	if (alloc_flag)
1219 		__print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1220 	print_slab_summary();
1221 }
1222 
1223 static void print_page_result(struct perf_session *session)
1224 {
1225 	if (caller_flag || alloc_flag)
1226 		print_gfp_flags();
1227 	if (caller_flag)
1228 		__print_page_caller_result(session, caller_lines);
1229 	if (alloc_flag)
1230 		__print_page_alloc_result(session, alloc_lines);
1231 	print_page_summary();
1232 }
1233 
1234 static void print_result(struct perf_session *session)
1235 {
1236 	if (kmem_slab)
1237 		print_slab_result(session);
1238 	if (kmem_page)
1239 		print_page_result(session);
1240 }
1241 
1242 static LIST_HEAD(slab_caller_sort);
1243 static LIST_HEAD(slab_alloc_sort);
1244 static LIST_HEAD(page_caller_sort);
1245 static LIST_HEAD(page_alloc_sort);
1246 
1247 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1248 			     struct list_head *sort_list)
1249 {
1250 	struct rb_node **new = &(root->rb_node);
1251 	struct rb_node *parent = NULL;
1252 	struct sort_dimension *sort;
1253 
1254 	while (*new) {
1255 		struct alloc_stat *this;
1256 		int cmp = 0;
1257 
1258 		this = rb_entry(*new, struct alloc_stat, node);
1259 		parent = *new;
1260 
1261 		list_for_each_entry(sort, sort_list, list) {
1262 			cmp = sort->cmp(data, this);
1263 			if (cmp)
1264 				break;
1265 		}
1266 
1267 		if (cmp > 0)
1268 			new = &((*new)->rb_left);
1269 		else
1270 			new = &((*new)->rb_right);
1271 	}
1272 
1273 	rb_link_node(&data->node, parent, new);
1274 	rb_insert_color(&data->node, root);
1275 }
1276 
1277 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1278 			       struct list_head *sort_list)
1279 {
1280 	struct rb_node *node;
1281 	struct alloc_stat *data;
1282 
1283 	for (;;) {
1284 		node = rb_first(root);
1285 		if (!node)
1286 			break;
1287 
1288 		rb_erase(node, root);
1289 		data = rb_entry(node, struct alloc_stat, node);
1290 		sort_slab_insert(root_sorted, data, sort_list);
1291 	}
1292 }
1293 
1294 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1295 			     struct list_head *sort_list)
1296 {
1297 	struct rb_node **new = &root->rb_node;
1298 	struct rb_node *parent = NULL;
1299 	struct sort_dimension *sort;
1300 
1301 	while (*new) {
1302 		struct page_stat *this;
1303 		int cmp = 0;
1304 
1305 		this = rb_entry(*new, struct page_stat, node);
1306 		parent = *new;
1307 
1308 		list_for_each_entry(sort, sort_list, list) {
1309 			cmp = sort->cmp(data, this);
1310 			if (cmp)
1311 				break;
1312 		}
1313 
1314 		if (cmp > 0)
1315 			new = &parent->rb_left;
1316 		else
1317 			new = &parent->rb_right;
1318 	}
1319 
1320 	rb_link_node(&data->node, parent, new);
1321 	rb_insert_color(&data->node, root);
1322 }
1323 
1324 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1325 			       struct list_head *sort_list)
1326 {
1327 	struct rb_node *node;
1328 	struct page_stat *data;
1329 
1330 	for (;;) {
1331 		node = rb_first(root);
1332 		if (!node)
1333 			break;
1334 
1335 		rb_erase(node, root);
1336 		data = rb_entry(node, struct page_stat, node);
1337 		sort_page_insert(root_sorted, data, sort_list);
1338 	}
1339 }
1340 
1341 static void sort_result(void)
1342 {
1343 	if (kmem_slab) {
1344 		__sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1345 				   &slab_alloc_sort);
1346 		__sort_slab_result(&root_caller_stat, &root_caller_sorted,
1347 				   &slab_caller_sort);
1348 	}
1349 	if (kmem_page) {
1350 		if (live_page)
1351 			__sort_page_result(&page_live_tree, &page_alloc_sorted,
1352 					   &page_alloc_sort);
1353 		else
1354 			__sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1355 					   &page_alloc_sort);
1356 
1357 		__sort_page_result(&page_caller_tree, &page_caller_sorted,
1358 				   &page_caller_sort);
1359 	}
1360 }
1361 
1362 static int __cmd_kmem(struct perf_session *session)
1363 {
1364 	int err = -EINVAL;
1365 	struct perf_evsel *evsel;
1366 	const struct perf_evsel_str_handler kmem_tracepoints[] = {
1367 		/* slab allocator */
1368 		{ "kmem:kmalloc",		perf_evsel__process_alloc_event, },
1369     		{ "kmem:kmem_cache_alloc",	perf_evsel__process_alloc_event, },
1370 		{ "kmem:kmalloc_node",		perf_evsel__process_alloc_node_event, },
1371     		{ "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
1372 		{ "kmem:kfree",			perf_evsel__process_free_event, },
1373     		{ "kmem:kmem_cache_free",	perf_evsel__process_free_event, },
1374 		/* page allocator */
1375 		{ "kmem:mm_page_alloc",		perf_evsel__process_page_alloc_event, },
1376 		{ "kmem:mm_page_free",		perf_evsel__process_page_free_event, },
1377 	};
1378 
1379 	if (!perf_session__has_traces(session, "kmem record"))
1380 		goto out;
1381 
1382 	if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1383 		pr_err("Initializing perf session tracepoint handlers failed\n");
1384 		goto out;
1385 	}
1386 
1387 	evlist__for_each_entry(session->evlist, evsel) {
1388 		if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") &&
1389 		    perf_evsel__field(evsel, "pfn")) {
1390 			use_pfn = true;
1391 			break;
1392 		}
1393 	}
1394 
1395 	setup_pager();
1396 	err = perf_session__process_events(session);
1397 	if (err != 0) {
1398 		pr_err("error during process events: %d\n", err);
1399 		goto out;
1400 	}
1401 	sort_result();
1402 	print_result(session);
1403 out:
1404 	return err;
1405 }
1406 
1407 /* slab sort keys */
1408 static int ptr_cmp(void *a, void *b)
1409 {
1410 	struct alloc_stat *l = a;
1411 	struct alloc_stat *r = b;
1412 
1413 	if (l->ptr < r->ptr)
1414 		return -1;
1415 	else if (l->ptr > r->ptr)
1416 		return 1;
1417 	return 0;
1418 }
1419 
1420 static struct sort_dimension ptr_sort_dimension = {
1421 	.name	= "ptr",
1422 	.cmp	= ptr_cmp,
1423 };
1424 
1425 static int slab_callsite_cmp(void *a, void *b)
1426 {
1427 	struct alloc_stat *l = a;
1428 	struct alloc_stat *r = b;
1429 
1430 	if (l->call_site < r->call_site)
1431 		return -1;
1432 	else if (l->call_site > r->call_site)
1433 		return 1;
1434 	return 0;
1435 }
1436 
1437 static struct sort_dimension callsite_sort_dimension = {
1438 	.name	= "callsite",
1439 	.cmp	= slab_callsite_cmp,
1440 };
1441 
1442 static int hit_cmp(void *a, void *b)
1443 {
1444 	struct alloc_stat *l = a;
1445 	struct alloc_stat *r = b;
1446 
1447 	if (l->hit < r->hit)
1448 		return -1;
1449 	else if (l->hit > r->hit)
1450 		return 1;
1451 	return 0;
1452 }
1453 
1454 static struct sort_dimension hit_sort_dimension = {
1455 	.name	= "hit",
1456 	.cmp	= hit_cmp,
1457 };
1458 
1459 static int bytes_cmp(void *a, void *b)
1460 {
1461 	struct alloc_stat *l = a;
1462 	struct alloc_stat *r = b;
1463 
1464 	if (l->bytes_alloc < r->bytes_alloc)
1465 		return -1;
1466 	else if (l->bytes_alloc > r->bytes_alloc)
1467 		return 1;
1468 	return 0;
1469 }
1470 
1471 static struct sort_dimension bytes_sort_dimension = {
1472 	.name	= "bytes",
1473 	.cmp	= bytes_cmp,
1474 };
1475 
1476 static int frag_cmp(void *a, void *b)
1477 {
1478 	double x, y;
1479 	struct alloc_stat *l = a;
1480 	struct alloc_stat *r = b;
1481 
1482 	x = fragmentation(l->bytes_req, l->bytes_alloc);
1483 	y = fragmentation(r->bytes_req, r->bytes_alloc);
1484 
1485 	if (x < y)
1486 		return -1;
1487 	else if (x > y)
1488 		return 1;
1489 	return 0;
1490 }
1491 
1492 static struct sort_dimension frag_sort_dimension = {
1493 	.name	= "frag",
1494 	.cmp	= frag_cmp,
1495 };
1496 
1497 static int pingpong_cmp(void *a, void *b)
1498 {
1499 	struct alloc_stat *l = a;
1500 	struct alloc_stat *r = b;
1501 
1502 	if (l->pingpong < r->pingpong)
1503 		return -1;
1504 	else if (l->pingpong > r->pingpong)
1505 		return 1;
1506 	return 0;
1507 }
1508 
1509 static struct sort_dimension pingpong_sort_dimension = {
1510 	.name	= "pingpong",
1511 	.cmp	= pingpong_cmp,
1512 };
1513 
1514 /* page sort keys */
1515 static int page_cmp(void *a, void *b)
1516 {
1517 	struct page_stat *l = a;
1518 	struct page_stat *r = b;
1519 
1520 	if (l->page < r->page)
1521 		return -1;
1522 	else if (l->page > r->page)
1523 		return 1;
1524 	return 0;
1525 }
1526 
1527 static struct sort_dimension page_sort_dimension = {
1528 	.name	= "page",
1529 	.cmp	= page_cmp,
1530 };
1531 
1532 static int page_callsite_cmp(void *a, void *b)
1533 {
1534 	struct page_stat *l = a;
1535 	struct page_stat *r = b;
1536 
1537 	if (l->callsite < r->callsite)
1538 		return -1;
1539 	else if (l->callsite > r->callsite)
1540 		return 1;
1541 	return 0;
1542 }
1543 
1544 static struct sort_dimension page_callsite_sort_dimension = {
1545 	.name	= "callsite",
1546 	.cmp	= page_callsite_cmp,
1547 };
1548 
1549 static int page_hit_cmp(void *a, void *b)
1550 {
1551 	struct page_stat *l = a;
1552 	struct page_stat *r = b;
1553 
1554 	if (l->nr_alloc < r->nr_alloc)
1555 		return -1;
1556 	else if (l->nr_alloc > r->nr_alloc)
1557 		return 1;
1558 	return 0;
1559 }
1560 
1561 static struct sort_dimension page_hit_sort_dimension = {
1562 	.name	= "hit",
1563 	.cmp	= page_hit_cmp,
1564 };
1565 
1566 static int page_bytes_cmp(void *a, void *b)
1567 {
1568 	struct page_stat *l = a;
1569 	struct page_stat *r = b;
1570 
1571 	if (l->alloc_bytes < r->alloc_bytes)
1572 		return -1;
1573 	else if (l->alloc_bytes > r->alloc_bytes)
1574 		return 1;
1575 	return 0;
1576 }
1577 
1578 static struct sort_dimension page_bytes_sort_dimension = {
1579 	.name	= "bytes",
1580 	.cmp	= page_bytes_cmp,
1581 };
1582 
1583 static int page_order_cmp(void *a, void *b)
1584 {
1585 	struct page_stat *l = a;
1586 	struct page_stat *r = b;
1587 
1588 	if (l->order < r->order)
1589 		return -1;
1590 	else if (l->order > r->order)
1591 		return 1;
1592 	return 0;
1593 }
1594 
1595 static struct sort_dimension page_order_sort_dimension = {
1596 	.name	= "order",
1597 	.cmp	= page_order_cmp,
1598 };
1599 
1600 static int migrate_type_cmp(void *a, void *b)
1601 {
1602 	struct page_stat *l = a;
1603 	struct page_stat *r = b;
1604 
1605 	/* for internal use to find free'd page */
1606 	if (l->migrate_type == -1U)
1607 		return 0;
1608 
1609 	if (l->migrate_type < r->migrate_type)
1610 		return -1;
1611 	else if (l->migrate_type > r->migrate_type)
1612 		return 1;
1613 	return 0;
1614 }
1615 
1616 static struct sort_dimension migrate_type_sort_dimension = {
1617 	.name	= "migtype",
1618 	.cmp	= migrate_type_cmp,
1619 };
1620 
1621 static int gfp_flags_cmp(void *a, void *b)
1622 {
1623 	struct page_stat *l = a;
1624 	struct page_stat *r = b;
1625 
1626 	/* for internal use to find free'd page */
1627 	if (l->gfp_flags == -1U)
1628 		return 0;
1629 
1630 	if (l->gfp_flags < r->gfp_flags)
1631 		return -1;
1632 	else if (l->gfp_flags > r->gfp_flags)
1633 		return 1;
1634 	return 0;
1635 }
1636 
1637 static struct sort_dimension gfp_flags_sort_dimension = {
1638 	.name	= "gfp",
1639 	.cmp	= gfp_flags_cmp,
1640 };
1641 
1642 static struct sort_dimension *slab_sorts[] = {
1643 	&ptr_sort_dimension,
1644 	&callsite_sort_dimension,
1645 	&hit_sort_dimension,
1646 	&bytes_sort_dimension,
1647 	&frag_sort_dimension,
1648 	&pingpong_sort_dimension,
1649 };
1650 
1651 static struct sort_dimension *page_sorts[] = {
1652 	&page_sort_dimension,
1653 	&page_callsite_sort_dimension,
1654 	&page_hit_sort_dimension,
1655 	&page_bytes_sort_dimension,
1656 	&page_order_sort_dimension,
1657 	&migrate_type_sort_dimension,
1658 	&gfp_flags_sort_dimension,
1659 };
1660 
1661 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1662 {
1663 	struct sort_dimension *sort;
1664 	int i;
1665 
1666 	for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1667 		if (!strcmp(slab_sorts[i]->name, tok)) {
1668 			sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1669 			if (!sort) {
1670 				pr_err("%s: memdup failed\n", __func__);
1671 				return -1;
1672 			}
1673 			list_add_tail(&sort->list, list);
1674 			return 0;
1675 		}
1676 	}
1677 
1678 	return -1;
1679 }
1680 
1681 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1682 {
1683 	struct sort_dimension *sort;
1684 	int i;
1685 
1686 	for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1687 		if (!strcmp(page_sorts[i]->name, tok)) {
1688 			sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1689 			if (!sort) {
1690 				pr_err("%s: memdup failed\n", __func__);
1691 				return -1;
1692 			}
1693 			list_add_tail(&sort->list, list);
1694 			return 0;
1695 		}
1696 	}
1697 
1698 	return -1;
1699 }
1700 
1701 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1702 {
1703 	char *tok;
1704 	char *str = strdup(arg);
1705 	char *pos = str;
1706 
1707 	if (!str) {
1708 		pr_err("%s: strdup failed\n", __func__);
1709 		return -1;
1710 	}
1711 
1712 	while (true) {
1713 		tok = strsep(&pos, ",");
1714 		if (!tok)
1715 			break;
1716 		if (slab_sort_dimension__add(tok, sort_list) < 0) {
1717 			pr_err("Unknown slab --sort key: '%s'", tok);
1718 			free(str);
1719 			return -1;
1720 		}
1721 	}
1722 
1723 	free(str);
1724 	return 0;
1725 }
1726 
1727 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1728 {
1729 	char *tok;
1730 	char *str = strdup(arg);
1731 	char *pos = str;
1732 
1733 	if (!str) {
1734 		pr_err("%s: strdup failed\n", __func__);
1735 		return -1;
1736 	}
1737 
1738 	while (true) {
1739 		tok = strsep(&pos, ",");
1740 		if (!tok)
1741 			break;
1742 		if (page_sort_dimension__add(tok, sort_list) < 0) {
1743 			pr_err("Unknown page --sort key: '%s'", tok);
1744 			free(str);
1745 			return -1;
1746 		}
1747 	}
1748 
1749 	free(str);
1750 	return 0;
1751 }
1752 
1753 static int parse_sort_opt(const struct option *opt __maybe_unused,
1754 			  const char *arg, int unset __maybe_unused)
1755 {
1756 	if (!arg)
1757 		return -1;
1758 
1759 	if (kmem_page > kmem_slab ||
1760 	    (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1761 		if (caller_flag > alloc_flag)
1762 			return setup_page_sorting(&page_caller_sort, arg);
1763 		else
1764 			return setup_page_sorting(&page_alloc_sort, arg);
1765 	} else {
1766 		if (caller_flag > alloc_flag)
1767 			return setup_slab_sorting(&slab_caller_sort, arg);
1768 		else
1769 			return setup_slab_sorting(&slab_alloc_sort, arg);
1770 	}
1771 
1772 	return 0;
1773 }
1774 
1775 static int parse_caller_opt(const struct option *opt __maybe_unused,
1776 			    const char *arg __maybe_unused,
1777 			    int unset __maybe_unused)
1778 {
1779 	caller_flag = (alloc_flag + 1);
1780 	return 0;
1781 }
1782 
1783 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1784 			   const char *arg __maybe_unused,
1785 			   int unset __maybe_unused)
1786 {
1787 	alloc_flag = (caller_flag + 1);
1788 	return 0;
1789 }
1790 
1791 static int parse_slab_opt(const struct option *opt __maybe_unused,
1792 			  const char *arg __maybe_unused,
1793 			  int unset __maybe_unused)
1794 {
1795 	kmem_slab = (kmem_page + 1);
1796 	return 0;
1797 }
1798 
1799 static int parse_page_opt(const struct option *opt __maybe_unused,
1800 			  const char *arg __maybe_unused,
1801 			  int unset __maybe_unused)
1802 {
1803 	kmem_page = (kmem_slab + 1);
1804 	return 0;
1805 }
1806 
1807 static int parse_line_opt(const struct option *opt __maybe_unused,
1808 			  const char *arg, int unset __maybe_unused)
1809 {
1810 	int lines;
1811 
1812 	if (!arg)
1813 		return -1;
1814 
1815 	lines = strtoul(arg, NULL, 10);
1816 
1817 	if (caller_flag > alloc_flag)
1818 		caller_lines = lines;
1819 	else
1820 		alloc_lines = lines;
1821 
1822 	return 0;
1823 }
1824 
1825 static int __cmd_record(int argc, const char **argv)
1826 {
1827 	const char * const record_args[] = {
1828 	"record", "-a", "-R", "-c", "1",
1829 	};
1830 	const char * const slab_events[] = {
1831 	"-e", "kmem:kmalloc",
1832 	"-e", "kmem:kmalloc_node",
1833 	"-e", "kmem:kfree",
1834 	"-e", "kmem:kmem_cache_alloc",
1835 	"-e", "kmem:kmem_cache_alloc_node",
1836 	"-e", "kmem:kmem_cache_free",
1837 	};
1838 	const char * const page_events[] = {
1839 	"-e", "kmem:mm_page_alloc",
1840 	"-e", "kmem:mm_page_free",
1841 	};
1842 	unsigned int rec_argc, i, j;
1843 	const char **rec_argv;
1844 
1845 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1846 	if (kmem_slab)
1847 		rec_argc += ARRAY_SIZE(slab_events);
1848 	if (kmem_page)
1849 		rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1850 
1851 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1852 
1853 	if (rec_argv == NULL)
1854 		return -ENOMEM;
1855 
1856 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1857 		rec_argv[i] = strdup(record_args[i]);
1858 
1859 	if (kmem_slab) {
1860 		for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1861 			rec_argv[i] = strdup(slab_events[j]);
1862 	}
1863 	if (kmem_page) {
1864 		rec_argv[i++] = strdup("-g");
1865 
1866 		for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1867 			rec_argv[i] = strdup(page_events[j]);
1868 	}
1869 
1870 	for (j = 1; j < (unsigned int)argc; j++, i++)
1871 		rec_argv[i] = argv[j];
1872 
1873 	return cmd_record(i, rec_argv);
1874 }
1875 
1876 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1877 {
1878 	if (!strcmp(var, "kmem.default")) {
1879 		if (!strcmp(value, "slab"))
1880 			kmem_default = KMEM_SLAB;
1881 		else if (!strcmp(value, "page"))
1882 			kmem_default = KMEM_PAGE;
1883 		else
1884 			pr_err("invalid default value ('slab' or 'page' required): %s\n",
1885 			       value);
1886 		return 0;
1887 	}
1888 
1889 	return 0;
1890 }
1891 
1892 int cmd_kmem(int argc, const char **argv)
1893 {
1894 	const char * const default_slab_sort = "frag,hit,bytes";
1895 	const char * const default_page_sort = "bytes,hit";
1896 	struct perf_data data = {
1897 		.mode = PERF_DATA_MODE_READ,
1898 	};
1899 	const struct option kmem_options[] = {
1900 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1901 	OPT_INCR('v', "verbose", &verbose,
1902 		    "be more verbose (show symbol address, etc)"),
1903 	OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1904 			   "show per-callsite statistics", parse_caller_opt),
1905 	OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1906 			   "show per-allocation statistics", parse_alloc_opt),
1907 	OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1908 		     "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1909 		     "page, order, migtype, gfp", parse_sort_opt),
1910 	OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1911 	OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1912 	OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
1913 	OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1914 			   parse_slab_opt),
1915 	OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1916 			   parse_page_opt),
1917 	OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1918 	OPT_STRING(0, "time", &time_str, "str",
1919 		   "Time span of interest (start,stop)"),
1920 	OPT_END()
1921 	};
1922 	const char *const kmem_subcommands[] = { "record", "stat", NULL };
1923 	const char *kmem_usage[] = {
1924 		NULL,
1925 		NULL
1926 	};
1927 	struct perf_session *session;
1928 	static const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1929 	int ret = perf_config(kmem_config, NULL);
1930 
1931 	if (ret)
1932 		return ret;
1933 
1934 	argc = parse_options_subcommand(argc, argv, kmem_options,
1935 					kmem_subcommands, kmem_usage, 0);
1936 
1937 	if (!argc)
1938 		usage_with_options(kmem_usage, kmem_options);
1939 
1940 	if (kmem_slab == 0 && kmem_page == 0) {
1941 		if (kmem_default == KMEM_SLAB)
1942 			kmem_slab = 1;
1943 		else
1944 			kmem_page = 1;
1945 	}
1946 
1947 	if (!strncmp(argv[0], "rec", 3)) {
1948 		symbol__init(NULL);
1949 		return __cmd_record(argc, argv);
1950 	}
1951 
1952 	data.path = input_name;
1953 
1954 	kmem_session = session = perf_session__new(&data, false, &perf_kmem);
1955 	if (session == NULL)
1956 		return -1;
1957 
1958 	ret = -1;
1959 
1960 	if (kmem_slab) {
1961 		if (!perf_evlist__find_tracepoint_by_name(session->evlist,
1962 							  "kmem:kmalloc")) {
1963 			pr_err(errmsg, "slab", "slab");
1964 			goto out_delete;
1965 		}
1966 	}
1967 
1968 	if (kmem_page) {
1969 		struct perf_evsel *evsel;
1970 
1971 		evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
1972 							     "kmem:mm_page_alloc");
1973 		if (evsel == NULL) {
1974 			pr_err(errmsg, "page", "page");
1975 			goto out_delete;
1976 		}
1977 
1978 		kmem_page_size = tep_get_page_size(evsel->tp_format->tep);
1979 		symbol_conf.use_callchain = true;
1980 	}
1981 
1982 	symbol__init(&session->header.env);
1983 
1984 	if (perf_time__parse_str(&ptime, time_str) != 0) {
1985 		pr_err("Invalid time string\n");
1986 		ret = -EINVAL;
1987 		goto out_delete;
1988 	}
1989 
1990 	if (!strcmp(argv[0], "stat")) {
1991 		setlocale(LC_ALL, "");
1992 
1993 		if (cpu__setup_cpunode_map())
1994 			goto out_delete;
1995 
1996 		if (list_empty(&slab_caller_sort))
1997 			setup_slab_sorting(&slab_caller_sort, default_slab_sort);
1998 		if (list_empty(&slab_alloc_sort))
1999 			setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
2000 		if (list_empty(&page_caller_sort))
2001 			setup_page_sorting(&page_caller_sort, default_page_sort);
2002 		if (list_empty(&page_alloc_sort))
2003 			setup_page_sorting(&page_alloc_sort, default_page_sort);
2004 
2005 		if (kmem_page) {
2006 			setup_page_sorting(&page_alloc_sort_input,
2007 					   "page,order,migtype,gfp");
2008 			setup_page_sorting(&page_caller_sort_input,
2009 					   "callsite,order,migtype,gfp");
2010 		}
2011 		ret = __cmd_kmem(session);
2012 	} else
2013 		usage_with_options(kmem_usage, kmem_options);
2014 
2015 out_delete:
2016 	perf_session__delete(session);
2017 
2018 	return ret;
2019 }
2020 
2021