xref: /linux/tools/perf/builtin-kmem.c (revision 122333d6bd229af279cdb35d1b874b71b3b9ccfb)
1 // SPDX-License-Identifier: GPL-2.0
2 #include "builtin.h"
3 
4 #include "util/dso.h"
5 #include "util/evlist.h"
6 #include "util/evsel.h"
7 #include "util/config.h"
8 #include "util/map.h"
9 #include "util/symbol.h"
10 #include "util/thread.h"
11 #include "util/header.h"
12 #include "util/session.h"
13 #include "util/tool.h"
14 #include "util/callchain.h"
15 #include "util/time-utils.h"
16 #include <linux/err.h>
17 
18 #include <subcmd/pager.h>
19 #include <subcmd/parse-options.h>
20 #include "util/trace-event.h"
21 #include "util/data.h"
22 #include "util/cpumap.h"
23 
24 #include "util/debug.h"
25 #include "util/string2.h"
26 #include "util/util.h"
27 
28 #include <linux/kernel.h>
29 #include <linux/numa.h>
30 #include <linux/rbtree.h>
31 #include <linux/string.h>
32 #include <linux/zalloc.h>
33 #include <errno.h>
34 #include <inttypes.h>
35 #include <locale.h>
36 #include <regex.h>
37 
38 #include <linux/ctype.h>
39 #include <traceevent/event-parse.h>
40 
41 static int	kmem_slab;
42 static int	kmem_page;
43 
44 static long	kmem_page_size;
45 static enum {
46 	KMEM_SLAB,
47 	KMEM_PAGE,
48 } kmem_default = KMEM_SLAB;  /* for backward compatibility */
49 
50 struct alloc_stat;
51 typedef int (*sort_fn_t)(void *, void *);
52 
53 static int			alloc_flag;
54 static int			caller_flag;
55 
56 static int			alloc_lines = -1;
57 static int			caller_lines = -1;
58 
59 static bool			raw_ip;
60 
61 struct alloc_stat {
62 	u64	call_site;
63 	u64	ptr;
64 	u64	bytes_req;
65 	u64	bytes_alloc;
66 	u64	last_alloc;
67 	u32	hit;
68 	u32	pingpong;
69 
70 	short	alloc_cpu;
71 
72 	struct rb_node node;
73 };
74 
75 static struct rb_root root_alloc_stat;
76 static struct rb_root root_alloc_sorted;
77 static struct rb_root root_caller_stat;
78 static struct rb_root root_caller_sorted;
79 
80 static unsigned long total_requested, total_allocated, total_freed;
81 static unsigned long nr_allocs, nr_cross_allocs;
82 
83 /* filters for controlling start and stop of time of analysis */
84 static struct perf_time_interval ptime;
85 const char *time_str;
86 
87 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
88 			     int bytes_req, int bytes_alloc, int cpu)
89 {
90 	struct rb_node **node = &root_alloc_stat.rb_node;
91 	struct rb_node *parent = NULL;
92 	struct alloc_stat *data = NULL;
93 
94 	while (*node) {
95 		parent = *node;
96 		data = rb_entry(*node, struct alloc_stat, node);
97 
98 		if (ptr > data->ptr)
99 			node = &(*node)->rb_right;
100 		else if (ptr < data->ptr)
101 			node = &(*node)->rb_left;
102 		else
103 			break;
104 	}
105 
106 	if (data && data->ptr == ptr) {
107 		data->hit++;
108 		data->bytes_req += bytes_req;
109 		data->bytes_alloc += bytes_alloc;
110 	} else {
111 		data = malloc(sizeof(*data));
112 		if (!data) {
113 			pr_err("%s: malloc failed\n", __func__);
114 			return -1;
115 		}
116 		data->ptr = ptr;
117 		data->pingpong = 0;
118 		data->hit = 1;
119 		data->bytes_req = bytes_req;
120 		data->bytes_alloc = bytes_alloc;
121 
122 		rb_link_node(&data->node, parent, node);
123 		rb_insert_color(&data->node, &root_alloc_stat);
124 	}
125 	data->call_site = call_site;
126 	data->alloc_cpu = cpu;
127 	data->last_alloc = bytes_alloc;
128 
129 	return 0;
130 }
131 
132 static int insert_caller_stat(unsigned long call_site,
133 			      int bytes_req, int bytes_alloc)
134 {
135 	struct rb_node **node = &root_caller_stat.rb_node;
136 	struct rb_node *parent = NULL;
137 	struct alloc_stat *data = NULL;
138 
139 	while (*node) {
140 		parent = *node;
141 		data = rb_entry(*node, struct alloc_stat, node);
142 
143 		if (call_site > data->call_site)
144 			node = &(*node)->rb_right;
145 		else if (call_site < data->call_site)
146 			node = &(*node)->rb_left;
147 		else
148 			break;
149 	}
150 
151 	if (data && data->call_site == call_site) {
152 		data->hit++;
153 		data->bytes_req += bytes_req;
154 		data->bytes_alloc += bytes_alloc;
155 	} else {
156 		data = malloc(sizeof(*data));
157 		if (!data) {
158 			pr_err("%s: malloc failed\n", __func__);
159 			return -1;
160 		}
161 		data->call_site = call_site;
162 		data->pingpong = 0;
163 		data->hit = 1;
164 		data->bytes_req = bytes_req;
165 		data->bytes_alloc = bytes_alloc;
166 
167 		rb_link_node(&data->node, parent, node);
168 		rb_insert_color(&data->node, &root_caller_stat);
169 	}
170 
171 	return 0;
172 }
173 
174 static int evsel__process_alloc_event(struct evsel *evsel, struct perf_sample *sample)
175 {
176 	unsigned long ptr = evsel__intval(evsel, sample, "ptr"),
177 		      call_site = evsel__intval(evsel, sample, "call_site");
178 	int bytes_req = evsel__intval(evsel, sample, "bytes_req"),
179 	    bytes_alloc = evsel__intval(evsel, sample, "bytes_alloc");
180 
181 	if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
182 	    insert_caller_stat(call_site, bytes_req, bytes_alloc))
183 		return -1;
184 
185 	total_requested += bytes_req;
186 	total_allocated += bytes_alloc;
187 
188 	nr_allocs++;
189 
190 	/*
191 	 * Commit 11e9734bcb6a ("mm/slab_common: unify NUMA and UMA
192 	 * version of tracepoints") adds the field "node" into the
193 	 * tracepoints 'kmalloc' and 'kmem_cache_alloc'.
194 	 *
195 	 * The legacy tracepoints 'kmalloc_node' and 'kmem_cache_alloc_node'
196 	 * also contain the field "node".
197 	 *
198 	 * If the tracepoint contains the field "node" the tool stats the
199 	 * cross allocation.
200 	 */
201 	if (evsel__field(evsel, "node")) {
202 		int node1, node2;
203 
204 		node1 = cpu__get_node((struct perf_cpu){.cpu = sample->cpu});
205 		node2 = evsel__intval(evsel, sample, "node");
206 
207 		/*
208 		 * If the field "node" is NUMA_NO_NODE (-1), we don't take it
209 		 * as a cross allocation.
210 		 */
211 		if ((node2 != NUMA_NO_NODE) && (node1 != node2))
212 			nr_cross_allocs++;
213 	}
214 
215 	return 0;
216 }
217 
218 static int ptr_cmp(void *, void *);
219 static int slab_callsite_cmp(void *, void *);
220 
221 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
222 					    unsigned long call_site,
223 					    struct rb_root *root,
224 					    sort_fn_t sort_fn)
225 {
226 	struct rb_node *node = root->rb_node;
227 	struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
228 
229 	while (node) {
230 		struct alloc_stat *data;
231 		int cmp;
232 
233 		data = rb_entry(node, struct alloc_stat, node);
234 
235 		cmp = sort_fn(&key, data);
236 		if (cmp < 0)
237 			node = node->rb_left;
238 		else if (cmp > 0)
239 			node = node->rb_right;
240 		else
241 			return data;
242 	}
243 	return NULL;
244 }
245 
246 static int evsel__process_free_event(struct evsel *evsel, struct perf_sample *sample)
247 {
248 	unsigned long ptr = evsel__intval(evsel, sample, "ptr");
249 	struct alloc_stat *s_alloc, *s_caller;
250 
251 	s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
252 	if (!s_alloc)
253 		return 0;
254 
255 	total_freed += s_alloc->last_alloc;
256 
257 	if ((short)sample->cpu != s_alloc->alloc_cpu) {
258 		s_alloc->pingpong++;
259 
260 		s_caller = search_alloc_stat(0, s_alloc->call_site,
261 					     &root_caller_stat,
262 					     slab_callsite_cmp);
263 		if (!s_caller)
264 			return -1;
265 		s_caller->pingpong++;
266 	}
267 	s_alloc->alloc_cpu = -1;
268 
269 	return 0;
270 }
271 
272 static u64 total_page_alloc_bytes;
273 static u64 total_page_free_bytes;
274 static u64 total_page_nomatch_bytes;
275 static u64 total_page_fail_bytes;
276 static unsigned long nr_page_allocs;
277 static unsigned long nr_page_frees;
278 static unsigned long nr_page_fails;
279 static unsigned long nr_page_nomatch;
280 
281 static bool use_pfn;
282 static bool live_page;
283 static struct perf_session *kmem_session;
284 
285 #define MAX_MIGRATE_TYPES  6
286 #define MAX_PAGE_ORDER     11
287 
288 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
289 
290 struct page_stat {
291 	struct rb_node 	node;
292 	u64 		page;
293 	u64 		callsite;
294 	int 		order;
295 	unsigned 	gfp_flags;
296 	unsigned 	migrate_type;
297 	u64		alloc_bytes;
298 	u64 		free_bytes;
299 	int 		nr_alloc;
300 	int 		nr_free;
301 };
302 
303 static struct rb_root page_live_tree;
304 static struct rb_root page_alloc_tree;
305 static struct rb_root page_alloc_sorted;
306 static struct rb_root page_caller_tree;
307 static struct rb_root page_caller_sorted;
308 
309 struct alloc_func {
310 	u64 start;
311 	u64 end;
312 	char *name;
313 };
314 
315 static int nr_alloc_funcs;
316 static struct alloc_func *alloc_func_list;
317 
318 static int funcmp(const void *a, const void *b)
319 {
320 	const struct alloc_func *fa = a;
321 	const struct alloc_func *fb = b;
322 
323 	if (fa->start > fb->start)
324 		return 1;
325 	else
326 		return -1;
327 }
328 
329 static int callcmp(const void *a, const void *b)
330 {
331 	const struct alloc_func *fa = a;
332 	const struct alloc_func *fb = b;
333 
334 	if (fb->start <= fa->start && fa->end < fb->end)
335 		return 0;
336 
337 	if (fa->start > fb->start)
338 		return 1;
339 	else
340 		return -1;
341 }
342 
343 static int build_alloc_func_list(void)
344 {
345 	int ret;
346 	struct map *kernel_map;
347 	struct symbol *sym;
348 	struct rb_node *node;
349 	struct alloc_func *func;
350 	struct machine *machine = &kmem_session->machines.host;
351 	regex_t alloc_func_regex;
352 	static const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
353 
354 	ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
355 	if (ret) {
356 		char err[BUFSIZ];
357 
358 		regerror(ret, &alloc_func_regex, err, sizeof(err));
359 		pr_err("Invalid regex: %s\n%s", pattern, err);
360 		return -EINVAL;
361 	}
362 
363 	kernel_map = machine__kernel_map(machine);
364 	if (map__load(kernel_map) < 0) {
365 		pr_err("cannot load kernel map\n");
366 		return -ENOENT;
367 	}
368 
369 	map__for_each_symbol(kernel_map, sym, node) {
370 		if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
371 			continue;
372 
373 		func = realloc(alloc_func_list,
374 			       (nr_alloc_funcs + 1) * sizeof(*func));
375 		if (func == NULL)
376 			return -ENOMEM;
377 
378 		pr_debug("alloc func: %s\n", sym->name);
379 		func[nr_alloc_funcs].start = sym->start;
380 		func[nr_alloc_funcs].end   = sym->end;
381 		func[nr_alloc_funcs].name  = sym->name;
382 
383 		alloc_func_list = func;
384 		nr_alloc_funcs++;
385 	}
386 
387 	qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
388 
389 	regfree(&alloc_func_regex);
390 	return 0;
391 }
392 
393 /*
394  * Find first non-memory allocation function from callchain.
395  * The allocation functions are in the 'alloc_func_list'.
396  */
397 static u64 find_callsite(struct evsel *evsel, struct perf_sample *sample)
398 {
399 	struct addr_location al;
400 	struct machine *machine = &kmem_session->machines.host;
401 	struct callchain_cursor_node *node;
402 
403 	if (alloc_func_list == NULL) {
404 		if (build_alloc_func_list() < 0)
405 			goto out;
406 	}
407 
408 	al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
409 	sample__resolve_callchain(sample, &callchain_cursor, NULL, evsel, &al, 16);
410 
411 	callchain_cursor_commit(&callchain_cursor);
412 	while (true) {
413 		struct alloc_func key, *caller;
414 		u64 addr;
415 
416 		node = callchain_cursor_current(&callchain_cursor);
417 		if (node == NULL)
418 			break;
419 
420 		key.start = key.end = node->ip;
421 		caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
422 				 sizeof(key), callcmp);
423 		if (!caller) {
424 			/* found */
425 			if (node->ms.map)
426 				addr = map__dso_unmap_ip(node->ms.map, node->ip);
427 			else
428 				addr = node->ip;
429 
430 			return addr;
431 		} else
432 			pr_debug3("skipping alloc function: %s\n", caller->name);
433 
434 		callchain_cursor_advance(&callchain_cursor);
435 	}
436 
437 out:
438 	pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
439 	return sample->ip;
440 }
441 
442 struct sort_dimension {
443 	const char		name[20];
444 	sort_fn_t		cmp;
445 	struct list_head	list;
446 };
447 
448 static LIST_HEAD(page_alloc_sort_input);
449 static LIST_HEAD(page_caller_sort_input);
450 
451 static struct page_stat *
452 __page_stat__findnew_page(struct page_stat *pstat, bool create)
453 {
454 	struct rb_node **node = &page_live_tree.rb_node;
455 	struct rb_node *parent = NULL;
456 	struct page_stat *data;
457 
458 	while (*node) {
459 		s64 cmp;
460 
461 		parent = *node;
462 		data = rb_entry(*node, struct page_stat, node);
463 
464 		cmp = data->page - pstat->page;
465 		if (cmp < 0)
466 			node = &parent->rb_left;
467 		else if (cmp > 0)
468 			node = &parent->rb_right;
469 		else
470 			return data;
471 	}
472 
473 	if (!create)
474 		return NULL;
475 
476 	data = zalloc(sizeof(*data));
477 	if (data != NULL) {
478 		data->page = pstat->page;
479 		data->order = pstat->order;
480 		data->gfp_flags = pstat->gfp_flags;
481 		data->migrate_type = pstat->migrate_type;
482 
483 		rb_link_node(&data->node, parent, node);
484 		rb_insert_color(&data->node, &page_live_tree);
485 	}
486 
487 	return data;
488 }
489 
490 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
491 {
492 	return __page_stat__findnew_page(pstat, false);
493 }
494 
495 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
496 {
497 	return __page_stat__findnew_page(pstat, true);
498 }
499 
500 static struct page_stat *
501 __page_stat__findnew_alloc(struct page_stat *pstat, bool create)
502 {
503 	struct rb_node **node = &page_alloc_tree.rb_node;
504 	struct rb_node *parent = NULL;
505 	struct page_stat *data;
506 	struct sort_dimension *sort;
507 
508 	while (*node) {
509 		int cmp = 0;
510 
511 		parent = *node;
512 		data = rb_entry(*node, struct page_stat, node);
513 
514 		list_for_each_entry(sort, &page_alloc_sort_input, list) {
515 			cmp = sort->cmp(pstat, data);
516 			if (cmp)
517 				break;
518 		}
519 
520 		if (cmp < 0)
521 			node = &parent->rb_left;
522 		else if (cmp > 0)
523 			node = &parent->rb_right;
524 		else
525 			return data;
526 	}
527 
528 	if (!create)
529 		return NULL;
530 
531 	data = zalloc(sizeof(*data));
532 	if (data != NULL) {
533 		data->page = pstat->page;
534 		data->order = pstat->order;
535 		data->gfp_flags = pstat->gfp_flags;
536 		data->migrate_type = pstat->migrate_type;
537 
538 		rb_link_node(&data->node, parent, node);
539 		rb_insert_color(&data->node, &page_alloc_tree);
540 	}
541 
542 	return data;
543 }
544 
545 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
546 {
547 	return __page_stat__findnew_alloc(pstat, false);
548 }
549 
550 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
551 {
552 	return __page_stat__findnew_alloc(pstat, true);
553 }
554 
555 static struct page_stat *
556 __page_stat__findnew_caller(struct page_stat *pstat, bool create)
557 {
558 	struct rb_node **node = &page_caller_tree.rb_node;
559 	struct rb_node *parent = NULL;
560 	struct page_stat *data;
561 	struct sort_dimension *sort;
562 
563 	while (*node) {
564 		int cmp = 0;
565 
566 		parent = *node;
567 		data = rb_entry(*node, struct page_stat, node);
568 
569 		list_for_each_entry(sort, &page_caller_sort_input, list) {
570 			cmp = sort->cmp(pstat, data);
571 			if (cmp)
572 				break;
573 		}
574 
575 		if (cmp < 0)
576 			node = &parent->rb_left;
577 		else if (cmp > 0)
578 			node = &parent->rb_right;
579 		else
580 			return data;
581 	}
582 
583 	if (!create)
584 		return NULL;
585 
586 	data = zalloc(sizeof(*data));
587 	if (data != NULL) {
588 		data->callsite = pstat->callsite;
589 		data->order = pstat->order;
590 		data->gfp_flags = pstat->gfp_flags;
591 		data->migrate_type = pstat->migrate_type;
592 
593 		rb_link_node(&data->node, parent, node);
594 		rb_insert_color(&data->node, &page_caller_tree);
595 	}
596 
597 	return data;
598 }
599 
600 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
601 {
602 	return __page_stat__findnew_caller(pstat, false);
603 }
604 
605 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
606 {
607 	return __page_stat__findnew_caller(pstat, true);
608 }
609 
610 static bool valid_page(u64 pfn_or_page)
611 {
612 	if (use_pfn && pfn_or_page == -1UL)
613 		return false;
614 	if (!use_pfn && pfn_or_page == 0)
615 		return false;
616 	return true;
617 }
618 
619 struct gfp_flag {
620 	unsigned int flags;
621 	char *compact_str;
622 	char *human_readable;
623 };
624 
625 static struct gfp_flag *gfps;
626 static int nr_gfps;
627 
628 static int gfpcmp(const void *a, const void *b)
629 {
630 	const struct gfp_flag *fa = a;
631 	const struct gfp_flag *fb = b;
632 
633 	return fa->flags - fb->flags;
634 }
635 
636 /* see include/trace/events/mmflags.h */
637 static const struct {
638 	const char *original;
639 	const char *compact;
640 } gfp_compact_table[] = {
641 	{ "GFP_TRANSHUGE",		"THP" },
642 	{ "GFP_TRANSHUGE_LIGHT",	"THL" },
643 	{ "GFP_HIGHUSER_MOVABLE",	"HUM" },
644 	{ "GFP_HIGHUSER",		"HU" },
645 	{ "GFP_USER",			"U" },
646 	{ "GFP_KERNEL_ACCOUNT",		"KAC" },
647 	{ "GFP_KERNEL",			"K" },
648 	{ "GFP_NOFS",			"NF" },
649 	{ "GFP_ATOMIC",			"A" },
650 	{ "GFP_NOIO",			"NI" },
651 	{ "GFP_NOWAIT",			"NW" },
652 	{ "GFP_DMA",			"D" },
653 	{ "__GFP_HIGHMEM",		"HM" },
654 	{ "GFP_DMA32",			"D32" },
655 	{ "__GFP_HIGH",			"H" },
656 	{ "__GFP_IO",			"I" },
657 	{ "__GFP_FS",			"F" },
658 	{ "__GFP_NOWARN",		"NWR" },
659 	{ "__GFP_RETRY_MAYFAIL",	"R" },
660 	{ "__GFP_NOFAIL",		"NF" },
661 	{ "__GFP_NORETRY",		"NR" },
662 	{ "__GFP_COMP",			"C" },
663 	{ "__GFP_ZERO",			"Z" },
664 	{ "__GFP_NOMEMALLOC",		"NMA" },
665 	{ "__GFP_MEMALLOC",		"MA" },
666 	{ "__GFP_HARDWALL",		"HW" },
667 	{ "__GFP_THISNODE",		"TN" },
668 	{ "__GFP_RECLAIMABLE",		"RC" },
669 	{ "__GFP_MOVABLE",		"M" },
670 	{ "__GFP_ACCOUNT",		"AC" },
671 	{ "__GFP_WRITE",		"WR" },
672 	{ "__GFP_RECLAIM",		"R" },
673 	{ "__GFP_DIRECT_RECLAIM",	"DR" },
674 	{ "__GFP_KSWAPD_RECLAIM",	"KR" },
675 };
676 
677 static size_t max_gfp_len;
678 
679 static char *compact_gfp_flags(char *gfp_flags)
680 {
681 	char *orig_flags = strdup(gfp_flags);
682 	char *new_flags = NULL;
683 	char *str, *pos = NULL;
684 	size_t len = 0;
685 
686 	if (orig_flags == NULL)
687 		return NULL;
688 
689 	str = strtok_r(orig_flags, "|", &pos);
690 	while (str) {
691 		size_t i;
692 		char *new;
693 		const char *cpt;
694 
695 		for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
696 			if (strcmp(gfp_compact_table[i].original, str))
697 				continue;
698 
699 			cpt = gfp_compact_table[i].compact;
700 			new = realloc(new_flags, len + strlen(cpt) + 2);
701 			if (new == NULL) {
702 				free(new_flags);
703 				free(orig_flags);
704 				return NULL;
705 			}
706 
707 			new_flags = new;
708 
709 			if (!len) {
710 				strcpy(new_flags, cpt);
711 			} else {
712 				strcat(new_flags, "|");
713 				strcat(new_flags, cpt);
714 				len++;
715 			}
716 
717 			len += strlen(cpt);
718 		}
719 
720 		str = strtok_r(NULL, "|", &pos);
721 	}
722 
723 	if (max_gfp_len < len)
724 		max_gfp_len = len;
725 
726 	free(orig_flags);
727 	return new_flags;
728 }
729 
730 static char *compact_gfp_string(unsigned long gfp_flags)
731 {
732 	struct gfp_flag key = {
733 		.flags = gfp_flags,
734 	};
735 	struct gfp_flag *gfp;
736 
737 	gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
738 	if (gfp)
739 		return gfp->compact_str;
740 
741 	return NULL;
742 }
743 
744 static int parse_gfp_flags(struct evsel *evsel, struct perf_sample *sample,
745 			   unsigned int gfp_flags)
746 {
747 	struct tep_record record = {
748 		.cpu = sample->cpu,
749 		.data = sample->raw_data,
750 		.size = sample->raw_size,
751 	};
752 	struct trace_seq seq;
753 	char *str, *pos = NULL;
754 
755 	if (nr_gfps) {
756 		struct gfp_flag key = {
757 			.flags = gfp_flags,
758 		};
759 
760 		if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
761 			return 0;
762 	}
763 
764 	trace_seq_init(&seq);
765 	tep_print_event(evsel->tp_format->tep,
766 			&seq, &record, "%s", TEP_PRINT_INFO);
767 
768 	str = strtok_r(seq.buffer, " ", &pos);
769 	while (str) {
770 		if (!strncmp(str, "gfp_flags=", 10)) {
771 			struct gfp_flag *new;
772 
773 			new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
774 			if (new == NULL)
775 				return -ENOMEM;
776 
777 			gfps = new;
778 			new += nr_gfps++;
779 
780 			new->flags = gfp_flags;
781 			new->human_readable = strdup(str + 10);
782 			new->compact_str = compact_gfp_flags(str + 10);
783 			if (!new->human_readable || !new->compact_str)
784 				return -ENOMEM;
785 
786 			qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
787 		}
788 
789 		str = strtok_r(NULL, " ", &pos);
790 	}
791 
792 	trace_seq_destroy(&seq);
793 	return 0;
794 }
795 
796 static int evsel__process_page_alloc_event(struct evsel *evsel, struct perf_sample *sample)
797 {
798 	u64 page;
799 	unsigned int order = evsel__intval(evsel, sample, "order");
800 	unsigned int gfp_flags = evsel__intval(evsel, sample, "gfp_flags");
801 	unsigned int migrate_type = evsel__intval(evsel, sample,
802 						       "migratetype");
803 	u64 bytes = kmem_page_size << order;
804 	u64 callsite;
805 	struct page_stat *pstat;
806 	struct page_stat this = {
807 		.order = order,
808 		.gfp_flags = gfp_flags,
809 		.migrate_type = migrate_type,
810 	};
811 
812 	if (use_pfn)
813 		page = evsel__intval(evsel, sample, "pfn");
814 	else
815 		page = evsel__intval(evsel, sample, "page");
816 
817 	nr_page_allocs++;
818 	total_page_alloc_bytes += bytes;
819 
820 	if (!valid_page(page)) {
821 		nr_page_fails++;
822 		total_page_fail_bytes += bytes;
823 
824 		return 0;
825 	}
826 
827 	if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
828 		return -1;
829 
830 	callsite = find_callsite(evsel, sample);
831 
832 	/*
833 	 * This is to find the current page (with correct gfp flags and
834 	 * migrate type) at free event.
835 	 */
836 	this.page = page;
837 	pstat = page_stat__findnew_page(&this);
838 	if (pstat == NULL)
839 		return -ENOMEM;
840 
841 	pstat->nr_alloc++;
842 	pstat->alloc_bytes += bytes;
843 	pstat->callsite = callsite;
844 
845 	if (!live_page) {
846 		pstat = page_stat__findnew_alloc(&this);
847 		if (pstat == NULL)
848 			return -ENOMEM;
849 
850 		pstat->nr_alloc++;
851 		pstat->alloc_bytes += bytes;
852 		pstat->callsite = callsite;
853 	}
854 
855 	this.callsite = callsite;
856 	pstat = page_stat__findnew_caller(&this);
857 	if (pstat == NULL)
858 		return -ENOMEM;
859 
860 	pstat->nr_alloc++;
861 	pstat->alloc_bytes += bytes;
862 
863 	order_stats[order][migrate_type]++;
864 
865 	return 0;
866 }
867 
868 static int evsel__process_page_free_event(struct evsel *evsel, struct perf_sample *sample)
869 {
870 	u64 page;
871 	unsigned int order = evsel__intval(evsel, sample, "order");
872 	u64 bytes = kmem_page_size << order;
873 	struct page_stat *pstat;
874 	struct page_stat this = {
875 		.order = order,
876 	};
877 
878 	if (use_pfn)
879 		page = evsel__intval(evsel, sample, "pfn");
880 	else
881 		page = evsel__intval(evsel, sample, "page");
882 
883 	nr_page_frees++;
884 	total_page_free_bytes += bytes;
885 
886 	this.page = page;
887 	pstat = page_stat__find_page(&this);
888 	if (pstat == NULL) {
889 		pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
890 			  page, order);
891 
892 		nr_page_nomatch++;
893 		total_page_nomatch_bytes += bytes;
894 
895 		return 0;
896 	}
897 
898 	this.gfp_flags = pstat->gfp_flags;
899 	this.migrate_type = pstat->migrate_type;
900 	this.callsite = pstat->callsite;
901 
902 	rb_erase(&pstat->node, &page_live_tree);
903 	free(pstat);
904 
905 	if (live_page) {
906 		order_stats[this.order][this.migrate_type]--;
907 	} else {
908 		pstat = page_stat__find_alloc(&this);
909 		if (pstat == NULL)
910 			return -ENOMEM;
911 
912 		pstat->nr_free++;
913 		pstat->free_bytes += bytes;
914 	}
915 
916 	pstat = page_stat__find_caller(&this);
917 	if (pstat == NULL)
918 		return -ENOENT;
919 
920 	pstat->nr_free++;
921 	pstat->free_bytes += bytes;
922 
923 	if (live_page) {
924 		pstat->nr_alloc--;
925 		pstat->alloc_bytes -= bytes;
926 
927 		if (pstat->nr_alloc == 0) {
928 			rb_erase(&pstat->node, &page_caller_tree);
929 			free(pstat);
930 		}
931 	}
932 
933 	return 0;
934 }
935 
936 static bool perf_kmem__skip_sample(struct perf_sample *sample)
937 {
938 	/* skip sample based on time? */
939 	if (perf_time__skip_sample(&ptime, sample->time))
940 		return true;
941 
942 	return false;
943 }
944 
945 typedef int (*tracepoint_handler)(struct evsel *evsel,
946 				  struct perf_sample *sample);
947 
948 static int process_sample_event(struct perf_tool *tool __maybe_unused,
949 				union perf_event *event,
950 				struct perf_sample *sample,
951 				struct evsel *evsel,
952 				struct machine *machine)
953 {
954 	int err = 0;
955 	struct thread *thread = machine__findnew_thread(machine, sample->pid,
956 							sample->tid);
957 
958 	if (thread == NULL) {
959 		pr_debug("problem processing %d event, skipping it.\n",
960 			 event->header.type);
961 		return -1;
962 	}
963 
964 	if (perf_kmem__skip_sample(sample))
965 		return 0;
966 
967 	dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
968 
969 	if (evsel->handler != NULL) {
970 		tracepoint_handler f = evsel->handler;
971 		err = f(evsel, sample);
972 	}
973 
974 	thread__put(thread);
975 
976 	return err;
977 }
978 
979 static struct perf_tool perf_kmem = {
980 	.sample		 = process_sample_event,
981 	.comm		 = perf_event__process_comm,
982 	.mmap		 = perf_event__process_mmap,
983 	.mmap2		 = perf_event__process_mmap2,
984 	.namespaces	 = perf_event__process_namespaces,
985 	.ordered_events	 = true,
986 };
987 
988 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
989 {
990 	if (n_alloc == 0)
991 		return 0.0;
992 	else
993 		return 100.0 - (100.0 * n_req / n_alloc);
994 }
995 
996 static void __print_slab_result(struct rb_root *root,
997 				struct perf_session *session,
998 				int n_lines, int is_caller)
999 {
1000 	struct rb_node *next;
1001 	struct machine *machine = &session->machines.host;
1002 
1003 	printf("%.105s\n", graph_dotted_line);
1004 	printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
1005 	printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
1006 	printf("%.105s\n", graph_dotted_line);
1007 
1008 	next = rb_first(root);
1009 
1010 	while (next && n_lines--) {
1011 		struct alloc_stat *data = rb_entry(next, struct alloc_stat,
1012 						   node);
1013 		struct symbol *sym = NULL;
1014 		struct map *map;
1015 		char buf[BUFSIZ];
1016 		u64 addr;
1017 
1018 		if (is_caller) {
1019 			addr = data->call_site;
1020 			if (!raw_ip)
1021 				sym = machine__find_kernel_symbol(machine, addr, &map);
1022 		} else
1023 			addr = data->ptr;
1024 
1025 		if (sym != NULL)
1026 			snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1027 				 addr - map__unmap_ip(map, sym->start));
1028 		else
1029 			snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1030 		printf(" %-34s |", buf);
1031 
1032 		printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1033 		       (unsigned long long)data->bytes_alloc,
1034 		       (unsigned long)data->bytes_alloc / data->hit,
1035 		       (unsigned long long)data->bytes_req,
1036 		       (unsigned long)data->bytes_req / data->hit,
1037 		       (unsigned long)data->hit,
1038 		       (unsigned long)data->pingpong,
1039 		       fragmentation(data->bytes_req, data->bytes_alloc));
1040 
1041 		next = rb_next(next);
1042 	}
1043 
1044 	if (n_lines == -1)
1045 		printf(" ...                                | ...             | ...             | ...      | ...       | ...   \n");
1046 
1047 	printf("%.105s\n", graph_dotted_line);
1048 }
1049 
1050 static const char * const migrate_type_str[] = {
1051 	"UNMOVABL",
1052 	"RECLAIM",
1053 	"MOVABLE",
1054 	"RESERVED",
1055 	"CMA/ISLT",
1056 	"UNKNOWN",
1057 };
1058 
1059 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1060 {
1061 	struct rb_node *next = rb_first(&page_alloc_sorted);
1062 	struct machine *machine = &session->machines.host;
1063 	const char *format;
1064 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1065 
1066 	printf("\n%.105s\n", graph_dotted_line);
1067 	printf(" %-16s | %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1068 	       use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1069 	       gfp_len, "GFP flags");
1070 	printf("%.105s\n", graph_dotted_line);
1071 
1072 	if (use_pfn)
1073 		format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1074 	else
1075 		format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1076 
1077 	while (next && n_lines--) {
1078 		struct page_stat *data;
1079 		struct symbol *sym;
1080 		struct map *map;
1081 		char buf[32];
1082 		char *caller = buf;
1083 
1084 		data = rb_entry(next, struct page_stat, node);
1085 		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1086 		if (sym)
1087 			caller = sym->name;
1088 		else
1089 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1090 
1091 		printf(format, (unsigned long long)data->page,
1092 		       (unsigned long long)data->alloc_bytes / 1024,
1093 		       data->nr_alloc, data->order,
1094 		       migrate_type_str[data->migrate_type],
1095 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1096 
1097 		next = rb_next(next);
1098 	}
1099 
1100 	if (n_lines == -1) {
1101 		printf(" ...              | ...              | ...       | ...   | ...      | %-*s | ...\n",
1102 		       gfp_len, "...");
1103 	}
1104 
1105 	printf("%.105s\n", graph_dotted_line);
1106 }
1107 
1108 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1109 {
1110 	struct rb_node *next = rb_first(&page_caller_sorted);
1111 	struct machine *machine = &session->machines.host;
1112 	int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1113 
1114 	printf("\n%.105s\n", graph_dotted_line);
1115 	printf(" %5s alloc (KB) | Hits      | Order | Mig.type | %-*s | Callsite\n",
1116 	       live_page ? "Live" : "Total", gfp_len, "GFP flags");
1117 	printf("%.105s\n", graph_dotted_line);
1118 
1119 	while (next && n_lines--) {
1120 		struct page_stat *data;
1121 		struct symbol *sym;
1122 		struct map *map;
1123 		char buf[32];
1124 		char *caller = buf;
1125 
1126 		data = rb_entry(next, struct page_stat, node);
1127 		sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1128 		if (sym)
1129 			caller = sym->name;
1130 		else
1131 			scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1132 
1133 		printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1134 		       (unsigned long long)data->alloc_bytes / 1024,
1135 		       data->nr_alloc, data->order,
1136 		       migrate_type_str[data->migrate_type],
1137 		       gfp_len, compact_gfp_string(data->gfp_flags), caller);
1138 
1139 		next = rb_next(next);
1140 	}
1141 
1142 	if (n_lines == -1) {
1143 		printf(" ...              | ...       | ...   | ...      | %-*s | ...\n",
1144 		       gfp_len, "...");
1145 	}
1146 
1147 	printf("%.105s\n", graph_dotted_line);
1148 }
1149 
1150 static void print_gfp_flags(void)
1151 {
1152 	int i;
1153 
1154 	printf("#\n");
1155 	printf("# GFP flags\n");
1156 	printf("# ---------\n");
1157 	for (i = 0; i < nr_gfps; i++) {
1158 		printf("# %08x: %*s: %s\n", gfps[i].flags,
1159 		       (int) max_gfp_len, gfps[i].compact_str,
1160 		       gfps[i].human_readable);
1161 	}
1162 }
1163 
1164 static void print_slab_summary(void)
1165 {
1166 	printf("\nSUMMARY (SLAB allocator)");
1167 	printf("\n========================\n");
1168 	printf("Total bytes requested: %'lu\n", total_requested);
1169 	printf("Total bytes allocated: %'lu\n", total_allocated);
1170 	printf("Total bytes freed:     %'lu\n", total_freed);
1171 	if (total_allocated > total_freed) {
1172 		printf("Net total bytes allocated: %'lu\n",
1173 		total_allocated - total_freed);
1174 	}
1175 	printf("Total bytes wasted on internal fragmentation: %'lu\n",
1176 	       total_allocated - total_requested);
1177 	printf("Internal fragmentation: %f%%\n",
1178 	       fragmentation(total_requested, total_allocated));
1179 	printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1180 }
1181 
1182 static void print_page_summary(void)
1183 {
1184 	int o, m;
1185 	u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1186 	u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1187 
1188 	printf("\nSUMMARY (page allocator)");
1189 	printf("\n========================\n");
1190 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1191 	       nr_page_allocs, total_page_alloc_bytes / 1024);
1192 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free requests",
1193 	       nr_page_frees, total_page_free_bytes / 1024);
1194 	printf("\n");
1195 
1196 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1197 	       nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1198 	printf("%-30s: %'16"PRIu64"   [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1199 	       nr_page_allocs - nr_alloc_freed,
1200 	       (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1201 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1202 	       nr_page_nomatch, total_page_nomatch_bytes / 1024);
1203 	printf("\n");
1204 
1205 	printf("%-30s: %'16lu   [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1206 	       nr_page_fails, total_page_fail_bytes / 1024);
1207 	printf("\n");
1208 
1209 	printf("%5s  %12s  %12s  %12s  %12s  %12s\n", "Order",  "Unmovable",
1210 	       "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1211 	printf("%.5s  %.12s  %.12s  %.12s  %.12s  %.12s\n", graph_dotted_line,
1212 	       graph_dotted_line, graph_dotted_line, graph_dotted_line,
1213 	       graph_dotted_line, graph_dotted_line);
1214 
1215 	for (o = 0; o < MAX_PAGE_ORDER; o++) {
1216 		printf("%5d", o);
1217 		for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1218 			if (order_stats[o][m])
1219 				printf("  %'12d", order_stats[o][m]);
1220 			else
1221 				printf("  %12c", '.');
1222 		}
1223 		printf("\n");
1224 	}
1225 }
1226 
1227 static void print_slab_result(struct perf_session *session)
1228 {
1229 	if (caller_flag)
1230 		__print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1231 	if (alloc_flag)
1232 		__print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1233 	print_slab_summary();
1234 }
1235 
1236 static void print_page_result(struct perf_session *session)
1237 {
1238 	if (caller_flag || alloc_flag)
1239 		print_gfp_flags();
1240 	if (caller_flag)
1241 		__print_page_caller_result(session, caller_lines);
1242 	if (alloc_flag)
1243 		__print_page_alloc_result(session, alloc_lines);
1244 	print_page_summary();
1245 }
1246 
1247 static void print_result(struct perf_session *session)
1248 {
1249 	if (kmem_slab)
1250 		print_slab_result(session);
1251 	if (kmem_page)
1252 		print_page_result(session);
1253 }
1254 
1255 static LIST_HEAD(slab_caller_sort);
1256 static LIST_HEAD(slab_alloc_sort);
1257 static LIST_HEAD(page_caller_sort);
1258 static LIST_HEAD(page_alloc_sort);
1259 
1260 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1261 			     struct list_head *sort_list)
1262 {
1263 	struct rb_node **new = &(root->rb_node);
1264 	struct rb_node *parent = NULL;
1265 	struct sort_dimension *sort;
1266 
1267 	while (*new) {
1268 		struct alloc_stat *this;
1269 		int cmp = 0;
1270 
1271 		this = rb_entry(*new, struct alloc_stat, node);
1272 		parent = *new;
1273 
1274 		list_for_each_entry(sort, sort_list, list) {
1275 			cmp = sort->cmp(data, this);
1276 			if (cmp)
1277 				break;
1278 		}
1279 
1280 		if (cmp > 0)
1281 			new = &((*new)->rb_left);
1282 		else
1283 			new = &((*new)->rb_right);
1284 	}
1285 
1286 	rb_link_node(&data->node, parent, new);
1287 	rb_insert_color(&data->node, root);
1288 }
1289 
1290 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1291 			       struct list_head *sort_list)
1292 {
1293 	struct rb_node *node;
1294 	struct alloc_stat *data;
1295 
1296 	for (;;) {
1297 		node = rb_first(root);
1298 		if (!node)
1299 			break;
1300 
1301 		rb_erase(node, root);
1302 		data = rb_entry(node, struct alloc_stat, node);
1303 		sort_slab_insert(root_sorted, data, sort_list);
1304 	}
1305 }
1306 
1307 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1308 			     struct list_head *sort_list)
1309 {
1310 	struct rb_node **new = &root->rb_node;
1311 	struct rb_node *parent = NULL;
1312 	struct sort_dimension *sort;
1313 
1314 	while (*new) {
1315 		struct page_stat *this;
1316 		int cmp = 0;
1317 
1318 		this = rb_entry(*new, struct page_stat, node);
1319 		parent = *new;
1320 
1321 		list_for_each_entry(sort, sort_list, list) {
1322 			cmp = sort->cmp(data, this);
1323 			if (cmp)
1324 				break;
1325 		}
1326 
1327 		if (cmp > 0)
1328 			new = &parent->rb_left;
1329 		else
1330 			new = &parent->rb_right;
1331 	}
1332 
1333 	rb_link_node(&data->node, parent, new);
1334 	rb_insert_color(&data->node, root);
1335 }
1336 
1337 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1338 			       struct list_head *sort_list)
1339 {
1340 	struct rb_node *node;
1341 	struct page_stat *data;
1342 
1343 	for (;;) {
1344 		node = rb_first(root);
1345 		if (!node)
1346 			break;
1347 
1348 		rb_erase(node, root);
1349 		data = rb_entry(node, struct page_stat, node);
1350 		sort_page_insert(root_sorted, data, sort_list);
1351 	}
1352 }
1353 
1354 static void sort_result(void)
1355 {
1356 	if (kmem_slab) {
1357 		__sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1358 				   &slab_alloc_sort);
1359 		__sort_slab_result(&root_caller_stat, &root_caller_sorted,
1360 				   &slab_caller_sort);
1361 	}
1362 	if (kmem_page) {
1363 		if (live_page)
1364 			__sort_page_result(&page_live_tree, &page_alloc_sorted,
1365 					   &page_alloc_sort);
1366 		else
1367 			__sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1368 					   &page_alloc_sort);
1369 
1370 		__sort_page_result(&page_caller_tree, &page_caller_sorted,
1371 				   &page_caller_sort);
1372 	}
1373 }
1374 
1375 static int __cmd_kmem(struct perf_session *session)
1376 {
1377 	int err = -EINVAL;
1378 	struct evsel *evsel;
1379 	const struct evsel_str_handler kmem_tracepoints[] = {
1380 		/* slab allocator */
1381 		{ "kmem:kmalloc",		evsel__process_alloc_event, },
1382 		{ "kmem:kmem_cache_alloc",	evsel__process_alloc_event, },
1383 		{ "kmem:kmalloc_node",		evsel__process_alloc_event, },
1384 		{ "kmem:kmem_cache_alloc_node", evsel__process_alloc_event, },
1385 		{ "kmem:kfree",			evsel__process_free_event, },
1386 		{ "kmem:kmem_cache_free",	evsel__process_free_event, },
1387 		/* page allocator */
1388 		{ "kmem:mm_page_alloc",		evsel__process_page_alloc_event, },
1389 		{ "kmem:mm_page_free",		evsel__process_page_free_event, },
1390 	};
1391 
1392 	if (!perf_session__has_traces(session, "kmem record"))
1393 		goto out;
1394 
1395 	if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1396 		pr_err("Initializing perf session tracepoint handlers failed\n");
1397 		goto out;
1398 	}
1399 
1400 	evlist__for_each_entry(session->evlist, evsel) {
1401 		if (!strcmp(evsel__name(evsel), "kmem:mm_page_alloc") &&
1402 		    evsel__field(evsel, "pfn")) {
1403 			use_pfn = true;
1404 			break;
1405 		}
1406 	}
1407 
1408 	setup_pager();
1409 	err = perf_session__process_events(session);
1410 	if (err != 0) {
1411 		pr_err("error during process events: %d\n", err);
1412 		goto out;
1413 	}
1414 	sort_result();
1415 	print_result(session);
1416 out:
1417 	return err;
1418 }
1419 
1420 /* slab sort keys */
1421 static int ptr_cmp(void *a, void *b)
1422 {
1423 	struct alloc_stat *l = a;
1424 	struct alloc_stat *r = b;
1425 
1426 	if (l->ptr < r->ptr)
1427 		return -1;
1428 	else if (l->ptr > r->ptr)
1429 		return 1;
1430 	return 0;
1431 }
1432 
1433 static struct sort_dimension ptr_sort_dimension = {
1434 	.name	= "ptr",
1435 	.cmp	= ptr_cmp,
1436 };
1437 
1438 static int slab_callsite_cmp(void *a, void *b)
1439 {
1440 	struct alloc_stat *l = a;
1441 	struct alloc_stat *r = b;
1442 
1443 	if (l->call_site < r->call_site)
1444 		return -1;
1445 	else if (l->call_site > r->call_site)
1446 		return 1;
1447 	return 0;
1448 }
1449 
1450 static struct sort_dimension callsite_sort_dimension = {
1451 	.name	= "callsite",
1452 	.cmp	= slab_callsite_cmp,
1453 };
1454 
1455 static int hit_cmp(void *a, void *b)
1456 {
1457 	struct alloc_stat *l = a;
1458 	struct alloc_stat *r = b;
1459 
1460 	if (l->hit < r->hit)
1461 		return -1;
1462 	else if (l->hit > r->hit)
1463 		return 1;
1464 	return 0;
1465 }
1466 
1467 static struct sort_dimension hit_sort_dimension = {
1468 	.name	= "hit",
1469 	.cmp	= hit_cmp,
1470 };
1471 
1472 static int bytes_cmp(void *a, void *b)
1473 {
1474 	struct alloc_stat *l = a;
1475 	struct alloc_stat *r = b;
1476 
1477 	if (l->bytes_alloc < r->bytes_alloc)
1478 		return -1;
1479 	else if (l->bytes_alloc > r->bytes_alloc)
1480 		return 1;
1481 	return 0;
1482 }
1483 
1484 static struct sort_dimension bytes_sort_dimension = {
1485 	.name	= "bytes",
1486 	.cmp	= bytes_cmp,
1487 };
1488 
1489 static int frag_cmp(void *a, void *b)
1490 {
1491 	double x, y;
1492 	struct alloc_stat *l = a;
1493 	struct alloc_stat *r = b;
1494 
1495 	x = fragmentation(l->bytes_req, l->bytes_alloc);
1496 	y = fragmentation(r->bytes_req, r->bytes_alloc);
1497 
1498 	if (x < y)
1499 		return -1;
1500 	else if (x > y)
1501 		return 1;
1502 	return 0;
1503 }
1504 
1505 static struct sort_dimension frag_sort_dimension = {
1506 	.name	= "frag",
1507 	.cmp	= frag_cmp,
1508 };
1509 
1510 static int pingpong_cmp(void *a, void *b)
1511 {
1512 	struct alloc_stat *l = a;
1513 	struct alloc_stat *r = b;
1514 
1515 	if (l->pingpong < r->pingpong)
1516 		return -1;
1517 	else if (l->pingpong > r->pingpong)
1518 		return 1;
1519 	return 0;
1520 }
1521 
1522 static struct sort_dimension pingpong_sort_dimension = {
1523 	.name	= "pingpong",
1524 	.cmp	= pingpong_cmp,
1525 };
1526 
1527 /* page sort keys */
1528 static int page_cmp(void *a, void *b)
1529 {
1530 	struct page_stat *l = a;
1531 	struct page_stat *r = b;
1532 
1533 	if (l->page < r->page)
1534 		return -1;
1535 	else if (l->page > r->page)
1536 		return 1;
1537 	return 0;
1538 }
1539 
1540 static struct sort_dimension page_sort_dimension = {
1541 	.name	= "page",
1542 	.cmp	= page_cmp,
1543 };
1544 
1545 static int page_callsite_cmp(void *a, void *b)
1546 {
1547 	struct page_stat *l = a;
1548 	struct page_stat *r = b;
1549 
1550 	if (l->callsite < r->callsite)
1551 		return -1;
1552 	else if (l->callsite > r->callsite)
1553 		return 1;
1554 	return 0;
1555 }
1556 
1557 static struct sort_dimension page_callsite_sort_dimension = {
1558 	.name	= "callsite",
1559 	.cmp	= page_callsite_cmp,
1560 };
1561 
1562 static int page_hit_cmp(void *a, void *b)
1563 {
1564 	struct page_stat *l = a;
1565 	struct page_stat *r = b;
1566 
1567 	if (l->nr_alloc < r->nr_alloc)
1568 		return -1;
1569 	else if (l->nr_alloc > r->nr_alloc)
1570 		return 1;
1571 	return 0;
1572 }
1573 
1574 static struct sort_dimension page_hit_sort_dimension = {
1575 	.name	= "hit",
1576 	.cmp	= page_hit_cmp,
1577 };
1578 
1579 static int page_bytes_cmp(void *a, void *b)
1580 {
1581 	struct page_stat *l = a;
1582 	struct page_stat *r = b;
1583 
1584 	if (l->alloc_bytes < r->alloc_bytes)
1585 		return -1;
1586 	else if (l->alloc_bytes > r->alloc_bytes)
1587 		return 1;
1588 	return 0;
1589 }
1590 
1591 static struct sort_dimension page_bytes_sort_dimension = {
1592 	.name	= "bytes",
1593 	.cmp	= page_bytes_cmp,
1594 };
1595 
1596 static int page_order_cmp(void *a, void *b)
1597 {
1598 	struct page_stat *l = a;
1599 	struct page_stat *r = b;
1600 
1601 	if (l->order < r->order)
1602 		return -1;
1603 	else if (l->order > r->order)
1604 		return 1;
1605 	return 0;
1606 }
1607 
1608 static struct sort_dimension page_order_sort_dimension = {
1609 	.name	= "order",
1610 	.cmp	= page_order_cmp,
1611 };
1612 
1613 static int migrate_type_cmp(void *a, void *b)
1614 {
1615 	struct page_stat *l = a;
1616 	struct page_stat *r = b;
1617 
1618 	/* for internal use to find free'd page */
1619 	if (l->migrate_type == -1U)
1620 		return 0;
1621 
1622 	if (l->migrate_type < r->migrate_type)
1623 		return -1;
1624 	else if (l->migrate_type > r->migrate_type)
1625 		return 1;
1626 	return 0;
1627 }
1628 
1629 static struct sort_dimension migrate_type_sort_dimension = {
1630 	.name	= "migtype",
1631 	.cmp	= migrate_type_cmp,
1632 };
1633 
1634 static int gfp_flags_cmp(void *a, void *b)
1635 {
1636 	struct page_stat *l = a;
1637 	struct page_stat *r = b;
1638 
1639 	/* for internal use to find free'd page */
1640 	if (l->gfp_flags == -1U)
1641 		return 0;
1642 
1643 	if (l->gfp_flags < r->gfp_flags)
1644 		return -1;
1645 	else if (l->gfp_flags > r->gfp_flags)
1646 		return 1;
1647 	return 0;
1648 }
1649 
1650 static struct sort_dimension gfp_flags_sort_dimension = {
1651 	.name	= "gfp",
1652 	.cmp	= gfp_flags_cmp,
1653 };
1654 
1655 static struct sort_dimension *slab_sorts[] = {
1656 	&ptr_sort_dimension,
1657 	&callsite_sort_dimension,
1658 	&hit_sort_dimension,
1659 	&bytes_sort_dimension,
1660 	&frag_sort_dimension,
1661 	&pingpong_sort_dimension,
1662 };
1663 
1664 static struct sort_dimension *page_sorts[] = {
1665 	&page_sort_dimension,
1666 	&page_callsite_sort_dimension,
1667 	&page_hit_sort_dimension,
1668 	&page_bytes_sort_dimension,
1669 	&page_order_sort_dimension,
1670 	&migrate_type_sort_dimension,
1671 	&gfp_flags_sort_dimension,
1672 };
1673 
1674 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1675 {
1676 	struct sort_dimension *sort;
1677 	int i;
1678 
1679 	for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1680 		if (!strcmp(slab_sorts[i]->name, tok)) {
1681 			sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1682 			if (!sort) {
1683 				pr_err("%s: memdup failed\n", __func__);
1684 				return -1;
1685 			}
1686 			list_add_tail(&sort->list, list);
1687 			return 0;
1688 		}
1689 	}
1690 
1691 	return -1;
1692 }
1693 
1694 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1695 {
1696 	struct sort_dimension *sort;
1697 	int i;
1698 
1699 	for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1700 		if (!strcmp(page_sorts[i]->name, tok)) {
1701 			sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1702 			if (!sort) {
1703 				pr_err("%s: memdup failed\n", __func__);
1704 				return -1;
1705 			}
1706 			list_add_tail(&sort->list, list);
1707 			return 0;
1708 		}
1709 	}
1710 
1711 	return -1;
1712 }
1713 
1714 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1715 {
1716 	char *tok;
1717 	char *str = strdup(arg);
1718 	char *pos = str;
1719 
1720 	if (!str) {
1721 		pr_err("%s: strdup failed\n", __func__);
1722 		return -1;
1723 	}
1724 
1725 	while (true) {
1726 		tok = strsep(&pos, ",");
1727 		if (!tok)
1728 			break;
1729 		if (slab_sort_dimension__add(tok, sort_list) < 0) {
1730 			pr_err("Unknown slab --sort key: '%s'", tok);
1731 			free(str);
1732 			return -1;
1733 		}
1734 	}
1735 
1736 	free(str);
1737 	return 0;
1738 }
1739 
1740 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1741 {
1742 	char *tok;
1743 	char *str = strdup(arg);
1744 	char *pos = str;
1745 
1746 	if (!str) {
1747 		pr_err("%s: strdup failed\n", __func__);
1748 		return -1;
1749 	}
1750 
1751 	while (true) {
1752 		tok = strsep(&pos, ",");
1753 		if (!tok)
1754 			break;
1755 		if (page_sort_dimension__add(tok, sort_list) < 0) {
1756 			pr_err("Unknown page --sort key: '%s'", tok);
1757 			free(str);
1758 			return -1;
1759 		}
1760 	}
1761 
1762 	free(str);
1763 	return 0;
1764 }
1765 
1766 static int parse_sort_opt(const struct option *opt __maybe_unused,
1767 			  const char *arg, int unset __maybe_unused)
1768 {
1769 	if (!arg)
1770 		return -1;
1771 
1772 	if (kmem_page > kmem_slab ||
1773 	    (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1774 		if (caller_flag > alloc_flag)
1775 			return setup_page_sorting(&page_caller_sort, arg);
1776 		else
1777 			return setup_page_sorting(&page_alloc_sort, arg);
1778 	} else {
1779 		if (caller_flag > alloc_flag)
1780 			return setup_slab_sorting(&slab_caller_sort, arg);
1781 		else
1782 			return setup_slab_sorting(&slab_alloc_sort, arg);
1783 	}
1784 
1785 	return 0;
1786 }
1787 
1788 static int parse_caller_opt(const struct option *opt __maybe_unused,
1789 			    const char *arg __maybe_unused,
1790 			    int unset __maybe_unused)
1791 {
1792 	caller_flag = (alloc_flag + 1);
1793 	return 0;
1794 }
1795 
1796 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1797 			   const char *arg __maybe_unused,
1798 			   int unset __maybe_unused)
1799 {
1800 	alloc_flag = (caller_flag + 1);
1801 	return 0;
1802 }
1803 
1804 static int parse_slab_opt(const struct option *opt __maybe_unused,
1805 			  const char *arg __maybe_unused,
1806 			  int unset __maybe_unused)
1807 {
1808 	kmem_slab = (kmem_page + 1);
1809 	return 0;
1810 }
1811 
1812 static int parse_page_opt(const struct option *opt __maybe_unused,
1813 			  const char *arg __maybe_unused,
1814 			  int unset __maybe_unused)
1815 {
1816 	kmem_page = (kmem_slab + 1);
1817 	return 0;
1818 }
1819 
1820 static int parse_line_opt(const struct option *opt __maybe_unused,
1821 			  const char *arg, int unset __maybe_unused)
1822 {
1823 	int lines;
1824 
1825 	if (!arg)
1826 		return -1;
1827 
1828 	lines = strtoul(arg, NULL, 10);
1829 
1830 	if (caller_flag > alloc_flag)
1831 		caller_lines = lines;
1832 	else
1833 		alloc_lines = lines;
1834 
1835 	return 0;
1836 }
1837 
1838 static bool slab_legacy_tp_is_exposed(void)
1839 {
1840 	/*
1841 	 * The tracepoints "kmem:kmalloc_node" and
1842 	 * "kmem:kmem_cache_alloc_node" have been removed on the latest
1843 	 * kernel, if the tracepoint "kmem:kmalloc_node" is existed it
1844 	 * means the tool is running on an old kernel, we need to
1845 	 * rollback to support these legacy tracepoints.
1846 	 */
1847 	return IS_ERR(trace_event__tp_format("kmem", "kmalloc_node")) ?
1848 		false : true;
1849 }
1850 
1851 static int __cmd_record(int argc, const char **argv)
1852 {
1853 	const char * const record_args[] = {
1854 	"record", "-a", "-R", "-c", "1",
1855 	};
1856 	const char * const slab_events[] = {
1857 	"-e", "kmem:kmalloc",
1858 	"-e", "kmem:kfree",
1859 	"-e", "kmem:kmem_cache_alloc",
1860 	"-e", "kmem:kmem_cache_free",
1861 	};
1862 	const char * const slab_legacy_events[] = {
1863 	"-e", "kmem:kmalloc_node",
1864 	"-e", "kmem:kmem_cache_alloc_node",
1865 	};
1866 	const char * const page_events[] = {
1867 	"-e", "kmem:mm_page_alloc",
1868 	"-e", "kmem:mm_page_free",
1869 	};
1870 	unsigned int rec_argc, i, j;
1871 	const char **rec_argv;
1872 	unsigned int slab_legacy_tp_exposed = slab_legacy_tp_is_exposed();
1873 
1874 	rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1875 	if (kmem_slab) {
1876 		rec_argc += ARRAY_SIZE(slab_events);
1877 		if (slab_legacy_tp_exposed)
1878 			rec_argc += ARRAY_SIZE(slab_legacy_events);
1879 	}
1880 	if (kmem_page)
1881 		rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1882 
1883 	rec_argv = calloc(rec_argc + 1, sizeof(char *));
1884 
1885 	if (rec_argv == NULL)
1886 		return -ENOMEM;
1887 
1888 	for (i = 0; i < ARRAY_SIZE(record_args); i++)
1889 		rec_argv[i] = strdup(record_args[i]);
1890 
1891 	if (kmem_slab) {
1892 		for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1893 			rec_argv[i] = strdup(slab_events[j]);
1894 		if (slab_legacy_tp_exposed) {
1895 			for (j = 0; j < ARRAY_SIZE(slab_legacy_events); j++, i++)
1896 				rec_argv[i] = strdup(slab_legacy_events[j]);
1897 		}
1898 	}
1899 	if (kmem_page) {
1900 		rec_argv[i++] = strdup("-g");
1901 
1902 		for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1903 			rec_argv[i] = strdup(page_events[j]);
1904 	}
1905 
1906 	for (j = 1; j < (unsigned int)argc; j++, i++)
1907 		rec_argv[i] = argv[j];
1908 
1909 	return cmd_record(i, rec_argv);
1910 }
1911 
1912 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1913 {
1914 	if (!strcmp(var, "kmem.default")) {
1915 		if (!strcmp(value, "slab"))
1916 			kmem_default = KMEM_SLAB;
1917 		else if (!strcmp(value, "page"))
1918 			kmem_default = KMEM_PAGE;
1919 		else
1920 			pr_err("invalid default value ('slab' or 'page' required): %s\n",
1921 			       value);
1922 		return 0;
1923 	}
1924 
1925 	return 0;
1926 }
1927 
1928 int cmd_kmem(int argc, const char **argv)
1929 {
1930 	const char * const default_slab_sort = "frag,hit,bytes";
1931 	const char * const default_page_sort = "bytes,hit";
1932 	struct perf_data data = {
1933 		.mode = PERF_DATA_MODE_READ,
1934 	};
1935 	const struct option kmem_options[] = {
1936 	OPT_STRING('i', "input", &input_name, "file", "input file name"),
1937 	OPT_INCR('v', "verbose", &verbose,
1938 		    "be more verbose (show symbol address, etc)"),
1939 	OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1940 			   "show per-callsite statistics", parse_caller_opt),
1941 	OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1942 			   "show per-allocation statistics", parse_alloc_opt),
1943 	OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1944 		     "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1945 		     "page, order, migtype, gfp", parse_sort_opt),
1946 	OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1947 	OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1948 	OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
1949 	OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1950 			   parse_slab_opt),
1951 	OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1952 			   parse_page_opt),
1953 	OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1954 	OPT_STRING(0, "time", &time_str, "str",
1955 		   "Time span of interest (start,stop)"),
1956 	OPT_END()
1957 	};
1958 	const char *const kmem_subcommands[] = { "record", "stat", NULL };
1959 	const char *kmem_usage[] = {
1960 		NULL,
1961 		NULL
1962 	};
1963 	struct perf_session *session;
1964 	static const char errmsg[] = "No %s allocation events found.  Have you run 'perf kmem record --%s'?\n";
1965 	int ret = perf_config(kmem_config, NULL);
1966 
1967 	if (ret)
1968 		return ret;
1969 
1970 	argc = parse_options_subcommand(argc, argv, kmem_options,
1971 					kmem_subcommands, kmem_usage,
1972 					PARSE_OPT_STOP_AT_NON_OPTION);
1973 
1974 	if (!argc)
1975 		usage_with_options(kmem_usage, kmem_options);
1976 
1977 	if (kmem_slab == 0 && kmem_page == 0) {
1978 		if (kmem_default == KMEM_SLAB)
1979 			kmem_slab = 1;
1980 		else
1981 			kmem_page = 1;
1982 	}
1983 
1984 	if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
1985 		symbol__init(NULL);
1986 		return __cmd_record(argc, argv);
1987 	}
1988 
1989 	data.path = input_name;
1990 
1991 	kmem_session = session = perf_session__new(&data, &perf_kmem);
1992 	if (IS_ERR(session))
1993 		return PTR_ERR(session);
1994 
1995 	ret = -1;
1996 
1997 	if (kmem_slab) {
1998 		if (!evlist__find_tracepoint_by_name(session->evlist, "kmem:kmalloc")) {
1999 			pr_err(errmsg, "slab", "slab");
2000 			goto out_delete;
2001 		}
2002 	}
2003 
2004 	if (kmem_page) {
2005 		struct evsel *evsel = evlist__find_tracepoint_by_name(session->evlist, "kmem:mm_page_alloc");
2006 
2007 		if (evsel == NULL) {
2008 			pr_err(errmsg, "page", "page");
2009 			goto out_delete;
2010 		}
2011 
2012 		kmem_page_size = tep_get_page_size(evsel->tp_format->tep);
2013 		symbol_conf.use_callchain = true;
2014 	}
2015 
2016 	symbol__init(&session->header.env);
2017 
2018 	if (perf_time__parse_str(&ptime, time_str) != 0) {
2019 		pr_err("Invalid time string\n");
2020 		ret = -EINVAL;
2021 		goto out_delete;
2022 	}
2023 
2024 	if (!strcmp(argv[0], "stat")) {
2025 		setlocale(LC_ALL, "");
2026 
2027 		if (cpu__setup_cpunode_map())
2028 			goto out_delete;
2029 
2030 		if (list_empty(&slab_caller_sort))
2031 			setup_slab_sorting(&slab_caller_sort, default_slab_sort);
2032 		if (list_empty(&slab_alloc_sort))
2033 			setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
2034 		if (list_empty(&page_caller_sort))
2035 			setup_page_sorting(&page_caller_sort, default_page_sort);
2036 		if (list_empty(&page_alloc_sort))
2037 			setup_page_sorting(&page_alloc_sort, default_page_sort);
2038 
2039 		if (kmem_page) {
2040 			setup_page_sorting(&page_alloc_sort_input,
2041 					   "page,order,migtype,gfp");
2042 			setup_page_sorting(&page_caller_sort_input,
2043 					   "callsite,order,migtype,gfp");
2044 		}
2045 		ret = __cmd_kmem(session);
2046 	} else
2047 		usage_with_options(kmem_usage, kmem_options);
2048 
2049 out_delete:
2050 	perf_session__delete(session);
2051 
2052 	return ret;
2053 }
2054 
2055