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