perf: Make common SAMPLE_EVENT parser
[deliverable/linux.git] / tools / perf / builtin-kmem.c
1 #include "builtin.h"
2 #include "perf.h"
3
4 #include "util/util.h"
5 #include "util/cache.h"
6 #include "util/symbol.h"
7 #include "util/thread.h"
8 #include "util/header.h"
9
10 #include "util/parse-options.h"
11 #include "util/trace-event.h"
12
13 #include "util/debug.h"
14 #include "util/data_map.h"
15
16 #include <linux/rbtree.h>
17
18 struct alloc_stat;
19 typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
20
21 static char const *input_name = "perf.data";
22
23 static struct perf_header *header;
24 static u64 sample_type;
25
26 static int alloc_flag;
27 static int caller_flag;
28
29 static int alloc_lines = -1;
30 static int caller_lines = -1;
31
32 static bool raw_ip;
33
34 static char default_sort_order[] = "frag,hit,bytes";
35
36 static int *cpunode_map;
37 static int max_cpu_num;
38
39 struct alloc_stat {
40 u64 call_site;
41 u64 ptr;
42 u64 bytes_req;
43 u64 bytes_alloc;
44 u32 hit;
45 u32 pingpong;
46
47 short alloc_cpu;
48
49 struct rb_node node;
50 };
51
52 static struct rb_root root_alloc_stat;
53 static struct rb_root root_alloc_sorted;
54 static struct rb_root root_caller_stat;
55 static struct rb_root root_caller_sorted;
56
57 static unsigned long total_requested, total_allocated;
58 static unsigned long nr_allocs, nr_cross_allocs;
59
60 struct raw_event_sample {
61 u32 size;
62 char data[0];
63 };
64
65 #define PATH_SYS_NODE "/sys/devices/system/node"
66
67 static void init_cpunode_map(void)
68 {
69 FILE *fp;
70 int i;
71
72 fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
73 if (!fp) {
74 max_cpu_num = 4096;
75 return;
76 }
77
78 if (fscanf(fp, "%d", &max_cpu_num) < 1)
79 die("Failed to read 'kernel_max' from sysfs");
80 max_cpu_num++;
81
82 cpunode_map = calloc(max_cpu_num, sizeof(int));
83 if (!cpunode_map)
84 die("calloc");
85 for (i = 0; i < max_cpu_num; i++)
86 cpunode_map[i] = -1;
87 fclose(fp);
88 }
89
90 static void setup_cpunode_map(void)
91 {
92 struct dirent *dent1, *dent2;
93 DIR *dir1, *dir2;
94 unsigned int cpu, mem;
95 char buf[PATH_MAX];
96
97 init_cpunode_map();
98
99 dir1 = opendir(PATH_SYS_NODE);
100 if (!dir1)
101 return;
102
103 while (true) {
104 dent1 = readdir(dir1);
105 if (!dent1)
106 break;
107
108 if (sscanf(dent1->d_name, "node%u", &mem) < 1)
109 continue;
110
111 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
112 dir2 = opendir(buf);
113 if (!dir2)
114 continue;
115 while (true) {
116 dent2 = readdir(dir2);
117 if (!dent2)
118 break;
119 if (sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
120 continue;
121 cpunode_map[cpu] = mem;
122 }
123 }
124 }
125
126 static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
127 int bytes_req, int bytes_alloc, int cpu)
128 {
129 struct rb_node **node = &root_alloc_stat.rb_node;
130 struct rb_node *parent = NULL;
131 struct alloc_stat *data = NULL;
132
133 while (*node) {
134 parent = *node;
135 data = rb_entry(*node, struct alloc_stat, node);
136
137 if (ptr > data->ptr)
138 node = &(*node)->rb_right;
139 else if (ptr < data->ptr)
140 node = &(*node)->rb_left;
141 else
142 break;
143 }
144
145 if (data && data->ptr == ptr) {
146 data->hit++;
147 data->bytes_req += bytes_req;
148 data->bytes_alloc += bytes_req;
149 } else {
150 data = malloc(sizeof(*data));
151 if (!data)
152 die("malloc");
153 data->ptr = ptr;
154 data->pingpong = 0;
155 data->hit = 1;
156 data->bytes_req = bytes_req;
157 data->bytes_alloc = bytes_alloc;
158
159 rb_link_node(&data->node, parent, node);
160 rb_insert_color(&data->node, &root_alloc_stat);
161 }
162 data->call_site = call_site;
163 data->alloc_cpu = cpu;
164 }
165
166 static void insert_caller_stat(unsigned long call_site,
167 int bytes_req, int bytes_alloc)
168 {
169 struct rb_node **node = &root_caller_stat.rb_node;
170 struct rb_node *parent = NULL;
171 struct alloc_stat *data = NULL;
172
173 while (*node) {
174 parent = *node;
175 data = rb_entry(*node, struct alloc_stat, node);
176
177 if (call_site > data->call_site)
178 node = &(*node)->rb_right;
179 else if (call_site < data->call_site)
180 node = &(*node)->rb_left;
181 else
182 break;
183 }
184
185 if (data && data->call_site == call_site) {
186 data->hit++;
187 data->bytes_req += bytes_req;
188 data->bytes_alloc += bytes_req;
189 } else {
190 data = malloc(sizeof(*data));
191 if (!data)
192 die("malloc");
193 data->call_site = call_site;
194 data->pingpong = 0;
195 data->hit = 1;
196 data->bytes_req = bytes_req;
197 data->bytes_alloc = bytes_alloc;
198
199 rb_link_node(&data->node, parent, node);
200 rb_insert_color(&data->node, &root_caller_stat);
201 }
202 }
203
204 static void process_alloc_event(struct raw_event_sample *raw,
205 struct event *event,
206 int cpu,
207 u64 timestamp __used,
208 struct thread *thread __used,
209 int node)
210 {
211 unsigned long call_site;
212 unsigned long ptr;
213 int bytes_req;
214 int bytes_alloc;
215 int node1, node2;
216
217 ptr = raw_field_value(event, "ptr", raw->data);
218 call_site = raw_field_value(event, "call_site", raw->data);
219 bytes_req = raw_field_value(event, "bytes_req", raw->data);
220 bytes_alloc = raw_field_value(event, "bytes_alloc", raw->data);
221
222 insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
223 insert_caller_stat(call_site, bytes_req, bytes_alloc);
224
225 total_requested += bytes_req;
226 total_allocated += bytes_alloc;
227
228 if (node) {
229 node1 = cpunode_map[cpu];
230 node2 = raw_field_value(event, "node", raw->data);
231 if (node1 != node2)
232 nr_cross_allocs++;
233 }
234 nr_allocs++;
235 }
236
237 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
238 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
239
240 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
241 unsigned long call_site,
242 struct rb_root *root,
243 sort_fn_t sort_fn)
244 {
245 struct rb_node *node = root->rb_node;
246 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
247
248 while (node) {
249 struct alloc_stat *data;
250 int cmp;
251
252 data = rb_entry(node, struct alloc_stat, node);
253
254 cmp = sort_fn(&key, data);
255 if (cmp < 0)
256 node = node->rb_left;
257 else if (cmp > 0)
258 node = node->rb_right;
259 else
260 return data;
261 }
262 return NULL;
263 }
264
265 static void process_free_event(struct raw_event_sample *raw,
266 struct event *event,
267 int cpu,
268 u64 timestamp __used,
269 struct thread *thread __used)
270 {
271 unsigned long ptr;
272 struct alloc_stat *s_alloc, *s_caller;
273
274 ptr = raw_field_value(event, "ptr", raw->data);
275
276 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
277 if (!s_alloc)
278 return;
279
280 if (cpu != s_alloc->alloc_cpu) {
281 s_alloc->pingpong++;
282
283 s_caller = search_alloc_stat(0, s_alloc->call_site,
284 &root_caller_stat, callsite_cmp);
285 assert(s_caller);
286 s_caller->pingpong++;
287 }
288 s_alloc->alloc_cpu = -1;
289 }
290
291 static void
292 process_raw_event(event_t *raw_event __used, void *more_data,
293 int cpu, u64 timestamp, struct thread *thread)
294 {
295 struct raw_event_sample *raw = more_data;
296 struct event *event;
297 int type;
298
299 type = trace_parse_common_type(raw->data);
300 event = trace_find_event(type);
301
302 if (!strcmp(event->name, "kmalloc") ||
303 !strcmp(event->name, "kmem_cache_alloc")) {
304 process_alloc_event(raw, event, cpu, timestamp, thread, 0);
305 return;
306 }
307
308 if (!strcmp(event->name, "kmalloc_node") ||
309 !strcmp(event->name, "kmem_cache_alloc_node")) {
310 process_alloc_event(raw, event, cpu, timestamp, thread, 1);
311 return;
312 }
313
314 if (!strcmp(event->name, "kfree") ||
315 !strcmp(event->name, "kmem_cache_free")) {
316 process_free_event(raw, event, cpu, timestamp, thread);
317 return;
318 }
319 }
320
321 static int process_sample_event(event_t *event)
322 {
323 struct sample_data data;
324 struct thread *thread;
325
326 memset(&data, 0, sizeof(data));
327 data.time = -1;
328 data.cpu = -1;
329 data.period = 1;
330
331 event__parse_sample(event, sample_type, &data);
332
333 dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
334 event->header.misc,
335 data.pid, data.tid,
336 (void *)(long)data.ip,
337 (long long)data.period);
338
339 thread = threads__findnew(event->ip.pid);
340 if (thread == NULL) {
341 pr_debug("problem processing %d event, skipping it.\n",
342 event->header.type);
343 return -1;
344 }
345
346 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
347
348 process_raw_event(event, data.raw_data, data.cpu, data.time, thread);
349
350 return 0;
351 }
352
353 static int sample_type_check(u64 type)
354 {
355 sample_type = type;
356
357 if (!(sample_type & PERF_SAMPLE_RAW)) {
358 fprintf(stderr,
359 "No trace sample to read. Did you call perf record "
360 "without -R?");
361 return -1;
362 }
363
364 return 0;
365 }
366
367 static struct perf_file_handler file_handler = {
368 .process_sample_event = process_sample_event,
369 .process_comm_event = event__process_comm,
370 .sample_type_check = sample_type_check,
371 };
372
373 static int read_events(void)
374 {
375 register_idle_thread();
376 register_perf_file_handler(&file_handler);
377
378 return mmap_dispatch_perf_file(&header, input_name, 0, 0,
379 &event__cwdlen, &event__cwd);
380 }
381
382 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
383 {
384 if (n_alloc == 0)
385 return 0.0;
386 else
387 return 100.0 - (100.0 * n_req / n_alloc);
388 }
389
390 static void __print_result(struct rb_root *root, int n_lines, int is_caller)
391 {
392 struct rb_node *next;
393
394 printf("%.102s\n", graph_dotted_line);
395 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
396 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
397 printf("%.102s\n", graph_dotted_line);
398
399 next = rb_first(root);
400
401 while (next && n_lines--) {
402 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
403 node);
404 struct symbol *sym = NULL;
405 char buf[BUFSIZ];
406 u64 addr;
407
408 if (is_caller) {
409 addr = data->call_site;
410 if (!raw_ip)
411 sym = thread__find_function(kthread, addr, NULL);
412 } else
413 addr = data->ptr;
414
415 if (sym != NULL)
416 snprintf(buf, sizeof(buf), "%s+%Lx", sym->name,
417 addr - sym->start);
418 else
419 snprintf(buf, sizeof(buf), "%#Lx", addr);
420 printf(" %-34s |", buf);
421
422 printf(" %9llu/%-5lu | %9llu/%-5lu | %6lu | %8lu | %6.3f%%\n",
423 (unsigned long long)data->bytes_alloc,
424 (unsigned long)data->bytes_alloc / data->hit,
425 (unsigned long long)data->bytes_req,
426 (unsigned long)data->bytes_req / data->hit,
427 (unsigned long)data->hit,
428 (unsigned long)data->pingpong,
429 fragmentation(data->bytes_req, data->bytes_alloc));
430
431 next = rb_next(next);
432 }
433
434 if (n_lines == -1)
435 printf(" ... | ... | ... | ... | ... | ... \n");
436
437 printf("%.102s\n", graph_dotted_line);
438 }
439
440 static void print_summary(void)
441 {
442 printf("\nSUMMARY\n=======\n");
443 printf("Total bytes requested: %lu\n", total_requested);
444 printf("Total bytes allocated: %lu\n", total_allocated);
445 printf("Total bytes wasted on internal fragmentation: %lu\n",
446 total_allocated - total_requested);
447 printf("Internal fragmentation: %f%%\n",
448 fragmentation(total_requested, total_allocated));
449 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
450 }
451
452 static void print_result(void)
453 {
454 if (caller_flag)
455 __print_result(&root_caller_sorted, caller_lines, 1);
456 if (alloc_flag)
457 __print_result(&root_alloc_sorted, alloc_lines, 0);
458 print_summary();
459 }
460
461 struct sort_dimension {
462 const char name[20];
463 sort_fn_t cmp;
464 struct list_head list;
465 };
466
467 static LIST_HEAD(caller_sort);
468 static LIST_HEAD(alloc_sort);
469
470 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
471 struct list_head *sort_list)
472 {
473 struct rb_node **new = &(root->rb_node);
474 struct rb_node *parent = NULL;
475 struct sort_dimension *sort;
476
477 while (*new) {
478 struct alloc_stat *this;
479 int cmp = 0;
480
481 this = rb_entry(*new, struct alloc_stat, node);
482 parent = *new;
483
484 list_for_each_entry(sort, sort_list, list) {
485 cmp = sort->cmp(data, this);
486 if (cmp)
487 break;
488 }
489
490 if (cmp > 0)
491 new = &((*new)->rb_left);
492 else
493 new = &((*new)->rb_right);
494 }
495
496 rb_link_node(&data->node, parent, new);
497 rb_insert_color(&data->node, root);
498 }
499
500 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
501 struct list_head *sort_list)
502 {
503 struct rb_node *node;
504 struct alloc_stat *data;
505
506 for (;;) {
507 node = rb_first(root);
508 if (!node)
509 break;
510
511 rb_erase(node, root);
512 data = rb_entry(node, struct alloc_stat, node);
513 sort_insert(root_sorted, data, sort_list);
514 }
515 }
516
517 static void sort_result(void)
518 {
519 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
520 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
521 }
522
523 static int __cmd_kmem(void)
524 {
525 setup_pager();
526 read_events();
527 sort_result();
528 print_result();
529
530 return 0;
531 }
532
533 static const char * const kmem_usage[] = {
534 "perf kmem [<options>] {record}",
535 NULL
536 };
537
538 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
539 {
540 if (l->ptr < r->ptr)
541 return -1;
542 else if (l->ptr > r->ptr)
543 return 1;
544 return 0;
545 }
546
547 static struct sort_dimension ptr_sort_dimension = {
548 .name = "ptr",
549 .cmp = ptr_cmp,
550 };
551
552 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
553 {
554 if (l->call_site < r->call_site)
555 return -1;
556 else if (l->call_site > r->call_site)
557 return 1;
558 return 0;
559 }
560
561 static struct sort_dimension callsite_sort_dimension = {
562 .name = "callsite",
563 .cmp = callsite_cmp,
564 };
565
566 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
567 {
568 if (l->hit < r->hit)
569 return -1;
570 else if (l->hit > r->hit)
571 return 1;
572 return 0;
573 }
574
575 static struct sort_dimension hit_sort_dimension = {
576 .name = "hit",
577 .cmp = hit_cmp,
578 };
579
580 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
581 {
582 if (l->bytes_alloc < r->bytes_alloc)
583 return -1;
584 else if (l->bytes_alloc > r->bytes_alloc)
585 return 1;
586 return 0;
587 }
588
589 static struct sort_dimension bytes_sort_dimension = {
590 .name = "bytes",
591 .cmp = bytes_cmp,
592 };
593
594 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
595 {
596 double x, y;
597
598 x = fragmentation(l->bytes_req, l->bytes_alloc);
599 y = fragmentation(r->bytes_req, r->bytes_alloc);
600
601 if (x < y)
602 return -1;
603 else if (x > y)
604 return 1;
605 return 0;
606 }
607
608 static struct sort_dimension frag_sort_dimension = {
609 .name = "frag",
610 .cmp = frag_cmp,
611 };
612
613 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
614 {
615 if (l->pingpong < r->pingpong)
616 return -1;
617 else if (l->pingpong > r->pingpong)
618 return 1;
619 return 0;
620 }
621
622 static struct sort_dimension pingpong_sort_dimension = {
623 .name = "pingpong",
624 .cmp = pingpong_cmp,
625 };
626
627 static struct sort_dimension *avail_sorts[] = {
628 &ptr_sort_dimension,
629 &callsite_sort_dimension,
630 &hit_sort_dimension,
631 &bytes_sort_dimension,
632 &frag_sort_dimension,
633 &pingpong_sort_dimension,
634 };
635
636 #define NUM_AVAIL_SORTS \
637 (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
638
639 static int sort_dimension__add(const char *tok, struct list_head *list)
640 {
641 struct sort_dimension *sort;
642 int i;
643
644 for (i = 0; i < NUM_AVAIL_SORTS; i++) {
645 if (!strcmp(avail_sorts[i]->name, tok)) {
646 sort = malloc(sizeof(*sort));
647 if (!sort)
648 die("malloc");
649 memcpy(sort, avail_sorts[i], sizeof(*sort));
650 list_add_tail(&sort->list, list);
651 return 0;
652 }
653 }
654
655 return -1;
656 }
657
658 static int setup_sorting(struct list_head *sort_list, const char *arg)
659 {
660 char *tok;
661 char *str = strdup(arg);
662
663 if (!str)
664 die("strdup");
665
666 while (true) {
667 tok = strsep(&str, ",");
668 if (!tok)
669 break;
670 if (sort_dimension__add(tok, sort_list) < 0) {
671 error("Unknown --sort key: '%s'", tok);
672 return -1;
673 }
674 }
675
676 free(str);
677 return 0;
678 }
679
680 static int parse_sort_opt(const struct option *opt __used,
681 const char *arg, int unset __used)
682 {
683 if (!arg)
684 return -1;
685
686 if (caller_flag > alloc_flag)
687 return setup_sorting(&caller_sort, arg);
688 else
689 return setup_sorting(&alloc_sort, arg);
690
691 return 0;
692 }
693
694 static int parse_stat_opt(const struct option *opt __used,
695 const char *arg, int unset __used)
696 {
697 if (!arg)
698 return -1;
699
700 if (strcmp(arg, "alloc") == 0)
701 alloc_flag = (caller_flag + 1);
702 else if (strcmp(arg, "caller") == 0)
703 caller_flag = (alloc_flag + 1);
704 else
705 return -1;
706 return 0;
707 }
708
709 static int parse_line_opt(const struct option *opt __used,
710 const char *arg, int unset __used)
711 {
712 int lines;
713
714 if (!arg)
715 return -1;
716
717 lines = strtoul(arg, NULL, 10);
718
719 if (caller_flag > alloc_flag)
720 caller_lines = lines;
721 else
722 alloc_lines = lines;
723
724 return 0;
725 }
726
727 static const struct option kmem_options[] = {
728 OPT_STRING('i', "input", &input_name, "file",
729 "input file name"),
730 OPT_CALLBACK(0, "stat", NULL, "<alloc>|<caller>",
731 "stat selector, Pass 'alloc' or 'caller'.",
732 parse_stat_opt),
733 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
734 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
735 parse_sort_opt),
736 OPT_CALLBACK('l', "line", NULL, "num",
737 "show n lins",
738 parse_line_opt),
739 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
740 OPT_END()
741 };
742
743 static const char *record_args[] = {
744 "record",
745 "-a",
746 "-R",
747 "-M",
748 "-f",
749 "-c", "1",
750 "-e", "kmem:kmalloc",
751 "-e", "kmem:kmalloc_node",
752 "-e", "kmem:kfree",
753 "-e", "kmem:kmem_cache_alloc",
754 "-e", "kmem:kmem_cache_alloc_node",
755 "-e", "kmem:kmem_cache_free",
756 };
757
758 static int __cmd_record(int argc, const char **argv)
759 {
760 unsigned int rec_argc, i, j;
761 const char **rec_argv;
762
763 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
764 rec_argv = calloc(rec_argc + 1, sizeof(char *));
765
766 for (i = 0; i < ARRAY_SIZE(record_args); i++)
767 rec_argv[i] = strdup(record_args[i]);
768
769 for (j = 1; j < (unsigned int)argc; j++, i++)
770 rec_argv[i] = argv[j];
771
772 return cmd_record(i, rec_argv, NULL);
773 }
774
775 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
776 {
777 symbol__init(0);
778
779 argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
780
781 if (argc && !strncmp(argv[0], "rec", 3))
782 return __cmd_record(argc, argv);
783 else if (argc)
784 usage_with_options(kmem_usage, kmem_options);
785
786 if (list_empty(&caller_sort))
787 setup_sorting(&caller_sort, default_sort_order);
788 if (list_empty(&alloc_sort))
789 setup_sorting(&alloc_sort, default_sort_order);
790
791 setup_cpunode_map();
792
793 return __cmd_kmem();
794 }
795
This page took 0.057278 seconds and 6 git commands to generate.