perf evsel: Cache associated event_format
[deliverable/linux.git] / tools / perf / builtin-kmem.c
1 #include "builtin.h"
2 #include "perf.h"
3
4 #include "util/evsel.h"
5 #include "util/util.h"
6 #include "util/cache.h"
7 #include "util/symbol.h"
8 #include "util/thread.h"
9 #include "util/header.h"
10 #include "util/session.h"
11 #include "util/tool.h"
12
13 #include "util/parse-options.h"
14 #include "util/trace-event.h"
15
16 #include "util/debug.h"
17
18 #include <linux/rbtree.h>
19
20 struct alloc_stat;
21 typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
22
23 static const char *input_name;
24
25 static int alloc_flag;
26 static int caller_flag;
27
28 static int alloc_lines = -1;
29 static int caller_lines = -1;
30
31 static bool raw_ip;
32
33 static char default_sort_order[] = "frag,hit,bytes";
34
35 static int *cpunode_map;
36 static int max_cpu_num;
37
38 struct alloc_stat {
39 u64 call_site;
40 u64 ptr;
41 u64 bytes_req;
42 u64 bytes_alloc;
43 u32 hit;
44 u32 pingpong;
45
46 short alloc_cpu;
47
48 struct rb_node node;
49 };
50
51 static struct rb_root root_alloc_stat;
52 static struct rb_root root_alloc_sorted;
53 static struct rb_root root_caller_stat;
54 static struct rb_root root_caller_sorted;
55
56 static unsigned long total_requested, total_allocated;
57 static unsigned long nr_allocs, nr_cross_allocs;
58
59 #define PATH_SYS_NODE "/sys/devices/system/node"
60
61 static void init_cpunode_map(void)
62 {
63 FILE *fp;
64 int i;
65
66 fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
67 if (!fp) {
68 max_cpu_num = 4096;
69 return;
70 }
71
72 if (fscanf(fp, "%d", &max_cpu_num) < 1)
73 die("Failed to read 'kernel_max' from sysfs");
74 max_cpu_num++;
75
76 cpunode_map = calloc(max_cpu_num, sizeof(int));
77 if (!cpunode_map)
78 die("calloc");
79 for (i = 0; i < max_cpu_num; i++)
80 cpunode_map[i] = -1;
81 fclose(fp);
82 }
83
84 static void setup_cpunode_map(void)
85 {
86 struct dirent *dent1, *dent2;
87 DIR *dir1, *dir2;
88 unsigned int cpu, mem;
89 char buf[PATH_MAX];
90
91 init_cpunode_map();
92
93 dir1 = opendir(PATH_SYS_NODE);
94 if (!dir1)
95 return;
96
97 while ((dent1 = readdir(dir1)) != NULL) {
98 if (dent1->d_type != DT_DIR ||
99 sscanf(dent1->d_name, "node%u", &mem) < 1)
100 continue;
101
102 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
103 dir2 = opendir(buf);
104 if (!dir2)
105 continue;
106 while ((dent2 = readdir(dir2)) != NULL) {
107 if (dent2->d_type != DT_LNK ||
108 sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
109 continue;
110 cpunode_map[cpu] = mem;
111 }
112 closedir(dir2);
113 }
114 closedir(dir1);
115 }
116
117 static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
118 int bytes_req, int bytes_alloc, int cpu)
119 {
120 struct rb_node **node = &root_alloc_stat.rb_node;
121 struct rb_node *parent = NULL;
122 struct alloc_stat *data = NULL;
123
124 while (*node) {
125 parent = *node;
126 data = rb_entry(*node, struct alloc_stat, node);
127
128 if (ptr > data->ptr)
129 node = &(*node)->rb_right;
130 else if (ptr < data->ptr)
131 node = &(*node)->rb_left;
132 else
133 break;
134 }
135
136 if (data && data->ptr == ptr) {
137 data->hit++;
138 data->bytes_req += bytes_req;
139 data->bytes_alloc += bytes_alloc;
140 } else {
141 data = malloc(sizeof(*data));
142 if (!data)
143 die("malloc");
144 data->ptr = ptr;
145 data->pingpong = 0;
146 data->hit = 1;
147 data->bytes_req = bytes_req;
148 data->bytes_alloc = bytes_alloc;
149
150 rb_link_node(&data->node, parent, node);
151 rb_insert_color(&data->node, &root_alloc_stat);
152 }
153 data->call_site = call_site;
154 data->alloc_cpu = cpu;
155 }
156
157 static void insert_caller_stat(unsigned long call_site,
158 int bytes_req, int bytes_alloc)
159 {
160 struct rb_node **node = &root_caller_stat.rb_node;
161 struct rb_node *parent = NULL;
162 struct alloc_stat *data = NULL;
163
164 while (*node) {
165 parent = *node;
166 data = rb_entry(*node, struct alloc_stat, node);
167
168 if (call_site > data->call_site)
169 node = &(*node)->rb_right;
170 else if (call_site < data->call_site)
171 node = &(*node)->rb_left;
172 else
173 break;
174 }
175
176 if (data && data->call_site == call_site) {
177 data->hit++;
178 data->bytes_req += bytes_req;
179 data->bytes_alloc += bytes_alloc;
180 } else {
181 data = malloc(sizeof(*data));
182 if (!data)
183 die("malloc");
184 data->call_site = call_site;
185 data->pingpong = 0;
186 data->hit = 1;
187 data->bytes_req = bytes_req;
188 data->bytes_alloc = bytes_alloc;
189
190 rb_link_node(&data->node, parent, node);
191 rb_insert_color(&data->node, &root_caller_stat);
192 }
193 }
194
195 static void process_alloc_event(void *data,
196 struct event_format *event,
197 int cpu,
198 u64 timestamp __used,
199 struct thread *thread __used,
200 int node)
201 {
202 unsigned long call_site;
203 unsigned long ptr;
204 int bytes_req;
205 int bytes_alloc;
206 int node1, node2;
207
208 ptr = raw_field_value(event, "ptr", data);
209 call_site = raw_field_value(event, "call_site", data);
210 bytes_req = raw_field_value(event, "bytes_req", data);
211 bytes_alloc = raw_field_value(event, "bytes_alloc", data);
212
213 insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
214 insert_caller_stat(call_site, bytes_req, bytes_alloc);
215
216 total_requested += bytes_req;
217 total_allocated += bytes_alloc;
218
219 if (node) {
220 node1 = cpunode_map[cpu];
221 node2 = raw_field_value(event, "node", data);
222 if (node1 != node2)
223 nr_cross_allocs++;
224 }
225 nr_allocs++;
226 }
227
228 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
229 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
230
231 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
232 unsigned long call_site,
233 struct rb_root *root,
234 sort_fn_t sort_fn)
235 {
236 struct rb_node *node = root->rb_node;
237 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
238
239 while (node) {
240 struct alloc_stat *data;
241 int cmp;
242
243 data = rb_entry(node, struct alloc_stat, node);
244
245 cmp = sort_fn(&key, data);
246 if (cmp < 0)
247 node = node->rb_left;
248 else if (cmp > 0)
249 node = node->rb_right;
250 else
251 return data;
252 }
253 return NULL;
254 }
255
256 static void process_free_event(void *data,
257 struct event_format *event,
258 int cpu,
259 u64 timestamp __used,
260 struct thread *thread __used)
261 {
262 unsigned long ptr;
263 struct alloc_stat *s_alloc, *s_caller;
264
265 ptr = raw_field_value(event, "ptr", data);
266
267 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
268 if (!s_alloc)
269 return;
270
271 if (cpu != s_alloc->alloc_cpu) {
272 s_alloc->pingpong++;
273
274 s_caller = search_alloc_stat(0, s_alloc->call_site,
275 &root_caller_stat, callsite_cmp);
276 assert(s_caller);
277 s_caller->pingpong++;
278 }
279 s_alloc->alloc_cpu = -1;
280 }
281
282 static void process_raw_event(struct perf_evsel *evsel, void *data,
283 int cpu, u64 timestamp, struct thread *thread)
284 {
285 struct event_format *event = evsel->tp_format;
286
287 if (!strcmp(event->name, "kmalloc") ||
288 !strcmp(event->name, "kmem_cache_alloc")) {
289 process_alloc_event(data, event, cpu, timestamp, thread, 0);
290 return;
291 }
292
293 if (!strcmp(event->name, "kmalloc_node") ||
294 !strcmp(event->name, "kmem_cache_alloc_node")) {
295 process_alloc_event(data, event, cpu, timestamp, thread, 1);
296 return;
297 }
298
299 if (!strcmp(event->name, "kfree") ||
300 !strcmp(event->name, "kmem_cache_free")) {
301 process_free_event(data, event, cpu, timestamp, thread);
302 return;
303 }
304 }
305
306 static int process_sample_event(struct perf_tool *tool __used,
307 union perf_event *event,
308 struct perf_sample *sample,
309 struct perf_evsel *evsel,
310 struct machine *machine)
311 {
312 struct thread *thread = machine__findnew_thread(machine, event->ip.pid);
313
314 if (thread == NULL) {
315 pr_debug("problem processing %d event, skipping it.\n",
316 event->header.type);
317 return -1;
318 }
319
320 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
321
322 process_raw_event(evsel, sample->raw_data, sample->cpu,
323 sample->time, thread);
324
325 return 0;
326 }
327
328 static struct perf_tool perf_kmem = {
329 .sample = process_sample_event,
330 .comm = perf_event__process_comm,
331 .ordered_samples = true,
332 };
333
334 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
335 {
336 if (n_alloc == 0)
337 return 0.0;
338 else
339 return 100.0 - (100.0 * n_req / n_alloc);
340 }
341
342 static void __print_result(struct rb_root *root, struct perf_session *session,
343 int n_lines, int is_caller)
344 {
345 struct rb_node *next;
346 struct machine *machine;
347
348 printf("%.102s\n", graph_dotted_line);
349 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
350 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
351 printf("%.102s\n", graph_dotted_line);
352
353 next = rb_first(root);
354
355 machine = perf_session__find_host_machine(session);
356 if (!machine) {
357 pr_err("__print_result: couldn't find kernel information\n");
358 return;
359 }
360 while (next && n_lines--) {
361 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
362 node);
363 struct symbol *sym = NULL;
364 struct map *map;
365 char buf[BUFSIZ];
366 u64 addr;
367
368 if (is_caller) {
369 addr = data->call_site;
370 if (!raw_ip)
371 sym = machine__find_kernel_function(machine, addr, &map, NULL);
372 } else
373 addr = data->ptr;
374
375 if (sym != NULL)
376 snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
377 addr - map->unmap_ip(map, sym->start));
378 else
379 snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
380 printf(" %-34s |", buf);
381
382 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
383 (unsigned long long)data->bytes_alloc,
384 (unsigned long)data->bytes_alloc / data->hit,
385 (unsigned long long)data->bytes_req,
386 (unsigned long)data->bytes_req / data->hit,
387 (unsigned long)data->hit,
388 (unsigned long)data->pingpong,
389 fragmentation(data->bytes_req, data->bytes_alloc));
390
391 next = rb_next(next);
392 }
393
394 if (n_lines == -1)
395 printf(" ... | ... | ... | ... | ... | ... \n");
396
397 printf("%.102s\n", graph_dotted_line);
398 }
399
400 static void print_summary(void)
401 {
402 printf("\nSUMMARY\n=======\n");
403 printf("Total bytes requested: %lu\n", total_requested);
404 printf("Total bytes allocated: %lu\n", total_allocated);
405 printf("Total bytes wasted on internal fragmentation: %lu\n",
406 total_allocated - total_requested);
407 printf("Internal fragmentation: %f%%\n",
408 fragmentation(total_requested, total_allocated));
409 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
410 }
411
412 static void print_result(struct perf_session *session)
413 {
414 if (caller_flag)
415 __print_result(&root_caller_sorted, session, caller_lines, 1);
416 if (alloc_flag)
417 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
418 print_summary();
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(caller_sort);
428 static LIST_HEAD(alloc_sort);
429
430 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
431 struct list_head *sort_list)
432 {
433 struct rb_node **new = &(root->rb_node);
434 struct rb_node *parent = NULL;
435 struct sort_dimension *sort;
436
437 while (*new) {
438 struct alloc_stat *this;
439 int cmp = 0;
440
441 this = rb_entry(*new, struct alloc_stat, node);
442 parent = *new;
443
444 list_for_each_entry(sort, sort_list, list) {
445 cmp = sort->cmp(data, this);
446 if (cmp)
447 break;
448 }
449
450 if (cmp > 0)
451 new = &((*new)->rb_left);
452 else
453 new = &((*new)->rb_right);
454 }
455
456 rb_link_node(&data->node, parent, new);
457 rb_insert_color(&data->node, root);
458 }
459
460 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
461 struct list_head *sort_list)
462 {
463 struct rb_node *node;
464 struct alloc_stat *data;
465
466 for (;;) {
467 node = rb_first(root);
468 if (!node)
469 break;
470
471 rb_erase(node, root);
472 data = rb_entry(node, struct alloc_stat, node);
473 sort_insert(root_sorted, data, sort_list);
474 }
475 }
476
477 static void sort_result(void)
478 {
479 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
480 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
481 }
482
483 static int __cmd_kmem(void)
484 {
485 int err = -EINVAL;
486 struct perf_session *session;
487
488 session = perf_session__new(input_name, O_RDONLY, 0, false, &perf_kmem);
489 if (session == NULL)
490 return -ENOMEM;
491
492 if (perf_session__create_kernel_maps(session) < 0)
493 goto out_delete;
494
495 if (!perf_session__has_traces(session, "kmem record"))
496 goto out_delete;
497
498 setup_pager();
499 err = perf_session__process_events(session, &perf_kmem);
500 if (err != 0)
501 goto out_delete;
502 sort_result();
503 print_result(session);
504 out_delete:
505 perf_session__delete(session);
506 return err;
507 }
508
509 static const char * const kmem_usage[] = {
510 "perf kmem [<options>] {record|stat}",
511 NULL
512 };
513
514 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
515 {
516 if (l->ptr < r->ptr)
517 return -1;
518 else if (l->ptr > r->ptr)
519 return 1;
520 return 0;
521 }
522
523 static struct sort_dimension ptr_sort_dimension = {
524 .name = "ptr",
525 .cmp = ptr_cmp,
526 };
527
528 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
529 {
530 if (l->call_site < r->call_site)
531 return -1;
532 else if (l->call_site > r->call_site)
533 return 1;
534 return 0;
535 }
536
537 static struct sort_dimension callsite_sort_dimension = {
538 .name = "callsite",
539 .cmp = callsite_cmp,
540 };
541
542 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
543 {
544 if (l->hit < r->hit)
545 return -1;
546 else if (l->hit > r->hit)
547 return 1;
548 return 0;
549 }
550
551 static struct sort_dimension hit_sort_dimension = {
552 .name = "hit",
553 .cmp = hit_cmp,
554 };
555
556 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
557 {
558 if (l->bytes_alloc < r->bytes_alloc)
559 return -1;
560 else if (l->bytes_alloc > r->bytes_alloc)
561 return 1;
562 return 0;
563 }
564
565 static struct sort_dimension bytes_sort_dimension = {
566 .name = "bytes",
567 .cmp = bytes_cmp,
568 };
569
570 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
571 {
572 double x, y;
573
574 x = fragmentation(l->bytes_req, l->bytes_alloc);
575 y = fragmentation(r->bytes_req, r->bytes_alloc);
576
577 if (x < y)
578 return -1;
579 else if (x > y)
580 return 1;
581 return 0;
582 }
583
584 static struct sort_dimension frag_sort_dimension = {
585 .name = "frag",
586 .cmp = frag_cmp,
587 };
588
589 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
590 {
591 if (l->pingpong < r->pingpong)
592 return -1;
593 else if (l->pingpong > r->pingpong)
594 return 1;
595 return 0;
596 }
597
598 static struct sort_dimension pingpong_sort_dimension = {
599 .name = "pingpong",
600 .cmp = pingpong_cmp,
601 };
602
603 static struct sort_dimension *avail_sorts[] = {
604 &ptr_sort_dimension,
605 &callsite_sort_dimension,
606 &hit_sort_dimension,
607 &bytes_sort_dimension,
608 &frag_sort_dimension,
609 &pingpong_sort_dimension,
610 };
611
612 #define NUM_AVAIL_SORTS \
613 (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
614
615 static int sort_dimension__add(const char *tok, struct list_head *list)
616 {
617 struct sort_dimension *sort;
618 int i;
619
620 for (i = 0; i < NUM_AVAIL_SORTS; i++) {
621 if (!strcmp(avail_sorts[i]->name, tok)) {
622 sort = malloc(sizeof(*sort));
623 if (!sort)
624 die("malloc");
625 memcpy(sort, avail_sorts[i], sizeof(*sort));
626 list_add_tail(&sort->list, list);
627 return 0;
628 }
629 }
630
631 return -1;
632 }
633
634 static int setup_sorting(struct list_head *sort_list, const char *arg)
635 {
636 char *tok;
637 char *str = strdup(arg);
638
639 if (!str)
640 die("strdup");
641
642 while (true) {
643 tok = strsep(&str, ",");
644 if (!tok)
645 break;
646 if (sort_dimension__add(tok, sort_list) < 0) {
647 error("Unknown --sort key: '%s'", tok);
648 free(str);
649 return -1;
650 }
651 }
652
653 free(str);
654 return 0;
655 }
656
657 static int parse_sort_opt(const struct option *opt __used,
658 const char *arg, int unset __used)
659 {
660 if (!arg)
661 return -1;
662
663 if (caller_flag > alloc_flag)
664 return setup_sorting(&caller_sort, arg);
665 else
666 return setup_sorting(&alloc_sort, arg);
667
668 return 0;
669 }
670
671 static int parse_caller_opt(const struct option *opt __used,
672 const char *arg __used, int unset __used)
673 {
674 caller_flag = (alloc_flag + 1);
675 return 0;
676 }
677
678 static int parse_alloc_opt(const struct option *opt __used,
679 const char *arg __used, int unset __used)
680 {
681 alloc_flag = (caller_flag + 1);
682 return 0;
683 }
684
685 static int parse_line_opt(const struct option *opt __used,
686 const char *arg, int unset __used)
687 {
688 int lines;
689
690 if (!arg)
691 return -1;
692
693 lines = strtoul(arg, NULL, 10);
694
695 if (caller_flag > alloc_flag)
696 caller_lines = lines;
697 else
698 alloc_lines = lines;
699
700 return 0;
701 }
702
703 static const struct option kmem_options[] = {
704 OPT_STRING('i', "input", &input_name, "file",
705 "input file name"),
706 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
707 "show per-callsite statistics",
708 parse_caller_opt),
709 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
710 "show per-allocation statistics",
711 parse_alloc_opt),
712 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
713 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
714 parse_sort_opt),
715 OPT_CALLBACK('l', "line", NULL, "num",
716 "show n lines",
717 parse_line_opt),
718 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
719 OPT_END()
720 };
721
722 static const char *record_args[] = {
723 "record",
724 "-a",
725 "-R",
726 "-f",
727 "-c", "1",
728 "-e", "kmem:kmalloc",
729 "-e", "kmem:kmalloc_node",
730 "-e", "kmem:kfree",
731 "-e", "kmem:kmem_cache_alloc",
732 "-e", "kmem:kmem_cache_alloc_node",
733 "-e", "kmem:kmem_cache_free",
734 };
735
736 static int __cmd_record(int argc, const char **argv)
737 {
738 unsigned int rec_argc, i, j;
739 const char **rec_argv;
740
741 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
742 rec_argv = calloc(rec_argc + 1, sizeof(char *));
743
744 if (rec_argv == NULL)
745 return -ENOMEM;
746
747 for (i = 0; i < ARRAY_SIZE(record_args); i++)
748 rec_argv[i] = strdup(record_args[i]);
749
750 for (j = 1; j < (unsigned int)argc; j++, i++)
751 rec_argv[i] = argv[j];
752
753 return cmd_record(i, rec_argv, NULL);
754 }
755
756 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
757 {
758 argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
759
760 if (!argc)
761 usage_with_options(kmem_usage, kmem_options);
762
763 symbol__init();
764
765 if (!strncmp(argv[0], "rec", 3)) {
766 return __cmd_record(argc, argv);
767 } else if (!strcmp(argv[0], "stat")) {
768 setup_cpunode_map();
769
770 if (list_empty(&caller_sort))
771 setup_sorting(&caller_sort, default_sort_order);
772 if (list_empty(&alloc_sort))
773 setup_sorting(&alloc_sort, default_sort_order);
774
775 return __cmd_kmem();
776 } else
777 usage_with_options(kmem_usage, kmem_options);
778
779 return 0;
780 }
781
This page took 0.04772 seconds and 5 git commands to generate.