projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
perf session: Parse sample earlier
[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 #include "util/session.h"
10
11 #include "util/parse-options.h"
12 #include "util/trace-event.h"
13
14 #include "util/debug.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 int alloc_flag;
24 static int caller_flag;
25
26 static int alloc_lines = -1;
27 static int caller_lines = -1;
28
29 static bool raw_ip;
30
31 static char default_sort_order[] = "frag,hit,bytes";
32
33 static int *cpunode_map;
34 static int max_cpu_num;
35
36 struct alloc_stat {
37 u64 call_site;
38 u64 ptr;
39 u64 bytes_req;
40 u64 bytes_alloc;
41 u32 hit;
42 u32 pingpong;
43
44 short alloc_cpu;
45
46 struct rb_node node;
47 };
48
49 static struct rb_root root_alloc_stat;
50 static struct rb_root root_alloc_sorted;
51 static struct rb_root root_caller_stat;
52 static struct rb_root root_caller_sorted;
53
54 static unsigned long total_requested, total_allocated;
55 static unsigned long nr_allocs, nr_cross_allocs;
56
57 #define PATH_SYS_NODE "/sys/devices/system/node"
58
59 static void init_cpunode_map(void)
60 {
61 FILE *fp;
62 int i;
63
64 fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
65 if (!fp) {
66 max_cpu_num = 4096;
67 return;
68 }
69
70 if (fscanf(fp, "%d", &max_cpu_num) < 1)
71 die("Failed to read 'kernel_max' from sysfs");
72 max_cpu_num++;
73
74 cpunode_map = calloc(max_cpu_num, sizeof(int));
75 if (!cpunode_map)
76 die("calloc");
77 for (i = 0; i < max_cpu_num; i++)
78 cpunode_map[i] = -1;
79 fclose(fp);
80 }
81
82 static void setup_cpunode_map(void)
83 {
84 struct dirent *dent1, *dent2;
85 DIR *dir1, *dir2;
86 unsigned int cpu, mem;
87 char buf[PATH_MAX];
88
89 init_cpunode_map();
90
91 dir1 = opendir(PATH_SYS_NODE);
92 if (!dir1)
93 return;
94
95 while ((dent1 = readdir(dir1)) != NULL) {
96 if (dent1->d_type != DT_DIR ||
97 sscanf(dent1->d_name, "node%u", &mem) < 1)
98 continue;
99
100 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
101 dir2 = opendir(buf);
102 if (!dir2)
103 continue;
104 while ((dent2 = readdir(dir2)) != NULL) {
105 if (dent2->d_type != DT_LNK ||
106 sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
107 continue;
108 cpunode_map[cpu] = mem;
109 }
110 }
111 }
112
113 static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
114 int bytes_req, int bytes_alloc, int cpu)
115 {
116 struct rb_node **node = &root_alloc_stat.rb_node;
117 struct rb_node *parent = NULL;
118 struct alloc_stat *data = NULL;
119
120 while (*node) {
121 parent = *node;
122 data = rb_entry(*node, struct alloc_stat, node);
123
124 if (ptr > data->ptr)
125 node = &(*node)->rb_right;
126 else if (ptr < data->ptr)
127 node = &(*node)->rb_left;
128 else
129 break;
130 }
131
132 if (data && data->ptr == ptr) {
133 data->hit++;
134 data->bytes_req += bytes_req;
135 data->bytes_alloc += bytes_alloc;
136 } else {
137 data = malloc(sizeof(*data));
138 if (!data)
139 die("malloc");
140 data->ptr = ptr;
141 data->pingpong = 0;
142 data->hit = 1;
143 data->bytes_req = bytes_req;
144 data->bytes_alloc = bytes_alloc;
145
146 rb_link_node(&data->node, parent, node);
147 rb_insert_color(&data->node, &root_alloc_stat);
148 }
149 data->call_site = call_site;
150 data->alloc_cpu = cpu;
151 }
152
153 static void insert_caller_stat(unsigned long call_site,
154 int bytes_req, int bytes_alloc)
155 {
156 struct rb_node **node = &root_caller_stat.rb_node;
157 struct rb_node *parent = NULL;
158 struct alloc_stat *data = NULL;
159
160 while (*node) {
161 parent = *node;
162 data = rb_entry(*node, struct alloc_stat, node);
163
164 if (call_site > data->call_site)
165 node = &(*node)->rb_right;
166 else if (call_site < data->call_site)
167 node = &(*node)->rb_left;
168 else
169 break;
170 }
171
172 if (data && data->call_site == call_site) {
173 data->hit++;
174 data->bytes_req += bytes_req;
175 data->bytes_alloc += bytes_alloc;
176 } else {
177 data = malloc(sizeof(*data));
178 if (!data)
179 die("malloc");
180 data->call_site = call_site;
181 data->pingpong = 0;
182 data->hit = 1;
183 data->bytes_req = bytes_req;
184 data->bytes_alloc = bytes_alloc;
185
186 rb_link_node(&data->node, parent, node);
187 rb_insert_color(&data->node, &root_caller_stat);
188 }
189 }
190
191 static void process_alloc_event(void *data,
192 struct event *event,
193 int cpu,
194 u64 timestamp __used,
195 struct thread *thread __used,
196 int node)
197 {
198 unsigned long call_site;
199 unsigned long ptr;
200 int bytes_req;
201 int bytes_alloc;
202 int node1, node2;
203
204 ptr = raw_field_value(event, "ptr", data);
205 call_site = raw_field_value(event, "call_site", data);
206 bytes_req = raw_field_value(event, "bytes_req", data);
207 bytes_alloc = raw_field_value(event, "bytes_alloc", data);
208
209 insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
210 insert_caller_stat(call_site, bytes_req, bytes_alloc);
211
212 total_requested += bytes_req;
213 total_allocated += bytes_alloc;
214
215 if (node) {
216 node1 = cpunode_map[cpu];
217 node2 = raw_field_value(event, "node", data);
218 if (node1 != node2)
219 nr_cross_allocs++;
220 }
221 nr_allocs++;
222 }
223
224 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
225 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
226
227 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
228 unsigned long call_site,
229 struct rb_root *root,
230 sort_fn_t sort_fn)
231 {
232 struct rb_node *node = root->rb_node;
233 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
234
235 while (node) {
236 struct alloc_stat *data;
237 int cmp;
238
239 data = rb_entry(node, struct alloc_stat, node);
240
241 cmp = sort_fn(&key, data);
242 if (cmp < 0)
243 node = node->rb_left;
244 else if (cmp > 0)
245 node = node->rb_right;
246 else
247 return data;
248 }
249 return NULL;
250 }
251
252 static void process_free_event(void *data,
253 struct event *event,
254 int cpu,
255 u64 timestamp __used,
256 struct thread *thread __used)
257 {
258 unsigned long ptr;
259 struct alloc_stat *s_alloc, *s_caller;
260
261 ptr = raw_field_value(event, "ptr", data);
262
263 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
264 if (!s_alloc)
265 return;
266
267 if (cpu != s_alloc->alloc_cpu) {
268 s_alloc->pingpong++;
269
270 s_caller = search_alloc_stat(0, s_alloc->call_site,
271 &root_caller_stat, callsite_cmp);
272 assert(s_caller);
273 s_caller->pingpong++;
274 }
275 s_alloc->alloc_cpu = -1;
276 }
277
278 static void
279 process_raw_event(event_t *raw_event __used, void *data,
280 int cpu, u64 timestamp, struct thread *thread)
281 {
282 struct event *event;
283 int type;
284
285 type = trace_parse_common_type(data);
286 event = trace_find_event(type);
287
288 if (!strcmp(event->name, "kmalloc") ||
289 !strcmp(event->name, "kmem_cache_alloc")) {
290 process_alloc_event(data, event, cpu, timestamp, thread, 0);
291 return;
292 }
293
294 if (!strcmp(event->name, "kmalloc_node") ||
295 !strcmp(event->name, "kmem_cache_alloc_node")) {
296 process_alloc_event(data, event, cpu, timestamp, thread, 1);
297 return;
298 }
299
300 if (!strcmp(event->name, "kfree") ||
301 !strcmp(event->name, "kmem_cache_free")) {
302 process_free_event(data, event, cpu, timestamp, thread);
303 return;
304 }
305 }
306
307 static int process_sample_event(event_t *event, struct sample_data *sample,
308 struct perf_session *session)
309 {
310 struct thread *thread = perf_session__findnew(session, event->ip.pid);
311
312 if (thread == NULL) {
313 pr_debug("problem processing %d event, skipping it.\n",
314 event->header.type);
315 return -1;
316 }
317
318 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
319
320 process_raw_event(event, sample->raw_data, sample->cpu,
321 sample->time, thread);
322
323 return 0;
324 }
325
326 static struct perf_event_ops event_ops = {
327 .sample = process_sample_event,
328 .comm = event__process_comm,
329 .ordered_samples = true,
330 };
331
332 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
333 {
334 if (n_alloc == 0)
335 return 0.0;
336 else
337 return 100.0 - (100.0 * n_req / n_alloc);
338 }
339
340 static void __print_result(struct rb_root *root, struct perf_session *session,
341 int n_lines, int is_caller)
342 {
343 struct rb_node *next;
344 struct machine *machine;
345
346 printf("%.102s\n", graph_dotted_line);
347 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
348 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
349 printf("%.102s\n", graph_dotted_line);
350
351 next = rb_first(root);
352
353 machine = perf_session__find_host_machine(session);
354 if (!machine) {
355 pr_err("__print_result: couldn't find kernel information\n");
356 return;
357 }
358 while (next && n_lines--) {
359 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
360 node);
361 struct symbol *sym = NULL;
362 struct map *map;
363 char buf[BUFSIZ];
364 u64 addr;
365
366 if (is_caller) {
367 addr = data->call_site;
368 if (!raw_ip)
369 sym = machine__find_kernel_function(machine, addr, &map, NULL);
370 } else
371 addr = data->ptr;
372
373 if (sym != NULL)
374 snprintf(buf, sizeof(buf), "%s+%Lx", sym->name,
375 addr - map->unmap_ip(map, sym->start));
376 else
377 snprintf(buf, sizeof(buf), "%#Lx", addr);
378 printf(" %-34s |", buf);
379
380 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
381 (unsigned long long)data->bytes_alloc,
382 (unsigned long)data->bytes_alloc / data->hit,
383 (unsigned long long)data->bytes_req,
384 (unsigned long)data->bytes_req / data->hit,
385 (unsigned long)data->hit,
386 (unsigned long)data->pingpong,
387 fragmentation(data->bytes_req, data->bytes_alloc));
388
389 next = rb_next(next);
390 }
391
392 if (n_lines == -1)
393 printf(" ... | ... | ... | ... | ... | ... \n");
394
395 printf("%.102s\n", graph_dotted_line);
396 }
397
398 static void print_summary(void)
399 {
400 printf("\nSUMMARY\n=======\n");
401 printf("Total bytes requested: %lu\n", total_requested);
402 printf("Total bytes allocated: %lu\n", total_allocated);
403 printf("Total bytes wasted on internal fragmentation: %lu\n",
404 total_allocated - total_requested);
405 printf("Internal fragmentation: %f%%\n",
406 fragmentation(total_requested, total_allocated));
407 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
408 }
409
410 static void print_result(struct perf_session *session)
411 {
412 if (caller_flag)
413 __print_result(&root_caller_sorted, session, caller_lines, 1);
414 if (alloc_flag)
415 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
416 print_summary();
417 }
418
419 struct sort_dimension {
420 const char name[20];
421 sort_fn_t cmp;
422 struct list_head list;
423 };
424
425 static LIST_HEAD(caller_sort);
426 static LIST_HEAD(alloc_sort);
427
428 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
429 struct list_head *sort_list)
430 {
431 struct rb_node **new = &(root->rb_node);
432 struct rb_node *parent = NULL;
433 struct sort_dimension *sort;
434
435 while (*new) {
436 struct alloc_stat *this;
437 int cmp = 0;
438
439 this = rb_entry(*new, struct alloc_stat, node);
440 parent = *new;
441
442 list_for_each_entry(sort, sort_list, list) {
443 cmp = sort->cmp(data, this);
444 if (cmp)
445 break;
446 }
447
448 if (cmp > 0)
449 new = &((*new)->rb_left);
450 else
451 new = &((*new)->rb_right);
452 }
453
454 rb_link_node(&data->node, parent, new);
455 rb_insert_color(&data->node, root);
456 }
457
458 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
459 struct list_head *sort_list)
460 {
461 struct rb_node *node;
462 struct alloc_stat *data;
463
464 for (;;) {
465 node = rb_first(root);
466 if (!node)
467 break;
468
469 rb_erase(node, root);
470 data = rb_entry(node, struct alloc_stat, node);
471 sort_insert(root_sorted, data, sort_list);
472 }
473 }
474
475 static void sort_result(void)
476 {
477 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
478 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
479 }
480
481 static int __cmd_kmem(void)
482 {
483 int err = -EINVAL;
484 struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0, false);
485 if (session == NULL)
486 return -ENOMEM;
487
488 if (perf_session__create_kernel_maps(session) < 0)
489 goto out_delete;
490
491 if (!perf_session__has_traces(session, "kmem record"))
492 goto out_delete;
493
494 setup_pager();
495 err = perf_session__process_events(session, &event_ops);
496 if (err != 0)
497 goto out_delete;
498 sort_result();
499 print_result(session);
500 out_delete:
501 perf_session__delete(session);
502 return err;
503 }
504
505 static const char * const kmem_usage[] = {
506 "perf kmem [<options>] {record|stat}",
507 NULL
508 };
509
510 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
511 {
512 if (l->ptr < r->ptr)
513 return -1;
514 else if (l->ptr > r->ptr)
515 return 1;
516 return 0;
517 }
518
519 static struct sort_dimension ptr_sort_dimension = {
520 .name = "ptr",
521 .cmp = ptr_cmp,
522 };
523
524 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
525 {
526 if (l->call_site < r->call_site)
527 return -1;
528 else if (l->call_site > r->call_site)
529 return 1;
530 return 0;
531 }
532
533 static struct sort_dimension callsite_sort_dimension = {
534 .name = "callsite",
535 .cmp = callsite_cmp,
536 };
537
538 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
539 {
540 if (l->hit < r->hit)
541 return -1;
542 else if (l->hit > r->hit)
543 return 1;
544 return 0;
545 }
546
547 static struct sort_dimension hit_sort_dimension = {
548 .name = "hit",
549 .cmp = hit_cmp,
550 };
551
552 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
553 {
554 if (l->bytes_alloc < r->bytes_alloc)
555 return -1;
556 else if (l->bytes_alloc > r->bytes_alloc)
557 return 1;
558 return 0;
559 }
560
561 static struct sort_dimension bytes_sort_dimension = {
562 .name = "bytes",
563 .cmp = bytes_cmp,
564 };
565
566 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
567 {
568 double x, y;
569
570 x = fragmentation(l->bytes_req, l->bytes_alloc);
571 y = fragmentation(r->bytes_req, r->bytes_alloc);
572
573 if (x < y)
574 return -1;
575 else if (x > y)
576 return 1;
577 return 0;
578 }
579
580 static struct sort_dimension frag_sort_dimension = {
581 .name = "frag",
582 .cmp = frag_cmp,
583 };
584
585 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
586 {
587 if (l->pingpong < r->pingpong)
588 return -1;
589 else if (l->pingpong > r->pingpong)
590 return 1;
591 return 0;
592 }
593
594 static struct sort_dimension pingpong_sort_dimension = {
595 .name = "pingpong",
596 .cmp = pingpong_cmp,
597 };
598
599 static struct sort_dimension *avail_sorts[] = {
600 &ptr_sort_dimension,
601 &callsite_sort_dimension,
602 &hit_sort_dimension,
603 &bytes_sort_dimension,
604 &frag_sort_dimension,
605 &pingpong_sort_dimension,
606 };
607
608 #define NUM_AVAIL_SORTS \
609 (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
610
611 static int sort_dimension__add(const char *tok, struct list_head *list)
612 {
613 struct sort_dimension *sort;
614 int i;
615
616 for (i = 0; i < NUM_AVAIL_SORTS; i++) {
617 if (!strcmp(avail_sorts[i]->name, tok)) {
618 sort = malloc(sizeof(*sort));
619 if (!sort)
620 die("malloc");
621 memcpy(sort, avail_sorts[i], sizeof(*sort));
622 list_add_tail(&sort->list, list);
623 return 0;
624 }
625 }
626
627 return -1;
628 }
629
630 static int setup_sorting(struct list_head *sort_list, const char *arg)
631 {
632 char *tok;
633 char *str = strdup(arg);
634
635 if (!str)
636 die("strdup");
637
638 while (true) {
639 tok = strsep(&str, ",");
640 if (!tok)
641 break;
642 if (sort_dimension__add(tok, sort_list) < 0) {
643 error("Unknown --sort key: '%s'", tok);
644 return -1;
645 }
646 }
647
648 free(str);
649 return 0;
650 }
651
652 static int parse_sort_opt(const struct option *opt __used,
653 const char *arg, int unset __used)
654 {
655 if (!arg)
656 return -1;
657
658 if (caller_flag > alloc_flag)
659 return setup_sorting(&caller_sort, arg);
660 else
661 return setup_sorting(&alloc_sort, arg);
662
663 return 0;
664 }
665
666 static int parse_caller_opt(const struct option *opt __used,
667 const char *arg __used, int unset __used)
668 {
669 caller_flag = (alloc_flag + 1);
670 return 0;
671 }
672
673 static int parse_alloc_opt(const struct option *opt __used,
674 const char *arg __used, int unset __used)
675 {
676 alloc_flag = (caller_flag + 1);
677 return 0;
678 }
679
680 static int parse_line_opt(const struct option *opt __used,
681 const char *arg, int unset __used)
682 {
683 int lines;
684
685 if (!arg)
686 return -1;
687
688 lines = strtoul(arg, NULL, 10);
689
690 if (caller_flag > alloc_flag)
691 caller_lines = lines;
692 else
693 alloc_lines = lines;
694
695 return 0;
696 }
697
698 static const struct option kmem_options[] = {
699 OPT_STRING('i', "input", &input_name, "file",
700 "input file name"),
701 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
702 "show per-callsite statistics",
703 parse_caller_opt),
704 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
705 "show per-allocation statistics",
706 parse_alloc_opt),
707 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
708 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
709 parse_sort_opt),
710 OPT_CALLBACK('l', "line", NULL, "num",
711 "show n lines",
712 parse_line_opt),
713 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
714 OPT_END()
715 };
716
717 static const char *record_args[] = {
718 "record",
719 "-a",
720 "-R",
721 "-f",
722 "-c", "1",
723 "-e", "kmem:kmalloc",
724 "-e", "kmem:kmalloc_node",
725 "-e", "kmem:kfree",
726 "-e", "kmem:kmem_cache_alloc",
727 "-e", "kmem:kmem_cache_alloc_node",
728 "-e", "kmem:kmem_cache_free",
729 };
730
731 static int __cmd_record(int argc, const char **argv)
732 {
733 unsigned int rec_argc, i, j;
734 const char **rec_argv;
735
736 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
737 rec_argv = calloc(rec_argc + 1, sizeof(char *));
738
739 for (i = 0; i < ARRAY_SIZE(record_args); i++)
740 rec_argv[i] = strdup(record_args[i]);
741
742 for (j = 1; j < (unsigned int)argc; j++, i++)
743 rec_argv[i] = argv[j];
744
745 return cmd_record(i, rec_argv, NULL);
746 }
747
748 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
749 {
750 argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
751
752 if (!argc)
753 usage_with_options(kmem_usage, kmem_options);
754
755 symbol__init();
756
757 if (!strncmp(argv[0], "rec", 3)) {
758 return __cmd_record(argc, argv);
759 } else if (!strcmp(argv[0], "stat")) {
760 setup_cpunode_map();
761
762 if (list_empty(&caller_sort))
763 setup_sorting(&caller_sort, default_sort_order);
764 if (list_empty(&alloc_sort))
765 setup_sorting(&alloc_sort, default_sort_order);
766
767 return __cmd_kmem();
768 } else
769 usage_with_options(kmem_usage, kmem_options);
770
771 return 0;
772 }
773
Linux kernel - Deliverables
This page took 0.072832 seconds and 5 git commands to generate.