Commit | Line | Data |
---|---|---|
ba77c9e1 LZ |
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 | sort_fn_t alloc_sort_fn; | |
30 | sort_fn_t caller_sort_fn; | |
31 | ||
32 | static int alloc_lines = -1; | |
33 | static int caller_lines = -1; | |
34 | ||
7707b6b6 LZ |
35 | static bool raw_ip; |
36 | ||
ba77c9e1 LZ |
37 | static char *cwd; |
38 | static int cwdlen; | |
39 | ||
40 | struct alloc_stat { | |
41 | union { | |
7707b6b6 | 42 | u64 call_site; |
ba77c9e1 LZ |
43 | u64 ptr; |
44 | }; | |
45 | u64 bytes_req; | |
46 | u64 bytes_alloc; | |
47 | u32 hit; | |
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 | ||
59 | struct raw_event_sample { | |
60 | u32 size; | |
61 | char data[0]; | |
62 | }; | |
63 | ||
64 | static int | |
65 | process_comm_event(event_t *event, unsigned long offset, unsigned long head) | |
66 | { | |
67 | struct thread *thread = threads__findnew(event->comm.pid); | |
68 | ||
69 | dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n", | |
70 | (void *)(offset + head), | |
71 | (void *)(long)(event->header.size), | |
72 | event->comm.comm, event->comm.pid); | |
73 | ||
74 | if (thread == NULL || | |
75 | thread__set_comm(thread, event->comm.comm)) { | |
76 | dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n"); | |
77 | return -1; | |
78 | } | |
79 | ||
80 | return 0; | |
81 | } | |
82 | ||
83 | static void insert_alloc_stat(unsigned long ptr, | |
84 | int bytes_req, int bytes_alloc) | |
85 | { | |
86 | struct rb_node **node = &root_alloc_stat.rb_node; | |
87 | struct rb_node *parent = NULL; | |
88 | struct alloc_stat *data = NULL; | |
89 | ||
90 | if (!alloc_flag) | |
91 | return; | |
92 | ||
93 | while (*node) { | |
94 | parent = *node; | |
95 | data = rb_entry(*node, struct alloc_stat, node); | |
96 | ||
97 | if (ptr > data->ptr) | |
98 | node = &(*node)->rb_right; | |
99 | else if (ptr < data->ptr) | |
100 | node = &(*node)->rb_left; | |
101 | else | |
102 | break; | |
103 | } | |
104 | ||
105 | if (data && data->ptr == ptr) { | |
106 | data->hit++; | |
107 | data->bytes_req += bytes_req; | |
108 | data->bytes_alloc += bytes_req; | |
109 | } else { | |
110 | data = malloc(sizeof(*data)); | |
111 | data->ptr = ptr; | |
112 | data->hit = 1; | |
113 | data->bytes_req = bytes_req; | |
114 | data->bytes_alloc = bytes_alloc; | |
115 | ||
116 | rb_link_node(&data->node, parent, node); | |
117 | rb_insert_color(&data->node, &root_alloc_stat); | |
118 | } | |
119 | } | |
120 | ||
121 | static void insert_caller_stat(unsigned long call_site, | |
122 | int bytes_req, int bytes_alloc) | |
123 | { | |
124 | struct rb_node **node = &root_caller_stat.rb_node; | |
125 | struct rb_node *parent = NULL; | |
126 | struct alloc_stat *data = NULL; | |
127 | ||
128 | if (!caller_flag) | |
129 | return; | |
130 | ||
131 | while (*node) { | |
132 | parent = *node; | |
133 | data = rb_entry(*node, struct alloc_stat, node); | |
134 | ||
135 | if (call_site > data->call_site) | |
136 | node = &(*node)->rb_right; | |
137 | else if (call_site < data->call_site) | |
138 | node = &(*node)->rb_left; | |
139 | else | |
140 | break; | |
141 | } | |
142 | ||
143 | if (data && data->call_site == call_site) { | |
144 | data->hit++; | |
145 | data->bytes_req += bytes_req; | |
146 | data->bytes_alloc += bytes_req; | |
147 | } else { | |
148 | data = malloc(sizeof(*data)); | |
149 | data->call_site = call_site; | |
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 | ||
159 | static void process_alloc_event(struct raw_event_sample *raw, | |
160 | struct event *event, | |
161 | int cpu __used, | |
162 | u64 timestamp __used, | |
163 | struct thread *thread __used, | |
164 | int node __used) | |
165 | { | |
166 | unsigned long call_site; | |
167 | unsigned long ptr; | |
168 | int bytes_req; | |
169 | int bytes_alloc; | |
170 | ||
171 | ptr = raw_field_value(event, "ptr", raw->data); | |
172 | call_site = raw_field_value(event, "call_site", raw->data); | |
173 | bytes_req = raw_field_value(event, "bytes_req", raw->data); | |
174 | bytes_alloc = raw_field_value(event, "bytes_alloc", raw->data); | |
175 | ||
176 | insert_alloc_stat(ptr, bytes_req, bytes_alloc); | |
177 | insert_caller_stat(call_site, bytes_req, bytes_alloc); | |
178 | ||
179 | total_requested += bytes_req; | |
180 | total_allocated += bytes_alloc; | |
181 | } | |
182 | ||
183 | static void process_free_event(struct raw_event_sample *raw __used, | |
184 | struct event *event __used, | |
185 | int cpu __used, | |
186 | u64 timestamp __used, | |
187 | struct thread *thread __used) | |
188 | { | |
189 | } | |
190 | ||
191 | static void | |
192 | process_raw_event(event_t *raw_event __used, void *more_data, | |
193 | int cpu, u64 timestamp, struct thread *thread) | |
194 | { | |
195 | struct raw_event_sample *raw = more_data; | |
196 | struct event *event; | |
197 | int type; | |
198 | ||
199 | type = trace_parse_common_type(raw->data); | |
200 | event = trace_find_event(type); | |
201 | ||
202 | if (!strcmp(event->name, "kmalloc") || | |
203 | !strcmp(event->name, "kmem_cache_alloc")) { | |
204 | process_alloc_event(raw, event, cpu, timestamp, thread, 0); | |
205 | return; | |
206 | } | |
207 | ||
208 | if (!strcmp(event->name, "kmalloc_node") || | |
209 | !strcmp(event->name, "kmem_cache_alloc_node")) { | |
210 | process_alloc_event(raw, event, cpu, timestamp, thread, 1); | |
211 | return; | |
212 | } | |
213 | ||
214 | if (!strcmp(event->name, "kfree") || | |
215 | !strcmp(event->name, "kmem_cache_free")) { | |
216 | process_free_event(raw, event, cpu, timestamp, thread); | |
217 | return; | |
218 | } | |
219 | } | |
220 | ||
221 | static int | |
222 | process_sample_event(event_t *event, unsigned long offset, unsigned long head) | |
223 | { | |
224 | u64 ip = event->ip.ip; | |
225 | u64 timestamp = -1; | |
226 | u32 cpu = -1; | |
227 | u64 period = 1; | |
228 | void *more_data = event->ip.__more_data; | |
229 | struct thread *thread = threads__findnew(event->ip.pid); | |
230 | ||
231 | if (sample_type & PERF_SAMPLE_TIME) { | |
232 | timestamp = *(u64 *)more_data; | |
233 | more_data += sizeof(u64); | |
234 | } | |
235 | ||
236 | if (sample_type & PERF_SAMPLE_CPU) { | |
237 | cpu = *(u32 *)more_data; | |
238 | more_data += sizeof(u32); | |
239 | more_data += sizeof(u32); /* reserved */ | |
240 | } | |
241 | ||
242 | if (sample_type & PERF_SAMPLE_PERIOD) { | |
243 | period = *(u64 *)more_data; | |
244 | more_data += sizeof(u64); | |
245 | } | |
246 | ||
247 | dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n", | |
248 | (void *)(offset + head), | |
249 | (void *)(long)(event->header.size), | |
250 | event->header.misc, | |
251 | event->ip.pid, event->ip.tid, | |
252 | (void *)(long)ip, | |
253 | (long long)period); | |
254 | ||
255 | if (thread == NULL) { | |
256 | pr_debug("problem processing %d event, skipping it.\n", | |
257 | event->header.type); | |
258 | return -1; | |
259 | } | |
260 | ||
261 | dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid); | |
262 | ||
263 | process_raw_event(event, more_data, cpu, timestamp, thread); | |
264 | ||
265 | return 0; | |
266 | } | |
267 | ||
268 | static int sample_type_check(u64 type) | |
269 | { | |
270 | sample_type = type; | |
271 | ||
272 | if (!(sample_type & PERF_SAMPLE_RAW)) { | |
273 | fprintf(stderr, | |
274 | "No trace sample to read. Did you call perf record " | |
275 | "without -R?"); | |
276 | return -1; | |
277 | } | |
278 | ||
279 | return 0; | |
280 | } | |
281 | ||
282 | static struct perf_file_handler file_handler = { | |
283 | .process_sample_event = process_sample_event, | |
284 | .process_comm_event = process_comm_event, | |
285 | .sample_type_check = sample_type_check, | |
286 | }; | |
287 | ||
288 | static int read_events(void) | |
289 | { | |
290 | register_idle_thread(); | |
291 | register_perf_file_handler(&file_handler); | |
292 | ||
cc612d81 | 293 | return mmap_dispatch_perf_file(&header, input_name, NULL, false, 0, 0, |
ba77c9e1 LZ |
294 | &cwdlen, &cwd); |
295 | } | |
296 | ||
297 | static double fragmentation(unsigned long n_req, unsigned long n_alloc) | |
298 | { | |
299 | if (n_alloc == 0) | |
300 | return 0.0; | |
301 | else | |
302 | return 100.0 - (100.0 * n_req / n_alloc); | |
303 | } | |
304 | ||
305 | static void __print_result(struct rb_root *root, int n_lines, int is_caller) | |
306 | { | |
307 | struct rb_node *next; | |
308 | ||
1b145ae5 ACM |
309 | printf("%.78s\n", graph_dotted_line); |
310 | printf("%-28s|", is_caller ? "Callsite": "Alloc Ptr"); | |
311 | printf("Total_alloc/Per | Total_req/Per | Hit | Frag\n"); | |
312 | printf("%.78s\n", graph_dotted_line); | |
ba77c9e1 LZ |
313 | |
314 | next = rb_first(root); | |
315 | ||
316 | while (next && n_lines--) { | |
1b145ae5 ACM |
317 | struct alloc_stat *data = rb_entry(next, struct alloc_stat, |
318 | node); | |
319 | struct symbol *sym = NULL; | |
320 | char bf[BUFSIZ]; | |
321 | u64 addr; | |
322 | ||
323 | if (is_caller) { | |
324 | addr = data->call_site; | |
7707b6b6 LZ |
325 | if (!raw_ip) |
326 | sym = kernel_maps__find_symbol(addr, | |
327 | NULL, NULL); | |
1b145ae5 ACM |
328 | } else |
329 | addr = data->ptr; | |
330 | ||
331 | if (sym != NULL) | |
7707b6b6 | 332 | snprintf(bf, sizeof(bf), "%s+%Lx", sym->name, |
1b145ae5 ACM |
333 | addr - sym->start); |
334 | else | |
335 | snprintf(bf, sizeof(bf), "%#Lx", addr); | |
ba77c9e1 | 336 | |
1b145ae5 ACM |
337 | printf("%-28s|%8llu/%-6lu |%8llu/%-6lu|%6lu|%8.3f%%\n", |
338 | bf, (unsigned long long)data->bytes_alloc, | |
ba77c9e1 LZ |
339 | (unsigned long)data->bytes_alloc / data->hit, |
340 | (unsigned long long)data->bytes_req, | |
341 | (unsigned long)data->bytes_req / data->hit, | |
342 | (unsigned long)data->hit, | |
343 | fragmentation(data->bytes_req, data->bytes_alloc)); | |
344 | ||
345 | next = rb_next(next); | |
346 | } | |
347 | ||
348 | if (n_lines == -1) | |
7707b6b6 | 349 | printf(" ... | ... | ... | ... | ... \n"); |
ba77c9e1 | 350 | |
7707b6b6 | 351 | printf("%.78s\n", graph_dotted_line); |
ba77c9e1 LZ |
352 | } |
353 | ||
354 | static void print_summary(void) | |
355 | { | |
356 | printf("\nSUMMARY\n=======\n"); | |
357 | printf("Total bytes requested: %lu\n", total_requested); | |
358 | printf("Total bytes allocated: %lu\n", total_allocated); | |
359 | printf("Total bytes wasted on internal fragmentation: %lu\n", | |
360 | total_allocated - total_requested); | |
361 | printf("Internal fragmentation: %f%%\n", | |
362 | fragmentation(total_requested, total_allocated)); | |
363 | } | |
364 | ||
365 | static void print_result(void) | |
366 | { | |
367 | if (caller_flag) | |
368 | __print_result(&root_caller_sorted, caller_lines, 1); | |
369 | if (alloc_flag) | |
370 | __print_result(&root_alloc_sorted, alloc_lines, 0); | |
371 | print_summary(); | |
372 | } | |
373 | ||
374 | static void sort_insert(struct rb_root *root, struct alloc_stat *data, | |
375 | sort_fn_t sort_fn) | |
376 | { | |
377 | struct rb_node **new = &(root->rb_node); | |
378 | struct rb_node *parent = NULL; | |
379 | ||
380 | while (*new) { | |
381 | struct alloc_stat *this; | |
382 | int cmp; | |
383 | ||
384 | this = rb_entry(*new, struct alloc_stat, node); | |
385 | parent = *new; | |
386 | ||
387 | cmp = sort_fn(data, this); | |
388 | ||
389 | if (cmp > 0) | |
390 | new = &((*new)->rb_left); | |
391 | else | |
392 | new = &((*new)->rb_right); | |
393 | } | |
394 | ||
395 | rb_link_node(&data->node, parent, new); | |
396 | rb_insert_color(&data->node, root); | |
397 | } | |
398 | ||
399 | static void __sort_result(struct rb_root *root, struct rb_root *root_sorted, | |
400 | sort_fn_t sort_fn) | |
401 | { | |
402 | struct rb_node *node; | |
403 | struct alloc_stat *data; | |
404 | ||
405 | for (;;) { | |
406 | node = rb_first(root); | |
407 | if (!node) | |
408 | break; | |
409 | ||
410 | rb_erase(node, root); | |
411 | data = rb_entry(node, struct alloc_stat, node); | |
412 | sort_insert(root_sorted, data, sort_fn); | |
413 | } | |
414 | } | |
415 | ||
416 | static void sort_result(void) | |
417 | { | |
418 | __sort_result(&root_alloc_stat, &root_alloc_sorted, alloc_sort_fn); | |
419 | __sort_result(&root_caller_stat, &root_caller_sorted, caller_sort_fn); | |
420 | } | |
421 | ||
422 | static int __cmd_kmem(void) | |
423 | { | |
424 | setup_pager(); | |
425 | read_events(); | |
426 | sort_result(); | |
427 | print_result(); | |
428 | ||
429 | return 0; | |
430 | } | |
431 | ||
432 | static const char * const kmem_usage[] = { | |
433 | "perf kmem [<options>] {record}", | |
434 | NULL | |
435 | }; | |
436 | ||
437 | ||
438 | static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r) | |
439 | { | |
440 | if (l->ptr < r->ptr) | |
441 | return -1; | |
442 | else if (l->ptr > r->ptr) | |
443 | return 1; | |
444 | return 0; | |
445 | } | |
446 | ||
447 | static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r) | |
448 | { | |
449 | if (l->call_site < r->call_site) | |
450 | return -1; | |
451 | else if (l->call_site > r->call_site) | |
452 | return 1; | |
453 | return 0; | |
454 | } | |
455 | ||
f3ced7cd PE |
456 | static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r) |
457 | { | |
458 | if (l->hit < r->hit) | |
459 | return -1; | |
460 | else if (l->hit > r->hit) | |
461 | return 1; | |
462 | return 0; | |
463 | } | |
464 | ||
ba77c9e1 LZ |
465 | static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r) |
466 | { | |
467 | if (l->bytes_alloc < r->bytes_alloc) | |
468 | return -1; | |
469 | else if (l->bytes_alloc > r->bytes_alloc) | |
470 | return 1; | |
471 | return 0; | |
472 | } | |
473 | ||
f3ced7cd PE |
474 | static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r) |
475 | { | |
476 | double x, y; | |
477 | ||
478 | x = fragmentation(l->bytes_req, l->bytes_alloc); | |
479 | y = fragmentation(r->bytes_req, r->bytes_alloc); | |
480 | ||
481 | if (x < y) | |
482 | return -1; | |
483 | else if (x > y) | |
484 | return 1; | |
485 | return 0; | |
486 | } | |
487 | ||
ba77c9e1 LZ |
488 | static int parse_sort_opt(const struct option *opt __used, |
489 | const char *arg, int unset __used) | |
490 | { | |
491 | sort_fn_t sort_fn; | |
492 | ||
493 | if (!arg) | |
494 | return -1; | |
495 | ||
496 | if (strcmp(arg, "ptr") == 0) | |
497 | sort_fn = ptr_cmp; | |
498 | else if (strcmp(arg, "call_site") == 0) | |
499 | sort_fn = callsite_cmp; | |
f3ced7cd PE |
500 | else if (strcmp(arg, "hit") == 0) |
501 | sort_fn = hit_cmp; | |
ba77c9e1 LZ |
502 | else if (strcmp(arg, "bytes") == 0) |
503 | sort_fn = bytes_cmp; | |
f3ced7cd PE |
504 | else if (strcmp(arg, "frag") == 0) |
505 | sort_fn = frag_cmp; | |
ba77c9e1 LZ |
506 | else |
507 | return -1; | |
508 | ||
509 | if (caller_flag > alloc_flag) | |
510 | caller_sort_fn = sort_fn; | |
511 | else | |
512 | alloc_sort_fn = sort_fn; | |
513 | ||
514 | return 0; | |
515 | } | |
516 | ||
517 | static int parse_stat_opt(const struct option *opt __used, | |
518 | const char *arg, int unset __used) | |
519 | { | |
520 | if (!arg) | |
521 | return -1; | |
522 | ||
523 | if (strcmp(arg, "alloc") == 0) | |
524 | alloc_flag = (caller_flag + 1); | |
525 | else if (strcmp(arg, "caller") == 0) | |
526 | caller_flag = (alloc_flag + 1); | |
527 | else | |
528 | return -1; | |
529 | return 0; | |
530 | } | |
531 | ||
532 | static int parse_line_opt(const struct option *opt __used, | |
533 | const char *arg, int unset __used) | |
534 | { | |
535 | int lines; | |
536 | ||
537 | if (!arg) | |
538 | return -1; | |
539 | ||
540 | lines = strtoul(arg, NULL, 10); | |
541 | ||
542 | if (caller_flag > alloc_flag) | |
543 | caller_lines = lines; | |
544 | else | |
545 | alloc_lines = lines; | |
546 | ||
547 | return 0; | |
548 | } | |
549 | ||
550 | static const struct option kmem_options[] = { | |
551 | OPT_STRING('i', "input", &input_name, "file", | |
552 | "input file name"), | |
553 | OPT_CALLBACK(0, "stat", NULL, "<alloc>|<caller>", | |
554 | "stat selector, Pass 'alloc' or 'caller'.", | |
555 | parse_stat_opt), | |
556 | OPT_CALLBACK('s', "sort", NULL, "key", | |
f3ced7cd | 557 | "sort by key: ptr, call_site, hit, bytes, frag", |
ba77c9e1 LZ |
558 | parse_sort_opt), |
559 | OPT_CALLBACK('l', "line", NULL, "num", | |
560 | "show n lins", | |
561 | parse_line_opt), | |
7707b6b6 | 562 | OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), |
ba77c9e1 LZ |
563 | OPT_END() |
564 | }; | |
565 | ||
566 | static const char *record_args[] = { | |
567 | "record", | |
568 | "-a", | |
569 | "-R", | |
570 | "-M", | |
571 | "-f", | |
572 | "-c", "1", | |
573 | "-e", "kmem:kmalloc", | |
574 | "-e", "kmem:kmalloc_node", | |
575 | "-e", "kmem:kfree", | |
576 | "-e", "kmem:kmem_cache_alloc", | |
577 | "-e", "kmem:kmem_cache_alloc_node", | |
578 | "-e", "kmem:kmem_cache_free", | |
579 | }; | |
580 | ||
581 | static int __cmd_record(int argc, const char **argv) | |
582 | { | |
583 | unsigned int rec_argc, i, j; | |
584 | const char **rec_argv; | |
585 | ||
586 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
587 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); | |
588 | ||
589 | for (i = 0; i < ARRAY_SIZE(record_args); i++) | |
590 | rec_argv[i] = strdup(record_args[i]); | |
591 | ||
592 | for (j = 1; j < (unsigned int)argc; j++, i++) | |
593 | rec_argv[i] = argv[j]; | |
594 | ||
595 | return cmd_record(i, rec_argv, NULL); | |
596 | } | |
597 | ||
598 | int cmd_kmem(int argc, const char **argv, const char *prefix __used) | |
599 | { | |
600 | symbol__init(0); | |
601 | ||
602 | argc = parse_options(argc, argv, kmem_options, kmem_usage, 0); | |
603 | ||
604 | if (argc && !strncmp(argv[0], "rec", 3)) | |
605 | return __cmd_record(argc, argv); | |
606 | else if (argc) | |
607 | usage_with_options(kmem_usage, kmem_options); | |
608 | ||
609 | if (!alloc_sort_fn) | |
610 | alloc_sort_fn = bytes_cmp; | |
611 | if (!caller_sort_fn) | |
612 | caller_sort_fn = bytes_cmp; | |
613 | ||
614 | return __cmd_kmem(); | |
615 | } | |
616 |