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" | |
94c744b6 | 9 | #include "util/session.h" |
ba77c9e1 LZ |
10 | |
11 | #include "util/parse-options.h" | |
12 | #include "util/trace-event.h" | |
13 | ||
14 | #include "util/debug.h" | |
15 | #include "util/data_map.h" | |
16 | ||
17 | #include <linux/rbtree.h> | |
18 | ||
19 | struct alloc_stat; | |
20 | typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *); | |
21 | ||
22 | static char const *input_name = "perf.data"; | |
23 | ||
ba77c9e1 LZ |
24 | static u64 sample_type; |
25 | ||
26 | static int alloc_flag; | |
27 | static int caller_flag; | |
28 | ||
ba77c9e1 LZ |
29 | static int alloc_lines = -1; |
30 | static int caller_lines = -1; | |
31 | ||
7707b6b6 LZ |
32 | static bool raw_ip; |
33 | ||
29b3e152 LZ |
34 | static char default_sort_order[] = "frag,hit,bytes"; |
35 | ||
7d0d3945 LZ |
36 | static int *cpunode_map; |
37 | static int max_cpu_num; | |
38 | ||
ba77c9e1 | 39 | struct alloc_stat { |
079d3f65 LZ |
40 | u64 call_site; |
41 | u64 ptr; | |
ba77c9e1 LZ |
42 | u64 bytes_req; |
43 | u64 bytes_alloc; | |
44 | u32 hit; | |
079d3f65 LZ |
45 | u32 pingpong; |
46 | ||
47 | short alloc_cpu; | |
ba77c9e1 LZ |
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; | |
7d0d3945 | 58 | static unsigned long nr_allocs, nr_cross_allocs; |
ba77c9e1 | 59 | |
7d0d3945 LZ |
60 | #define PATH_SYS_NODE "/sys/devices/system/node" |
61 | ||
62 | static void init_cpunode_map(void) | |
63 | { | |
64 | FILE *fp; | |
65 | int i; | |
66 | ||
67 | fp = fopen("/sys/devices/system/cpu/kernel_max", "r"); | |
68 | if (!fp) { | |
69 | max_cpu_num = 4096; | |
70 | return; | |
71 | } | |
72 | ||
73 | if (fscanf(fp, "%d", &max_cpu_num) < 1) | |
74 | die("Failed to read 'kernel_max' from sysfs"); | |
75 | max_cpu_num++; | |
76 | ||
77 | cpunode_map = calloc(max_cpu_num, sizeof(int)); | |
78 | if (!cpunode_map) | |
79 | die("calloc"); | |
80 | for (i = 0; i < max_cpu_num; i++) | |
81 | cpunode_map[i] = -1; | |
82 | fclose(fp); | |
83 | } | |
84 | ||
85 | static void setup_cpunode_map(void) | |
86 | { | |
87 | struct dirent *dent1, *dent2; | |
88 | DIR *dir1, *dir2; | |
89 | unsigned int cpu, mem; | |
90 | char buf[PATH_MAX]; | |
91 | ||
92 | init_cpunode_map(); | |
93 | ||
94 | dir1 = opendir(PATH_SYS_NODE); | |
95 | if (!dir1) | |
96 | return; | |
97 | ||
98 | while (true) { | |
99 | dent1 = readdir(dir1); | |
100 | if (!dent1) | |
101 | break; | |
102 | ||
103 | if (sscanf(dent1->d_name, "node%u", &mem) < 1) | |
104 | continue; | |
105 | ||
106 | snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name); | |
107 | dir2 = opendir(buf); | |
108 | if (!dir2) | |
109 | continue; | |
110 | while (true) { | |
111 | dent2 = readdir(dir2); | |
112 | if (!dent2) | |
113 | break; | |
114 | if (sscanf(dent2->d_name, "cpu%u", &cpu) < 1) | |
115 | continue; | |
116 | cpunode_map[cpu] = mem; | |
117 | } | |
118 | } | |
119 | } | |
120 | ||
079d3f65 LZ |
121 | static void insert_alloc_stat(unsigned long call_site, unsigned long ptr, |
122 | int bytes_req, int bytes_alloc, int cpu) | |
ba77c9e1 LZ |
123 | { |
124 | struct rb_node **node = &root_alloc_stat.rb_node; | |
125 | struct rb_node *parent = NULL; | |
126 | struct alloc_stat *data = NULL; | |
127 | ||
ba77c9e1 LZ |
128 | while (*node) { |
129 | parent = *node; | |
130 | data = rb_entry(*node, struct alloc_stat, node); | |
131 | ||
132 | if (ptr > data->ptr) | |
133 | node = &(*node)->rb_right; | |
134 | else if (ptr < data->ptr) | |
135 | node = &(*node)->rb_left; | |
136 | else | |
137 | break; | |
138 | } | |
139 | ||
140 | if (data && data->ptr == ptr) { | |
141 | data->hit++; | |
142 | data->bytes_req += bytes_req; | |
143 | data->bytes_alloc += bytes_req; | |
144 | } else { | |
145 | data = malloc(sizeof(*data)); | |
079d3f65 LZ |
146 | if (!data) |
147 | die("malloc"); | |
ba77c9e1 | 148 | data->ptr = ptr; |
079d3f65 | 149 | data->pingpong = 0; |
ba77c9e1 LZ |
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_alloc_stat); | |
156 | } | |
079d3f65 LZ |
157 | data->call_site = call_site; |
158 | data->alloc_cpu = cpu; | |
ba77c9e1 LZ |
159 | } |
160 | ||
161 | static void insert_caller_stat(unsigned long call_site, | |
162 | int bytes_req, int bytes_alloc) | |
163 | { | |
164 | struct rb_node **node = &root_caller_stat.rb_node; | |
165 | struct rb_node *parent = NULL; | |
166 | struct alloc_stat *data = NULL; | |
167 | ||
ba77c9e1 LZ |
168 | while (*node) { |
169 | parent = *node; | |
170 | data = rb_entry(*node, struct alloc_stat, node); | |
171 | ||
172 | if (call_site > data->call_site) | |
173 | node = &(*node)->rb_right; | |
174 | else if (call_site < data->call_site) | |
175 | node = &(*node)->rb_left; | |
176 | else | |
177 | break; | |
178 | } | |
179 | ||
180 | if (data && data->call_site == call_site) { | |
181 | data->hit++; | |
182 | data->bytes_req += bytes_req; | |
183 | data->bytes_alloc += bytes_req; | |
184 | } else { | |
185 | data = malloc(sizeof(*data)); | |
079d3f65 LZ |
186 | if (!data) |
187 | die("malloc"); | |
ba77c9e1 | 188 | data->call_site = call_site; |
079d3f65 | 189 | data->pingpong = 0; |
ba77c9e1 LZ |
190 | data->hit = 1; |
191 | data->bytes_req = bytes_req; | |
192 | data->bytes_alloc = bytes_alloc; | |
193 | ||
194 | rb_link_node(&data->node, parent, node); | |
195 | rb_insert_color(&data->node, &root_caller_stat); | |
196 | } | |
197 | } | |
198 | ||
f48f669d | 199 | static void process_alloc_event(void *data, |
ba77c9e1 | 200 | struct event *event, |
7d0d3945 | 201 | int cpu, |
ba77c9e1 LZ |
202 | u64 timestamp __used, |
203 | struct thread *thread __used, | |
7d0d3945 | 204 | int node) |
ba77c9e1 LZ |
205 | { |
206 | unsigned long call_site; | |
207 | unsigned long ptr; | |
208 | int bytes_req; | |
209 | int bytes_alloc; | |
7d0d3945 | 210 | int node1, node2; |
ba77c9e1 | 211 | |
f48f669d XG |
212 | ptr = raw_field_value(event, "ptr", data); |
213 | call_site = raw_field_value(event, "call_site", data); | |
214 | bytes_req = raw_field_value(event, "bytes_req", data); | |
215 | bytes_alloc = raw_field_value(event, "bytes_alloc", data); | |
ba77c9e1 | 216 | |
079d3f65 | 217 | insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu); |
ba77c9e1 LZ |
218 | insert_caller_stat(call_site, bytes_req, bytes_alloc); |
219 | ||
220 | total_requested += bytes_req; | |
221 | total_allocated += bytes_alloc; | |
7d0d3945 LZ |
222 | |
223 | if (node) { | |
224 | node1 = cpunode_map[cpu]; | |
f48f669d | 225 | node2 = raw_field_value(event, "node", data); |
7d0d3945 LZ |
226 | if (node1 != node2) |
227 | nr_cross_allocs++; | |
228 | } | |
229 | nr_allocs++; | |
ba77c9e1 LZ |
230 | } |
231 | ||
079d3f65 LZ |
232 | static int ptr_cmp(struct alloc_stat *, struct alloc_stat *); |
233 | static int callsite_cmp(struct alloc_stat *, struct alloc_stat *); | |
234 | ||
235 | static struct alloc_stat *search_alloc_stat(unsigned long ptr, | |
236 | unsigned long call_site, | |
237 | struct rb_root *root, | |
238 | sort_fn_t sort_fn) | |
239 | { | |
240 | struct rb_node *node = root->rb_node; | |
241 | struct alloc_stat key = { .ptr = ptr, .call_site = call_site }; | |
242 | ||
243 | while (node) { | |
244 | struct alloc_stat *data; | |
245 | int cmp; | |
246 | ||
247 | data = rb_entry(node, struct alloc_stat, node); | |
248 | ||
249 | cmp = sort_fn(&key, data); | |
250 | if (cmp < 0) | |
251 | node = node->rb_left; | |
252 | else if (cmp > 0) | |
253 | node = node->rb_right; | |
254 | else | |
255 | return data; | |
256 | } | |
257 | return NULL; | |
258 | } | |
259 | ||
f48f669d | 260 | static void process_free_event(void *data, |
079d3f65 LZ |
261 | struct event *event, |
262 | int cpu, | |
ba77c9e1 LZ |
263 | u64 timestamp __used, |
264 | struct thread *thread __used) | |
265 | { | |
079d3f65 LZ |
266 | unsigned long ptr; |
267 | struct alloc_stat *s_alloc, *s_caller; | |
268 | ||
f48f669d | 269 | ptr = raw_field_value(event, "ptr", data); |
079d3f65 LZ |
270 | |
271 | s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp); | |
272 | if (!s_alloc) | |
273 | return; | |
274 | ||
275 | if (cpu != s_alloc->alloc_cpu) { | |
276 | s_alloc->pingpong++; | |
277 | ||
278 | s_caller = search_alloc_stat(0, s_alloc->call_site, | |
279 | &root_caller_stat, callsite_cmp); | |
280 | assert(s_caller); | |
281 | s_caller->pingpong++; | |
282 | } | |
283 | s_alloc->alloc_cpu = -1; | |
ba77c9e1 LZ |
284 | } |
285 | ||
286 | static void | |
f48f669d | 287 | process_raw_event(event_t *raw_event __used, void *data, |
ba77c9e1 LZ |
288 | int cpu, u64 timestamp, struct thread *thread) |
289 | { | |
ba77c9e1 LZ |
290 | struct event *event; |
291 | int type; | |
292 | ||
f48f669d | 293 | type = trace_parse_common_type(data); |
ba77c9e1 LZ |
294 | event = trace_find_event(type); |
295 | ||
296 | if (!strcmp(event->name, "kmalloc") || | |
297 | !strcmp(event->name, "kmem_cache_alloc")) { | |
f48f669d | 298 | process_alloc_event(data, event, cpu, timestamp, thread, 0); |
ba77c9e1 LZ |
299 | return; |
300 | } | |
301 | ||
302 | if (!strcmp(event->name, "kmalloc_node") || | |
303 | !strcmp(event->name, "kmem_cache_alloc_node")) { | |
f48f669d | 304 | process_alloc_event(data, event, cpu, timestamp, thread, 1); |
ba77c9e1 LZ |
305 | return; |
306 | } | |
307 | ||
308 | if (!strcmp(event->name, "kfree") || | |
309 | !strcmp(event->name, "kmem_cache_free")) { | |
f48f669d | 310 | process_free_event(data, event, cpu, timestamp, thread); |
ba77c9e1 LZ |
311 | return; |
312 | } | |
313 | } | |
314 | ||
d8f66248 | 315 | static int process_sample_event(event_t *event, struct perf_session *session __used) |
ba77c9e1 | 316 | { |
180f95e2 OH |
317 | struct sample_data data; |
318 | struct thread *thread; | |
ba77c9e1 | 319 | |
180f95e2 OH |
320 | memset(&data, 0, sizeof(data)); |
321 | data.time = -1; | |
322 | data.cpu = -1; | |
323 | data.period = 1; | |
ba77c9e1 | 324 | |
180f95e2 | 325 | event__parse_sample(event, sample_type, &data); |
ba77c9e1 | 326 | |
62daacb5 | 327 | dump_printf("(IP, %d): %d/%d: %p period: %Ld\n", |
ba77c9e1 | 328 | event->header.misc, |
180f95e2 OH |
329 | data.pid, data.tid, |
330 | (void *)(long)data.ip, | |
331 | (long long)data.period); | |
ba77c9e1 | 332 | |
180f95e2 | 333 | thread = threads__findnew(event->ip.pid); |
ba77c9e1 LZ |
334 | if (thread == NULL) { |
335 | pr_debug("problem processing %d event, skipping it.\n", | |
336 | event->header.type); | |
337 | return -1; | |
338 | } | |
339 | ||
340 | dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid); | |
341 | ||
f48f669d | 342 | process_raw_event(event, data.raw_data, data.cpu, |
d8bd9e0a | 343 | data.time, thread); |
ba77c9e1 LZ |
344 | |
345 | return 0; | |
346 | } | |
347 | ||
348 | static int sample_type_check(u64 type) | |
349 | { | |
350 | sample_type = type; | |
351 | ||
352 | if (!(sample_type & PERF_SAMPLE_RAW)) { | |
353 | fprintf(stderr, | |
354 | "No trace sample to read. Did you call perf record " | |
355 | "without -R?"); | |
356 | return -1; | |
357 | } | |
358 | ||
359 | return 0; | |
360 | } | |
361 | ||
362 | static struct perf_file_handler file_handler = { | |
363 | .process_sample_event = process_sample_event, | |
62daacb5 | 364 | .process_comm_event = event__process_comm, |
ba77c9e1 LZ |
365 | .sample_type_check = sample_type_check, |
366 | }; | |
367 | ||
368 | static int read_events(void) | |
369 | { | |
94c744b6 ACM |
370 | int err; |
371 | struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0); | |
372 | ||
373 | if (session == NULL) | |
374 | return -ENOMEM; | |
375 | ||
ba77c9e1 LZ |
376 | register_idle_thread(); |
377 | register_perf_file_handler(&file_handler); | |
378 | ||
94c744b6 ACM |
379 | err = perf_session__process_events(session, 0, &event__cwdlen, &event__cwd); |
380 | perf_session__delete(session); | |
381 | return err; | |
ba77c9e1 LZ |
382 | } |
383 | ||
384 | static double fragmentation(unsigned long n_req, unsigned long n_alloc) | |
385 | { | |
386 | if (n_alloc == 0) | |
387 | return 0.0; | |
388 | else | |
389 | return 100.0 - (100.0 * n_req / n_alloc); | |
390 | } | |
391 | ||
392 | static void __print_result(struct rb_root *root, int n_lines, int is_caller) | |
393 | { | |
394 | struct rb_node *next; | |
395 | ||
079d3f65 LZ |
396 | printf("%.102s\n", graph_dotted_line); |
397 | printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr"); | |
398 | printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n"); | |
399 | printf("%.102s\n", graph_dotted_line); | |
ba77c9e1 LZ |
400 | |
401 | next = rb_first(root); | |
402 | ||
403 | while (next && n_lines--) { | |
1b145ae5 ACM |
404 | struct alloc_stat *data = rb_entry(next, struct alloc_stat, |
405 | node); | |
406 | struct symbol *sym = NULL; | |
079d3f65 | 407 | char buf[BUFSIZ]; |
1b145ae5 ACM |
408 | u64 addr; |
409 | ||
410 | if (is_caller) { | |
411 | addr = data->call_site; | |
7707b6b6 | 412 | if (!raw_ip) |
9958e1f0 | 413 | sym = map_groups__find_function(kmaps, addr, NULL); |
1b145ae5 ACM |
414 | } else |
415 | addr = data->ptr; | |
416 | ||
417 | if (sym != NULL) | |
079d3f65 | 418 | snprintf(buf, sizeof(buf), "%s+%Lx", sym->name, |
1b145ae5 ACM |
419 | addr - sym->start); |
420 | else | |
079d3f65 LZ |
421 | snprintf(buf, sizeof(buf), "%#Lx", addr); |
422 | printf(" %-34s |", buf); | |
ba77c9e1 | 423 | |
079d3f65 LZ |
424 | printf(" %9llu/%-5lu | %9llu/%-5lu | %6lu | %8lu | %6.3f%%\n", |
425 | (unsigned long long)data->bytes_alloc, | |
ba77c9e1 LZ |
426 | (unsigned long)data->bytes_alloc / data->hit, |
427 | (unsigned long long)data->bytes_req, | |
428 | (unsigned long)data->bytes_req / data->hit, | |
429 | (unsigned long)data->hit, | |
079d3f65 | 430 | (unsigned long)data->pingpong, |
ba77c9e1 LZ |
431 | fragmentation(data->bytes_req, data->bytes_alloc)); |
432 | ||
433 | next = rb_next(next); | |
434 | } | |
435 | ||
436 | if (n_lines == -1) | |
079d3f65 | 437 | printf(" ... | ... | ... | ... | ... | ... \n"); |
ba77c9e1 | 438 | |
079d3f65 | 439 | printf("%.102s\n", graph_dotted_line); |
ba77c9e1 LZ |
440 | } |
441 | ||
442 | static void print_summary(void) | |
443 | { | |
444 | printf("\nSUMMARY\n=======\n"); | |
445 | printf("Total bytes requested: %lu\n", total_requested); | |
446 | printf("Total bytes allocated: %lu\n", total_allocated); | |
447 | printf("Total bytes wasted on internal fragmentation: %lu\n", | |
448 | total_allocated - total_requested); | |
449 | printf("Internal fragmentation: %f%%\n", | |
450 | fragmentation(total_requested, total_allocated)); | |
7d0d3945 | 451 | printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs); |
ba77c9e1 LZ |
452 | } |
453 | ||
454 | static void print_result(void) | |
455 | { | |
456 | if (caller_flag) | |
457 | __print_result(&root_caller_sorted, caller_lines, 1); | |
458 | if (alloc_flag) | |
459 | __print_result(&root_alloc_sorted, alloc_lines, 0); | |
460 | print_summary(); | |
461 | } | |
462 | ||
29b3e152 LZ |
463 | struct sort_dimension { |
464 | const char name[20]; | |
465 | sort_fn_t cmp; | |
466 | struct list_head list; | |
467 | }; | |
468 | ||
469 | static LIST_HEAD(caller_sort); | |
470 | static LIST_HEAD(alloc_sort); | |
471 | ||
ba77c9e1 | 472 | static void sort_insert(struct rb_root *root, struct alloc_stat *data, |
29b3e152 | 473 | struct list_head *sort_list) |
ba77c9e1 LZ |
474 | { |
475 | struct rb_node **new = &(root->rb_node); | |
476 | struct rb_node *parent = NULL; | |
29b3e152 | 477 | struct sort_dimension *sort; |
ba77c9e1 LZ |
478 | |
479 | while (*new) { | |
480 | struct alloc_stat *this; | |
29b3e152 | 481 | int cmp = 0; |
ba77c9e1 LZ |
482 | |
483 | this = rb_entry(*new, struct alloc_stat, node); | |
484 | parent = *new; | |
485 | ||
29b3e152 LZ |
486 | list_for_each_entry(sort, sort_list, list) { |
487 | cmp = sort->cmp(data, this); | |
488 | if (cmp) | |
489 | break; | |
490 | } | |
ba77c9e1 LZ |
491 | |
492 | if (cmp > 0) | |
493 | new = &((*new)->rb_left); | |
494 | else | |
495 | new = &((*new)->rb_right); | |
496 | } | |
497 | ||
498 | rb_link_node(&data->node, parent, new); | |
499 | rb_insert_color(&data->node, root); | |
500 | } | |
501 | ||
502 | static void __sort_result(struct rb_root *root, struct rb_root *root_sorted, | |
29b3e152 | 503 | struct list_head *sort_list) |
ba77c9e1 LZ |
504 | { |
505 | struct rb_node *node; | |
506 | struct alloc_stat *data; | |
507 | ||
508 | for (;;) { | |
509 | node = rb_first(root); | |
510 | if (!node) | |
511 | break; | |
512 | ||
513 | rb_erase(node, root); | |
514 | data = rb_entry(node, struct alloc_stat, node); | |
29b3e152 | 515 | sort_insert(root_sorted, data, sort_list); |
ba77c9e1 LZ |
516 | } |
517 | } | |
518 | ||
519 | static void sort_result(void) | |
520 | { | |
29b3e152 LZ |
521 | __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort); |
522 | __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort); | |
ba77c9e1 LZ |
523 | } |
524 | ||
525 | static int __cmd_kmem(void) | |
526 | { | |
527 | setup_pager(); | |
528 | read_events(); | |
529 | sort_result(); | |
530 | print_result(); | |
531 | ||
532 | return 0; | |
533 | } | |
534 | ||
535 | static const char * const kmem_usage[] = { | |
90b86a9f | 536 | "perf kmem [<options>] {record|stat}", |
ba77c9e1 LZ |
537 | NULL |
538 | }; | |
539 | ||
ba77c9e1 LZ |
540 | static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r) |
541 | { | |
542 | if (l->ptr < r->ptr) | |
543 | return -1; | |
544 | else if (l->ptr > r->ptr) | |
545 | return 1; | |
546 | return 0; | |
547 | } | |
548 | ||
29b3e152 LZ |
549 | static struct sort_dimension ptr_sort_dimension = { |
550 | .name = "ptr", | |
551 | .cmp = ptr_cmp, | |
552 | }; | |
553 | ||
ba77c9e1 LZ |
554 | static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r) |
555 | { | |
556 | if (l->call_site < r->call_site) | |
557 | return -1; | |
558 | else if (l->call_site > r->call_site) | |
559 | return 1; | |
560 | return 0; | |
561 | } | |
562 | ||
29b3e152 LZ |
563 | static struct sort_dimension callsite_sort_dimension = { |
564 | .name = "callsite", | |
565 | .cmp = callsite_cmp, | |
566 | }; | |
567 | ||
f3ced7cd PE |
568 | static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r) |
569 | { | |
570 | if (l->hit < r->hit) | |
571 | return -1; | |
572 | else if (l->hit > r->hit) | |
573 | return 1; | |
574 | return 0; | |
575 | } | |
576 | ||
29b3e152 LZ |
577 | static struct sort_dimension hit_sort_dimension = { |
578 | .name = "hit", | |
579 | .cmp = hit_cmp, | |
580 | }; | |
581 | ||
ba77c9e1 LZ |
582 | static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r) |
583 | { | |
584 | if (l->bytes_alloc < r->bytes_alloc) | |
585 | return -1; | |
586 | else if (l->bytes_alloc > r->bytes_alloc) | |
587 | return 1; | |
588 | return 0; | |
589 | } | |
590 | ||
29b3e152 LZ |
591 | static struct sort_dimension bytes_sort_dimension = { |
592 | .name = "bytes", | |
593 | .cmp = bytes_cmp, | |
594 | }; | |
595 | ||
f3ced7cd PE |
596 | static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r) |
597 | { | |
598 | double x, y; | |
599 | ||
600 | x = fragmentation(l->bytes_req, l->bytes_alloc); | |
601 | y = fragmentation(r->bytes_req, r->bytes_alloc); | |
602 | ||
603 | if (x < y) | |
604 | return -1; | |
605 | else if (x > y) | |
606 | return 1; | |
607 | return 0; | |
608 | } | |
609 | ||
29b3e152 LZ |
610 | static struct sort_dimension frag_sort_dimension = { |
611 | .name = "frag", | |
612 | .cmp = frag_cmp, | |
613 | }; | |
614 | ||
079d3f65 LZ |
615 | static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r) |
616 | { | |
617 | if (l->pingpong < r->pingpong) | |
618 | return -1; | |
619 | else if (l->pingpong > r->pingpong) | |
620 | return 1; | |
621 | return 0; | |
622 | } | |
623 | ||
624 | static struct sort_dimension pingpong_sort_dimension = { | |
625 | .name = "pingpong", | |
626 | .cmp = pingpong_cmp, | |
627 | }; | |
628 | ||
29b3e152 LZ |
629 | static struct sort_dimension *avail_sorts[] = { |
630 | &ptr_sort_dimension, | |
631 | &callsite_sort_dimension, | |
632 | &hit_sort_dimension, | |
633 | &bytes_sort_dimension, | |
634 | &frag_sort_dimension, | |
079d3f65 | 635 | &pingpong_sort_dimension, |
29b3e152 LZ |
636 | }; |
637 | ||
638 | #define NUM_AVAIL_SORTS \ | |
639 | (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *)) | |
640 | ||
641 | static int sort_dimension__add(const char *tok, struct list_head *list) | |
642 | { | |
643 | struct sort_dimension *sort; | |
644 | int i; | |
645 | ||
646 | for (i = 0; i < NUM_AVAIL_SORTS; i++) { | |
647 | if (!strcmp(avail_sorts[i]->name, tok)) { | |
648 | sort = malloc(sizeof(*sort)); | |
649 | if (!sort) | |
650 | die("malloc"); | |
651 | memcpy(sort, avail_sorts[i], sizeof(*sort)); | |
652 | list_add_tail(&sort->list, list); | |
653 | return 0; | |
654 | } | |
655 | } | |
656 | ||
657 | return -1; | |
658 | } | |
659 | ||
660 | static int setup_sorting(struct list_head *sort_list, const char *arg) | |
661 | { | |
662 | char *tok; | |
663 | char *str = strdup(arg); | |
664 | ||
665 | if (!str) | |
666 | die("strdup"); | |
667 | ||
668 | while (true) { | |
669 | tok = strsep(&str, ","); | |
670 | if (!tok) | |
671 | break; | |
672 | if (sort_dimension__add(tok, sort_list) < 0) { | |
673 | error("Unknown --sort key: '%s'", tok); | |
674 | return -1; | |
675 | } | |
676 | } | |
677 | ||
678 | free(str); | |
679 | return 0; | |
680 | } | |
681 | ||
ba77c9e1 LZ |
682 | static int parse_sort_opt(const struct option *opt __used, |
683 | const char *arg, int unset __used) | |
684 | { | |
ba77c9e1 LZ |
685 | if (!arg) |
686 | return -1; | |
687 | ||
ba77c9e1 | 688 | if (caller_flag > alloc_flag) |
29b3e152 | 689 | return setup_sorting(&caller_sort, arg); |
ba77c9e1 | 690 | else |
29b3e152 | 691 | return setup_sorting(&alloc_sort, arg); |
ba77c9e1 LZ |
692 | |
693 | return 0; | |
694 | } | |
695 | ||
90b86a9f | 696 | static int parse_caller_opt(const struct option *opt __used, |
79312416 | 697 | const char *arg __used, int unset __used) |
ba77c9e1 | 698 | { |
90b86a9f LZ |
699 | caller_flag = (alloc_flag + 1); |
700 | return 0; | |
701 | } | |
ba77c9e1 | 702 | |
90b86a9f | 703 | static int parse_alloc_opt(const struct option *opt __used, |
79312416 | 704 | const char *arg __used, int unset __used) |
90b86a9f LZ |
705 | { |
706 | alloc_flag = (caller_flag + 1); | |
ba77c9e1 LZ |
707 | return 0; |
708 | } | |
709 | ||
710 | static int parse_line_opt(const struct option *opt __used, | |
711 | const char *arg, int unset __used) | |
712 | { | |
713 | int lines; | |
714 | ||
715 | if (!arg) | |
716 | return -1; | |
717 | ||
718 | lines = strtoul(arg, NULL, 10); | |
719 | ||
720 | if (caller_flag > alloc_flag) | |
721 | caller_lines = lines; | |
722 | else | |
723 | alloc_lines = lines; | |
724 | ||
725 | return 0; | |
726 | } | |
727 | ||
728 | static const struct option kmem_options[] = { | |
729 | OPT_STRING('i', "input", &input_name, "file", | |
730 | "input file name"), | |
90b86a9f LZ |
731 | OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL, |
732 | "show per-callsite statistics", | |
733 | parse_caller_opt), | |
734 | OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL, | |
735 | "show per-allocation statistics", | |
736 | parse_alloc_opt), | |
29b3e152 | 737 | OPT_CALLBACK('s', "sort", NULL, "key[,key2...]", |
079d3f65 | 738 | "sort by keys: ptr, call_site, bytes, hit, pingpong, frag", |
ba77c9e1 LZ |
739 | parse_sort_opt), |
740 | OPT_CALLBACK('l', "line", NULL, "num", | |
90b86a9f | 741 | "show n lines", |
ba77c9e1 | 742 | parse_line_opt), |
7707b6b6 | 743 | OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), |
ba77c9e1 LZ |
744 | OPT_END() |
745 | }; | |
746 | ||
747 | static const char *record_args[] = { | |
748 | "record", | |
749 | "-a", | |
750 | "-R", | |
751 | "-M", | |
752 | "-f", | |
753 | "-c", "1", | |
754 | "-e", "kmem:kmalloc", | |
755 | "-e", "kmem:kmalloc_node", | |
756 | "-e", "kmem:kfree", | |
757 | "-e", "kmem:kmem_cache_alloc", | |
758 | "-e", "kmem:kmem_cache_alloc_node", | |
759 | "-e", "kmem:kmem_cache_free", | |
760 | }; | |
761 | ||
762 | static int __cmd_record(int argc, const char **argv) | |
763 | { | |
764 | unsigned int rec_argc, i, j; | |
765 | const char **rec_argv; | |
766 | ||
767 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
768 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); | |
769 | ||
770 | for (i = 0; i < ARRAY_SIZE(record_args); i++) | |
771 | rec_argv[i] = strdup(record_args[i]); | |
772 | ||
773 | for (j = 1; j < (unsigned int)argc; j++, i++) | |
774 | rec_argv[i] = argv[j]; | |
775 | ||
776 | return cmd_record(i, rec_argv, NULL); | |
777 | } | |
778 | ||
779 | int cmd_kmem(int argc, const char **argv, const char *prefix __used) | |
780 | { | |
781 | symbol__init(0); | |
782 | ||
783 | argc = parse_options(argc, argv, kmem_options, kmem_usage, 0); | |
784 | ||
90b86a9f | 785 | if (!argc) |
ba77c9e1 LZ |
786 | usage_with_options(kmem_usage, kmem_options); |
787 | ||
90b86a9f LZ |
788 | if (!strncmp(argv[0], "rec", 3)) { |
789 | return __cmd_record(argc, argv); | |
790 | } else if (!strcmp(argv[0], "stat")) { | |
791 | setup_cpunode_map(); | |
792 | ||
793 | if (list_empty(&caller_sort)) | |
794 | setup_sorting(&caller_sort, default_sort_order); | |
795 | if (list_empty(&alloc_sort)) | |
796 | setup_sorting(&alloc_sort, default_sort_order); | |
ba77c9e1 | 797 | |
90b86a9f LZ |
798 | return __cmd_kmem(); |
799 | } | |
7d0d3945 | 800 | |
90b86a9f | 801 | return 0; |
ba77c9e1 LZ |
802 | } |
803 |