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" | |
ba77c9e1 LZ |
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 | ||
ba77c9e1 LZ |
23 | static int alloc_flag; |
24 | static int caller_flag; | |
25 | ||
ba77c9e1 LZ |
26 | static int alloc_lines = -1; |
27 | static int caller_lines = -1; | |
28 | ||
7707b6b6 LZ |
29 | static bool raw_ip; |
30 | ||
29b3e152 LZ |
31 | static char default_sort_order[] = "frag,hit,bytes"; |
32 | ||
7d0d3945 LZ |
33 | static int *cpunode_map; |
34 | static int max_cpu_num; | |
35 | ||
ba77c9e1 | 36 | struct alloc_stat { |
079d3f65 LZ |
37 | u64 call_site; |
38 | u64 ptr; | |
ba77c9e1 LZ |
39 | u64 bytes_req; |
40 | u64 bytes_alloc; | |
41 | u32 hit; | |
079d3f65 LZ |
42 | u32 pingpong; |
43 | ||
44 | short alloc_cpu; | |
ba77c9e1 LZ |
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; | |
7d0d3945 | 55 | static unsigned long nr_allocs, nr_cross_allocs; |
ba77c9e1 | 56 | |
7d0d3945 LZ |
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 (true) { | |
96 | dent1 = readdir(dir1); | |
97 | if (!dent1) | |
98 | break; | |
99 | ||
100 | if (sscanf(dent1->d_name, "node%u", &mem) < 1) | |
101 | continue; | |
102 | ||
103 | snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name); | |
104 | dir2 = opendir(buf); | |
105 | if (!dir2) | |
106 | continue; | |
107 | while (true) { | |
108 | dent2 = readdir(dir2); | |
109 | if (!dent2) | |
110 | break; | |
111 | if (sscanf(dent2->d_name, "cpu%u", &cpu) < 1) | |
112 | continue; | |
113 | cpunode_map[cpu] = mem; | |
114 | } | |
115 | } | |
116 | } | |
117 | ||
079d3f65 LZ |
118 | static void insert_alloc_stat(unsigned long call_site, unsigned long ptr, |
119 | int bytes_req, int bytes_alloc, int cpu) | |
ba77c9e1 LZ |
120 | { |
121 | struct rb_node **node = &root_alloc_stat.rb_node; | |
122 | struct rb_node *parent = NULL; | |
123 | struct alloc_stat *data = NULL; | |
124 | ||
ba77c9e1 LZ |
125 | while (*node) { |
126 | parent = *node; | |
127 | data = rb_entry(*node, struct alloc_stat, node); | |
128 | ||
129 | if (ptr > data->ptr) | |
130 | node = &(*node)->rb_right; | |
131 | else if (ptr < data->ptr) | |
132 | node = &(*node)->rb_left; | |
133 | else | |
134 | break; | |
135 | } | |
136 | ||
137 | if (data && data->ptr == ptr) { | |
138 | data->hit++; | |
139 | data->bytes_req += bytes_req; | |
140 | data->bytes_alloc += bytes_req; | |
141 | } else { | |
142 | data = malloc(sizeof(*data)); | |
079d3f65 LZ |
143 | if (!data) |
144 | die("malloc"); | |
ba77c9e1 | 145 | data->ptr = ptr; |
079d3f65 | 146 | data->pingpong = 0; |
ba77c9e1 LZ |
147 | data->hit = 1; |
148 | data->bytes_req = bytes_req; | |
149 | data->bytes_alloc = bytes_alloc; | |
150 | ||
151 | rb_link_node(&data->node, parent, node); | |
152 | rb_insert_color(&data->node, &root_alloc_stat); | |
153 | } | |
079d3f65 LZ |
154 | data->call_site = call_site; |
155 | data->alloc_cpu = cpu; | |
ba77c9e1 LZ |
156 | } |
157 | ||
158 | static void insert_caller_stat(unsigned long call_site, | |
159 | int bytes_req, int bytes_alloc) | |
160 | { | |
161 | struct rb_node **node = &root_caller_stat.rb_node; | |
162 | struct rb_node *parent = NULL; | |
163 | struct alloc_stat *data = NULL; | |
164 | ||
ba77c9e1 LZ |
165 | while (*node) { |
166 | parent = *node; | |
167 | data = rb_entry(*node, struct alloc_stat, node); | |
168 | ||
169 | if (call_site > data->call_site) | |
170 | node = &(*node)->rb_right; | |
171 | else if (call_site < data->call_site) | |
172 | node = &(*node)->rb_left; | |
173 | else | |
174 | break; | |
175 | } | |
176 | ||
177 | if (data && data->call_site == call_site) { | |
178 | data->hit++; | |
179 | data->bytes_req += bytes_req; | |
180 | data->bytes_alloc += bytes_req; | |
181 | } else { | |
182 | data = malloc(sizeof(*data)); | |
079d3f65 LZ |
183 | if (!data) |
184 | die("malloc"); | |
ba77c9e1 | 185 | data->call_site = call_site; |
079d3f65 | 186 | data->pingpong = 0; |
ba77c9e1 LZ |
187 | data->hit = 1; |
188 | data->bytes_req = bytes_req; | |
189 | data->bytes_alloc = bytes_alloc; | |
190 | ||
191 | rb_link_node(&data->node, parent, node); | |
192 | rb_insert_color(&data->node, &root_caller_stat); | |
193 | } | |
194 | } | |
195 | ||
f48f669d | 196 | static void process_alloc_event(void *data, |
ba77c9e1 | 197 | struct event *event, |
7d0d3945 | 198 | int cpu, |
ba77c9e1 LZ |
199 | u64 timestamp __used, |
200 | struct thread *thread __used, | |
7d0d3945 | 201 | int node) |
ba77c9e1 LZ |
202 | { |
203 | unsigned long call_site; | |
204 | unsigned long ptr; | |
205 | int bytes_req; | |
206 | int bytes_alloc; | |
7d0d3945 | 207 | int node1, node2; |
ba77c9e1 | 208 | |
f48f669d XG |
209 | ptr = raw_field_value(event, "ptr", data); |
210 | call_site = raw_field_value(event, "call_site", data); | |
211 | bytes_req = raw_field_value(event, "bytes_req", data); | |
212 | bytes_alloc = raw_field_value(event, "bytes_alloc", data); | |
ba77c9e1 | 213 | |
079d3f65 | 214 | insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu); |
ba77c9e1 LZ |
215 | insert_caller_stat(call_site, bytes_req, bytes_alloc); |
216 | ||
217 | total_requested += bytes_req; | |
218 | total_allocated += bytes_alloc; | |
7d0d3945 LZ |
219 | |
220 | if (node) { | |
221 | node1 = cpunode_map[cpu]; | |
f48f669d | 222 | node2 = raw_field_value(event, "node", data); |
7d0d3945 LZ |
223 | if (node1 != node2) |
224 | nr_cross_allocs++; | |
225 | } | |
226 | nr_allocs++; | |
ba77c9e1 LZ |
227 | } |
228 | ||
079d3f65 LZ |
229 | static int ptr_cmp(struct alloc_stat *, struct alloc_stat *); |
230 | static int callsite_cmp(struct alloc_stat *, struct alloc_stat *); | |
231 | ||
232 | static struct alloc_stat *search_alloc_stat(unsigned long ptr, | |
233 | unsigned long call_site, | |
234 | struct rb_root *root, | |
235 | sort_fn_t sort_fn) | |
236 | { | |
237 | struct rb_node *node = root->rb_node; | |
238 | struct alloc_stat key = { .ptr = ptr, .call_site = call_site }; | |
239 | ||
240 | while (node) { | |
241 | struct alloc_stat *data; | |
242 | int cmp; | |
243 | ||
244 | data = rb_entry(node, struct alloc_stat, node); | |
245 | ||
246 | cmp = sort_fn(&key, data); | |
247 | if (cmp < 0) | |
248 | node = node->rb_left; | |
249 | else if (cmp > 0) | |
250 | node = node->rb_right; | |
251 | else | |
252 | return data; | |
253 | } | |
254 | return NULL; | |
255 | } | |
256 | ||
f48f669d | 257 | static void process_free_event(void *data, |
079d3f65 LZ |
258 | struct event *event, |
259 | int cpu, | |
ba77c9e1 LZ |
260 | u64 timestamp __used, |
261 | struct thread *thread __used) | |
262 | { | |
079d3f65 LZ |
263 | unsigned long ptr; |
264 | struct alloc_stat *s_alloc, *s_caller; | |
265 | ||
f48f669d | 266 | ptr = raw_field_value(event, "ptr", data); |
079d3f65 LZ |
267 | |
268 | s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp); | |
269 | if (!s_alloc) | |
270 | return; | |
271 | ||
272 | if (cpu != s_alloc->alloc_cpu) { | |
273 | s_alloc->pingpong++; | |
274 | ||
275 | s_caller = search_alloc_stat(0, s_alloc->call_site, | |
276 | &root_caller_stat, callsite_cmp); | |
277 | assert(s_caller); | |
278 | s_caller->pingpong++; | |
279 | } | |
280 | s_alloc->alloc_cpu = -1; | |
ba77c9e1 LZ |
281 | } |
282 | ||
283 | static void | |
f48f669d | 284 | process_raw_event(event_t *raw_event __used, void *data, |
ba77c9e1 LZ |
285 | int cpu, u64 timestamp, struct thread *thread) |
286 | { | |
ba77c9e1 LZ |
287 | struct event *event; |
288 | int type; | |
289 | ||
f48f669d | 290 | type = trace_parse_common_type(data); |
ba77c9e1 LZ |
291 | event = trace_find_event(type); |
292 | ||
293 | if (!strcmp(event->name, "kmalloc") || | |
294 | !strcmp(event->name, "kmem_cache_alloc")) { | |
f48f669d | 295 | process_alloc_event(data, event, cpu, timestamp, thread, 0); |
ba77c9e1 LZ |
296 | return; |
297 | } | |
298 | ||
299 | if (!strcmp(event->name, "kmalloc_node") || | |
300 | !strcmp(event->name, "kmem_cache_alloc_node")) { | |
f48f669d | 301 | process_alloc_event(data, event, cpu, timestamp, thread, 1); |
ba77c9e1 LZ |
302 | return; |
303 | } | |
304 | ||
305 | if (!strcmp(event->name, "kfree") || | |
306 | !strcmp(event->name, "kmem_cache_free")) { | |
f48f669d | 307 | process_free_event(data, event, cpu, timestamp, thread); |
ba77c9e1 LZ |
308 | return; |
309 | } | |
310 | } | |
311 | ||
b3165f41 | 312 | static int process_sample_event(event_t *event, struct perf_session *session) |
ba77c9e1 | 313 | { |
180f95e2 OH |
314 | struct sample_data data; |
315 | struct thread *thread; | |
ba77c9e1 | 316 | |
180f95e2 OH |
317 | memset(&data, 0, sizeof(data)); |
318 | data.time = -1; | |
319 | data.cpu = -1; | |
320 | data.period = 1; | |
ba77c9e1 | 321 | |
c019879b | 322 | event__parse_sample(event, session->sample_type, &data); |
ba77c9e1 | 323 | |
62daacb5 | 324 | dump_printf("(IP, %d): %d/%d: %p period: %Ld\n", |
ba77c9e1 | 325 | event->header.misc, |
180f95e2 OH |
326 | data.pid, data.tid, |
327 | (void *)(long)data.ip, | |
328 | (long long)data.period); | |
ba77c9e1 | 329 | |
b3165f41 | 330 | thread = perf_session__findnew(session, event->ip.pid); |
ba77c9e1 LZ |
331 | if (thread == NULL) { |
332 | pr_debug("problem processing %d event, skipping it.\n", | |
333 | event->header.type); | |
334 | return -1; | |
335 | } | |
336 | ||
337 | dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid); | |
338 | ||
f48f669d | 339 | process_raw_event(event, data.raw_data, data.cpu, |
d8bd9e0a | 340 | data.time, thread); |
ba77c9e1 LZ |
341 | |
342 | return 0; | |
343 | } | |
344 | ||
c019879b | 345 | static int sample_type_check(struct perf_session *session) |
ba77c9e1 | 346 | { |
c019879b | 347 | if (!(session->sample_type & PERF_SAMPLE_RAW)) { |
ba77c9e1 LZ |
348 | fprintf(stderr, |
349 | "No trace sample to read. Did you call perf record " | |
350 | "without -R?"); | |
351 | return -1; | |
352 | } | |
353 | ||
354 | return 0; | |
355 | } | |
356 | ||
301a0b02 | 357 | static struct perf_event_ops event_ops = { |
ba77c9e1 | 358 | .process_sample_event = process_sample_event, |
62daacb5 | 359 | .process_comm_event = event__process_comm, |
ba77c9e1 LZ |
360 | .sample_type_check = sample_type_check, |
361 | }; | |
362 | ||
ba77c9e1 LZ |
363 | static double fragmentation(unsigned long n_req, unsigned long n_alloc) |
364 | { | |
365 | if (n_alloc == 0) | |
366 | return 0.0; | |
367 | else | |
368 | return 100.0 - (100.0 * n_req / n_alloc); | |
369 | } | |
370 | ||
4aa65636 ACM |
371 | static void __print_result(struct rb_root *root, struct perf_session *session, |
372 | int n_lines, int is_caller) | |
ba77c9e1 LZ |
373 | { |
374 | struct rb_node *next; | |
375 | ||
079d3f65 LZ |
376 | printf("%.102s\n", graph_dotted_line); |
377 | printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr"); | |
378 | printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n"); | |
379 | printf("%.102s\n", graph_dotted_line); | |
ba77c9e1 LZ |
380 | |
381 | next = rb_first(root); | |
382 | ||
383 | while (next && n_lines--) { | |
1b145ae5 ACM |
384 | struct alloc_stat *data = rb_entry(next, struct alloc_stat, |
385 | node); | |
386 | struct symbol *sym = NULL; | |
079d3f65 | 387 | char buf[BUFSIZ]; |
1b145ae5 ACM |
388 | u64 addr; |
389 | ||
390 | if (is_caller) { | |
391 | addr = data->call_site; | |
7707b6b6 | 392 | if (!raw_ip) |
4aa65636 | 393 | sym = map_groups__find_function(&session->kmaps, session, addr, NULL); |
1b145ae5 ACM |
394 | } else |
395 | addr = data->ptr; | |
396 | ||
397 | if (sym != NULL) | |
079d3f65 | 398 | snprintf(buf, sizeof(buf), "%s+%Lx", sym->name, |
1b145ae5 ACM |
399 | addr - sym->start); |
400 | else | |
079d3f65 LZ |
401 | snprintf(buf, sizeof(buf), "%#Lx", addr); |
402 | printf(" %-34s |", buf); | |
ba77c9e1 | 403 | |
079d3f65 LZ |
404 | printf(" %9llu/%-5lu | %9llu/%-5lu | %6lu | %8lu | %6.3f%%\n", |
405 | (unsigned long long)data->bytes_alloc, | |
ba77c9e1 LZ |
406 | (unsigned long)data->bytes_alloc / data->hit, |
407 | (unsigned long long)data->bytes_req, | |
408 | (unsigned long)data->bytes_req / data->hit, | |
409 | (unsigned long)data->hit, | |
079d3f65 | 410 | (unsigned long)data->pingpong, |
ba77c9e1 LZ |
411 | fragmentation(data->bytes_req, data->bytes_alloc)); |
412 | ||
413 | next = rb_next(next); | |
414 | } | |
415 | ||
416 | if (n_lines == -1) | |
079d3f65 | 417 | printf(" ... | ... | ... | ... | ... | ... \n"); |
ba77c9e1 | 418 | |
079d3f65 | 419 | printf("%.102s\n", graph_dotted_line); |
ba77c9e1 LZ |
420 | } |
421 | ||
422 | static void print_summary(void) | |
423 | { | |
424 | printf("\nSUMMARY\n=======\n"); | |
425 | printf("Total bytes requested: %lu\n", total_requested); | |
426 | printf("Total bytes allocated: %lu\n", total_allocated); | |
427 | printf("Total bytes wasted on internal fragmentation: %lu\n", | |
428 | total_allocated - total_requested); | |
429 | printf("Internal fragmentation: %f%%\n", | |
430 | fragmentation(total_requested, total_allocated)); | |
7d0d3945 | 431 | printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs); |
ba77c9e1 LZ |
432 | } |
433 | ||
4aa65636 | 434 | static void print_result(struct perf_session *session) |
ba77c9e1 LZ |
435 | { |
436 | if (caller_flag) | |
4aa65636 | 437 | __print_result(&root_caller_sorted, session, caller_lines, 1); |
ba77c9e1 | 438 | if (alloc_flag) |
4aa65636 | 439 | __print_result(&root_alloc_sorted, session, alloc_lines, 0); |
ba77c9e1 LZ |
440 | print_summary(); |
441 | } | |
442 | ||
29b3e152 LZ |
443 | struct sort_dimension { |
444 | const char name[20]; | |
445 | sort_fn_t cmp; | |
446 | struct list_head list; | |
447 | }; | |
448 | ||
449 | static LIST_HEAD(caller_sort); | |
450 | static LIST_HEAD(alloc_sort); | |
451 | ||
ba77c9e1 | 452 | static void sort_insert(struct rb_root *root, struct alloc_stat *data, |
29b3e152 | 453 | struct list_head *sort_list) |
ba77c9e1 LZ |
454 | { |
455 | struct rb_node **new = &(root->rb_node); | |
456 | struct rb_node *parent = NULL; | |
29b3e152 | 457 | struct sort_dimension *sort; |
ba77c9e1 LZ |
458 | |
459 | while (*new) { | |
460 | struct alloc_stat *this; | |
29b3e152 | 461 | int cmp = 0; |
ba77c9e1 LZ |
462 | |
463 | this = rb_entry(*new, struct alloc_stat, node); | |
464 | parent = *new; | |
465 | ||
29b3e152 LZ |
466 | list_for_each_entry(sort, sort_list, list) { |
467 | cmp = sort->cmp(data, this); | |
468 | if (cmp) | |
469 | break; | |
470 | } | |
ba77c9e1 LZ |
471 | |
472 | if (cmp > 0) | |
473 | new = &((*new)->rb_left); | |
474 | else | |
475 | new = &((*new)->rb_right); | |
476 | } | |
477 | ||
478 | rb_link_node(&data->node, parent, new); | |
479 | rb_insert_color(&data->node, root); | |
480 | } | |
481 | ||
482 | static void __sort_result(struct rb_root *root, struct rb_root *root_sorted, | |
29b3e152 | 483 | struct list_head *sort_list) |
ba77c9e1 LZ |
484 | { |
485 | struct rb_node *node; | |
486 | struct alloc_stat *data; | |
487 | ||
488 | for (;;) { | |
489 | node = rb_first(root); | |
490 | if (!node) | |
491 | break; | |
492 | ||
493 | rb_erase(node, root); | |
494 | data = rb_entry(node, struct alloc_stat, node); | |
29b3e152 | 495 | sort_insert(root_sorted, data, sort_list); |
ba77c9e1 LZ |
496 | } |
497 | } | |
498 | ||
499 | static void sort_result(void) | |
500 | { | |
29b3e152 LZ |
501 | __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort); |
502 | __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort); | |
ba77c9e1 LZ |
503 | } |
504 | ||
505 | static int __cmd_kmem(void) | |
506 | { | |
4aa65636 ACM |
507 | int err; |
508 | struct perf_session *session = perf_session__new(input_name, O_RDONLY, | |
509 | 0, NULL); | |
510 | if (session == NULL) | |
511 | return -ENOMEM; | |
512 | ||
ba77c9e1 | 513 | setup_pager(); |
4aa65636 ACM |
514 | err = perf_session__process_events(session, &event_ops); |
515 | if (err != 0) | |
516 | goto out_delete; | |
ba77c9e1 | 517 | sort_result(); |
4aa65636 ACM |
518 | print_result(session); |
519 | out_delete: | |
520 | perf_session__delete(session); | |
521 | return err; | |
ba77c9e1 LZ |
522 | } |
523 | ||
524 | static const char * const kmem_usage[] = { | |
90b86a9f | 525 | "perf kmem [<options>] {record|stat}", |
ba77c9e1 LZ |
526 | NULL |
527 | }; | |
528 | ||
ba77c9e1 LZ |
529 | static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r) |
530 | { | |
531 | if (l->ptr < r->ptr) | |
532 | return -1; | |
533 | else if (l->ptr > r->ptr) | |
534 | return 1; | |
535 | return 0; | |
536 | } | |
537 | ||
29b3e152 LZ |
538 | static struct sort_dimension ptr_sort_dimension = { |
539 | .name = "ptr", | |
540 | .cmp = ptr_cmp, | |
541 | }; | |
542 | ||
ba77c9e1 LZ |
543 | static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r) |
544 | { | |
545 | if (l->call_site < r->call_site) | |
546 | return -1; | |
547 | else if (l->call_site > r->call_site) | |
548 | return 1; | |
549 | return 0; | |
550 | } | |
551 | ||
29b3e152 LZ |
552 | static struct sort_dimension callsite_sort_dimension = { |
553 | .name = "callsite", | |
554 | .cmp = callsite_cmp, | |
555 | }; | |
556 | ||
f3ced7cd PE |
557 | static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r) |
558 | { | |
559 | if (l->hit < r->hit) | |
560 | return -1; | |
561 | else if (l->hit > r->hit) | |
562 | return 1; | |
563 | return 0; | |
564 | } | |
565 | ||
29b3e152 LZ |
566 | static struct sort_dimension hit_sort_dimension = { |
567 | .name = "hit", | |
568 | .cmp = hit_cmp, | |
569 | }; | |
570 | ||
ba77c9e1 LZ |
571 | static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r) |
572 | { | |
573 | if (l->bytes_alloc < r->bytes_alloc) | |
574 | return -1; | |
575 | else if (l->bytes_alloc > r->bytes_alloc) | |
576 | return 1; | |
577 | return 0; | |
578 | } | |
579 | ||
29b3e152 LZ |
580 | static struct sort_dimension bytes_sort_dimension = { |
581 | .name = "bytes", | |
582 | .cmp = bytes_cmp, | |
583 | }; | |
584 | ||
f3ced7cd PE |
585 | static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r) |
586 | { | |
587 | double x, y; | |
588 | ||
589 | x = fragmentation(l->bytes_req, l->bytes_alloc); | |
590 | y = fragmentation(r->bytes_req, r->bytes_alloc); | |
591 | ||
592 | if (x < y) | |
593 | return -1; | |
594 | else if (x > y) | |
595 | return 1; | |
596 | return 0; | |
597 | } | |
598 | ||
29b3e152 LZ |
599 | static struct sort_dimension frag_sort_dimension = { |
600 | .name = "frag", | |
601 | .cmp = frag_cmp, | |
602 | }; | |
603 | ||
079d3f65 LZ |
604 | static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r) |
605 | { | |
606 | if (l->pingpong < r->pingpong) | |
607 | return -1; | |
608 | else if (l->pingpong > r->pingpong) | |
609 | return 1; | |
610 | return 0; | |
611 | } | |
612 | ||
613 | static struct sort_dimension pingpong_sort_dimension = { | |
614 | .name = "pingpong", | |
615 | .cmp = pingpong_cmp, | |
616 | }; | |
617 | ||
29b3e152 LZ |
618 | static struct sort_dimension *avail_sorts[] = { |
619 | &ptr_sort_dimension, | |
620 | &callsite_sort_dimension, | |
621 | &hit_sort_dimension, | |
622 | &bytes_sort_dimension, | |
623 | &frag_sort_dimension, | |
079d3f65 | 624 | &pingpong_sort_dimension, |
29b3e152 LZ |
625 | }; |
626 | ||
627 | #define NUM_AVAIL_SORTS \ | |
628 | (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *)) | |
629 | ||
630 | static int sort_dimension__add(const char *tok, struct list_head *list) | |
631 | { | |
632 | struct sort_dimension *sort; | |
633 | int i; | |
634 | ||
635 | for (i = 0; i < NUM_AVAIL_SORTS; i++) { | |
636 | if (!strcmp(avail_sorts[i]->name, tok)) { | |
637 | sort = malloc(sizeof(*sort)); | |
638 | if (!sort) | |
639 | die("malloc"); | |
640 | memcpy(sort, avail_sorts[i], sizeof(*sort)); | |
641 | list_add_tail(&sort->list, list); | |
642 | return 0; | |
643 | } | |
644 | } | |
645 | ||
646 | return -1; | |
647 | } | |
648 | ||
649 | static int setup_sorting(struct list_head *sort_list, const char *arg) | |
650 | { | |
651 | char *tok; | |
652 | char *str = strdup(arg); | |
653 | ||
654 | if (!str) | |
655 | die("strdup"); | |
656 | ||
657 | while (true) { | |
658 | tok = strsep(&str, ","); | |
659 | if (!tok) | |
660 | break; | |
661 | if (sort_dimension__add(tok, sort_list) < 0) { | |
662 | error("Unknown --sort key: '%s'", tok); | |
663 | return -1; | |
664 | } | |
665 | } | |
666 | ||
667 | free(str); | |
668 | return 0; | |
669 | } | |
670 | ||
ba77c9e1 LZ |
671 | static int parse_sort_opt(const struct option *opt __used, |
672 | const char *arg, int unset __used) | |
673 | { | |
ba77c9e1 LZ |
674 | if (!arg) |
675 | return -1; | |
676 | ||
ba77c9e1 | 677 | if (caller_flag > alloc_flag) |
29b3e152 | 678 | return setup_sorting(&caller_sort, arg); |
ba77c9e1 | 679 | else |
29b3e152 | 680 | return setup_sorting(&alloc_sort, arg); |
ba77c9e1 LZ |
681 | |
682 | return 0; | |
683 | } | |
684 | ||
90b86a9f | 685 | static int parse_caller_opt(const struct option *opt __used, |
79312416 | 686 | const char *arg __used, int unset __used) |
ba77c9e1 | 687 | { |
90b86a9f LZ |
688 | caller_flag = (alloc_flag + 1); |
689 | return 0; | |
690 | } | |
ba77c9e1 | 691 | |
90b86a9f | 692 | static int parse_alloc_opt(const struct option *opt __used, |
79312416 | 693 | const char *arg __used, int unset __used) |
90b86a9f LZ |
694 | { |
695 | alloc_flag = (caller_flag + 1); | |
ba77c9e1 LZ |
696 | return 0; |
697 | } | |
698 | ||
699 | static int parse_line_opt(const struct option *opt __used, | |
700 | const char *arg, int unset __used) | |
701 | { | |
702 | int lines; | |
703 | ||
704 | if (!arg) | |
705 | return -1; | |
706 | ||
707 | lines = strtoul(arg, NULL, 10); | |
708 | ||
709 | if (caller_flag > alloc_flag) | |
710 | caller_lines = lines; | |
711 | else | |
712 | alloc_lines = lines; | |
713 | ||
714 | return 0; | |
715 | } | |
716 | ||
717 | static const struct option kmem_options[] = { | |
718 | OPT_STRING('i', "input", &input_name, "file", | |
719 | "input file name"), | |
90b86a9f LZ |
720 | OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL, |
721 | "show per-callsite statistics", | |
722 | parse_caller_opt), | |
723 | OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL, | |
724 | "show per-allocation statistics", | |
725 | parse_alloc_opt), | |
29b3e152 | 726 | OPT_CALLBACK('s', "sort", NULL, "key[,key2...]", |
079d3f65 | 727 | "sort by keys: ptr, call_site, bytes, hit, pingpong, frag", |
ba77c9e1 LZ |
728 | parse_sort_opt), |
729 | OPT_CALLBACK('l', "line", NULL, "num", | |
90b86a9f | 730 | "show n lines", |
ba77c9e1 | 731 | parse_line_opt), |
7707b6b6 | 732 | OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), |
ba77c9e1 LZ |
733 | OPT_END() |
734 | }; | |
735 | ||
736 | static const char *record_args[] = { | |
737 | "record", | |
738 | "-a", | |
739 | "-R", | |
740 | "-M", | |
741 | "-f", | |
742 | "-c", "1", | |
743 | "-e", "kmem:kmalloc", | |
744 | "-e", "kmem:kmalloc_node", | |
745 | "-e", "kmem:kfree", | |
746 | "-e", "kmem:kmem_cache_alloc", | |
747 | "-e", "kmem:kmem_cache_alloc_node", | |
748 | "-e", "kmem:kmem_cache_free", | |
749 | }; | |
750 | ||
751 | static int __cmd_record(int argc, const char **argv) | |
752 | { | |
753 | unsigned int rec_argc, i, j; | |
754 | const char **rec_argv; | |
755 | ||
756 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
757 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); | |
758 | ||
759 | for (i = 0; i < ARRAY_SIZE(record_args); i++) | |
760 | rec_argv[i] = strdup(record_args[i]); | |
761 | ||
762 | for (j = 1; j < (unsigned int)argc; j++, i++) | |
763 | rec_argv[i] = argv[j]; | |
764 | ||
765 | return cmd_record(i, rec_argv, NULL); | |
766 | } | |
767 | ||
768 | int cmd_kmem(int argc, const char **argv, const char *prefix __used) | |
769 | { | |
770 | symbol__init(0); | |
771 | ||
772 | argc = parse_options(argc, argv, kmem_options, kmem_usage, 0); | |
773 | ||
90b86a9f | 774 | if (!argc) |
ba77c9e1 LZ |
775 | usage_with_options(kmem_usage, kmem_options); |
776 | ||
90b86a9f LZ |
777 | if (!strncmp(argv[0], "rec", 3)) { |
778 | return __cmd_record(argc, argv); | |
779 | } else if (!strcmp(argv[0], "stat")) { | |
780 | setup_cpunode_map(); | |
781 | ||
782 | if (list_empty(&caller_sort)) | |
783 | setup_sorting(&caller_sort, default_sort_order); | |
784 | if (list_empty(&alloc_sort)) | |
785 | setup_sorting(&alloc_sort, default_sort_order); | |
ba77c9e1 | 786 | |
90b86a9f LZ |
787 | return __cmd_kmem(); |
788 | } | |
7d0d3945 | 789 | |
90b86a9f | 790 | return 0; |
ba77c9e1 LZ |
791 | } |
792 |