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 | ||
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 LZ |
59 | |
60 | struct raw_event_sample { | |
61 | u32 size; | |
62 | char data[0]; | |
63 | }; | |
64 | ||
7d0d3945 LZ |
65 | #define PATH_SYS_NODE "/sys/devices/system/node" |
66 | ||
67 | static void init_cpunode_map(void) | |
68 | { | |
69 | FILE *fp; | |
70 | int i; | |
71 | ||
72 | fp = fopen("/sys/devices/system/cpu/kernel_max", "r"); | |
73 | if (!fp) { | |
74 | max_cpu_num = 4096; | |
75 | return; | |
76 | } | |
77 | ||
78 | if (fscanf(fp, "%d", &max_cpu_num) < 1) | |
79 | die("Failed to read 'kernel_max' from sysfs"); | |
80 | max_cpu_num++; | |
81 | ||
82 | cpunode_map = calloc(max_cpu_num, sizeof(int)); | |
83 | if (!cpunode_map) | |
84 | die("calloc"); | |
85 | for (i = 0; i < max_cpu_num; i++) | |
86 | cpunode_map[i] = -1; | |
87 | fclose(fp); | |
88 | } | |
89 | ||
90 | static void setup_cpunode_map(void) | |
91 | { | |
92 | struct dirent *dent1, *dent2; | |
93 | DIR *dir1, *dir2; | |
94 | unsigned int cpu, mem; | |
95 | char buf[PATH_MAX]; | |
96 | ||
97 | init_cpunode_map(); | |
98 | ||
99 | dir1 = opendir(PATH_SYS_NODE); | |
100 | if (!dir1) | |
101 | return; | |
102 | ||
103 | while (true) { | |
104 | dent1 = readdir(dir1); | |
105 | if (!dent1) | |
106 | break; | |
107 | ||
108 | if (sscanf(dent1->d_name, "node%u", &mem) < 1) | |
109 | continue; | |
110 | ||
111 | snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name); | |
112 | dir2 = opendir(buf); | |
113 | if (!dir2) | |
114 | continue; | |
115 | while (true) { | |
116 | dent2 = readdir(dir2); | |
117 | if (!dent2) | |
118 | break; | |
119 | if (sscanf(dent2->d_name, "cpu%u", &cpu) < 1) | |
120 | continue; | |
121 | cpunode_map[cpu] = mem; | |
122 | } | |
123 | } | |
124 | } | |
125 | ||
079d3f65 LZ |
126 | static void insert_alloc_stat(unsigned long call_site, unsigned long ptr, |
127 | int bytes_req, int bytes_alloc, int cpu) | |
ba77c9e1 LZ |
128 | { |
129 | struct rb_node **node = &root_alloc_stat.rb_node; | |
130 | struct rb_node *parent = NULL; | |
131 | struct alloc_stat *data = NULL; | |
132 | ||
ba77c9e1 LZ |
133 | while (*node) { |
134 | parent = *node; | |
135 | data = rb_entry(*node, struct alloc_stat, node); | |
136 | ||
137 | if (ptr > data->ptr) | |
138 | node = &(*node)->rb_right; | |
139 | else if (ptr < data->ptr) | |
140 | node = &(*node)->rb_left; | |
141 | else | |
142 | break; | |
143 | } | |
144 | ||
145 | if (data && data->ptr == ptr) { | |
146 | data->hit++; | |
147 | data->bytes_req += bytes_req; | |
148 | data->bytes_alloc += bytes_req; | |
149 | } else { | |
150 | data = malloc(sizeof(*data)); | |
079d3f65 LZ |
151 | if (!data) |
152 | die("malloc"); | |
ba77c9e1 | 153 | data->ptr = ptr; |
079d3f65 | 154 | data->pingpong = 0; |
ba77c9e1 LZ |
155 | data->hit = 1; |
156 | data->bytes_req = bytes_req; | |
157 | data->bytes_alloc = bytes_alloc; | |
158 | ||
159 | rb_link_node(&data->node, parent, node); | |
160 | rb_insert_color(&data->node, &root_alloc_stat); | |
161 | } | |
079d3f65 LZ |
162 | data->call_site = call_site; |
163 | data->alloc_cpu = cpu; | |
ba77c9e1 LZ |
164 | } |
165 | ||
166 | static void insert_caller_stat(unsigned long call_site, | |
167 | int bytes_req, int bytes_alloc) | |
168 | { | |
169 | struct rb_node **node = &root_caller_stat.rb_node; | |
170 | struct rb_node *parent = NULL; | |
171 | struct alloc_stat *data = NULL; | |
172 | ||
ba77c9e1 LZ |
173 | while (*node) { |
174 | parent = *node; | |
175 | data = rb_entry(*node, struct alloc_stat, node); | |
176 | ||
177 | if (call_site > data->call_site) | |
178 | node = &(*node)->rb_right; | |
179 | else if (call_site < data->call_site) | |
180 | node = &(*node)->rb_left; | |
181 | else | |
182 | break; | |
183 | } | |
184 | ||
185 | if (data && data->call_site == call_site) { | |
186 | data->hit++; | |
187 | data->bytes_req += bytes_req; | |
188 | data->bytes_alloc += bytes_req; | |
189 | } else { | |
190 | data = malloc(sizeof(*data)); | |
079d3f65 LZ |
191 | if (!data) |
192 | die("malloc"); | |
ba77c9e1 | 193 | data->call_site = call_site; |
079d3f65 | 194 | data->pingpong = 0; |
ba77c9e1 LZ |
195 | data->hit = 1; |
196 | data->bytes_req = bytes_req; | |
197 | data->bytes_alloc = bytes_alloc; | |
198 | ||
199 | rb_link_node(&data->node, parent, node); | |
200 | rb_insert_color(&data->node, &root_caller_stat); | |
201 | } | |
202 | } | |
203 | ||
204 | static void process_alloc_event(struct raw_event_sample *raw, | |
205 | struct event *event, | |
7d0d3945 | 206 | int cpu, |
ba77c9e1 LZ |
207 | u64 timestamp __used, |
208 | struct thread *thread __used, | |
7d0d3945 | 209 | int node) |
ba77c9e1 LZ |
210 | { |
211 | unsigned long call_site; | |
212 | unsigned long ptr; | |
213 | int bytes_req; | |
214 | int bytes_alloc; | |
7d0d3945 | 215 | int node1, node2; |
ba77c9e1 LZ |
216 | |
217 | ptr = raw_field_value(event, "ptr", raw->data); | |
218 | call_site = raw_field_value(event, "call_site", raw->data); | |
219 | bytes_req = raw_field_value(event, "bytes_req", raw->data); | |
220 | bytes_alloc = raw_field_value(event, "bytes_alloc", raw->data); | |
221 | ||
079d3f65 | 222 | insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu); |
ba77c9e1 LZ |
223 | insert_caller_stat(call_site, bytes_req, bytes_alloc); |
224 | ||
225 | total_requested += bytes_req; | |
226 | total_allocated += bytes_alloc; | |
7d0d3945 LZ |
227 | |
228 | if (node) { | |
229 | node1 = cpunode_map[cpu]; | |
230 | node2 = raw_field_value(event, "node", raw->data); | |
231 | if (node1 != node2) | |
232 | nr_cross_allocs++; | |
233 | } | |
234 | nr_allocs++; | |
ba77c9e1 LZ |
235 | } |
236 | ||
079d3f65 LZ |
237 | static int ptr_cmp(struct alloc_stat *, struct alloc_stat *); |
238 | static int callsite_cmp(struct alloc_stat *, struct alloc_stat *); | |
239 | ||
240 | static struct alloc_stat *search_alloc_stat(unsigned long ptr, | |
241 | unsigned long call_site, | |
242 | struct rb_root *root, | |
243 | sort_fn_t sort_fn) | |
244 | { | |
245 | struct rb_node *node = root->rb_node; | |
246 | struct alloc_stat key = { .ptr = ptr, .call_site = call_site }; | |
247 | ||
248 | while (node) { | |
249 | struct alloc_stat *data; | |
250 | int cmp; | |
251 | ||
252 | data = rb_entry(node, struct alloc_stat, node); | |
253 | ||
254 | cmp = sort_fn(&key, data); | |
255 | if (cmp < 0) | |
256 | node = node->rb_left; | |
257 | else if (cmp > 0) | |
258 | node = node->rb_right; | |
259 | else | |
260 | return data; | |
261 | } | |
262 | return NULL; | |
263 | } | |
264 | ||
265 | static void process_free_event(struct raw_event_sample *raw, | |
266 | struct event *event, | |
267 | int cpu, | |
ba77c9e1 LZ |
268 | u64 timestamp __used, |
269 | struct thread *thread __used) | |
270 | { | |
079d3f65 LZ |
271 | unsigned long ptr; |
272 | struct alloc_stat *s_alloc, *s_caller; | |
273 | ||
274 | ptr = raw_field_value(event, "ptr", raw->data); | |
275 | ||
276 | s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp); | |
277 | if (!s_alloc) | |
278 | return; | |
279 | ||
280 | if (cpu != s_alloc->alloc_cpu) { | |
281 | s_alloc->pingpong++; | |
282 | ||
283 | s_caller = search_alloc_stat(0, s_alloc->call_site, | |
284 | &root_caller_stat, callsite_cmp); | |
285 | assert(s_caller); | |
286 | s_caller->pingpong++; | |
287 | } | |
288 | s_alloc->alloc_cpu = -1; | |
ba77c9e1 LZ |
289 | } |
290 | ||
291 | static void | |
292 | process_raw_event(event_t *raw_event __used, void *more_data, | |
293 | int cpu, u64 timestamp, struct thread *thread) | |
294 | { | |
295 | struct raw_event_sample *raw = more_data; | |
296 | struct event *event; | |
297 | int type; | |
298 | ||
299 | type = trace_parse_common_type(raw->data); | |
300 | event = trace_find_event(type); | |
301 | ||
302 | if (!strcmp(event->name, "kmalloc") || | |
303 | !strcmp(event->name, "kmem_cache_alloc")) { | |
304 | process_alloc_event(raw, event, cpu, timestamp, thread, 0); | |
305 | return; | |
306 | } | |
307 | ||
308 | if (!strcmp(event->name, "kmalloc_node") || | |
309 | !strcmp(event->name, "kmem_cache_alloc_node")) { | |
310 | process_alloc_event(raw, event, cpu, timestamp, thread, 1); | |
311 | return; | |
312 | } | |
313 | ||
314 | if (!strcmp(event->name, "kfree") || | |
315 | !strcmp(event->name, "kmem_cache_free")) { | |
316 | process_free_event(raw, event, cpu, timestamp, thread); | |
317 | return; | |
318 | } | |
319 | } | |
320 | ||
62daacb5 | 321 | static int process_sample_event(event_t *event) |
ba77c9e1 LZ |
322 | { |
323 | u64 ip = event->ip.ip; | |
324 | u64 timestamp = -1; | |
325 | u32 cpu = -1; | |
326 | u64 period = 1; | |
327 | void *more_data = event->ip.__more_data; | |
328 | struct thread *thread = threads__findnew(event->ip.pid); | |
329 | ||
330 | if (sample_type & PERF_SAMPLE_TIME) { | |
331 | timestamp = *(u64 *)more_data; | |
332 | more_data += sizeof(u64); | |
333 | } | |
334 | ||
335 | if (sample_type & PERF_SAMPLE_CPU) { | |
336 | cpu = *(u32 *)more_data; | |
337 | more_data += sizeof(u32); | |
338 | more_data += sizeof(u32); /* reserved */ | |
339 | } | |
340 | ||
341 | if (sample_type & PERF_SAMPLE_PERIOD) { | |
342 | period = *(u64 *)more_data; | |
343 | more_data += sizeof(u64); | |
344 | } | |
345 | ||
62daacb5 | 346 | dump_printf("(IP, %d): %d/%d: %p period: %Ld\n", |
ba77c9e1 LZ |
347 | event->header.misc, |
348 | event->ip.pid, event->ip.tid, | |
349 | (void *)(long)ip, | |
350 | (long long)period); | |
351 | ||
352 | if (thread == NULL) { | |
353 | pr_debug("problem processing %d event, skipping it.\n", | |
354 | event->header.type); | |
355 | return -1; | |
356 | } | |
357 | ||
358 | dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid); | |
359 | ||
360 | process_raw_event(event, more_data, cpu, timestamp, thread); | |
361 | ||
362 | return 0; | |
363 | } | |
364 | ||
365 | static int sample_type_check(u64 type) | |
366 | { | |
367 | sample_type = type; | |
368 | ||
369 | if (!(sample_type & PERF_SAMPLE_RAW)) { | |
370 | fprintf(stderr, | |
371 | "No trace sample to read. Did you call perf record " | |
372 | "without -R?"); | |
373 | return -1; | |
374 | } | |
375 | ||
376 | return 0; | |
377 | } | |
378 | ||
379 | static struct perf_file_handler file_handler = { | |
380 | .process_sample_event = process_sample_event, | |
62daacb5 | 381 | .process_comm_event = event__process_comm, |
ba77c9e1 LZ |
382 | .sample_type_check = sample_type_check, |
383 | }; | |
384 | ||
385 | static int read_events(void) | |
386 | { | |
387 | register_idle_thread(); | |
388 | register_perf_file_handler(&file_handler); | |
389 | ||
b32d133a | 390 | return mmap_dispatch_perf_file(&header, input_name, 0, 0, |
62daacb5 | 391 | &event__cwdlen, &event__cwd); |
ba77c9e1 LZ |
392 | } |
393 | ||
394 | static double fragmentation(unsigned long n_req, unsigned long n_alloc) | |
395 | { | |
396 | if (n_alloc == 0) | |
397 | return 0.0; | |
398 | else | |
399 | return 100.0 - (100.0 * n_req / n_alloc); | |
400 | } | |
401 | ||
402 | static void __print_result(struct rb_root *root, int n_lines, int is_caller) | |
403 | { | |
404 | struct rb_node *next; | |
405 | ||
079d3f65 LZ |
406 | printf("%.102s\n", graph_dotted_line); |
407 | printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr"); | |
408 | printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n"); | |
409 | printf("%.102s\n", graph_dotted_line); | |
ba77c9e1 LZ |
410 | |
411 | next = rb_first(root); | |
412 | ||
413 | while (next && n_lines--) { | |
1b145ae5 ACM |
414 | struct alloc_stat *data = rb_entry(next, struct alloc_stat, |
415 | node); | |
416 | struct symbol *sym = NULL; | |
079d3f65 | 417 | char buf[BUFSIZ]; |
1b145ae5 ACM |
418 | u64 addr; |
419 | ||
420 | if (is_caller) { | |
421 | addr = data->call_site; | |
7707b6b6 | 422 | if (!raw_ip) |
1ed091c4 | 423 | sym = thread__find_function(kthread, addr, NULL); |
1b145ae5 ACM |
424 | } else |
425 | addr = data->ptr; | |
426 | ||
427 | if (sym != NULL) | |
079d3f65 | 428 | snprintf(buf, sizeof(buf), "%s+%Lx", sym->name, |
1b145ae5 ACM |
429 | addr - sym->start); |
430 | else | |
079d3f65 LZ |
431 | snprintf(buf, sizeof(buf), "%#Lx", addr); |
432 | printf(" %-34s |", buf); | |
ba77c9e1 | 433 | |
079d3f65 LZ |
434 | printf(" %9llu/%-5lu | %9llu/%-5lu | %6lu | %8lu | %6.3f%%\n", |
435 | (unsigned long long)data->bytes_alloc, | |
ba77c9e1 LZ |
436 | (unsigned long)data->bytes_alloc / data->hit, |
437 | (unsigned long long)data->bytes_req, | |
438 | (unsigned long)data->bytes_req / data->hit, | |
439 | (unsigned long)data->hit, | |
079d3f65 | 440 | (unsigned long)data->pingpong, |
ba77c9e1 LZ |
441 | fragmentation(data->bytes_req, data->bytes_alloc)); |
442 | ||
443 | next = rb_next(next); | |
444 | } | |
445 | ||
446 | if (n_lines == -1) | |
079d3f65 | 447 | printf(" ... | ... | ... | ... | ... | ... \n"); |
ba77c9e1 | 448 | |
079d3f65 | 449 | printf("%.102s\n", graph_dotted_line); |
ba77c9e1 LZ |
450 | } |
451 | ||
452 | static void print_summary(void) | |
453 | { | |
454 | printf("\nSUMMARY\n=======\n"); | |
455 | printf("Total bytes requested: %lu\n", total_requested); | |
456 | printf("Total bytes allocated: %lu\n", total_allocated); | |
457 | printf("Total bytes wasted on internal fragmentation: %lu\n", | |
458 | total_allocated - total_requested); | |
459 | printf("Internal fragmentation: %f%%\n", | |
460 | fragmentation(total_requested, total_allocated)); | |
7d0d3945 | 461 | printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs); |
ba77c9e1 LZ |
462 | } |
463 | ||
464 | static void print_result(void) | |
465 | { | |
466 | if (caller_flag) | |
467 | __print_result(&root_caller_sorted, caller_lines, 1); | |
468 | if (alloc_flag) | |
469 | __print_result(&root_alloc_sorted, alloc_lines, 0); | |
470 | print_summary(); | |
471 | } | |
472 | ||
29b3e152 LZ |
473 | struct sort_dimension { |
474 | const char name[20]; | |
475 | sort_fn_t cmp; | |
476 | struct list_head list; | |
477 | }; | |
478 | ||
479 | static LIST_HEAD(caller_sort); | |
480 | static LIST_HEAD(alloc_sort); | |
481 | ||
ba77c9e1 | 482 | static void sort_insert(struct rb_root *root, struct alloc_stat *data, |
29b3e152 | 483 | struct list_head *sort_list) |
ba77c9e1 LZ |
484 | { |
485 | struct rb_node **new = &(root->rb_node); | |
486 | struct rb_node *parent = NULL; | |
29b3e152 | 487 | struct sort_dimension *sort; |
ba77c9e1 LZ |
488 | |
489 | while (*new) { | |
490 | struct alloc_stat *this; | |
29b3e152 | 491 | int cmp = 0; |
ba77c9e1 LZ |
492 | |
493 | this = rb_entry(*new, struct alloc_stat, node); | |
494 | parent = *new; | |
495 | ||
29b3e152 LZ |
496 | list_for_each_entry(sort, sort_list, list) { |
497 | cmp = sort->cmp(data, this); | |
498 | if (cmp) | |
499 | break; | |
500 | } | |
ba77c9e1 LZ |
501 | |
502 | if (cmp > 0) | |
503 | new = &((*new)->rb_left); | |
504 | else | |
505 | new = &((*new)->rb_right); | |
506 | } | |
507 | ||
508 | rb_link_node(&data->node, parent, new); | |
509 | rb_insert_color(&data->node, root); | |
510 | } | |
511 | ||
512 | static void __sort_result(struct rb_root *root, struct rb_root *root_sorted, | |
29b3e152 | 513 | struct list_head *sort_list) |
ba77c9e1 LZ |
514 | { |
515 | struct rb_node *node; | |
516 | struct alloc_stat *data; | |
517 | ||
518 | for (;;) { | |
519 | node = rb_first(root); | |
520 | if (!node) | |
521 | break; | |
522 | ||
523 | rb_erase(node, root); | |
524 | data = rb_entry(node, struct alloc_stat, node); | |
29b3e152 | 525 | sort_insert(root_sorted, data, sort_list); |
ba77c9e1 LZ |
526 | } |
527 | } | |
528 | ||
529 | static void sort_result(void) | |
530 | { | |
29b3e152 LZ |
531 | __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort); |
532 | __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort); | |
ba77c9e1 LZ |
533 | } |
534 | ||
535 | static int __cmd_kmem(void) | |
536 | { | |
537 | setup_pager(); | |
538 | read_events(); | |
539 | sort_result(); | |
540 | print_result(); | |
541 | ||
542 | return 0; | |
543 | } | |
544 | ||
545 | static const char * const kmem_usage[] = { | |
546 | "perf kmem [<options>] {record}", | |
547 | NULL | |
548 | }; | |
549 | ||
ba77c9e1 LZ |
550 | static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r) |
551 | { | |
552 | if (l->ptr < r->ptr) | |
553 | return -1; | |
554 | else if (l->ptr > r->ptr) | |
555 | return 1; | |
556 | return 0; | |
557 | } | |
558 | ||
29b3e152 LZ |
559 | static struct sort_dimension ptr_sort_dimension = { |
560 | .name = "ptr", | |
561 | .cmp = ptr_cmp, | |
562 | }; | |
563 | ||
ba77c9e1 LZ |
564 | static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r) |
565 | { | |
566 | if (l->call_site < r->call_site) | |
567 | return -1; | |
568 | else if (l->call_site > r->call_site) | |
569 | return 1; | |
570 | return 0; | |
571 | } | |
572 | ||
29b3e152 LZ |
573 | static struct sort_dimension callsite_sort_dimension = { |
574 | .name = "callsite", | |
575 | .cmp = callsite_cmp, | |
576 | }; | |
577 | ||
f3ced7cd PE |
578 | static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r) |
579 | { | |
580 | if (l->hit < r->hit) | |
581 | return -1; | |
582 | else if (l->hit > r->hit) | |
583 | return 1; | |
584 | return 0; | |
585 | } | |
586 | ||
29b3e152 LZ |
587 | static struct sort_dimension hit_sort_dimension = { |
588 | .name = "hit", | |
589 | .cmp = hit_cmp, | |
590 | }; | |
591 | ||
ba77c9e1 LZ |
592 | static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r) |
593 | { | |
594 | if (l->bytes_alloc < r->bytes_alloc) | |
595 | return -1; | |
596 | else if (l->bytes_alloc > r->bytes_alloc) | |
597 | return 1; | |
598 | return 0; | |
599 | } | |
600 | ||
29b3e152 LZ |
601 | static struct sort_dimension bytes_sort_dimension = { |
602 | .name = "bytes", | |
603 | .cmp = bytes_cmp, | |
604 | }; | |
605 | ||
f3ced7cd PE |
606 | static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r) |
607 | { | |
608 | double x, y; | |
609 | ||
610 | x = fragmentation(l->bytes_req, l->bytes_alloc); | |
611 | y = fragmentation(r->bytes_req, r->bytes_alloc); | |
612 | ||
613 | if (x < y) | |
614 | return -1; | |
615 | else if (x > y) | |
616 | return 1; | |
617 | return 0; | |
618 | } | |
619 | ||
29b3e152 LZ |
620 | static struct sort_dimension frag_sort_dimension = { |
621 | .name = "frag", | |
622 | .cmp = frag_cmp, | |
623 | }; | |
624 | ||
079d3f65 LZ |
625 | static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r) |
626 | { | |
627 | if (l->pingpong < r->pingpong) | |
628 | return -1; | |
629 | else if (l->pingpong > r->pingpong) | |
630 | return 1; | |
631 | return 0; | |
632 | } | |
633 | ||
634 | static struct sort_dimension pingpong_sort_dimension = { | |
635 | .name = "pingpong", | |
636 | .cmp = pingpong_cmp, | |
637 | }; | |
638 | ||
29b3e152 LZ |
639 | static struct sort_dimension *avail_sorts[] = { |
640 | &ptr_sort_dimension, | |
641 | &callsite_sort_dimension, | |
642 | &hit_sort_dimension, | |
643 | &bytes_sort_dimension, | |
644 | &frag_sort_dimension, | |
079d3f65 | 645 | &pingpong_sort_dimension, |
29b3e152 LZ |
646 | }; |
647 | ||
648 | #define NUM_AVAIL_SORTS \ | |
649 | (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *)) | |
650 | ||
651 | static int sort_dimension__add(const char *tok, struct list_head *list) | |
652 | { | |
653 | struct sort_dimension *sort; | |
654 | int i; | |
655 | ||
656 | for (i = 0; i < NUM_AVAIL_SORTS; i++) { | |
657 | if (!strcmp(avail_sorts[i]->name, tok)) { | |
658 | sort = malloc(sizeof(*sort)); | |
659 | if (!sort) | |
660 | die("malloc"); | |
661 | memcpy(sort, avail_sorts[i], sizeof(*sort)); | |
662 | list_add_tail(&sort->list, list); | |
663 | return 0; | |
664 | } | |
665 | } | |
666 | ||
667 | return -1; | |
668 | } | |
669 | ||
670 | static int setup_sorting(struct list_head *sort_list, const char *arg) | |
671 | { | |
672 | char *tok; | |
673 | char *str = strdup(arg); | |
674 | ||
675 | if (!str) | |
676 | die("strdup"); | |
677 | ||
678 | while (true) { | |
679 | tok = strsep(&str, ","); | |
680 | if (!tok) | |
681 | break; | |
682 | if (sort_dimension__add(tok, sort_list) < 0) { | |
683 | error("Unknown --sort key: '%s'", tok); | |
684 | return -1; | |
685 | } | |
686 | } | |
687 | ||
688 | free(str); | |
689 | return 0; | |
690 | } | |
691 | ||
ba77c9e1 LZ |
692 | static int parse_sort_opt(const struct option *opt __used, |
693 | const char *arg, int unset __used) | |
694 | { | |
ba77c9e1 LZ |
695 | if (!arg) |
696 | return -1; | |
697 | ||
ba77c9e1 | 698 | if (caller_flag > alloc_flag) |
29b3e152 | 699 | return setup_sorting(&caller_sort, arg); |
ba77c9e1 | 700 | else |
29b3e152 | 701 | return setup_sorting(&alloc_sort, arg); |
ba77c9e1 LZ |
702 | |
703 | return 0; | |
704 | } | |
705 | ||
706 | static int parse_stat_opt(const struct option *opt __used, | |
707 | const char *arg, int unset __used) | |
708 | { | |
709 | if (!arg) | |
710 | return -1; | |
711 | ||
712 | if (strcmp(arg, "alloc") == 0) | |
713 | alloc_flag = (caller_flag + 1); | |
714 | else if (strcmp(arg, "caller") == 0) | |
715 | caller_flag = (alloc_flag + 1); | |
716 | else | |
717 | return -1; | |
718 | return 0; | |
719 | } | |
720 | ||
721 | static int parse_line_opt(const struct option *opt __used, | |
722 | const char *arg, int unset __used) | |
723 | { | |
724 | int lines; | |
725 | ||
726 | if (!arg) | |
727 | return -1; | |
728 | ||
729 | lines = strtoul(arg, NULL, 10); | |
730 | ||
731 | if (caller_flag > alloc_flag) | |
732 | caller_lines = lines; | |
733 | else | |
734 | alloc_lines = lines; | |
735 | ||
736 | return 0; | |
737 | } | |
738 | ||
739 | static const struct option kmem_options[] = { | |
740 | OPT_STRING('i', "input", &input_name, "file", | |
741 | "input file name"), | |
742 | OPT_CALLBACK(0, "stat", NULL, "<alloc>|<caller>", | |
743 | "stat selector, Pass 'alloc' or 'caller'.", | |
744 | parse_stat_opt), | |
29b3e152 | 745 | OPT_CALLBACK('s', "sort", NULL, "key[,key2...]", |
079d3f65 | 746 | "sort by keys: ptr, call_site, bytes, hit, pingpong, frag", |
ba77c9e1 LZ |
747 | parse_sort_opt), |
748 | OPT_CALLBACK('l', "line", NULL, "num", | |
749 | "show n lins", | |
750 | parse_line_opt), | |
7707b6b6 | 751 | OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), |
ba77c9e1 LZ |
752 | OPT_END() |
753 | }; | |
754 | ||
755 | static const char *record_args[] = { | |
756 | "record", | |
757 | "-a", | |
758 | "-R", | |
759 | "-M", | |
760 | "-f", | |
761 | "-c", "1", | |
762 | "-e", "kmem:kmalloc", | |
763 | "-e", "kmem:kmalloc_node", | |
764 | "-e", "kmem:kfree", | |
765 | "-e", "kmem:kmem_cache_alloc", | |
766 | "-e", "kmem:kmem_cache_alloc_node", | |
767 | "-e", "kmem:kmem_cache_free", | |
768 | }; | |
769 | ||
770 | static int __cmd_record(int argc, const char **argv) | |
771 | { | |
772 | unsigned int rec_argc, i, j; | |
773 | const char **rec_argv; | |
774 | ||
775 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
776 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); | |
777 | ||
778 | for (i = 0; i < ARRAY_SIZE(record_args); i++) | |
779 | rec_argv[i] = strdup(record_args[i]); | |
780 | ||
781 | for (j = 1; j < (unsigned int)argc; j++, i++) | |
782 | rec_argv[i] = argv[j]; | |
783 | ||
784 | return cmd_record(i, rec_argv, NULL); | |
785 | } | |
786 | ||
787 | int cmd_kmem(int argc, const char **argv, const char *prefix __used) | |
788 | { | |
789 | symbol__init(0); | |
790 | ||
791 | argc = parse_options(argc, argv, kmem_options, kmem_usage, 0); | |
792 | ||
793 | if (argc && !strncmp(argv[0], "rec", 3)) | |
794 | return __cmd_record(argc, argv); | |
795 | else if (argc) | |
796 | usage_with_options(kmem_usage, kmem_options); | |
797 | ||
29b3e152 LZ |
798 | if (list_empty(&caller_sort)) |
799 | setup_sorting(&caller_sort, default_sort_order); | |
800 | if (list_empty(&alloc_sort)) | |
801 | setup_sorting(&alloc_sort, default_sort_order); | |
ba77c9e1 | 802 | |
7d0d3945 LZ |
803 | setup_cpunode_map(); |
804 | ||
ba77c9e1 LZ |
805 | return __cmd_kmem(); |
806 | } | |
807 |