]>
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" |
45694aa7 | 10 | #include "util/tool.h" |
ba77c9e1 LZ |
11 | |
12 | #include "util/parse-options.h" | |
13 | #include "util/trace-event.h" | |
14 | ||
15 | #include "util/debug.h" | |
ba77c9e1 LZ |
16 | |
17 | #include <linux/rbtree.h> | |
18 | ||
19 | struct alloc_stat; | |
20 | typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *); | |
21 | ||
efad1415 | 22 | static const char *input_name; |
ba77c9e1 | 23 | |
ba77c9e1 LZ |
24 | static int alloc_flag; |
25 | static int caller_flag; | |
26 | ||
ba77c9e1 LZ |
27 | static int alloc_lines = -1; |
28 | static int caller_lines = -1; | |
29 | ||
7707b6b6 LZ |
30 | static bool raw_ip; |
31 | ||
29b3e152 LZ |
32 | static char default_sort_order[] = "frag,hit,bytes"; |
33 | ||
7d0d3945 LZ |
34 | static int *cpunode_map; |
35 | static int max_cpu_num; | |
36 | ||
ba77c9e1 | 37 | struct alloc_stat { |
079d3f65 LZ |
38 | u64 call_site; |
39 | u64 ptr; | |
ba77c9e1 LZ |
40 | u64 bytes_req; |
41 | u64 bytes_alloc; | |
42 | u32 hit; | |
079d3f65 LZ |
43 | u32 pingpong; |
44 | ||
45 | short alloc_cpu; | |
ba77c9e1 LZ |
46 | |
47 | struct rb_node node; | |
48 | }; | |
49 | ||
50 | static struct rb_root root_alloc_stat; | |
51 | static struct rb_root root_alloc_sorted; | |
52 | static struct rb_root root_caller_stat; | |
53 | static struct rb_root root_caller_sorted; | |
54 | ||
55 | static unsigned long total_requested, total_allocated; | |
7d0d3945 | 56 | static unsigned long nr_allocs, nr_cross_allocs; |
ba77c9e1 | 57 | |
7d0d3945 LZ |
58 | #define PATH_SYS_NODE "/sys/devices/system/node" |
59 | ||
60 | static void init_cpunode_map(void) | |
61 | { | |
62 | FILE *fp; | |
63 | int i; | |
64 | ||
65 | fp = fopen("/sys/devices/system/cpu/kernel_max", "r"); | |
66 | if (!fp) { | |
67 | max_cpu_num = 4096; | |
68 | return; | |
69 | } | |
70 | ||
71 | if (fscanf(fp, "%d", &max_cpu_num) < 1) | |
72 | die("Failed to read 'kernel_max' from sysfs"); | |
73 | max_cpu_num++; | |
74 | ||
75 | cpunode_map = calloc(max_cpu_num, sizeof(int)); | |
76 | if (!cpunode_map) | |
77 | die("calloc"); | |
78 | for (i = 0; i < max_cpu_num; i++) | |
79 | cpunode_map[i] = -1; | |
80 | fclose(fp); | |
81 | } | |
82 | ||
83 | static void setup_cpunode_map(void) | |
84 | { | |
85 | struct dirent *dent1, *dent2; | |
86 | DIR *dir1, *dir2; | |
87 | unsigned int cpu, mem; | |
88 | char buf[PATH_MAX]; | |
89 | ||
90 | init_cpunode_map(); | |
91 | ||
92 | dir1 = opendir(PATH_SYS_NODE); | |
93 | if (!dir1) | |
94 | return; | |
95 | ||
659d8cfb UD |
96 | while ((dent1 = readdir(dir1)) != NULL) { |
97 | if (dent1->d_type != DT_DIR || | |
98 | sscanf(dent1->d_name, "node%u", &mem) < 1) | |
7d0d3945 LZ |
99 | continue; |
100 | ||
101 | snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name); | |
102 | dir2 = opendir(buf); | |
103 | if (!dir2) | |
104 | continue; | |
659d8cfb UD |
105 | while ((dent2 = readdir(dir2)) != NULL) { |
106 | if (dent2->d_type != DT_LNK || | |
107 | sscanf(dent2->d_name, "cpu%u", &cpu) < 1) | |
7d0d3945 LZ |
108 | continue; |
109 | cpunode_map[cpu] = mem; | |
110 | } | |
8442da1d | 111 | closedir(dir2); |
7d0d3945 | 112 | } |
8442da1d | 113 | closedir(dir1); |
7d0d3945 LZ |
114 | } |
115 | ||
079d3f65 LZ |
116 | static void insert_alloc_stat(unsigned long call_site, unsigned long ptr, |
117 | int bytes_req, int bytes_alloc, int cpu) | |
ba77c9e1 LZ |
118 | { |
119 | struct rb_node **node = &root_alloc_stat.rb_node; | |
120 | struct rb_node *parent = NULL; | |
121 | struct alloc_stat *data = NULL; | |
122 | ||
ba77c9e1 LZ |
123 | while (*node) { |
124 | parent = *node; | |
125 | data = rb_entry(*node, struct alloc_stat, node); | |
126 | ||
127 | if (ptr > data->ptr) | |
128 | node = &(*node)->rb_right; | |
129 | else if (ptr < data->ptr) | |
130 | node = &(*node)->rb_left; | |
131 | else | |
132 | break; | |
133 | } | |
134 | ||
135 | if (data && data->ptr == ptr) { | |
136 | data->hit++; | |
137 | data->bytes_req += bytes_req; | |
4efb5290 | 138 | data->bytes_alloc += bytes_alloc; |
ba77c9e1 LZ |
139 | } else { |
140 | data = malloc(sizeof(*data)); | |
079d3f65 LZ |
141 | if (!data) |
142 | die("malloc"); | |
ba77c9e1 | 143 | data->ptr = ptr; |
079d3f65 | 144 | data->pingpong = 0; |
ba77c9e1 LZ |
145 | data->hit = 1; |
146 | data->bytes_req = bytes_req; | |
147 | data->bytes_alloc = bytes_alloc; | |
148 | ||
149 | rb_link_node(&data->node, parent, node); | |
150 | rb_insert_color(&data->node, &root_alloc_stat); | |
151 | } | |
079d3f65 LZ |
152 | data->call_site = call_site; |
153 | data->alloc_cpu = cpu; | |
ba77c9e1 LZ |
154 | } |
155 | ||
156 | static void insert_caller_stat(unsigned long call_site, | |
157 | int bytes_req, int bytes_alloc) | |
158 | { | |
159 | struct rb_node **node = &root_caller_stat.rb_node; | |
160 | struct rb_node *parent = NULL; | |
161 | struct alloc_stat *data = NULL; | |
162 | ||
ba77c9e1 LZ |
163 | while (*node) { |
164 | parent = *node; | |
165 | data = rb_entry(*node, struct alloc_stat, node); | |
166 | ||
167 | if (call_site > data->call_site) | |
168 | node = &(*node)->rb_right; | |
169 | else if (call_site < data->call_site) | |
170 | node = &(*node)->rb_left; | |
171 | else | |
172 | break; | |
173 | } | |
174 | ||
175 | if (data && data->call_site == call_site) { | |
176 | data->hit++; | |
177 | data->bytes_req += bytes_req; | |
4efb5290 | 178 | data->bytes_alloc += bytes_alloc; |
ba77c9e1 LZ |
179 | } else { |
180 | data = malloc(sizeof(*data)); | |
079d3f65 LZ |
181 | if (!data) |
182 | die("malloc"); | |
ba77c9e1 | 183 | data->call_site = call_site; |
079d3f65 | 184 | data->pingpong = 0; |
ba77c9e1 LZ |
185 | data->hit = 1; |
186 | data->bytes_req = bytes_req; | |
187 | data->bytes_alloc = bytes_alloc; | |
188 | ||
189 | rb_link_node(&data->node, parent, node); | |
190 | rb_insert_color(&data->node, &root_caller_stat); | |
191 | } | |
192 | } | |
193 | ||
f48f669d | 194 | static void process_alloc_event(void *data, |
aaf045f7 | 195 | struct event_format *event, |
7d0d3945 | 196 | int cpu, |
ba77c9e1 LZ |
197 | u64 timestamp __used, |
198 | struct thread *thread __used, | |
7d0d3945 | 199 | int node) |
ba77c9e1 LZ |
200 | { |
201 | unsigned long call_site; | |
202 | unsigned long ptr; | |
203 | int bytes_req; | |
204 | int bytes_alloc; | |
7d0d3945 | 205 | int node1, node2; |
ba77c9e1 | 206 | |
f48f669d XG |
207 | ptr = raw_field_value(event, "ptr", data); |
208 | call_site = raw_field_value(event, "call_site", data); | |
209 | bytes_req = raw_field_value(event, "bytes_req", data); | |
210 | bytes_alloc = raw_field_value(event, "bytes_alloc", data); | |
ba77c9e1 | 211 | |
079d3f65 | 212 | insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu); |
ba77c9e1 LZ |
213 | insert_caller_stat(call_site, bytes_req, bytes_alloc); |
214 | ||
215 | total_requested += bytes_req; | |
216 | total_allocated += bytes_alloc; | |
7d0d3945 LZ |
217 | |
218 | if (node) { | |
219 | node1 = cpunode_map[cpu]; | |
f48f669d | 220 | node2 = raw_field_value(event, "node", data); |
7d0d3945 LZ |
221 | if (node1 != node2) |
222 | nr_cross_allocs++; | |
223 | } | |
224 | nr_allocs++; | |
ba77c9e1 LZ |
225 | } |
226 | ||
079d3f65 LZ |
227 | static int ptr_cmp(struct alloc_stat *, struct alloc_stat *); |
228 | static int callsite_cmp(struct alloc_stat *, struct alloc_stat *); | |
229 | ||
230 | static struct alloc_stat *search_alloc_stat(unsigned long ptr, | |
231 | unsigned long call_site, | |
232 | struct rb_root *root, | |
233 | sort_fn_t sort_fn) | |
234 | { | |
235 | struct rb_node *node = root->rb_node; | |
236 | struct alloc_stat key = { .ptr = ptr, .call_site = call_site }; | |
237 | ||
238 | while (node) { | |
239 | struct alloc_stat *data; | |
240 | int cmp; | |
241 | ||
242 | data = rb_entry(node, struct alloc_stat, node); | |
243 | ||
244 | cmp = sort_fn(&key, data); | |
245 | if (cmp < 0) | |
246 | node = node->rb_left; | |
247 | else if (cmp > 0) | |
248 | node = node->rb_right; | |
249 | else | |
250 | return data; | |
251 | } | |
252 | return NULL; | |
253 | } | |
254 | ||
f48f669d | 255 | static void process_free_event(void *data, |
aaf045f7 | 256 | struct event_format *event, |
079d3f65 | 257 | int cpu, |
ba77c9e1 LZ |
258 | u64 timestamp __used, |
259 | struct thread *thread __used) | |
260 | { | |
079d3f65 LZ |
261 | unsigned long ptr; |
262 | struct alloc_stat *s_alloc, *s_caller; | |
263 | ||
f48f669d | 264 | ptr = raw_field_value(event, "ptr", data); |
079d3f65 LZ |
265 | |
266 | s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp); | |
267 | if (!s_alloc) | |
268 | return; | |
269 | ||
270 | if (cpu != s_alloc->alloc_cpu) { | |
271 | s_alloc->pingpong++; | |
272 | ||
273 | s_caller = search_alloc_stat(0, s_alloc->call_site, | |
274 | &root_caller_stat, callsite_cmp); | |
275 | assert(s_caller); | |
276 | s_caller->pingpong++; | |
277 | } | |
278 | s_alloc->alloc_cpu = -1; | |
ba77c9e1 LZ |
279 | } |
280 | ||
8115d60c ACM |
281 | static void process_raw_event(union perf_event *raw_event __used, void *data, |
282 | int cpu, u64 timestamp, struct thread *thread) | |
ba77c9e1 | 283 | { |
aaf045f7 | 284 | struct event_format *event; |
ba77c9e1 LZ |
285 | int type; |
286 | ||
f48f669d | 287 | type = trace_parse_common_type(data); |
ba77c9e1 LZ |
288 | event = trace_find_event(type); |
289 | ||
290 | if (!strcmp(event->name, "kmalloc") || | |
291 | !strcmp(event->name, "kmem_cache_alloc")) { | |
f48f669d | 292 | process_alloc_event(data, event, cpu, timestamp, thread, 0); |
ba77c9e1 LZ |
293 | return; |
294 | } | |
295 | ||
296 | if (!strcmp(event->name, "kmalloc_node") || | |
297 | !strcmp(event->name, "kmem_cache_alloc_node")) { | |
f48f669d | 298 | process_alloc_event(data, event, cpu, timestamp, thread, 1); |
ba77c9e1 LZ |
299 | return; |
300 | } | |
301 | ||
302 | if (!strcmp(event->name, "kfree") || | |
303 | !strcmp(event->name, "kmem_cache_free")) { | |
f48f669d | 304 | process_free_event(data, event, cpu, timestamp, thread); |
ba77c9e1 LZ |
305 | return; |
306 | } | |
307 | } | |
308 | ||
45694aa7 | 309 | static int process_sample_event(struct perf_tool *tool __used, |
d20deb64 | 310 | union perf_event *event, |
8115d60c | 311 | struct perf_sample *sample, |
9e69c210 | 312 | struct perf_evsel *evsel __used, |
743eb868 | 313 | struct machine *machine) |
ba77c9e1 | 314 | { |
743eb868 | 315 | struct thread *thread = machine__findnew_thread(machine, event->ip.pid); |
ba77c9e1 | 316 | |
ba77c9e1 LZ |
317 | if (thread == NULL) { |
318 | pr_debug("problem processing %d event, skipping it.\n", | |
319 | event->header.type); | |
320 | return -1; | |
321 | } | |
322 | ||
323 | dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid); | |
324 | ||
640c03ce ACM |
325 | process_raw_event(event, sample->raw_data, sample->cpu, |
326 | sample->time, thread); | |
ba77c9e1 LZ |
327 | |
328 | return 0; | |
329 | } | |
330 | ||
45694aa7 | 331 | static struct perf_tool perf_kmem = { |
587570d4 | 332 | .sample = process_sample_event, |
8115d60c | 333 | .comm = perf_event__process_comm, |
587570d4 | 334 | .ordered_samples = true, |
ba77c9e1 LZ |
335 | }; |
336 | ||
ba77c9e1 LZ |
337 | static double fragmentation(unsigned long n_req, unsigned long n_alloc) |
338 | { | |
339 | if (n_alloc == 0) | |
340 | return 0.0; | |
341 | else | |
342 | return 100.0 - (100.0 * n_req / n_alloc); | |
343 | } | |
344 | ||
4aa65636 ACM |
345 | static void __print_result(struct rb_root *root, struct perf_session *session, |
346 | int n_lines, int is_caller) | |
ba77c9e1 LZ |
347 | { |
348 | struct rb_node *next; | |
23346f21 | 349 | struct machine *machine; |
ba77c9e1 | 350 | |
079d3f65 LZ |
351 | printf("%.102s\n", graph_dotted_line); |
352 | printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr"); | |
47103277 | 353 | printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n"); |
079d3f65 | 354 | printf("%.102s\n", graph_dotted_line); |
ba77c9e1 LZ |
355 | |
356 | next = rb_first(root); | |
357 | ||
23346f21 ACM |
358 | machine = perf_session__find_host_machine(session); |
359 | if (!machine) { | |
a1645ce1 ZY |
360 | pr_err("__print_result: couldn't find kernel information\n"); |
361 | return; | |
362 | } | |
ba77c9e1 | 363 | while (next && n_lines--) { |
1b145ae5 ACM |
364 | struct alloc_stat *data = rb_entry(next, struct alloc_stat, |
365 | node); | |
366 | struct symbol *sym = NULL; | |
71cf8b8f | 367 | struct map *map; |
079d3f65 | 368 | char buf[BUFSIZ]; |
1b145ae5 ACM |
369 | u64 addr; |
370 | ||
371 | if (is_caller) { | |
372 | addr = data->call_site; | |
7707b6b6 | 373 | if (!raw_ip) |
5c0541d5 | 374 | sym = machine__find_kernel_function(machine, addr, &map, NULL); |
1b145ae5 ACM |
375 | } else |
376 | addr = data->ptr; | |
377 | ||
378 | if (sym != NULL) | |
9486aa38 | 379 | snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name, |
71cf8b8f | 380 | addr - map->unmap_ip(map, sym->start)); |
1b145ae5 | 381 | else |
9486aa38 | 382 | snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr); |
079d3f65 | 383 | printf(" %-34s |", buf); |
ba77c9e1 | 384 | |
47103277 | 385 | printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n", |
079d3f65 | 386 | (unsigned long long)data->bytes_alloc, |
ba77c9e1 LZ |
387 | (unsigned long)data->bytes_alloc / data->hit, |
388 | (unsigned long long)data->bytes_req, | |
389 | (unsigned long)data->bytes_req / data->hit, | |
390 | (unsigned long)data->hit, | |
079d3f65 | 391 | (unsigned long)data->pingpong, |
ba77c9e1 LZ |
392 | fragmentation(data->bytes_req, data->bytes_alloc)); |
393 | ||
394 | next = rb_next(next); | |
395 | } | |
396 | ||
397 | if (n_lines == -1) | |
079d3f65 | 398 | printf(" ... | ... | ... | ... | ... | ... \n"); |
ba77c9e1 | 399 | |
079d3f65 | 400 | printf("%.102s\n", graph_dotted_line); |
ba77c9e1 LZ |
401 | } |
402 | ||
403 | static void print_summary(void) | |
404 | { | |
405 | printf("\nSUMMARY\n=======\n"); | |
406 | printf("Total bytes requested: %lu\n", total_requested); | |
407 | printf("Total bytes allocated: %lu\n", total_allocated); | |
408 | printf("Total bytes wasted on internal fragmentation: %lu\n", | |
409 | total_allocated - total_requested); | |
410 | printf("Internal fragmentation: %f%%\n", | |
411 | fragmentation(total_requested, total_allocated)); | |
7d0d3945 | 412 | printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs); |
ba77c9e1 LZ |
413 | } |
414 | ||
4aa65636 | 415 | static void print_result(struct perf_session *session) |
ba77c9e1 LZ |
416 | { |
417 | if (caller_flag) | |
4aa65636 | 418 | __print_result(&root_caller_sorted, session, caller_lines, 1); |
ba77c9e1 | 419 | if (alloc_flag) |
4aa65636 | 420 | __print_result(&root_alloc_sorted, session, alloc_lines, 0); |
ba77c9e1 LZ |
421 | print_summary(); |
422 | } | |
423 | ||
29b3e152 LZ |
424 | struct sort_dimension { |
425 | const char name[20]; | |
426 | sort_fn_t cmp; | |
427 | struct list_head list; | |
428 | }; | |
429 | ||
430 | static LIST_HEAD(caller_sort); | |
431 | static LIST_HEAD(alloc_sort); | |
432 | ||
ba77c9e1 | 433 | static void sort_insert(struct rb_root *root, struct alloc_stat *data, |
29b3e152 | 434 | struct list_head *sort_list) |
ba77c9e1 LZ |
435 | { |
436 | struct rb_node **new = &(root->rb_node); | |
437 | struct rb_node *parent = NULL; | |
29b3e152 | 438 | struct sort_dimension *sort; |
ba77c9e1 LZ |
439 | |
440 | while (*new) { | |
441 | struct alloc_stat *this; | |
29b3e152 | 442 | int cmp = 0; |
ba77c9e1 LZ |
443 | |
444 | this = rb_entry(*new, struct alloc_stat, node); | |
445 | parent = *new; | |
446 | ||
29b3e152 LZ |
447 | list_for_each_entry(sort, sort_list, list) { |
448 | cmp = sort->cmp(data, this); | |
449 | if (cmp) | |
450 | break; | |
451 | } | |
ba77c9e1 LZ |
452 | |
453 | if (cmp > 0) | |
454 | new = &((*new)->rb_left); | |
455 | else | |
456 | new = &((*new)->rb_right); | |
457 | } | |
458 | ||
459 | rb_link_node(&data->node, parent, new); | |
460 | rb_insert_color(&data->node, root); | |
461 | } | |
462 | ||
463 | static void __sort_result(struct rb_root *root, struct rb_root *root_sorted, | |
29b3e152 | 464 | struct list_head *sort_list) |
ba77c9e1 LZ |
465 | { |
466 | struct rb_node *node; | |
467 | struct alloc_stat *data; | |
468 | ||
469 | for (;;) { | |
470 | node = rb_first(root); | |
471 | if (!node) | |
472 | break; | |
473 | ||
474 | rb_erase(node, root); | |
475 | data = rb_entry(node, struct alloc_stat, node); | |
29b3e152 | 476 | sort_insert(root_sorted, data, sort_list); |
ba77c9e1 LZ |
477 | } |
478 | } | |
479 | ||
480 | static void sort_result(void) | |
481 | { | |
29b3e152 LZ |
482 | __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort); |
483 | __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort); | |
ba77c9e1 LZ |
484 | } |
485 | ||
486 | static int __cmd_kmem(void) | |
487 | { | |
d549c769 | 488 | int err = -EINVAL; |
21ef97f0 | 489 | struct perf_session *session = perf_session__new(input_name, O_RDONLY, |
45694aa7 | 490 | 0, false, &perf_kmem); |
4aa65636 ACM |
491 | if (session == NULL) |
492 | return -ENOMEM; | |
493 | ||
e727ca73 ACM |
494 | if (perf_session__create_kernel_maps(session) < 0) |
495 | goto out_delete; | |
496 | ||
d549c769 ACM |
497 | if (!perf_session__has_traces(session, "kmem record")) |
498 | goto out_delete; | |
499 | ||
ba77c9e1 | 500 | setup_pager(); |
45694aa7 | 501 | err = perf_session__process_events(session, &perf_kmem); |
4aa65636 ACM |
502 | if (err != 0) |
503 | goto out_delete; | |
ba77c9e1 | 504 | sort_result(); |
4aa65636 ACM |
505 | print_result(session); |
506 | out_delete: | |
507 | perf_session__delete(session); | |
508 | return err; | |
ba77c9e1 LZ |
509 | } |
510 | ||
511 | static const char * const kmem_usage[] = { | |
90b86a9f | 512 | "perf kmem [<options>] {record|stat}", |
ba77c9e1 LZ |
513 | NULL |
514 | }; | |
515 | ||
ba77c9e1 LZ |
516 | static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r) |
517 | { | |
518 | if (l->ptr < r->ptr) | |
519 | return -1; | |
520 | else if (l->ptr > r->ptr) | |
521 | return 1; | |
522 | return 0; | |
523 | } | |
524 | ||
29b3e152 LZ |
525 | static struct sort_dimension ptr_sort_dimension = { |
526 | .name = "ptr", | |
527 | .cmp = ptr_cmp, | |
528 | }; | |
529 | ||
ba77c9e1 LZ |
530 | static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r) |
531 | { | |
532 | if (l->call_site < r->call_site) | |
533 | return -1; | |
534 | else if (l->call_site > r->call_site) | |
535 | return 1; | |
536 | return 0; | |
537 | } | |
538 | ||
29b3e152 LZ |
539 | static struct sort_dimension callsite_sort_dimension = { |
540 | .name = "callsite", | |
541 | .cmp = callsite_cmp, | |
542 | }; | |
543 | ||
f3ced7cd PE |
544 | static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r) |
545 | { | |
546 | if (l->hit < r->hit) | |
547 | return -1; | |
548 | else if (l->hit > r->hit) | |
549 | return 1; | |
550 | return 0; | |
551 | } | |
552 | ||
29b3e152 LZ |
553 | static struct sort_dimension hit_sort_dimension = { |
554 | .name = "hit", | |
555 | .cmp = hit_cmp, | |
556 | }; | |
557 | ||
ba77c9e1 LZ |
558 | static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r) |
559 | { | |
560 | if (l->bytes_alloc < r->bytes_alloc) | |
561 | return -1; | |
562 | else if (l->bytes_alloc > r->bytes_alloc) | |
563 | return 1; | |
564 | return 0; | |
565 | } | |
566 | ||
29b3e152 LZ |
567 | static struct sort_dimension bytes_sort_dimension = { |
568 | .name = "bytes", | |
569 | .cmp = bytes_cmp, | |
570 | }; | |
571 | ||
f3ced7cd PE |
572 | static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r) |
573 | { | |
574 | double x, y; | |
575 | ||
576 | x = fragmentation(l->bytes_req, l->bytes_alloc); | |
577 | y = fragmentation(r->bytes_req, r->bytes_alloc); | |
578 | ||
579 | if (x < y) | |
580 | return -1; | |
581 | else if (x > y) | |
582 | return 1; | |
583 | return 0; | |
584 | } | |
585 | ||
29b3e152 LZ |
586 | static struct sort_dimension frag_sort_dimension = { |
587 | .name = "frag", | |
588 | .cmp = frag_cmp, | |
589 | }; | |
590 | ||
079d3f65 LZ |
591 | static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r) |
592 | { | |
593 | if (l->pingpong < r->pingpong) | |
594 | return -1; | |
595 | else if (l->pingpong > r->pingpong) | |
596 | return 1; | |
597 | return 0; | |
598 | } | |
599 | ||
600 | static struct sort_dimension pingpong_sort_dimension = { | |
601 | .name = "pingpong", | |
602 | .cmp = pingpong_cmp, | |
603 | }; | |
604 | ||
29b3e152 LZ |
605 | static struct sort_dimension *avail_sorts[] = { |
606 | &ptr_sort_dimension, | |
607 | &callsite_sort_dimension, | |
608 | &hit_sort_dimension, | |
609 | &bytes_sort_dimension, | |
610 | &frag_sort_dimension, | |
079d3f65 | 611 | &pingpong_sort_dimension, |
29b3e152 LZ |
612 | }; |
613 | ||
614 | #define NUM_AVAIL_SORTS \ | |
615 | (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *)) | |
616 | ||
617 | static int sort_dimension__add(const char *tok, struct list_head *list) | |
618 | { | |
619 | struct sort_dimension *sort; | |
620 | int i; | |
621 | ||
622 | for (i = 0; i < NUM_AVAIL_SORTS; i++) { | |
623 | if (!strcmp(avail_sorts[i]->name, tok)) { | |
624 | sort = malloc(sizeof(*sort)); | |
625 | if (!sort) | |
626 | die("malloc"); | |
627 | memcpy(sort, avail_sorts[i], sizeof(*sort)); | |
628 | list_add_tail(&sort->list, list); | |
629 | return 0; | |
630 | } | |
631 | } | |
632 | ||
633 | return -1; | |
634 | } | |
635 | ||
636 | static int setup_sorting(struct list_head *sort_list, const char *arg) | |
637 | { | |
638 | char *tok; | |
639 | char *str = strdup(arg); | |
640 | ||
641 | if (!str) | |
642 | die("strdup"); | |
643 | ||
644 | while (true) { | |
645 | tok = strsep(&str, ","); | |
646 | if (!tok) | |
647 | break; | |
648 | if (sort_dimension__add(tok, sort_list) < 0) { | |
649 | error("Unknown --sort key: '%s'", tok); | |
1b22859d | 650 | free(str); |
29b3e152 LZ |
651 | return -1; |
652 | } | |
653 | } | |
654 | ||
655 | free(str); | |
656 | return 0; | |
657 | } | |
658 | ||
ba77c9e1 LZ |
659 | static int parse_sort_opt(const struct option *opt __used, |
660 | const char *arg, int unset __used) | |
661 | { | |
ba77c9e1 LZ |
662 | if (!arg) |
663 | return -1; | |
664 | ||
ba77c9e1 | 665 | if (caller_flag > alloc_flag) |
29b3e152 | 666 | return setup_sorting(&caller_sort, arg); |
ba77c9e1 | 667 | else |
29b3e152 | 668 | return setup_sorting(&alloc_sort, arg); |
ba77c9e1 LZ |
669 | |
670 | return 0; | |
671 | } | |
672 | ||
90b86a9f | 673 | static int parse_caller_opt(const struct option *opt __used, |
79312416 | 674 | const char *arg __used, int unset __used) |
ba77c9e1 | 675 | { |
90b86a9f LZ |
676 | caller_flag = (alloc_flag + 1); |
677 | return 0; | |
678 | } | |
ba77c9e1 | 679 | |
90b86a9f | 680 | static int parse_alloc_opt(const struct option *opt __used, |
79312416 | 681 | const char *arg __used, int unset __used) |
90b86a9f LZ |
682 | { |
683 | alloc_flag = (caller_flag + 1); | |
ba77c9e1 LZ |
684 | return 0; |
685 | } | |
686 | ||
687 | static int parse_line_opt(const struct option *opt __used, | |
688 | const char *arg, int unset __used) | |
689 | { | |
690 | int lines; | |
691 | ||
692 | if (!arg) | |
693 | return -1; | |
694 | ||
695 | lines = strtoul(arg, NULL, 10); | |
696 | ||
697 | if (caller_flag > alloc_flag) | |
698 | caller_lines = lines; | |
699 | else | |
700 | alloc_lines = lines; | |
701 | ||
702 | return 0; | |
703 | } | |
704 | ||
705 | static const struct option kmem_options[] = { | |
706 | OPT_STRING('i', "input", &input_name, "file", | |
707 | "input file name"), | |
90b86a9f LZ |
708 | OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL, |
709 | "show per-callsite statistics", | |
710 | parse_caller_opt), | |
711 | OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL, | |
712 | "show per-allocation statistics", | |
713 | parse_alloc_opt), | |
29b3e152 | 714 | OPT_CALLBACK('s', "sort", NULL, "key[,key2...]", |
079d3f65 | 715 | "sort by keys: ptr, call_site, bytes, hit, pingpong, frag", |
ba77c9e1 LZ |
716 | parse_sort_opt), |
717 | OPT_CALLBACK('l', "line", NULL, "num", | |
90b86a9f | 718 | "show n lines", |
ba77c9e1 | 719 | parse_line_opt), |
7707b6b6 | 720 | OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), |
ba77c9e1 LZ |
721 | OPT_END() |
722 | }; | |
723 | ||
724 | static const char *record_args[] = { | |
725 | "record", | |
726 | "-a", | |
727 | "-R", | |
ba77c9e1 LZ |
728 | "-f", |
729 | "-c", "1", | |
730 | "-e", "kmem:kmalloc", | |
731 | "-e", "kmem:kmalloc_node", | |
732 | "-e", "kmem:kfree", | |
733 | "-e", "kmem:kmem_cache_alloc", | |
734 | "-e", "kmem:kmem_cache_alloc_node", | |
735 | "-e", "kmem:kmem_cache_free", | |
736 | }; | |
737 | ||
738 | static int __cmd_record(int argc, const char **argv) | |
739 | { | |
740 | unsigned int rec_argc, i, j; | |
741 | const char **rec_argv; | |
742 | ||
743 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
744 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); | |
745 | ||
ce47dc56 CS |
746 | if (rec_argv == NULL) |
747 | return -ENOMEM; | |
748 | ||
ba77c9e1 LZ |
749 | for (i = 0; i < ARRAY_SIZE(record_args); i++) |
750 | rec_argv[i] = strdup(record_args[i]); | |
751 | ||
752 | for (j = 1; j < (unsigned int)argc; j++, i++) | |
753 | rec_argv[i] = argv[j]; | |
754 | ||
755 | return cmd_record(i, rec_argv, NULL); | |
756 | } | |
757 | ||
758 | int cmd_kmem(int argc, const char **argv, const char *prefix __used) | |
759 | { | |
ba77c9e1 LZ |
760 | argc = parse_options(argc, argv, kmem_options, kmem_usage, 0); |
761 | ||
90b86a9f | 762 | if (!argc) |
ba77c9e1 LZ |
763 | usage_with_options(kmem_usage, kmem_options); |
764 | ||
655000e7 ACM |
765 | symbol__init(); |
766 | ||
90b86a9f LZ |
767 | if (!strncmp(argv[0], "rec", 3)) { |
768 | return __cmd_record(argc, argv); | |
769 | } else if (!strcmp(argv[0], "stat")) { | |
770 | setup_cpunode_map(); | |
771 | ||
772 | if (list_empty(&caller_sort)) | |
773 | setup_sorting(&caller_sort, default_sort_order); | |
774 | if (list_empty(&alloc_sort)) | |
775 | setup_sorting(&alloc_sort, default_sort_order); | |
ba77c9e1 | 776 | |
90b86a9f | 777 | return __cmd_kmem(); |
b00eca8c PE |
778 | } else |
779 | usage_with_options(kmem_usage, kmem_options); | |
7d0d3945 | 780 | |
90b86a9f | 781 | return 0; |
ba77c9e1 LZ |
782 | } |
783 |