]>
Commit | Line | Data |
---|---|---|
ba77c9e1 LZ |
1 | #include "builtin.h" |
2 | #include "perf.h" | |
3 | ||
0f7d2f1b | 4 | #include "util/evlist.h" |
fcf65bf1 | 5 | #include "util/evsel.h" |
ba77c9e1 LZ |
6 | #include "util/util.h" |
7 | #include "util/cache.h" | |
8 | #include "util/symbol.h" | |
9 | #include "util/thread.h" | |
10 | #include "util/header.h" | |
94c744b6 | 11 | #include "util/session.h" |
45694aa7 | 12 | #include "util/tool.h" |
ba77c9e1 LZ |
13 | |
14 | #include "util/parse-options.h" | |
15 | #include "util/trace-event.h" | |
f5fc1412 | 16 | #include "util/data.h" |
4b627957 | 17 | #include "util/cpumap.h" |
ba77c9e1 LZ |
18 | |
19 | #include "util/debug.h" | |
ba77c9e1 LZ |
20 | |
21 | #include <linux/rbtree.h> | |
8d9233f2 | 22 | #include <linux/string.h> |
77cfe388 | 23 | #include <locale.h> |
ba77c9e1 LZ |
24 | |
25 | struct alloc_stat; | |
26 | typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *); | |
27 | ||
ba77c9e1 LZ |
28 | static int alloc_flag; |
29 | static int caller_flag; | |
30 | ||
ba77c9e1 LZ |
31 | static int alloc_lines = -1; |
32 | static int caller_lines = -1; | |
33 | ||
7707b6b6 LZ |
34 | static bool raw_ip; |
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 | |
2814eb05 ACM |
57 | static int insert_alloc_stat(unsigned long call_site, unsigned long ptr, |
58 | int bytes_req, int bytes_alloc, int cpu) | |
ba77c9e1 LZ |
59 | { |
60 | struct rb_node **node = &root_alloc_stat.rb_node; | |
61 | struct rb_node *parent = NULL; | |
62 | struct alloc_stat *data = NULL; | |
63 | ||
ba77c9e1 LZ |
64 | while (*node) { |
65 | parent = *node; | |
66 | data = rb_entry(*node, struct alloc_stat, node); | |
67 | ||
68 | if (ptr > data->ptr) | |
69 | node = &(*node)->rb_right; | |
70 | else if (ptr < data->ptr) | |
71 | node = &(*node)->rb_left; | |
72 | else | |
73 | break; | |
74 | } | |
75 | ||
76 | if (data && data->ptr == ptr) { | |
77 | data->hit++; | |
78 | data->bytes_req += bytes_req; | |
4efb5290 | 79 | data->bytes_alloc += bytes_alloc; |
ba77c9e1 LZ |
80 | } else { |
81 | data = malloc(sizeof(*data)); | |
2814eb05 ACM |
82 | if (!data) { |
83 | pr_err("%s: malloc failed\n", __func__); | |
84 | return -1; | |
85 | } | |
ba77c9e1 | 86 | data->ptr = ptr; |
079d3f65 | 87 | data->pingpong = 0; |
ba77c9e1 LZ |
88 | data->hit = 1; |
89 | data->bytes_req = bytes_req; | |
90 | data->bytes_alloc = bytes_alloc; | |
91 | ||
92 | rb_link_node(&data->node, parent, node); | |
93 | rb_insert_color(&data->node, &root_alloc_stat); | |
94 | } | |
079d3f65 LZ |
95 | data->call_site = call_site; |
96 | data->alloc_cpu = cpu; | |
2814eb05 | 97 | return 0; |
ba77c9e1 LZ |
98 | } |
99 | ||
2814eb05 | 100 | static int insert_caller_stat(unsigned long call_site, |
ba77c9e1 LZ |
101 | int bytes_req, int bytes_alloc) |
102 | { | |
103 | struct rb_node **node = &root_caller_stat.rb_node; | |
104 | struct rb_node *parent = NULL; | |
105 | struct alloc_stat *data = NULL; | |
106 | ||
ba77c9e1 LZ |
107 | while (*node) { |
108 | parent = *node; | |
109 | data = rb_entry(*node, struct alloc_stat, node); | |
110 | ||
111 | if (call_site > data->call_site) | |
112 | node = &(*node)->rb_right; | |
113 | else if (call_site < data->call_site) | |
114 | node = &(*node)->rb_left; | |
115 | else | |
116 | break; | |
117 | } | |
118 | ||
119 | if (data && data->call_site == call_site) { | |
120 | data->hit++; | |
121 | data->bytes_req += bytes_req; | |
4efb5290 | 122 | data->bytes_alloc += bytes_alloc; |
ba77c9e1 LZ |
123 | } else { |
124 | data = malloc(sizeof(*data)); | |
2814eb05 ACM |
125 | if (!data) { |
126 | pr_err("%s: malloc failed\n", __func__); | |
127 | return -1; | |
128 | } | |
ba77c9e1 | 129 | data->call_site = call_site; |
079d3f65 | 130 | data->pingpong = 0; |
ba77c9e1 LZ |
131 | data->hit = 1; |
132 | data->bytes_req = bytes_req; | |
133 | data->bytes_alloc = bytes_alloc; | |
134 | ||
135 | rb_link_node(&data->node, parent, node); | |
136 | rb_insert_color(&data->node, &root_caller_stat); | |
137 | } | |
2814eb05 ACM |
138 | |
139 | return 0; | |
ba77c9e1 LZ |
140 | } |
141 | ||
2814eb05 | 142 | static int perf_evsel__process_alloc_event(struct perf_evsel *evsel, |
0f7d2f1b | 143 | struct perf_sample *sample) |
ba77c9e1 | 144 | { |
0f7d2f1b ACM |
145 | unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"), |
146 | call_site = perf_evsel__intval(evsel, sample, "call_site"); | |
147 | int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"), | |
148 | bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc"); | |
149 | ||
150 | if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) || | |
2814eb05 ACM |
151 | insert_caller_stat(call_site, bytes_req, bytes_alloc)) |
152 | return -1; | |
ba77c9e1 LZ |
153 | |
154 | total_requested += bytes_req; | |
155 | total_allocated += bytes_alloc; | |
7d0d3945 | 156 | |
0f7d2f1b ACM |
157 | nr_allocs++; |
158 | return 0; | |
159 | } | |
160 | ||
161 | static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel, | |
162 | struct perf_sample *sample) | |
163 | { | |
164 | int ret = perf_evsel__process_alloc_event(evsel, sample); | |
165 | ||
166 | if (!ret) { | |
4b627957 | 167 | int node1 = cpu__get_node(sample->cpu), |
0f7d2f1b ACM |
168 | node2 = perf_evsel__intval(evsel, sample, "node"); |
169 | ||
7d0d3945 LZ |
170 | if (node1 != node2) |
171 | nr_cross_allocs++; | |
172 | } | |
0f7d2f1b ACM |
173 | |
174 | return ret; | |
ba77c9e1 LZ |
175 | } |
176 | ||
079d3f65 LZ |
177 | static int ptr_cmp(struct alloc_stat *, struct alloc_stat *); |
178 | static int callsite_cmp(struct alloc_stat *, struct alloc_stat *); | |
179 | ||
180 | static struct alloc_stat *search_alloc_stat(unsigned long ptr, | |
181 | unsigned long call_site, | |
182 | struct rb_root *root, | |
183 | sort_fn_t sort_fn) | |
184 | { | |
185 | struct rb_node *node = root->rb_node; | |
186 | struct alloc_stat key = { .ptr = ptr, .call_site = call_site }; | |
187 | ||
188 | while (node) { | |
189 | struct alloc_stat *data; | |
190 | int cmp; | |
191 | ||
192 | data = rb_entry(node, struct alloc_stat, node); | |
193 | ||
194 | cmp = sort_fn(&key, data); | |
195 | if (cmp < 0) | |
196 | node = node->rb_left; | |
197 | else if (cmp > 0) | |
198 | node = node->rb_right; | |
199 | else | |
200 | return data; | |
201 | } | |
202 | return NULL; | |
203 | } | |
204 | ||
2814eb05 ACM |
205 | static int perf_evsel__process_free_event(struct perf_evsel *evsel, |
206 | struct perf_sample *sample) | |
ba77c9e1 | 207 | { |
0f7d2f1b | 208 | unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"); |
079d3f65 LZ |
209 | struct alloc_stat *s_alloc, *s_caller; |
210 | ||
079d3f65 LZ |
211 | s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp); |
212 | if (!s_alloc) | |
2814eb05 | 213 | return 0; |
079d3f65 | 214 | |
22ad798c | 215 | if ((short)sample->cpu != s_alloc->alloc_cpu) { |
079d3f65 LZ |
216 | s_alloc->pingpong++; |
217 | ||
218 | s_caller = search_alloc_stat(0, s_alloc->call_site, | |
219 | &root_caller_stat, callsite_cmp); | |
2814eb05 ACM |
220 | if (!s_caller) |
221 | return -1; | |
079d3f65 LZ |
222 | s_caller->pingpong++; |
223 | } | |
224 | s_alloc->alloc_cpu = -1; | |
2814eb05 ACM |
225 | |
226 | return 0; | |
ba77c9e1 LZ |
227 | } |
228 | ||
0f7d2f1b ACM |
229 | typedef int (*tracepoint_handler)(struct perf_evsel *evsel, |
230 | struct perf_sample *sample); | |
ba77c9e1 | 231 | |
1d037ca1 | 232 | static int process_sample_event(struct perf_tool *tool __maybe_unused, |
d20deb64 | 233 | union perf_event *event, |
8115d60c | 234 | struct perf_sample *sample, |
fcf65bf1 | 235 | struct perf_evsel *evsel, |
743eb868 | 236 | struct machine *machine) |
ba77c9e1 | 237 | { |
ef89325f | 238 | struct thread *thread = machine__findnew_thread(machine, sample->pid, |
13ce34df | 239 | sample->tid); |
ba77c9e1 | 240 | |
ba77c9e1 LZ |
241 | if (thread == NULL) { |
242 | pr_debug("problem processing %d event, skipping it.\n", | |
243 | event->header.type); | |
244 | return -1; | |
245 | } | |
246 | ||
b9c5143a | 247 | dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid); |
ba77c9e1 | 248 | |
744a9719 ACM |
249 | if (evsel->handler != NULL) { |
250 | tracepoint_handler f = evsel->handler; | |
0f7d2f1b ACM |
251 | return f(evsel, sample); |
252 | } | |
253 | ||
254 | return 0; | |
ba77c9e1 LZ |
255 | } |
256 | ||
fcf65bf1 ACM |
257 | static struct perf_tool perf_kmem = { |
258 | .sample = process_sample_event, | |
259 | .comm = perf_event__process_comm, | |
64c40908 NK |
260 | .mmap = perf_event__process_mmap, |
261 | .mmap2 = perf_event__process_mmap2, | |
0a8cb85c | 262 | .ordered_events = true, |
ba77c9e1 LZ |
263 | }; |
264 | ||
ba77c9e1 LZ |
265 | static double fragmentation(unsigned long n_req, unsigned long n_alloc) |
266 | { | |
267 | if (n_alloc == 0) | |
268 | return 0.0; | |
269 | else | |
270 | return 100.0 - (100.0 * n_req / n_alloc); | |
271 | } | |
272 | ||
4aa65636 ACM |
273 | static void __print_result(struct rb_root *root, struct perf_session *session, |
274 | int n_lines, int is_caller) | |
ba77c9e1 LZ |
275 | { |
276 | struct rb_node *next; | |
34ba5122 | 277 | struct machine *machine = &session->machines.host; |
ba77c9e1 | 278 | |
65f46e02 | 279 | printf("%.105s\n", graph_dotted_line); |
079d3f65 | 280 | printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr"); |
47103277 | 281 | printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n"); |
65f46e02 | 282 | printf("%.105s\n", graph_dotted_line); |
ba77c9e1 LZ |
283 | |
284 | next = rb_first(root); | |
285 | ||
286 | while (next && n_lines--) { | |
1b145ae5 ACM |
287 | struct alloc_stat *data = rb_entry(next, struct alloc_stat, |
288 | node); | |
289 | struct symbol *sym = NULL; | |
71cf8b8f | 290 | struct map *map; |
079d3f65 | 291 | char buf[BUFSIZ]; |
1b145ae5 ACM |
292 | u64 addr; |
293 | ||
294 | if (is_caller) { | |
295 | addr = data->call_site; | |
7707b6b6 | 296 | if (!raw_ip) |
5c0541d5 | 297 | sym = machine__find_kernel_function(machine, addr, &map, NULL); |
1b145ae5 ACM |
298 | } else |
299 | addr = data->ptr; | |
300 | ||
301 | if (sym != NULL) | |
9486aa38 | 302 | snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name, |
71cf8b8f | 303 | addr - map->unmap_ip(map, sym->start)); |
1b145ae5 | 304 | else |
9486aa38 | 305 | snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr); |
079d3f65 | 306 | printf(" %-34s |", buf); |
ba77c9e1 | 307 | |
65f46e02 | 308 | printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n", |
079d3f65 | 309 | (unsigned long long)data->bytes_alloc, |
ba77c9e1 LZ |
310 | (unsigned long)data->bytes_alloc / data->hit, |
311 | (unsigned long long)data->bytes_req, | |
312 | (unsigned long)data->bytes_req / data->hit, | |
313 | (unsigned long)data->hit, | |
079d3f65 | 314 | (unsigned long)data->pingpong, |
ba77c9e1 LZ |
315 | fragmentation(data->bytes_req, data->bytes_alloc)); |
316 | ||
317 | next = rb_next(next); | |
318 | } | |
319 | ||
320 | if (n_lines == -1) | |
65f46e02 | 321 | printf(" ... | ... | ... | ... | ... | ... \n"); |
ba77c9e1 | 322 | |
65f46e02 | 323 | printf("%.105s\n", graph_dotted_line); |
ba77c9e1 LZ |
324 | } |
325 | ||
326 | static void print_summary(void) | |
327 | { | |
328 | printf("\nSUMMARY\n=======\n"); | |
77cfe388 NK |
329 | printf("Total bytes requested: %'lu\n", total_requested); |
330 | printf("Total bytes allocated: %'lu\n", total_allocated); | |
331 | printf("Total bytes wasted on internal fragmentation: %'lu\n", | |
ba77c9e1 LZ |
332 | total_allocated - total_requested); |
333 | printf("Internal fragmentation: %f%%\n", | |
334 | fragmentation(total_requested, total_allocated)); | |
77cfe388 | 335 | printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs); |
ba77c9e1 LZ |
336 | } |
337 | ||
4aa65636 | 338 | static void print_result(struct perf_session *session) |
ba77c9e1 LZ |
339 | { |
340 | if (caller_flag) | |
4aa65636 | 341 | __print_result(&root_caller_sorted, session, caller_lines, 1); |
ba77c9e1 | 342 | if (alloc_flag) |
4aa65636 | 343 | __print_result(&root_alloc_sorted, session, alloc_lines, 0); |
ba77c9e1 LZ |
344 | print_summary(); |
345 | } | |
346 | ||
29b3e152 LZ |
347 | struct sort_dimension { |
348 | const char name[20]; | |
349 | sort_fn_t cmp; | |
350 | struct list_head list; | |
351 | }; | |
352 | ||
353 | static LIST_HEAD(caller_sort); | |
354 | static LIST_HEAD(alloc_sort); | |
355 | ||
ba77c9e1 | 356 | static void sort_insert(struct rb_root *root, struct alloc_stat *data, |
29b3e152 | 357 | struct list_head *sort_list) |
ba77c9e1 LZ |
358 | { |
359 | struct rb_node **new = &(root->rb_node); | |
360 | struct rb_node *parent = NULL; | |
29b3e152 | 361 | struct sort_dimension *sort; |
ba77c9e1 LZ |
362 | |
363 | while (*new) { | |
364 | struct alloc_stat *this; | |
29b3e152 | 365 | int cmp = 0; |
ba77c9e1 LZ |
366 | |
367 | this = rb_entry(*new, struct alloc_stat, node); | |
368 | parent = *new; | |
369 | ||
29b3e152 LZ |
370 | list_for_each_entry(sort, sort_list, list) { |
371 | cmp = sort->cmp(data, this); | |
372 | if (cmp) | |
373 | break; | |
374 | } | |
ba77c9e1 LZ |
375 | |
376 | if (cmp > 0) | |
377 | new = &((*new)->rb_left); | |
378 | else | |
379 | new = &((*new)->rb_right); | |
380 | } | |
381 | ||
382 | rb_link_node(&data->node, parent, new); | |
383 | rb_insert_color(&data->node, root); | |
384 | } | |
385 | ||
386 | static void __sort_result(struct rb_root *root, struct rb_root *root_sorted, | |
29b3e152 | 387 | struct list_head *sort_list) |
ba77c9e1 LZ |
388 | { |
389 | struct rb_node *node; | |
390 | struct alloc_stat *data; | |
391 | ||
392 | for (;;) { | |
393 | node = rb_first(root); | |
394 | if (!node) | |
395 | break; | |
396 | ||
397 | rb_erase(node, root); | |
398 | data = rb_entry(node, struct alloc_stat, node); | |
29b3e152 | 399 | sort_insert(root_sorted, data, sort_list); |
ba77c9e1 LZ |
400 | } |
401 | } | |
402 | ||
403 | static void sort_result(void) | |
404 | { | |
29b3e152 LZ |
405 | __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort); |
406 | __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort); | |
ba77c9e1 LZ |
407 | } |
408 | ||
2b2b2c68 | 409 | static int __cmd_kmem(struct perf_session *session) |
ba77c9e1 | 410 | { |
d549c769 | 411 | int err = -EINVAL; |
0f7d2f1b ACM |
412 | const struct perf_evsel_str_handler kmem_tracepoints[] = { |
413 | { "kmem:kmalloc", perf_evsel__process_alloc_event, }, | |
414 | { "kmem:kmem_cache_alloc", perf_evsel__process_alloc_event, }, | |
415 | { "kmem:kmalloc_node", perf_evsel__process_alloc_node_event, }, | |
416 | { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, }, | |
417 | { "kmem:kfree", perf_evsel__process_free_event, }, | |
418 | { "kmem:kmem_cache_free", perf_evsel__process_free_event, }, | |
419 | }; | |
4aa65636 | 420 | |
d549c769 | 421 | if (!perf_session__has_traces(session, "kmem record")) |
2b2b2c68 | 422 | goto out; |
d549c769 | 423 | |
0f7d2f1b ACM |
424 | if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) { |
425 | pr_err("Initializing perf session tracepoint handlers failed\n"); | |
2b2b2c68 | 426 | goto out; |
0f7d2f1b ACM |
427 | } |
428 | ||
ba77c9e1 | 429 | setup_pager(); |
b7b61cbe | 430 | err = perf_session__process_events(session); |
4aa65636 | 431 | if (err != 0) |
2b2b2c68 | 432 | goto out; |
ba77c9e1 | 433 | sort_result(); |
4aa65636 | 434 | print_result(session); |
2b2b2c68 | 435 | out: |
4aa65636 | 436 | return err; |
ba77c9e1 LZ |
437 | } |
438 | ||
ba77c9e1 LZ |
439 | static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r) |
440 | { | |
441 | if (l->ptr < r->ptr) | |
442 | return -1; | |
443 | else if (l->ptr > r->ptr) | |
444 | return 1; | |
445 | return 0; | |
446 | } | |
447 | ||
29b3e152 LZ |
448 | static struct sort_dimension ptr_sort_dimension = { |
449 | .name = "ptr", | |
450 | .cmp = ptr_cmp, | |
451 | }; | |
452 | ||
ba77c9e1 LZ |
453 | static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r) |
454 | { | |
455 | if (l->call_site < r->call_site) | |
456 | return -1; | |
457 | else if (l->call_site > r->call_site) | |
458 | return 1; | |
459 | return 0; | |
460 | } | |
461 | ||
29b3e152 LZ |
462 | static struct sort_dimension callsite_sort_dimension = { |
463 | .name = "callsite", | |
464 | .cmp = callsite_cmp, | |
465 | }; | |
466 | ||
f3ced7cd PE |
467 | static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r) |
468 | { | |
469 | if (l->hit < r->hit) | |
470 | return -1; | |
471 | else if (l->hit > r->hit) | |
472 | return 1; | |
473 | return 0; | |
474 | } | |
475 | ||
29b3e152 LZ |
476 | static struct sort_dimension hit_sort_dimension = { |
477 | .name = "hit", | |
478 | .cmp = hit_cmp, | |
479 | }; | |
480 | ||
ba77c9e1 LZ |
481 | static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r) |
482 | { | |
483 | if (l->bytes_alloc < r->bytes_alloc) | |
484 | return -1; | |
485 | else if (l->bytes_alloc > r->bytes_alloc) | |
486 | return 1; | |
487 | return 0; | |
488 | } | |
489 | ||
29b3e152 LZ |
490 | static struct sort_dimension bytes_sort_dimension = { |
491 | .name = "bytes", | |
492 | .cmp = bytes_cmp, | |
493 | }; | |
494 | ||
f3ced7cd PE |
495 | static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r) |
496 | { | |
497 | double x, y; | |
498 | ||
499 | x = fragmentation(l->bytes_req, l->bytes_alloc); | |
500 | y = fragmentation(r->bytes_req, r->bytes_alloc); | |
501 | ||
502 | if (x < y) | |
503 | return -1; | |
504 | else if (x > y) | |
505 | return 1; | |
506 | return 0; | |
507 | } | |
508 | ||
29b3e152 LZ |
509 | static struct sort_dimension frag_sort_dimension = { |
510 | .name = "frag", | |
511 | .cmp = frag_cmp, | |
512 | }; | |
513 | ||
079d3f65 LZ |
514 | static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r) |
515 | { | |
516 | if (l->pingpong < r->pingpong) | |
517 | return -1; | |
518 | else if (l->pingpong > r->pingpong) | |
519 | return 1; | |
520 | return 0; | |
521 | } | |
522 | ||
523 | static struct sort_dimension pingpong_sort_dimension = { | |
524 | .name = "pingpong", | |
525 | .cmp = pingpong_cmp, | |
526 | }; | |
527 | ||
29b3e152 LZ |
528 | static struct sort_dimension *avail_sorts[] = { |
529 | &ptr_sort_dimension, | |
530 | &callsite_sort_dimension, | |
531 | &hit_sort_dimension, | |
532 | &bytes_sort_dimension, | |
533 | &frag_sort_dimension, | |
079d3f65 | 534 | &pingpong_sort_dimension, |
29b3e152 LZ |
535 | }; |
536 | ||
49e4ba54 | 537 | #define NUM_AVAIL_SORTS ((int)ARRAY_SIZE(avail_sorts)) |
29b3e152 LZ |
538 | |
539 | static int sort_dimension__add(const char *tok, struct list_head *list) | |
540 | { | |
541 | struct sort_dimension *sort; | |
542 | int i; | |
543 | ||
544 | for (i = 0; i < NUM_AVAIL_SORTS; i++) { | |
545 | if (!strcmp(avail_sorts[i]->name, tok)) { | |
8d9233f2 | 546 | sort = memdup(avail_sorts[i], sizeof(*avail_sorts[i])); |
2814eb05 | 547 | if (!sort) { |
8d9233f2 | 548 | pr_err("%s: memdup failed\n", __func__); |
2814eb05 ACM |
549 | return -1; |
550 | } | |
29b3e152 LZ |
551 | list_add_tail(&sort->list, list); |
552 | return 0; | |
553 | } | |
554 | } | |
555 | ||
556 | return -1; | |
557 | } | |
558 | ||
559 | static int setup_sorting(struct list_head *sort_list, const char *arg) | |
560 | { | |
561 | char *tok; | |
562 | char *str = strdup(arg); | |
405f8755 | 563 | char *pos = str; |
29b3e152 | 564 | |
2814eb05 ACM |
565 | if (!str) { |
566 | pr_err("%s: strdup failed\n", __func__); | |
567 | return -1; | |
568 | } | |
29b3e152 LZ |
569 | |
570 | while (true) { | |
405f8755 | 571 | tok = strsep(&pos, ","); |
29b3e152 LZ |
572 | if (!tok) |
573 | break; | |
574 | if (sort_dimension__add(tok, sort_list) < 0) { | |
575 | error("Unknown --sort key: '%s'", tok); | |
1b22859d | 576 | free(str); |
29b3e152 LZ |
577 | return -1; |
578 | } | |
579 | } | |
580 | ||
581 | free(str); | |
582 | return 0; | |
583 | } | |
584 | ||
1d037ca1 IT |
585 | static int parse_sort_opt(const struct option *opt __maybe_unused, |
586 | const char *arg, int unset __maybe_unused) | |
ba77c9e1 | 587 | { |
ba77c9e1 LZ |
588 | if (!arg) |
589 | return -1; | |
590 | ||
ba77c9e1 | 591 | if (caller_flag > alloc_flag) |
29b3e152 | 592 | return setup_sorting(&caller_sort, arg); |
ba77c9e1 | 593 | else |
29b3e152 | 594 | return setup_sorting(&alloc_sort, arg); |
ba77c9e1 LZ |
595 | |
596 | return 0; | |
597 | } | |
598 | ||
1d037ca1 IT |
599 | static int parse_caller_opt(const struct option *opt __maybe_unused, |
600 | const char *arg __maybe_unused, | |
601 | int unset __maybe_unused) | |
ba77c9e1 | 602 | { |
90b86a9f LZ |
603 | caller_flag = (alloc_flag + 1); |
604 | return 0; | |
605 | } | |
ba77c9e1 | 606 | |
1d037ca1 IT |
607 | static int parse_alloc_opt(const struct option *opt __maybe_unused, |
608 | const char *arg __maybe_unused, | |
609 | int unset __maybe_unused) | |
90b86a9f LZ |
610 | { |
611 | alloc_flag = (caller_flag + 1); | |
ba77c9e1 LZ |
612 | return 0; |
613 | } | |
614 | ||
1d037ca1 IT |
615 | static int parse_line_opt(const struct option *opt __maybe_unused, |
616 | const char *arg, int unset __maybe_unused) | |
ba77c9e1 LZ |
617 | { |
618 | int lines; | |
619 | ||
620 | if (!arg) | |
621 | return -1; | |
622 | ||
623 | lines = strtoul(arg, NULL, 10); | |
624 | ||
625 | if (caller_flag > alloc_flag) | |
626 | caller_lines = lines; | |
627 | else | |
628 | alloc_lines = lines; | |
629 | ||
630 | return 0; | |
631 | } | |
632 | ||
0433ffbe ACM |
633 | static int __cmd_record(int argc, const char **argv) |
634 | { | |
635 | const char * const record_args[] = { | |
4a4d371a | 636 | "record", "-a", "-R", "-c", "1", |
ba77c9e1 LZ |
637 | "-e", "kmem:kmalloc", |
638 | "-e", "kmem:kmalloc_node", | |
639 | "-e", "kmem:kfree", | |
640 | "-e", "kmem:kmem_cache_alloc", | |
641 | "-e", "kmem:kmem_cache_alloc_node", | |
642 | "-e", "kmem:kmem_cache_free", | |
0433ffbe | 643 | }; |
ba77c9e1 LZ |
644 | unsigned int rec_argc, i, j; |
645 | const char **rec_argv; | |
646 | ||
647 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
648 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); | |
649 | ||
ce47dc56 CS |
650 | if (rec_argv == NULL) |
651 | return -ENOMEM; | |
652 | ||
ba77c9e1 LZ |
653 | for (i = 0; i < ARRAY_SIZE(record_args); i++) |
654 | rec_argv[i] = strdup(record_args[i]); | |
655 | ||
656 | for (j = 1; j < (unsigned int)argc; j++, i++) | |
657 | rec_argv[i] = argv[j]; | |
658 | ||
659 | return cmd_record(i, rec_argv, NULL); | |
660 | } | |
661 | ||
1d037ca1 | 662 | int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused) |
ba77c9e1 | 663 | { |
0433ffbe | 664 | const char * const default_sort_order = "frag,hit,bytes"; |
d1eeb77c | 665 | struct perf_data_file file = { |
d1eeb77c YS |
666 | .mode = PERF_DATA_MODE_READ, |
667 | }; | |
0433ffbe ACM |
668 | const struct option kmem_options[] = { |
669 | OPT_STRING('i', "input", &input_name, "file", "input file name"), | |
bd72a33e NK |
670 | OPT_INCR('v', "verbose", &verbose, |
671 | "be more verbose (show symbol address, etc)"), | |
0433ffbe ACM |
672 | OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL, |
673 | "show per-callsite statistics", parse_caller_opt), | |
674 | OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL, | |
675 | "show per-allocation statistics", parse_alloc_opt), | |
676 | OPT_CALLBACK('s', "sort", NULL, "key[,key2...]", | |
677 | "sort by keys: ptr, call_site, bytes, hit, pingpong, frag", | |
678 | parse_sort_opt), | |
679 | OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt), | |
680 | OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), | |
d1eeb77c | 681 | OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"), |
0433ffbe ACM |
682 | OPT_END() |
683 | }; | |
3bca2354 RR |
684 | const char *const kmem_subcommands[] = { "record", "stat", NULL }; |
685 | const char *kmem_usage[] = { | |
686 | NULL, | |
0433ffbe ACM |
687 | NULL |
688 | }; | |
2b2b2c68 | 689 | struct perf_session *session; |
2b2b2c68 NK |
690 | int ret = -1; |
691 | ||
3bca2354 RR |
692 | argc = parse_options_subcommand(argc, argv, kmem_options, |
693 | kmem_subcommands, kmem_usage, 0); | |
ba77c9e1 | 694 | |
90b86a9f | 695 | if (!argc) |
ba77c9e1 LZ |
696 | usage_with_options(kmem_usage, kmem_options); |
697 | ||
90b86a9f | 698 | if (!strncmp(argv[0], "rec", 3)) { |
0a7e6d1b | 699 | symbol__init(NULL); |
90b86a9f | 700 | return __cmd_record(argc, argv); |
2b2b2c68 NK |
701 | } |
702 | ||
28939e1a JO |
703 | file.path = input_name; |
704 | ||
2b2b2c68 NK |
705 | session = perf_session__new(&file, false, &perf_kmem); |
706 | if (session == NULL) | |
52e02834 | 707 | return -1; |
2b2b2c68 | 708 | |
0a7e6d1b | 709 | symbol__init(&session->header.env); |
2b2b2c68 NK |
710 | |
711 | if (!strcmp(argv[0], "stat")) { | |
77cfe388 NK |
712 | setlocale(LC_ALL, ""); |
713 | ||
4b627957 | 714 | if (cpu__setup_cpunode_map()) |
2b2b2c68 | 715 | goto out_delete; |
90b86a9f LZ |
716 | |
717 | if (list_empty(&caller_sort)) | |
718 | setup_sorting(&caller_sort, default_sort_order); | |
719 | if (list_empty(&alloc_sort)) | |
720 | setup_sorting(&alloc_sort, default_sort_order); | |
ba77c9e1 | 721 | |
2b2b2c68 | 722 | ret = __cmd_kmem(session); |
b00eca8c PE |
723 | } else |
724 | usage_with_options(kmem_usage, kmem_options); | |
7d0d3945 | 725 | |
2b2b2c68 NK |
726 | out_delete: |
727 | perf_session__delete(session); | |
728 | ||
729 | return ret; | |
ba77c9e1 LZ |
730 | } |
731 |