]>
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> |
ba77c9e1 LZ |
23 | |
24 | struct alloc_stat; | |
25 | typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *); | |
26 | ||
ba77c9e1 LZ |
27 | static int alloc_flag; |
28 | static int caller_flag; | |
29 | ||
ba77c9e1 LZ |
30 | static int alloc_lines = -1; |
31 | static int caller_lines = -1; | |
32 | ||
7707b6b6 LZ |
33 | static bool raw_ip; |
34 | ||
ba77c9e1 | 35 | struct alloc_stat { |
079d3f65 LZ |
36 | u64 call_site; |
37 | u64 ptr; | |
ba77c9e1 LZ |
38 | u64 bytes_req; |
39 | u64 bytes_alloc; | |
40 | u32 hit; | |
079d3f65 LZ |
41 | u32 pingpong; |
42 | ||
43 | short alloc_cpu; | |
ba77c9e1 LZ |
44 | |
45 | struct rb_node node; | |
46 | }; | |
47 | ||
48 | static struct rb_root root_alloc_stat; | |
49 | static struct rb_root root_alloc_sorted; | |
50 | static struct rb_root root_caller_stat; | |
51 | static struct rb_root root_caller_sorted; | |
52 | ||
53 | static unsigned long total_requested, total_allocated; | |
7d0d3945 | 54 | static unsigned long nr_allocs, nr_cross_allocs; |
ba77c9e1 | 55 | |
2814eb05 ACM |
56 | static int insert_alloc_stat(unsigned long call_site, unsigned long ptr, |
57 | int bytes_req, int bytes_alloc, int cpu) | |
ba77c9e1 LZ |
58 | { |
59 | struct rb_node **node = &root_alloc_stat.rb_node; | |
60 | struct rb_node *parent = NULL; | |
61 | struct alloc_stat *data = NULL; | |
62 | ||
ba77c9e1 LZ |
63 | while (*node) { |
64 | parent = *node; | |
65 | data = rb_entry(*node, struct alloc_stat, node); | |
66 | ||
67 | if (ptr > data->ptr) | |
68 | node = &(*node)->rb_right; | |
69 | else if (ptr < data->ptr) | |
70 | node = &(*node)->rb_left; | |
71 | else | |
72 | break; | |
73 | } | |
74 | ||
75 | if (data && data->ptr == ptr) { | |
76 | data->hit++; | |
77 | data->bytes_req += bytes_req; | |
4efb5290 | 78 | data->bytes_alloc += bytes_alloc; |
ba77c9e1 LZ |
79 | } else { |
80 | data = malloc(sizeof(*data)); | |
2814eb05 ACM |
81 | if (!data) { |
82 | pr_err("%s: malloc failed\n", __func__); | |
83 | return -1; | |
84 | } | |
ba77c9e1 | 85 | data->ptr = ptr; |
079d3f65 | 86 | data->pingpong = 0; |
ba77c9e1 LZ |
87 | data->hit = 1; |
88 | data->bytes_req = bytes_req; | |
89 | data->bytes_alloc = bytes_alloc; | |
90 | ||
91 | rb_link_node(&data->node, parent, node); | |
92 | rb_insert_color(&data->node, &root_alloc_stat); | |
93 | } | |
079d3f65 LZ |
94 | data->call_site = call_site; |
95 | data->alloc_cpu = cpu; | |
2814eb05 | 96 | return 0; |
ba77c9e1 LZ |
97 | } |
98 | ||
2814eb05 | 99 | static int insert_caller_stat(unsigned long call_site, |
ba77c9e1 LZ |
100 | int bytes_req, int bytes_alloc) |
101 | { | |
102 | struct rb_node **node = &root_caller_stat.rb_node; | |
103 | struct rb_node *parent = NULL; | |
104 | struct alloc_stat *data = NULL; | |
105 | ||
ba77c9e1 LZ |
106 | while (*node) { |
107 | parent = *node; | |
108 | data = rb_entry(*node, struct alloc_stat, node); | |
109 | ||
110 | if (call_site > data->call_site) | |
111 | node = &(*node)->rb_right; | |
112 | else if (call_site < data->call_site) | |
113 | node = &(*node)->rb_left; | |
114 | else | |
115 | break; | |
116 | } | |
117 | ||
118 | if (data && data->call_site == call_site) { | |
119 | data->hit++; | |
120 | data->bytes_req += bytes_req; | |
4efb5290 | 121 | data->bytes_alloc += bytes_alloc; |
ba77c9e1 LZ |
122 | } else { |
123 | data = malloc(sizeof(*data)); | |
2814eb05 ACM |
124 | if (!data) { |
125 | pr_err("%s: malloc failed\n", __func__); | |
126 | return -1; | |
127 | } | |
ba77c9e1 | 128 | data->call_site = call_site; |
079d3f65 | 129 | data->pingpong = 0; |
ba77c9e1 LZ |
130 | data->hit = 1; |
131 | data->bytes_req = bytes_req; | |
132 | data->bytes_alloc = bytes_alloc; | |
133 | ||
134 | rb_link_node(&data->node, parent, node); | |
135 | rb_insert_color(&data->node, &root_caller_stat); | |
136 | } | |
2814eb05 ACM |
137 | |
138 | return 0; | |
ba77c9e1 LZ |
139 | } |
140 | ||
2814eb05 | 141 | static int perf_evsel__process_alloc_event(struct perf_evsel *evsel, |
0f7d2f1b | 142 | struct perf_sample *sample) |
ba77c9e1 | 143 | { |
0f7d2f1b ACM |
144 | unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"), |
145 | call_site = perf_evsel__intval(evsel, sample, "call_site"); | |
146 | int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"), | |
147 | bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc"); | |
148 | ||
149 | if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) || | |
2814eb05 ACM |
150 | insert_caller_stat(call_site, bytes_req, bytes_alloc)) |
151 | return -1; | |
ba77c9e1 LZ |
152 | |
153 | total_requested += bytes_req; | |
154 | total_allocated += bytes_alloc; | |
7d0d3945 | 155 | |
0f7d2f1b ACM |
156 | nr_allocs++; |
157 | return 0; | |
158 | } | |
159 | ||
160 | static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel, | |
161 | struct perf_sample *sample) | |
162 | { | |
163 | int ret = perf_evsel__process_alloc_event(evsel, sample); | |
164 | ||
165 | if (!ret) { | |
4b627957 | 166 | int node1 = cpu__get_node(sample->cpu), |
0f7d2f1b ACM |
167 | node2 = perf_evsel__intval(evsel, sample, "node"); |
168 | ||
7d0d3945 LZ |
169 | if (node1 != node2) |
170 | nr_cross_allocs++; | |
171 | } | |
0f7d2f1b ACM |
172 | |
173 | return ret; | |
ba77c9e1 LZ |
174 | } |
175 | ||
079d3f65 LZ |
176 | static int ptr_cmp(struct alloc_stat *, struct alloc_stat *); |
177 | static int callsite_cmp(struct alloc_stat *, struct alloc_stat *); | |
178 | ||
179 | static struct alloc_stat *search_alloc_stat(unsigned long ptr, | |
180 | unsigned long call_site, | |
181 | struct rb_root *root, | |
182 | sort_fn_t sort_fn) | |
183 | { | |
184 | struct rb_node *node = root->rb_node; | |
185 | struct alloc_stat key = { .ptr = ptr, .call_site = call_site }; | |
186 | ||
187 | while (node) { | |
188 | struct alloc_stat *data; | |
189 | int cmp; | |
190 | ||
191 | data = rb_entry(node, struct alloc_stat, node); | |
192 | ||
193 | cmp = sort_fn(&key, data); | |
194 | if (cmp < 0) | |
195 | node = node->rb_left; | |
196 | else if (cmp > 0) | |
197 | node = node->rb_right; | |
198 | else | |
199 | return data; | |
200 | } | |
201 | return NULL; | |
202 | } | |
203 | ||
2814eb05 ACM |
204 | static int perf_evsel__process_free_event(struct perf_evsel *evsel, |
205 | struct perf_sample *sample) | |
ba77c9e1 | 206 | { |
0f7d2f1b | 207 | unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"); |
079d3f65 LZ |
208 | struct alloc_stat *s_alloc, *s_caller; |
209 | ||
079d3f65 LZ |
210 | s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp); |
211 | if (!s_alloc) | |
2814eb05 | 212 | return 0; |
079d3f65 | 213 | |
22ad798c | 214 | if ((short)sample->cpu != s_alloc->alloc_cpu) { |
079d3f65 LZ |
215 | s_alloc->pingpong++; |
216 | ||
217 | s_caller = search_alloc_stat(0, s_alloc->call_site, | |
218 | &root_caller_stat, callsite_cmp); | |
2814eb05 ACM |
219 | if (!s_caller) |
220 | return -1; | |
079d3f65 LZ |
221 | s_caller->pingpong++; |
222 | } | |
223 | s_alloc->alloc_cpu = -1; | |
2814eb05 ACM |
224 | |
225 | return 0; | |
ba77c9e1 LZ |
226 | } |
227 | ||
0f7d2f1b ACM |
228 | typedef int (*tracepoint_handler)(struct perf_evsel *evsel, |
229 | struct perf_sample *sample); | |
ba77c9e1 | 230 | |
1d037ca1 | 231 | static int process_sample_event(struct perf_tool *tool __maybe_unused, |
d20deb64 | 232 | union perf_event *event, |
8115d60c | 233 | struct perf_sample *sample, |
fcf65bf1 | 234 | struct perf_evsel *evsel, |
743eb868 | 235 | struct machine *machine) |
ba77c9e1 | 236 | { |
ef89325f | 237 | struct thread *thread = machine__findnew_thread(machine, sample->pid, |
13ce34df | 238 | sample->tid); |
ba77c9e1 | 239 | |
ba77c9e1 LZ |
240 | if (thread == NULL) { |
241 | pr_debug("problem processing %d event, skipping it.\n", | |
242 | event->header.type); | |
243 | return -1; | |
244 | } | |
245 | ||
b9c5143a | 246 | dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid); |
ba77c9e1 | 247 | |
744a9719 ACM |
248 | if (evsel->handler != NULL) { |
249 | tracepoint_handler f = evsel->handler; | |
0f7d2f1b ACM |
250 | return f(evsel, sample); |
251 | } | |
252 | ||
253 | return 0; | |
ba77c9e1 LZ |
254 | } |
255 | ||
fcf65bf1 ACM |
256 | static struct perf_tool perf_kmem = { |
257 | .sample = process_sample_event, | |
258 | .comm = perf_event__process_comm, | |
64c40908 NK |
259 | .mmap = perf_event__process_mmap, |
260 | .mmap2 = perf_event__process_mmap2, | |
0a8cb85c | 261 | .ordered_events = true, |
ba77c9e1 LZ |
262 | }; |
263 | ||
ba77c9e1 LZ |
264 | static double fragmentation(unsigned long n_req, unsigned long n_alloc) |
265 | { | |
266 | if (n_alloc == 0) | |
267 | return 0.0; | |
268 | else | |
269 | return 100.0 - (100.0 * n_req / n_alloc); | |
270 | } | |
271 | ||
4aa65636 ACM |
272 | static void __print_result(struct rb_root *root, struct perf_session *session, |
273 | int n_lines, int is_caller) | |
ba77c9e1 LZ |
274 | { |
275 | struct rb_node *next; | |
34ba5122 | 276 | struct machine *machine = &session->machines.host; |
ba77c9e1 | 277 | |
079d3f65 LZ |
278 | printf("%.102s\n", graph_dotted_line); |
279 | printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr"); | |
47103277 | 280 | printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n"); |
079d3f65 | 281 | printf("%.102s\n", graph_dotted_line); |
ba77c9e1 LZ |
282 | |
283 | next = rb_first(root); | |
284 | ||
285 | while (next && n_lines--) { | |
1b145ae5 ACM |
286 | struct alloc_stat *data = rb_entry(next, struct alloc_stat, |
287 | node); | |
288 | struct symbol *sym = NULL; | |
71cf8b8f | 289 | struct map *map; |
079d3f65 | 290 | char buf[BUFSIZ]; |
1b145ae5 ACM |
291 | u64 addr; |
292 | ||
293 | if (is_caller) { | |
294 | addr = data->call_site; | |
7707b6b6 | 295 | if (!raw_ip) |
5c0541d5 | 296 | sym = machine__find_kernel_function(machine, addr, &map, NULL); |
1b145ae5 ACM |
297 | } else |
298 | addr = data->ptr; | |
299 | ||
300 | if (sym != NULL) | |
9486aa38 | 301 | snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name, |
71cf8b8f | 302 | addr - map->unmap_ip(map, sym->start)); |
1b145ae5 | 303 | else |
9486aa38 | 304 | snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr); |
079d3f65 | 305 | printf(" %-34s |", buf); |
ba77c9e1 | 306 | |
47103277 | 307 | printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n", |
079d3f65 | 308 | (unsigned long long)data->bytes_alloc, |
ba77c9e1 LZ |
309 | (unsigned long)data->bytes_alloc / data->hit, |
310 | (unsigned long long)data->bytes_req, | |
311 | (unsigned long)data->bytes_req / data->hit, | |
312 | (unsigned long)data->hit, | |
079d3f65 | 313 | (unsigned long)data->pingpong, |
ba77c9e1 LZ |
314 | fragmentation(data->bytes_req, data->bytes_alloc)); |
315 | ||
316 | next = rb_next(next); | |
317 | } | |
318 | ||
319 | if (n_lines == -1) | |
079d3f65 | 320 | printf(" ... | ... | ... | ... | ... | ... \n"); |
ba77c9e1 | 321 | |
079d3f65 | 322 | printf("%.102s\n", graph_dotted_line); |
ba77c9e1 LZ |
323 | } |
324 | ||
325 | static void print_summary(void) | |
326 | { | |
327 | printf("\nSUMMARY\n=======\n"); | |
328 | printf("Total bytes requested: %lu\n", total_requested); | |
329 | printf("Total bytes allocated: %lu\n", total_allocated); | |
330 | printf("Total bytes wasted on internal fragmentation: %lu\n", | |
331 | total_allocated - total_requested); | |
332 | printf("Internal fragmentation: %f%%\n", | |
333 | fragmentation(total_requested, total_allocated)); | |
7d0d3945 | 334 | printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs); |
ba77c9e1 LZ |
335 | } |
336 | ||
4aa65636 | 337 | static void print_result(struct perf_session *session) |
ba77c9e1 LZ |
338 | { |
339 | if (caller_flag) | |
4aa65636 | 340 | __print_result(&root_caller_sorted, session, caller_lines, 1); |
ba77c9e1 | 341 | if (alloc_flag) |
4aa65636 | 342 | __print_result(&root_alloc_sorted, session, alloc_lines, 0); |
ba77c9e1 LZ |
343 | print_summary(); |
344 | } | |
345 | ||
29b3e152 LZ |
346 | struct sort_dimension { |
347 | const char name[20]; | |
348 | sort_fn_t cmp; | |
349 | struct list_head list; | |
350 | }; | |
351 | ||
352 | static LIST_HEAD(caller_sort); | |
353 | static LIST_HEAD(alloc_sort); | |
354 | ||
ba77c9e1 | 355 | static void sort_insert(struct rb_root *root, struct alloc_stat *data, |
29b3e152 | 356 | struct list_head *sort_list) |
ba77c9e1 LZ |
357 | { |
358 | struct rb_node **new = &(root->rb_node); | |
359 | struct rb_node *parent = NULL; | |
29b3e152 | 360 | struct sort_dimension *sort; |
ba77c9e1 LZ |
361 | |
362 | while (*new) { | |
363 | struct alloc_stat *this; | |
29b3e152 | 364 | int cmp = 0; |
ba77c9e1 LZ |
365 | |
366 | this = rb_entry(*new, struct alloc_stat, node); | |
367 | parent = *new; | |
368 | ||
29b3e152 LZ |
369 | list_for_each_entry(sort, sort_list, list) { |
370 | cmp = sort->cmp(data, this); | |
371 | if (cmp) | |
372 | break; | |
373 | } | |
ba77c9e1 LZ |
374 | |
375 | if (cmp > 0) | |
376 | new = &((*new)->rb_left); | |
377 | else | |
378 | new = &((*new)->rb_right); | |
379 | } | |
380 | ||
381 | rb_link_node(&data->node, parent, new); | |
382 | rb_insert_color(&data->node, root); | |
383 | } | |
384 | ||
385 | static void __sort_result(struct rb_root *root, struct rb_root *root_sorted, | |
29b3e152 | 386 | struct list_head *sort_list) |
ba77c9e1 LZ |
387 | { |
388 | struct rb_node *node; | |
389 | struct alloc_stat *data; | |
390 | ||
391 | for (;;) { | |
392 | node = rb_first(root); | |
393 | if (!node) | |
394 | break; | |
395 | ||
396 | rb_erase(node, root); | |
397 | data = rb_entry(node, struct alloc_stat, node); | |
29b3e152 | 398 | sort_insert(root_sorted, data, sort_list); |
ba77c9e1 LZ |
399 | } |
400 | } | |
401 | ||
402 | static void sort_result(void) | |
403 | { | |
29b3e152 LZ |
404 | __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort); |
405 | __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort); | |
ba77c9e1 LZ |
406 | } |
407 | ||
2b2b2c68 | 408 | static int __cmd_kmem(struct perf_session *session) |
ba77c9e1 | 409 | { |
d549c769 | 410 | int err = -EINVAL; |
0f7d2f1b ACM |
411 | const struct perf_evsel_str_handler kmem_tracepoints[] = { |
412 | { "kmem:kmalloc", perf_evsel__process_alloc_event, }, | |
413 | { "kmem:kmem_cache_alloc", perf_evsel__process_alloc_event, }, | |
414 | { "kmem:kmalloc_node", perf_evsel__process_alloc_node_event, }, | |
415 | { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, }, | |
416 | { "kmem:kfree", perf_evsel__process_free_event, }, | |
417 | { "kmem:kmem_cache_free", perf_evsel__process_free_event, }, | |
418 | }; | |
4aa65636 | 419 | |
d549c769 | 420 | if (!perf_session__has_traces(session, "kmem record")) |
2b2b2c68 | 421 | goto out; |
d549c769 | 422 | |
0f7d2f1b ACM |
423 | if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) { |
424 | pr_err("Initializing perf session tracepoint handlers failed\n"); | |
2b2b2c68 | 425 | goto out; |
0f7d2f1b ACM |
426 | } |
427 | ||
ba77c9e1 | 428 | setup_pager(); |
fcf65bf1 | 429 | err = perf_session__process_events(session, &perf_kmem); |
4aa65636 | 430 | if (err != 0) |
2b2b2c68 | 431 | goto out; |
ba77c9e1 | 432 | sort_result(); |
4aa65636 | 433 | print_result(session); |
2b2b2c68 | 434 | out: |
4aa65636 | 435 | return err; |
ba77c9e1 LZ |
436 | } |
437 | ||
ba77c9e1 LZ |
438 | static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r) |
439 | { | |
440 | if (l->ptr < r->ptr) | |
441 | return -1; | |
442 | else if (l->ptr > r->ptr) | |
443 | return 1; | |
444 | return 0; | |
445 | } | |
446 | ||
29b3e152 LZ |
447 | static struct sort_dimension ptr_sort_dimension = { |
448 | .name = "ptr", | |
449 | .cmp = ptr_cmp, | |
450 | }; | |
451 | ||
ba77c9e1 LZ |
452 | static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r) |
453 | { | |
454 | if (l->call_site < r->call_site) | |
455 | return -1; | |
456 | else if (l->call_site > r->call_site) | |
457 | return 1; | |
458 | return 0; | |
459 | } | |
460 | ||
29b3e152 LZ |
461 | static struct sort_dimension callsite_sort_dimension = { |
462 | .name = "callsite", | |
463 | .cmp = callsite_cmp, | |
464 | }; | |
465 | ||
f3ced7cd PE |
466 | static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r) |
467 | { | |
468 | if (l->hit < r->hit) | |
469 | return -1; | |
470 | else if (l->hit > r->hit) | |
471 | return 1; | |
472 | return 0; | |
473 | } | |
474 | ||
29b3e152 LZ |
475 | static struct sort_dimension hit_sort_dimension = { |
476 | .name = "hit", | |
477 | .cmp = hit_cmp, | |
478 | }; | |
479 | ||
ba77c9e1 LZ |
480 | static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r) |
481 | { | |
482 | if (l->bytes_alloc < r->bytes_alloc) | |
483 | return -1; | |
484 | else if (l->bytes_alloc > r->bytes_alloc) | |
485 | return 1; | |
486 | return 0; | |
487 | } | |
488 | ||
29b3e152 LZ |
489 | static struct sort_dimension bytes_sort_dimension = { |
490 | .name = "bytes", | |
491 | .cmp = bytes_cmp, | |
492 | }; | |
493 | ||
f3ced7cd PE |
494 | static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r) |
495 | { | |
496 | double x, y; | |
497 | ||
498 | x = fragmentation(l->bytes_req, l->bytes_alloc); | |
499 | y = fragmentation(r->bytes_req, r->bytes_alloc); | |
500 | ||
501 | if (x < y) | |
502 | return -1; | |
503 | else if (x > y) | |
504 | return 1; | |
505 | return 0; | |
506 | } | |
507 | ||
29b3e152 LZ |
508 | static struct sort_dimension frag_sort_dimension = { |
509 | .name = "frag", | |
510 | .cmp = frag_cmp, | |
511 | }; | |
512 | ||
079d3f65 LZ |
513 | static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r) |
514 | { | |
515 | if (l->pingpong < r->pingpong) | |
516 | return -1; | |
517 | else if (l->pingpong > r->pingpong) | |
518 | return 1; | |
519 | return 0; | |
520 | } | |
521 | ||
522 | static struct sort_dimension pingpong_sort_dimension = { | |
523 | .name = "pingpong", | |
524 | .cmp = pingpong_cmp, | |
525 | }; | |
526 | ||
29b3e152 LZ |
527 | static struct sort_dimension *avail_sorts[] = { |
528 | &ptr_sort_dimension, | |
529 | &callsite_sort_dimension, | |
530 | &hit_sort_dimension, | |
531 | &bytes_sort_dimension, | |
532 | &frag_sort_dimension, | |
079d3f65 | 533 | &pingpong_sort_dimension, |
29b3e152 LZ |
534 | }; |
535 | ||
49e4ba54 | 536 | #define NUM_AVAIL_SORTS ((int)ARRAY_SIZE(avail_sorts)) |
29b3e152 LZ |
537 | |
538 | static int sort_dimension__add(const char *tok, struct list_head *list) | |
539 | { | |
540 | struct sort_dimension *sort; | |
541 | int i; | |
542 | ||
543 | for (i = 0; i < NUM_AVAIL_SORTS; i++) { | |
544 | if (!strcmp(avail_sorts[i]->name, tok)) { | |
8d9233f2 | 545 | sort = memdup(avail_sorts[i], sizeof(*avail_sorts[i])); |
2814eb05 | 546 | if (!sort) { |
8d9233f2 | 547 | pr_err("%s: memdup failed\n", __func__); |
2814eb05 ACM |
548 | return -1; |
549 | } | |
29b3e152 LZ |
550 | list_add_tail(&sort->list, list); |
551 | return 0; | |
552 | } | |
553 | } | |
554 | ||
555 | return -1; | |
556 | } | |
557 | ||
558 | static int setup_sorting(struct list_head *sort_list, const char *arg) | |
559 | { | |
560 | char *tok; | |
561 | char *str = strdup(arg); | |
562 | ||
2814eb05 ACM |
563 | if (!str) { |
564 | pr_err("%s: strdup failed\n", __func__); | |
565 | return -1; | |
566 | } | |
29b3e152 LZ |
567 | |
568 | while (true) { | |
569 | tok = strsep(&str, ","); | |
570 | if (!tok) | |
571 | break; | |
572 | if (sort_dimension__add(tok, sort_list) < 0) { | |
573 | error("Unknown --sort key: '%s'", tok); | |
1b22859d | 574 | free(str); |
29b3e152 LZ |
575 | return -1; |
576 | } | |
577 | } | |
578 | ||
579 | free(str); | |
580 | return 0; | |
581 | } | |
582 | ||
1d037ca1 IT |
583 | static int parse_sort_opt(const struct option *opt __maybe_unused, |
584 | const char *arg, int unset __maybe_unused) | |
ba77c9e1 | 585 | { |
ba77c9e1 LZ |
586 | if (!arg) |
587 | return -1; | |
588 | ||
ba77c9e1 | 589 | if (caller_flag > alloc_flag) |
29b3e152 | 590 | return setup_sorting(&caller_sort, arg); |
ba77c9e1 | 591 | else |
29b3e152 | 592 | return setup_sorting(&alloc_sort, arg); |
ba77c9e1 LZ |
593 | |
594 | return 0; | |
595 | } | |
596 | ||
1d037ca1 IT |
597 | static int parse_caller_opt(const struct option *opt __maybe_unused, |
598 | const char *arg __maybe_unused, | |
599 | int unset __maybe_unused) | |
ba77c9e1 | 600 | { |
90b86a9f LZ |
601 | caller_flag = (alloc_flag + 1); |
602 | return 0; | |
603 | } | |
ba77c9e1 | 604 | |
1d037ca1 IT |
605 | static int parse_alloc_opt(const struct option *opt __maybe_unused, |
606 | const char *arg __maybe_unused, | |
607 | int unset __maybe_unused) | |
90b86a9f LZ |
608 | { |
609 | alloc_flag = (caller_flag + 1); | |
ba77c9e1 LZ |
610 | return 0; |
611 | } | |
612 | ||
1d037ca1 IT |
613 | static int parse_line_opt(const struct option *opt __maybe_unused, |
614 | const char *arg, int unset __maybe_unused) | |
ba77c9e1 LZ |
615 | { |
616 | int lines; | |
617 | ||
618 | if (!arg) | |
619 | return -1; | |
620 | ||
621 | lines = strtoul(arg, NULL, 10); | |
622 | ||
623 | if (caller_flag > alloc_flag) | |
624 | caller_lines = lines; | |
625 | else | |
626 | alloc_lines = lines; | |
627 | ||
628 | return 0; | |
629 | } | |
630 | ||
0433ffbe ACM |
631 | static int __cmd_record(int argc, const char **argv) |
632 | { | |
633 | const char * const record_args[] = { | |
4a4d371a | 634 | "record", "-a", "-R", "-c", "1", |
ba77c9e1 LZ |
635 | "-e", "kmem:kmalloc", |
636 | "-e", "kmem:kmalloc_node", | |
637 | "-e", "kmem:kfree", | |
638 | "-e", "kmem:kmem_cache_alloc", | |
639 | "-e", "kmem:kmem_cache_alloc_node", | |
640 | "-e", "kmem:kmem_cache_free", | |
0433ffbe | 641 | }; |
ba77c9e1 LZ |
642 | unsigned int rec_argc, i, j; |
643 | const char **rec_argv; | |
644 | ||
645 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
646 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); | |
647 | ||
ce47dc56 CS |
648 | if (rec_argv == NULL) |
649 | return -ENOMEM; | |
650 | ||
ba77c9e1 LZ |
651 | for (i = 0; i < ARRAY_SIZE(record_args); i++) |
652 | rec_argv[i] = strdup(record_args[i]); | |
653 | ||
654 | for (j = 1; j < (unsigned int)argc; j++, i++) | |
655 | rec_argv[i] = argv[j]; | |
656 | ||
657 | return cmd_record(i, rec_argv, NULL); | |
658 | } | |
659 | ||
1d037ca1 | 660 | int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused) |
ba77c9e1 | 661 | { |
0433ffbe | 662 | const char * const default_sort_order = "frag,hit,bytes"; |
0433ffbe ACM |
663 | const struct option kmem_options[] = { |
664 | OPT_STRING('i', "input", &input_name, "file", "input file name"), | |
665 | OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL, | |
666 | "show per-callsite statistics", parse_caller_opt), | |
667 | OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL, | |
668 | "show per-allocation statistics", parse_alloc_opt), | |
669 | OPT_CALLBACK('s', "sort", NULL, "key[,key2...]", | |
670 | "sort by keys: ptr, call_site, bytes, hit, pingpong, frag", | |
671 | parse_sort_opt), | |
672 | OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt), | |
673 | OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), | |
674 | OPT_END() | |
675 | }; | |
3bca2354 RR |
676 | const char *const kmem_subcommands[] = { "record", "stat", NULL }; |
677 | const char *kmem_usage[] = { | |
678 | NULL, | |
0433ffbe ACM |
679 | NULL |
680 | }; | |
2b2b2c68 NK |
681 | struct perf_session *session; |
682 | struct perf_data_file file = { | |
683 | .path = input_name, | |
684 | .mode = PERF_DATA_MODE_READ, | |
685 | }; | |
686 | int ret = -1; | |
687 | ||
3bca2354 RR |
688 | argc = parse_options_subcommand(argc, argv, kmem_options, |
689 | kmem_subcommands, kmem_usage, 0); | |
ba77c9e1 | 690 | |
90b86a9f | 691 | if (!argc) |
ba77c9e1 LZ |
692 | usage_with_options(kmem_usage, kmem_options); |
693 | ||
90b86a9f | 694 | if (!strncmp(argv[0], "rec", 3)) { |
0a7e6d1b | 695 | symbol__init(NULL); |
90b86a9f | 696 | return __cmd_record(argc, argv); |
2b2b2c68 NK |
697 | } |
698 | ||
699 | session = perf_session__new(&file, false, &perf_kmem); | |
700 | if (session == NULL) | |
701 | return -ENOMEM; | |
702 | ||
0a7e6d1b | 703 | symbol__init(&session->header.env); |
2b2b2c68 NK |
704 | |
705 | if (!strcmp(argv[0], "stat")) { | |
4b627957 | 706 | if (cpu__setup_cpunode_map()) |
2b2b2c68 | 707 | goto out_delete; |
90b86a9f LZ |
708 | |
709 | if (list_empty(&caller_sort)) | |
710 | setup_sorting(&caller_sort, default_sort_order); | |
711 | if (list_empty(&alloc_sort)) | |
712 | setup_sorting(&alloc_sort, default_sort_order); | |
ba77c9e1 | 713 | |
2b2b2c68 | 714 | ret = __cmd_kmem(session); |
b00eca8c PE |
715 | } else |
716 | usage_with_options(kmem_usage, kmem_options); | |
7d0d3945 | 717 | |
2b2b2c68 NK |
718 | out_delete: |
719 | perf_session__delete(session); | |
720 | ||
721 | return ret; | |
ba77c9e1 LZ |
722 | } |
723 |