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 | 6 | #include "util/util.h" |
41840d21 | 7 | #include "util/config.h" |
ba77c9e1 LZ |
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" |
c9758cc4 | 13 | #include "util/callchain.h" |
2a865bd8 | 14 | #include "util/time-utils.h" |
ba77c9e1 | 15 | |
4b6ab94e | 16 | #include <subcmd/parse-options.h> |
ba77c9e1 | 17 | #include "util/trace-event.h" |
f5fc1412 | 18 | #include "util/data.h" |
4b627957 | 19 | #include "util/cpumap.h" |
ba77c9e1 LZ |
20 | |
21 | #include "util/debug.h" | |
ba77c9e1 | 22 | |
877a7a11 | 23 | #include <linux/kernel.h> |
ba77c9e1 | 24 | #include <linux/rbtree.h> |
8d9233f2 | 25 | #include <linux/string.h> |
a43783ae | 26 | #include <errno.h> |
fd20e811 | 27 | #include <inttypes.h> |
77cfe388 | 28 | #include <locale.h> |
c9758cc4 | 29 | #include <regex.h> |
ba77c9e1 | 30 | |
3d689ed6 ACM |
31 | #include "sane_ctype.h" |
32 | ||
0d68bc92 NK |
33 | static int kmem_slab; |
34 | static int kmem_page; | |
35 | ||
36 | static long kmem_page_size; | |
0c160d49 NK |
37 | static enum { |
38 | KMEM_SLAB, | |
39 | KMEM_PAGE, | |
40 | } kmem_default = KMEM_SLAB; /* for backward compatibility */ | |
0d68bc92 | 41 | |
ba77c9e1 | 42 | struct alloc_stat; |
fb4f313d | 43 | typedef int (*sort_fn_t)(void *, void *); |
ba77c9e1 | 44 | |
ba77c9e1 LZ |
45 | static int alloc_flag; |
46 | static int caller_flag; | |
47 | ||
ba77c9e1 LZ |
48 | static int alloc_lines = -1; |
49 | static int caller_lines = -1; | |
50 | ||
7707b6b6 LZ |
51 | static bool raw_ip; |
52 | ||
ba77c9e1 | 53 | struct alloc_stat { |
079d3f65 LZ |
54 | u64 call_site; |
55 | u64 ptr; | |
ba77c9e1 LZ |
56 | u64 bytes_req; |
57 | u64 bytes_alloc; | |
aa58e9af | 58 | u64 last_alloc; |
ba77c9e1 | 59 | u32 hit; |
079d3f65 LZ |
60 | u32 pingpong; |
61 | ||
62 | short alloc_cpu; | |
ba77c9e1 LZ |
63 | |
64 | struct rb_node node; | |
65 | }; | |
66 | ||
67 | static struct rb_root root_alloc_stat; | |
68 | static struct rb_root root_alloc_sorted; | |
69 | static struct rb_root root_caller_stat; | |
70 | static struct rb_root root_caller_sorted; | |
71 | ||
aa58e9af | 72 | static unsigned long total_requested, total_allocated, total_freed; |
7d0d3945 | 73 | static unsigned long nr_allocs, nr_cross_allocs; |
ba77c9e1 | 74 | |
2a865bd8 DA |
75 | /* filters for controlling start and stop of time of analysis */ |
76 | static struct perf_time_interval ptime; | |
77 | const char *time_str; | |
78 | ||
2814eb05 ACM |
79 | static int insert_alloc_stat(unsigned long call_site, unsigned long ptr, |
80 | int bytes_req, int bytes_alloc, int cpu) | |
ba77c9e1 LZ |
81 | { |
82 | struct rb_node **node = &root_alloc_stat.rb_node; | |
83 | struct rb_node *parent = NULL; | |
84 | struct alloc_stat *data = NULL; | |
85 | ||
ba77c9e1 LZ |
86 | while (*node) { |
87 | parent = *node; | |
88 | data = rb_entry(*node, struct alloc_stat, node); | |
89 | ||
90 | if (ptr > data->ptr) | |
91 | node = &(*node)->rb_right; | |
92 | else if (ptr < data->ptr) | |
93 | node = &(*node)->rb_left; | |
94 | else | |
95 | break; | |
96 | } | |
97 | ||
98 | if (data && data->ptr == ptr) { | |
99 | data->hit++; | |
100 | data->bytes_req += bytes_req; | |
4efb5290 | 101 | data->bytes_alloc += bytes_alloc; |
ba77c9e1 LZ |
102 | } else { |
103 | data = malloc(sizeof(*data)); | |
2814eb05 ACM |
104 | if (!data) { |
105 | pr_err("%s: malloc failed\n", __func__); | |
106 | return -1; | |
107 | } | |
ba77c9e1 | 108 | data->ptr = ptr; |
079d3f65 | 109 | data->pingpong = 0; |
ba77c9e1 LZ |
110 | data->hit = 1; |
111 | data->bytes_req = bytes_req; | |
112 | data->bytes_alloc = bytes_alloc; | |
113 | ||
114 | rb_link_node(&data->node, parent, node); | |
115 | rb_insert_color(&data->node, &root_alloc_stat); | |
116 | } | |
079d3f65 LZ |
117 | data->call_site = call_site; |
118 | data->alloc_cpu = cpu; | |
aa58e9af DA |
119 | data->last_alloc = bytes_alloc; |
120 | ||
2814eb05 | 121 | return 0; |
ba77c9e1 LZ |
122 | } |
123 | ||
2814eb05 | 124 | static int insert_caller_stat(unsigned long call_site, |
ba77c9e1 LZ |
125 | int bytes_req, int bytes_alloc) |
126 | { | |
127 | struct rb_node **node = &root_caller_stat.rb_node; | |
128 | struct rb_node *parent = NULL; | |
129 | struct alloc_stat *data = NULL; | |
130 | ||
ba77c9e1 LZ |
131 | while (*node) { |
132 | parent = *node; | |
133 | data = rb_entry(*node, struct alloc_stat, node); | |
134 | ||
135 | if (call_site > data->call_site) | |
136 | node = &(*node)->rb_right; | |
137 | else if (call_site < data->call_site) | |
138 | node = &(*node)->rb_left; | |
139 | else | |
140 | break; | |
141 | } | |
142 | ||
143 | if (data && data->call_site == call_site) { | |
144 | data->hit++; | |
145 | data->bytes_req += bytes_req; | |
4efb5290 | 146 | data->bytes_alloc += bytes_alloc; |
ba77c9e1 LZ |
147 | } else { |
148 | data = malloc(sizeof(*data)); | |
2814eb05 ACM |
149 | if (!data) { |
150 | pr_err("%s: malloc failed\n", __func__); | |
151 | return -1; | |
152 | } | |
ba77c9e1 | 153 | data->call_site = call_site; |
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_caller_stat); | |
161 | } | |
2814eb05 ACM |
162 | |
163 | return 0; | |
ba77c9e1 LZ |
164 | } |
165 | ||
2814eb05 | 166 | static int perf_evsel__process_alloc_event(struct perf_evsel *evsel, |
0f7d2f1b | 167 | struct perf_sample *sample) |
ba77c9e1 | 168 | { |
0f7d2f1b ACM |
169 | unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"), |
170 | call_site = perf_evsel__intval(evsel, sample, "call_site"); | |
171 | int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"), | |
172 | bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc"); | |
173 | ||
174 | if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) || | |
2814eb05 ACM |
175 | insert_caller_stat(call_site, bytes_req, bytes_alloc)) |
176 | return -1; | |
ba77c9e1 LZ |
177 | |
178 | total_requested += bytes_req; | |
179 | total_allocated += bytes_alloc; | |
7d0d3945 | 180 | |
0f7d2f1b ACM |
181 | nr_allocs++; |
182 | return 0; | |
183 | } | |
184 | ||
185 | static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel, | |
186 | struct perf_sample *sample) | |
187 | { | |
188 | int ret = perf_evsel__process_alloc_event(evsel, sample); | |
189 | ||
190 | if (!ret) { | |
4b627957 | 191 | int node1 = cpu__get_node(sample->cpu), |
0f7d2f1b ACM |
192 | node2 = perf_evsel__intval(evsel, sample, "node"); |
193 | ||
7d0d3945 LZ |
194 | if (node1 != node2) |
195 | nr_cross_allocs++; | |
196 | } | |
0f7d2f1b ACM |
197 | |
198 | return ret; | |
ba77c9e1 LZ |
199 | } |
200 | ||
fb4f313d NK |
201 | static int ptr_cmp(void *, void *); |
202 | static int slab_callsite_cmp(void *, void *); | |
079d3f65 LZ |
203 | |
204 | static struct alloc_stat *search_alloc_stat(unsigned long ptr, | |
205 | unsigned long call_site, | |
206 | struct rb_root *root, | |
207 | sort_fn_t sort_fn) | |
208 | { | |
209 | struct rb_node *node = root->rb_node; | |
210 | struct alloc_stat key = { .ptr = ptr, .call_site = call_site }; | |
211 | ||
212 | while (node) { | |
213 | struct alloc_stat *data; | |
214 | int cmp; | |
215 | ||
216 | data = rb_entry(node, struct alloc_stat, node); | |
217 | ||
218 | cmp = sort_fn(&key, data); | |
219 | if (cmp < 0) | |
220 | node = node->rb_left; | |
221 | else if (cmp > 0) | |
222 | node = node->rb_right; | |
223 | else | |
224 | return data; | |
225 | } | |
226 | return NULL; | |
227 | } | |
228 | ||
2814eb05 ACM |
229 | static int perf_evsel__process_free_event(struct perf_evsel *evsel, |
230 | struct perf_sample *sample) | |
ba77c9e1 | 231 | { |
0f7d2f1b | 232 | unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"); |
079d3f65 LZ |
233 | struct alloc_stat *s_alloc, *s_caller; |
234 | ||
079d3f65 LZ |
235 | s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp); |
236 | if (!s_alloc) | |
2814eb05 | 237 | return 0; |
079d3f65 | 238 | |
aa58e9af DA |
239 | total_freed += s_alloc->last_alloc; |
240 | ||
22ad798c | 241 | if ((short)sample->cpu != s_alloc->alloc_cpu) { |
079d3f65 LZ |
242 | s_alloc->pingpong++; |
243 | ||
244 | s_caller = search_alloc_stat(0, s_alloc->call_site, | |
fb4f313d NK |
245 | &root_caller_stat, |
246 | slab_callsite_cmp); | |
2814eb05 ACM |
247 | if (!s_caller) |
248 | return -1; | |
079d3f65 LZ |
249 | s_caller->pingpong++; |
250 | } | |
251 | s_alloc->alloc_cpu = -1; | |
2814eb05 ACM |
252 | |
253 | return 0; | |
ba77c9e1 LZ |
254 | } |
255 | ||
0d68bc92 NK |
256 | static u64 total_page_alloc_bytes; |
257 | static u64 total_page_free_bytes; | |
258 | static u64 total_page_nomatch_bytes; | |
259 | static u64 total_page_fail_bytes; | |
260 | static unsigned long nr_page_allocs; | |
261 | static unsigned long nr_page_frees; | |
262 | static unsigned long nr_page_fails; | |
263 | static unsigned long nr_page_nomatch; | |
264 | ||
265 | static bool use_pfn; | |
2a7ef02c | 266 | static bool live_page; |
c9758cc4 | 267 | static struct perf_session *kmem_session; |
0d68bc92 NK |
268 | |
269 | #define MAX_MIGRATE_TYPES 6 | |
270 | #define MAX_PAGE_ORDER 11 | |
271 | ||
272 | static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES]; | |
273 | ||
274 | struct page_stat { | |
275 | struct rb_node node; | |
276 | u64 page; | |
c9758cc4 | 277 | u64 callsite; |
0d68bc92 NK |
278 | int order; |
279 | unsigned gfp_flags; | |
280 | unsigned migrate_type; | |
281 | u64 alloc_bytes; | |
282 | u64 free_bytes; | |
283 | int nr_alloc; | |
284 | int nr_free; | |
285 | }; | |
286 | ||
2a7ef02c | 287 | static struct rb_root page_live_tree; |
0d68bc92 NK |
288 | static struct rb_root page_alloc_tree; |
289 | static struct rb_root page_alloc_sorted; | |
c9758cc4 NK |
290 | static struct rb_root page_caller_tree; |
291 | static struct rb_root page_caller_sorted; | |
0d68bc92 | 292 | |
c9758cc4 NK |
293 | struct alloc_func { |
294 | u64 start; | |
295 | u64 end; | |
296 | char *name; | |
297 | }; | |
298 | ||
299 | static int nr_alloc_funcs; | |
300 | static struct alloc_func *alloc_func_list; | |
301 | ||
302 | static int funcmp(const void *a, const void *b) | |
303 | { | |
304 | const struct alloc_func *fa = a; | |
305 | const struct alloc_func *fb = b; | |
306 | ||
307 | if (fa->start > fb->start) | |
308 | return 1; | |
309 | else | |
310 | return -1; | |
311 | } | |
312 | ||
313 | static int callcmp(const void *a, const void *b) | |
314 | { | |
315 | const struct alloc_func *fa = a; | |
316 | const struct alloc_func *fb = b; | |
317 | ||
318 | if (fb->start <= fa->start && fa->end < fb->end) | |
319 | return 0; | |
320 | ||
321 | if (fa->start > fb->start) | |
322 | return 1; | |
323 | else | |
324 | return -1; | |
325 | } | |
326 | ||
327 | static int build_alloc_func_list(void) | |
328 | { | |
329 | int ret; | |
330 | struct map *kernel_map; | |
331 | struct symbol *sym; | |
332 | struct rb_node *node; | |
333 | struct alloc_func *func; | |
334 | struct machine *machine = &kmem_session->machines.host; | |
335 | regex_t alloc_func_regex; | |
336 | const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?"; | |
337 | ||
338 | ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED); | |
339 | if (ret) { | |
340 | char err[BUFSIZ]; | |
341 | ||
342 | regerror(ret, &alloc_func_regex, err, sizeof(err)); | |
343 | pr_err("Invalid regex: %s\n%s", pattern, err); | |
344 | return -EINVAL; | |
345 | } | |
346 | ||
a5e813c6 | 347 | kernel_map = machine__kernel_map(machine); |
be39db9f | 348 | if (map__load(kernel_map) < 0) { |
c9758cc4 NK |
349 | pr_err("cannot load kernel map\n"); |
350 | return -ENOENT; | |
351 | } | |
352 | ||
353 | map__for_each_symbol(kernel_map, sym, node) { | |
354 | if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0)) | |
355 | continue; | |
356 | ||
357 | func = realloc(alloc_func_list, | |
358 | (nr_alloc_funcs + 1) * sizeof(*func)); | |
359 | if (func == NULL) | |
360 | return -ENOMEM; | |
361 | ||
362 | pr_debug("alloc func: %s\n", sym->name); | |
363 | func[nr_alloc_funcs].start = sym->start; | |
364 | func[nr_alloc_funcs].end = sym->end; | |
365 | func[nr_alloc_funcs].name = sym->name; | |
366 | ||
367 | alloc_func_list = func; | |
368 | nr_alloc_funcs++; | |
369 | } | |
370 | ||
371 | qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp); | |
372 | ||
373 | regfree(&alloc_func_regex); | |
374 | return 0; | |
375 | } | |
376 | ||
377 | /* | |
378 | * Find first non-memory allocation function from callchain. | |
379 | * The allocation functions are in the 'alloc_func_list'. | |
380 | */ | |
381 | static u64 find_callsite(struct perf_evsel *evsel, struct perf_sample *sample) | |
382 | { | |
383 | struct addr_location al; | |
384 | struct machine *machine = &kmem_session->machines.host; | |
385 | struct callchain_cursor_node *node; | |
386 | ||
387 | if (alloc_func_list == NULL) { | |
388 | if (build_alloc_func_list() < 0) | |
389 | goto out; | |
390 | } | |
391 | ||
392 | al.thread = machine__findnew_thread(machine, sample->pid, sample->tid); | |
91d7b2de | 393 | sample__resolve_callchain(sample, &callchain_cursor, NULL, evsel, &al, 16); |
c9758cc4 NK |
394 | |
395 | callchain_cursor_commit(&callchain_cursor); | |
396 | while (true) { | |
397 | struct alloc_func key, *caller; | |
398 | u64 addr; | |
399 | ||
400 | node = callchain_cursor_current(&callchain_cursor); | |
401 | if (node == NULL) | |
402 | break; | |
403 | ||
404 | key.start = key.end = node->ip; | |
405 | caller = bsearch(&key, alloc_func_list, nr_alloc_funcs, | |
406 | sizeof(key), callcmp); | |
407 | if (!caller) { | |
408 | /* found */ | |
409 | if (node->map) | |
410 | addr = map__unmap_ip(node->map, node->ip); | |
411 | else | |
412 | addr = node->ip; | |
413 | ||
414 | return addr; | |
415 | } else | |
416 | pr_debug3("skipping alloc function: %s\n", caller->name); | |
417 | ||
418 | callchain_cursor_advance(&callchain_cursor); | |
419 | } | |
420 | ||
421 | out: | |
422 | pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip); | |
423 | return sample->ip; | |
424 | } | |
425 | ||
2a7ef02c NK |
426 | struct sort_dimension { |
427 | const char name[20]; | |
428 | sort_fn_t cmp; | |
429 | struct list_head list; | |
430 | }; | |
431 | ||
432 | static LIST_HEAD(page_alloc_sort_input); | |
433 | static LIST_HEAD(page_caller_sort_input); | |
434 | ||
c9758cc4 | 435 | static struct page_stat * |
2a7ef02c | 436 | __page_stat__findnew_page(struct page_stat *pstat, bool create) |
0d68bc92 | 437 | { |
2a7ef02c | 438 | struct rb_node **node = &page_live_tree.rb_node; |
0d68bc92 NK |
439 | struct rb_node *parent = NULL; |
440 | struct page_stat *data; | |
441 | ||
442 | while (*node) { | |
443 | s64 cmp; | |
444 | ||
445 | parent = *node; | |
446 | data = rb_entry(*node, struct page_stat, node); | |
447 | ||
2a7ef02c | 448 | cmp = data->page - pstat->page; |
0d68bc92 NK |
449 | if (cmp < 0) |
450 | node = &parent->rb_left; | |
451 | else if (cmp > 0) | |
452 | node = &parent->rb_right; | |
453 | else | |
454 | return data; | |
455 | } | |
456 | ||
457 | if (!create) | |
458 | return NULL; | |
459 | ||
460 | data = zalloc(sizeof(*data)); | |
461 | if (data != NULL) { | |
2a7ef02c NK |
462 | data->page = pstat->page; |
463 | data->order = pstat->order; | |
464 | data->gfp_flags = pstat->gfp_flags; | |
465 | data->migrate_type = pstat->migrate_type; | |
0d68bc92 NK |
466 | |
467 | rb_link_node(&data->node, parent, node); | |
2a7ef02c | 468 | rb_insert_color(&data->node, &page_live_tree); |
0d68bc92 NK |
469 | } |
470 | ||
471 | return data; | |
472 | } | |
473 | ||
2a7ef02c | 474 | static struct page_stat *page_stat__find_page(struct page_stat *pstat) |
c9758cc4 | 475 | { |
2a7ef02c | 476 | return __page_stat__findnew_page(pstat, false); |
c9758cc4 NK |
477 | } |
478 | ||
2a7ef02c | 479 | static struct page_stat *page_stat__findnew_page(struct page_stat *pstat) |
c9758cc4 | 480 | { |
2a7ef02c | 481 | return __page_stat__findnew_page(pstat, true); |
c9758cc4 NK |
482 | } |
483 | ||
c9758cc4 NK |
484 | static struct page_stat * |
485 | __page_stat__findnew_alloc(struct page_stat *pstat, bool create) | |
0d68bc92 NK |
486 | { |
487 | struct rb_node **node = &page_alloc_tree.rb_node; | |
488 | struct rb_node *parent = NULL; | |
489 | struct page_stat *data; | |
fb4f313d | 490 | struct sort_dimension *sort; |
0d68bc92 NK |
491 | |
492 | while (*node) { | |
fb4f313d | 493 | int cmp = 0; |
0d68bc92 NK |
494 | |
495 | parent = *node; | |
496 | data = rb_entry(*node, struct page_stat, node); | |
497 | ||
fb4f313d NK |
498 | list_for_each_entry(sort, &page_alloc_sort_input, list) { |
499 | cmp = sort->cmp(pstat, data); | |
500 | if (cmp) | |
501 | break; | |
502 | } | |
503 | ||
0d68bc92 NK |
504 | if (cmp < 0) |
505 | node = &parent->rb_left; | |
506 | else if (cmp > 0) | |
507 | node = &parent->rb_right; | |
508 | else | |
509 | return data; | |
510 | } | |
511 | ||
512 | if (!create) | |
513 | return NULL; | |
514 | ||
515 | data = zalloc(sizeof(*data)); | |
516 | if (data != NULL) { | |
6b1a2752 DA |
517 | data->page = pstat->page; |
518 | data->order = pstat->order; | |
519 | data->gfp_flags = pstat->gfp_flags; | |
520 | data->migrate_type = pstat->migrate_type; | |
0d68bc92 NK |
521 | |
522 | rb_link_node(&data->node, parent, node); | |
523 | rb_insert_color(&data->node, &page_alloc_tree); | |
524 | } | |
525 | ||
526 | return data; | |
527 | } | |
528 | ||
c9758cc4 NK |
529 | static struct page_stat *page_stat__find_alloc(struct page_stat *pstat) |
530 | { | |
531 | return __page_stat__findnew_alloc(pstat, false); | |
532 | } | |
533 | ||
534 | static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat) | |
535 | { | |
536 | return __page_stat__findnew_alloc(pstat, true); | |
537 | } | |
538 | ||
539 | static struct page_stat * | |
fb4f313d | 540 | __page_stat__findnew_caller(struct page_stat *pstat, bool create) |
c9758cc4 NK |
541 | { |
542 | struct rb_node **node = &page_caller_tree.rb_node; | |
543 | struct rb_node *parent = NULL; | |
544 | struct page_stat *data; | |
fb4f313d | 545 | struct sort_dimension *sort; |
c9758cc4 NK |
546 | |
547 | while (*node) { | |
fb4f313d | 548 | int cmp = 0; |
c9758cc4 NK |
549 | |
550 | parent = *node; | |
551 | data = rb_entry(*node, struct page_stat, node); | |
552 | ||
fb4f313d NK |
553 | list_for_each_entry(sort, &page_caller_sort_input, list) { |
554 | cmp = sort->cmp(pstat, data); | |
555 | if (cmp) | |
556 | break; | |
557 | } | |
558 | ||
c9758cc4 NK |
559 | if (cmp < 0) |
560 | node = &parent->rb_left; | |
561 | else if (cmp > 0) | |
562 | node = &parent->rb_right; | |
563 | else | |
564 | return data; | |
565 | } | |
566 | ||
567 | if (!create) | |
568 | return NULL; | |
569 | ||
570 | data = zalloc(sizeof(*data)); | |
571 | if (data != NULL) { | |
fb4f313d NK |
572 | data->callsite = pstat->callsite; |
573 | data->order = pstat->order; | |
574 | data->gfp_flags = pstat->gfp_flags; | |
575 | data->migrate_type = pstat->migrate_type; | |
c9758cc4 NK |
576 | |
577 | rb_link_node(&data->node, parent, node); | |
578 | rb_insert_color(&data->node, &page_caller_tree); | |
579 | } | |
580 | ||
581 | return data; | |
582 | } | |
583 | ||
fb4f313d | 584 | static struct page_stat *page_stat__find_caller(struct page_stat *pstat) |
c9758cc4 | 585 | { |
fb4f313d | 586 | return __page_stat__findnew_caller(pstat, false); |
c9758cc4 NK |
587 | } |
588 | ||
fb4f313d | 589 | static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat) |
c9758cc4 | 590 | { |
fb4f313d | 591 | return __page_stat__findnew_caller(pstat, true); |
c9758cc4 NK |
592 | } |
593 | ||
0d68bc92 NK |
594 | static bool valid_page(u64 pfn_or_page) |
595 | { | |
596 | if (use_pfn && pfn_or_page == -1UL) | |
597 | return false; | |
598 | if (!use_pfn && pfn_or_page == 0) | |
599 | return false; | |
600 | return true; | |
601 | } | |
602 | ||
0e111156 NK |
603 | struct gfp_flag { |
604 | unsigned int flags; | |
605 | char *compact_str; | |
606 | char *human_readable; | |
607 | }; | |
608 | ||
609 | static struct gfp_flag *gfps; | |
610 | static int nr_gfps; | |
611 | ||
612 | static int gfpcmp(const void *a, const void *b) | |
613 | { | |
614 | const struct gfp_flag *fa = a; | |
615 | const struct gfp_flag *fb = b; | |
616 | ||
617 | return fa->flags - fb->flags; | |
618 | } | |
619 | ||
420adbe9 | 620 | /* see include/trace/events/mmflags.h */ |
0e111156 NK |
621 | static const struct { |
622 | const char *original; | |
623 | const char *compact; | |
624 | } gfp_compact_table[] = { | |
625 | { "GFP_TRANSHUGE", "THP" }, | |
25160354 | 626 | { "GFP_TRANSHUGE_LIGHT", "THL" }, |
0e111156 NK |
627 | { "GFP_HIGHUSER_MOVABLE", "HUM" }, |
628 | { "GFP_HIGHUSER", "HU" }, | |
629 | { "GFP_USER", "U" }, | |
630 | { "GFP_TEMPORARY", "TMP" }, | |
14e0a214 | 631 | { "GFP_KERNEL_ACCOUNT", "KAC" }, |
0e111156 NK |
632 | { "GFP_KERNEL", "K" }, |
633 | { "GFP_NOFS", "NF" }, | |
634 | { "GFP_ATOMIC", "A" }, | |
635 | { "GFP_NOIO", "NI" }, | |
0e111156 | 636 | { "GFP_NOWAIT", "NW" }, |
14e0a214 VB |
637 | { "GFP_DMA", "D" }, |
638 | { "__GFP_HIGHMEM", "HM" }, | |
639 | { "GFP_DMA32", "D32" }, | |
640 | { "__GFP_HIGH", "H" }, | |
641 | { "__GFP_ATOMIC", "_A" }, | |
642 | { "__GFP_IO", "I" }, | |
643 | { "__GFP_FS", "F" }, | |
644 | { "__GFP_COLD", "CO" }, | |
645 | { "__GFP_NOWARN", "NWR" }, | |
dcda9b04 | 646 | { "__GFP_RETRY_MAYFAIL", "R" }, |
14e0a214 VB |
647 | { "__GFP_NOFAIL", "NF" }, |
648 | { "__GFP_NORETRY", "NR" }, | |
649 | { "__GFP_COMP", "C" }, | |
650 | { "__GFP_ZERO", "Z" }, | |
651 | { "__GFP_NOMEMALLOC", "NMA" }, | |
652 | { "__GFP_MEMALLOC", "MA" }, | |
653 | { "__GFP_HARDWALL", "HW" }, | |
654 | { "__GFP_THISNODE", "TN" }, | |
655 | { "__GFP_RECLAIMABLE", "RC" }, | |
656 | { "__GFP_MOVABLE", "M" }, | |
657 | { "__GFP_ACCOUNT", "AC" }, | |
658 | { "__GFP_NOTRACK", "NT" }, | |
659 | { "__GFP_WRITE", "WR" }, | |
660 | { "__GFP_RECLAIM", "R" }, | |
661 | { "__GFP_DIRECT_RECLAIM", "DR" }, | |
662 | { "__GFP_KSWAPD_RECLAIM", "KR" }, | |
0e111156 NK |
663 | }; |
664 | ||
665 | static size_t max_gfp_len; | |
666 | ||
667 | static char *compact_gfp_flags(char *gfp_flags) | |
668 | { | |
669 | char *orig_flags = strdup(gfp_flags); | |
670 | char *new_flags = NULL; | |
b2365122 | 671 | char *str, *pos = NULL; |
0e111156 NK |
672 | size_t len = 0; |
673 | ||
674 | if (orig_flags == NULL) | |
675 | return NULL; | |
676 | ||
677 | str = strtok_r(orig_flags, "|", &pos); | |
678 | while (str) { | |
679 | size_t i; | |
680 | char *new; | |
681 | const char *cpt; | |
682 | ||
683 | for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) { | |
684 | if (strcmp(gfp_compact_table[i].original, str)) | |
685 | continue; | |
686 | ||
687 | cpt = gfp_compact_table[i].compact; | |
688 | new = realloc(new_flags, len + strlen(cpt) + 2); | |
689 | if (new == NULL) { | |
690 | free(new_flags); | |
691 | return NULL; | |
692 | } | |
693 | ||
694 | new_flags = new; | |
695 | ||
696 | if (!len) { | |
697 | strcpy(new_flags, cpt); | |
698 | } else { | |
699 | strcat(new_flags, "|"); | |
700 | strcat(new_flags, cpt); | |
701 | len++; | |
702 | } | |
703 | ||
704 | len += strlen(cpt); | |
705 | } | |
706 | ||
707 | str = strtok_r(NULL, "|", &pos); | |
708 | } | |
709 | ||
710 | if (max_gfp_len < len) | |
711 | max_gfp_len = len; | |
712 | ||
713 | free(orig_flags); | |
714 | return new_flags; | |
715 | } | |
716 | ||
717 | static char *compact_gfp_string(unsigned long gfp_flags) | |
718 | { | |
719 | struct gfp_flag key = { | |
720 | .flags = gfp_flags, | |
721 | }; | |
722 | struct gfp_flag *gfp; | |
723 | ||
724 | gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp); | |
725 | if (gfp) | |
726 | return gfp->compact_str; | |
727 | ||
728 | return NULL; | |
729 | } | |
730 | ||
731 | static int parse_gfp_flags(struct perf_evsel *evsel, struct perf_sample *sample, | |
732 | unsigned int gfp_flags) | |
733 | { | |
734 | struct pevent_record record = { | |
735 | .cpu = sample->cpu, | |
736 | .data = sample->raw_data, | |
737 | .size = sample->raw_size, | |
738 | }; | |
739 | struct trace_seq seq; | |
08a9b985 | 740 | char *str, *pos = NULL; |
0e111156 NK |
741 | |
742 | if (nr_gfps) { | |
743 | struct gfp_flag key = { | |
744 | .flags = gfp_flags, | |
745 | }; | |
746 | ||
747 | if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp)) | |
748 | return 0; | |
749 | } | |
750 | ||
751 | trace_seq_init(&seq); | |
752 | pevent_event_info(&seq, evsel->tp_format, &record); | |
753 | ||
754 | str = strtok_r(seq.buffer, " ", &pos); | |
755 | while (str) { | |
756 | if (!strncmp(str, "gfp_flags=", 10)) { | |
757 | struct gfp_flag *new; | |
758 | ||
759 | new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps)); | |
760 | if (new == NULL) | |
761 | return -ENOMEM; | |
762 | ||
763 | gfps = new; | |
764 | new += nr_gfps++; | |
765 | ||
766 | new->flags = gfp_flags; | |
767 | new->human_readable = strdup(str + 10); | |
768 | new->compact_str = compact_gfp_flags(str + 10); | |
769 | if (!new->human_readable || !new->compact_str) | |
770 | return -ENOMEM; | |
771 | ||
772 | qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp); | |
773 | } | |
774 | ||
775 | str = strtok_r(NULL, " ", &pos); | |
776 | } | |
777 | ||
778 | trace_seq_destroy(&seq); | |
779 | return 0; | |
780 | } | |
781 | ||
0d68bc92 NK |
782 | static int perf_evsel__process_page_alloc_event(struct perf_evsel *evsel, |
783 | struct perf_sample *sample) | |
784 | { | |
785 | u64 page; | |
786 | unsigned int order = perf_evsel__intval(evsel, sample, "order"); | |
787 | unsigned int gfp_flags = perf_evsel__intval(evsel, sample, "gfp_flags"); | |
788 | unsigned int migrate_type = perf_evsel__intval(evsel, sample, | |
789 | "migratetype"); | |
790 | u64 bytes = kmem_page_size << order; | |
c9758cc4 | 791 | u64 callsite; |
6b1a2752 | 792 | struct page_stat *pstat; |
0d68bc92 NK |
793 | struct page_stat this = { |
794 | .order = order, | |
795 | .gfp_flags = gfp_flags, | |
796 | .migrate_type = migrate_type, | |
797 | }; | |
798 | ||
799 | if (use_pfn) | |
800 | page = perf_evsel__intval(evsel, sample, "pfn"); | |
801 | else | |
802 | page = perf_evsel__intval(evsel, sample, "page"); | |
803 | ||
804 | nr_page_allocs++; | |
805 | total_page_alloc_bytes += bytes; | |
806 | ||
807 | if (!valid_page(page)) { | |
808 | nr_page_fails++; | |
809 | total_page_fail_bytes += bytes; | |
810 | ||
811 | return 0; | |
812 | } | |
813 | ||
0e111156 NK |
814 | if (parse_gfp_flags(evsel, sample, gfp_flags) < 0) |
815 | return -1; | |
816 | ||
c9758cc4 NK |
817 | callsite = find_callsite(evsel, sample); |
818 | ||
0d68bc92 NK |
819 | /* |
820 | * This is to find the current page (with correct gfp flags and | |
821 | * migrate type) at free event. | |
822 | */ | |
0d68bc92 | 823 | this.page = page; |
2a7ef02c | 824 | pstat = page_stat__findnew_page(&this); |
6b1a2752 | 825 | if (pstat == NULL) |
0d68bc92 NK |
826 | return -ENOMEM; |
827 | ||
c9758cc4 NK |
828 | pstat->nr_alloc++; |
829 | pstat->alloc_bytes += bytes; | |
830 | pstat->callsite = callsite; | |
831 | ||
2a7ef02c NK |
832 | if (!live_page) { |
833 | pstat = page_stat__findnew_alloc(&this); | |
834 | if (pstat == NULL) | |
835 | return -ENOMEM; | |
836 | ||
837 | pstat->nr_alloc++; | |
838 | pstat->alloc_bytes += bytes; | |
839 | pstat->callsite = callsite; | |
840 | } | |
841 | ||
fb4f313d NK |
842 | this.callsite = callsite; |
843 | pstat = page_stat__findnew_caller(&this); | |
c9758cc4 NK |
844 | if (pstat == NULL) |
845 | return -ENOMEM; | |
846 | ||
6b1a2752 DA |
847 | pstat->nr_alloc++; |
848 | pstat->alloc_bytes += bytes; | |
0d68bc92 NK |
849 | |
850 | order_stats[order][migrate_type]++; | |
851 | ||
852 | return 0; | |
853 | } | |
854 | ||
855 | static int perf_evsel__process_page_free_event(struct perf_evsel *evsel, | |
856 | struct perf_sample *sample) | |
857 | { | |
858 | u64 page; | |
859 | unsigned int order = perf_evsel__intval(evsel, sample, "order"); | |
860 | u64 bytes = kmem_page_size << order; | |
6b1a2752 | 861 | struct page_stat *pstat; |
0d68bc92 NK |
862 | struct page_stat this = { |
863 | .order = order, | |
864 | }; | |
865 | ||
866 | if (use_pfn) | |
867 | page = perf_evsel__intval(evsel, sample, "pfn"); | |
868 | else | |
869 | page = perf_evsel__intval(evsel, sample, "page"); | |
870 | ||
871 | nr_page_frees++; | |
872 | total_page_free_bytes += bytes; | |
873 | ||
2a7ef02c NK |
874 | this.page = page; |
875 | pstat = page_stat__find_page(&this); | |
6b1a2752 | 876 | if (pstat == NULL) { |
0d68bc92 NK |
877 | pr_debug2("missing free at page %"PRIx64" (order: %d)\n", |
878 | page, order); | |
879 | ||
880 | nr_page_nomatch++; | |
881 | total_page_nomatch_bytes += bytes; | |
882 | ||
883 | return 0; | |
884 | } | |
885 | ||
6b1a2752 DA |
886 | this.gfp_flags = pstat->gfp_flags; |
887 | this.migrate_type = pstat->migrate_type; | |
c9758cc4 | 888 | this.callsite = pstat->callsite; |
0d68bc92 | 889 | |
2a7ef02c | 890 | rb_erase(&pstat->node, &page_live_tree); |
6b1a2752 | 891 | free(pstat); |
0d68bc92 | 892 | |
2a7ef02c NK |
893 | if (live_page) { |
894 | order_stats[this.order][this.migrate_type]--; | |
895 | } else { | |
896 | pstat = page_stat__find_alloc(&this); | |
897 | if (pstat == NULL) | |
898 | return -ENOMEM; | |
c9758cc4 | 899 | |
2a7ef02c NK |
900 | pstat->nr_free++; |
901 | pstat->free_bytes += bytes; | |
902 | } | |
c9758cc4 | 903 | |
fb4f313d | 904 | pstat = page_stat__find_caller(&this); |
6b1a2752 | 905 | if (pstat == NULL) |
0d68bc92 NK |
906 | return -ENOENT; |
907 | ||
6b1a2752 DA |
908 | pstat->nr_free++; |
909 | pstat->free_bytes += bytes; | |
0d68bc92 | 910 | |
2a7ef02c NK |
911 | if (live_page) { |
912 | pstat->nr_alloc--; | |
913 | pstat->alloc_bytes -= bytes; | |
914 | ||
915 | if (pstat->nr_alloc == 0) { | |
916 | rb_erase(&pstat->node, &page_caller_tree); | |
917 | free(pstat); | |
918 | } | |
919 | } | |
920 | ||
0d68bc92 NK |
921 | return 0; |
922 | } | |
923 | ||
2a865bd8 DA |
924 | static bool perf_kmem__skip_sample(struct perf_sample *sample) |
925 | { | |
926 | /* skip sample based on time? */ | |
927 | if (perf_time__skip_sample(&ptime, sample->time)) | |
928 | return true; | |
929 | ||
930 | return false; | |
931 | } | |
932 | ||
0f7d2f1b ACM |
933 | typedef int (*tracepoint_handler)(struct perf_evsel *evsel, |
934 | struct perf_sample *sample); | |
ba77c9e1 | 935 | |
1d037ca1 | 936 | static int process_sample_event(struct perf_tool *tool __maybe_unused, |
d20deb64 | 937 | union perf_event *event, |
8115d60c | 938 | struct perf_sample *sample, |
fcf65bf1 | 939 | struct perf_evsel *evsel, |
743eb868 | 940 | struct machine *machine) |
ba77c9e1 | 941 | { |
b91fc39f | 942 | int err = 0; |
ef89325f | 943 | struct thread *thread = machine__findnew_thread(machine, sample->pid, |
13ce34df | 944 | sample->tid); |
ba77c9e1 | 945 | |
ba77c9e1 LZ |
946 | if (thread == NULL) { |
947 | pr_debug("problem processing %d event, skipping it.\n", | |
948 | event->header.type); | |
949 | return -1; | |
950 | } | |
951 | ||
2a865bd8 DA |
952 | if (perf_kmem__skip_sample(sample)) |
953 | return 0; | |
954 | ||
b9c5143a | 955 | dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid); |
ba77c9e1 | 956 | |
744a9719 ACM |
957 | if (evsel->handler != NULL) { |
958 | tracepoint_handler f = evsel->handler; | |
b91fc39f | 959 | err = f(evsel, sample); |
0f7d2f1b ACM |
960 | } |
961 | ||
b91fc39f ACM |
962 | thread__put(thread); |
963 | ||
964 | return err; | |
ba77c9e1 LZ |
965 | } |
966 | ||
fcf65bf1 ACM |
967 | static struct perf_tool perf_kmem = { |
968 | .sample = process_sample_event, | |
969 | .comm = perf_event__process_comm, | |
64c40908 NK |
970 | .mmap = perf_event__process_mmap, |
971 | .mmap2 = perf_event__process_mmap2, | |
f3b3614a | 972 | .namespaces = perf_event__process_namespaces, |
0a8cb85c | 973 | .ordered_events = true, |
ba77c9e1 LZ |
974 | }; |
975 | ||
ba77c9e1 LZ |
976 | static double fragmentation(unsigned long n_req, unsigned long n_alloc) |
977 | { | |
978 | if (n_alloc == 0) | |
979 | return 0.0; | |
980 | else | |
981 | return 100.0 - (100.0 * n_req / n_alloc); | |
982 | } | |
983 | ||
0d68bc92 NK |
984 | static void __print_slab_result(struct rb_root *root, |
985 | struct perf_session *session, | |
986 | int n_lines, int is_caller) | |
ba77c9e1 LZ |
987 | { |
988 | struct rb_node *next; | |
34ba5122 | 989 | struct machine *machine = &session->machines.host; |
ba77c9e1 | 990 | |
65f46e02 | 991 | printf("%.105s\n", graph_dotted_line); |
079d3f65 | 992 | printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr"); |
47103277 | 993 | printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n"); |
65f46e02 | 994 | printf("%.105s\n", graph_dotted_line); |
ba77c9e1 LZ |
995 | |
996 | next = rb_first(root); | |
997 | ||
998 | while (next && n_lines--) { | |
1b145ae5 ACM |
999 | struct alloc_stat *data = rb_entry(next, struct alloc_stat, |
1000 | node); | |
1001 | struct symbol *sym = NULL; | |
71cf8b8f | 1002 | struct map *map; |
079d3f65 | 1003 | char buf[BUFSIZ]; |
1b145ae5 ACM |
1004 | u64 addr; |
1005 | ||
1006 | if (is_caller) { | |
1007 | addr = data->call_site; | |
7707b6b6 | 1008 | if (!raw_ip) |
be39db9f | 1009 | sym = machine__find_kernel_function(machine, addr, &map); |
1b145ae5 ACM |
1010 | } else |
1011 | addr = data->ptr; | |
1012 | ||
1013 | if (sym != NULL) | |
9486aa38 | 1014 | snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name, |
71cf8b8f | 1015 | addr - map->unmap_ip(map, sym->start)); |
1b145ae5 | 1016 | else |
9486aa38 | 1017 | snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr); |
079d3f65 | 1018 | printf(" %-34s |", buf); |
ba77c9e1 | 1019 | |
65f46e02 | 1020 | printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n", |
079d3f65 | 1021 | (unsigned long long)data->bytes_alloc, |
ba77c9e1 LZ |
1022 | (unsigned long)data->bytes_alloc / data->hit, |
1023 | (unsigned long long)data->bytes_req, | |
1024 | (unsigned long)data->bytes_req / data->hit, | |
1025 | (unsigned long)data->hit, | |
079d3f65 | 1026 | (unsigned long)data->pingpong, |
ba77c9e1 LZ |
1027 | fragmentation(data->bytes_req, data->bytes_alloc)); |
1028 | ||
1029 | next = rb_next(next); | |
1030 | } | |
1031 | ||
1032 | if (n_lines == -1) | |
65f46e02 | 1033 | printf(" ... | ... | ... | ... | ... | ... \n"); |
ba77c9e1 | 1034 | |
65f46e02 | 1035 | printf("%.105s\n", graph_dotted_line); |
ba77c9e1 LZ |
1036 | } |
1037 | ||
0d68bc92 NK |
1038 | static const char * const migrate_type_str[] = { |
1039 | "UNMOVABL", | |
1040 | "RECLAIM", | |
1041 | "MOVABLE", | |
1042 | "RESERVED", | |
1043 | "CMA/ISLT", | |
1044 | "UNKNOWN", | |
1045 | }; | |
1046 | ||
c9758cc4 | 1047 | static void __print_page_alloc_result(struct perf_session *session, int n_lines) |
0d68bc92 | 1048 | { |
c9758cc4 NK |
1049 | struct rb_node *next = rb_first(&page_alloc_sorted); |
1050 | struct machine *machine = &session->machines.host; | |
0d68bc92 | 1051 | const char *format; |
0e111156 | 1052 | int gfp_len = max(strlen("GFP flags"), max_gfp_len); |
0d68bc92 | 1053 | |
c9758cc4 | 1054 | printf("\n%.105s\n", graph_dotted_line); |
0e111156 NK |
1055 | printf(" %-16s | %5s alloc (KB) | Hits | Order | Mig.type | %-*s | Callsite\n", |
1056 | use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total", | |
1057 | gfp_len, "GFP flags"); | |
c9758cc4 | 1058 | printf("%.105s\n", graph_dotted_line); |
0d68bc92 NK |
1059 | |
1060 | if (use_pfn) | |
0e111156 | 1061 | format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n"; |
0d68bc92 | 1062 | else |
0e111156 | 1063 | format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n"; |
0d68bc92 NK |
1064 | |
1065 | while (next && n_lines--) { | |
1066 | struct page_stat *data; | |
c9758cc4 NK |
1067 | struct symbol *sym; |
1068 | struct map *map; | |
1069 | char buf[32]; | |
1070 | char *caller = buf; | |
0d68bc92 NK |
1071 | |
1072 | data = rb_entry(next, struct page_stat, node); | |
be39db9f | 1073 | sym = machine__find_kernel_function(machine, data->callsite, &map); |
a7c3899c | 1074 | if (sym) |
c9758cc4 NK |
1075 | caller = sym->name; |
1076 | else | |
1077 | scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite); | |
0d68bc92 NK |
1078 | |
1079 | printf(format, (unsigned long long)data->page, | |
1080 | (unsigned long long)data->alloc_bytes / 1024, | |
1081 | data->nr_alloc, data->order, | |
1082 | migrate_type_str[data->migrate_type], | |
0e111156 | 1083 | gfp_len, compact_gfp_string(data->gfp_flags), caller); |
0d68bc92 NK |
1084 | |
1085 | next = rb_next(next); | |
1086 | } | |
1087 | ||
0e111156 NK |
1088 | if (n_lines == -1) { |
1089 | printf(" ... | ... | ... | ... | ... | %-*s | ...\n", | |
1090 | gfp_len, "..."); | |
1091 | } | |
0d68bc92 | 1092 | |
c9758cc4 NK |
1093 | printf("%.105s\n", graph_dotted_line); |
1094 | } | |
1095 | ||
1096 | static void __print_page_caller_result(struct perf_session *session, int n_lines) | |
1097 | { | |
1098 | struct rb_node *next = rb_first(&page_caller_sorted); | |
1099 | struct machine *machine = &session->machines.host; | |
0e111156 | 1100 | int gfp_len = max(strlen("GFP flags"), max_gfp_len); |
c9758cc4 NK |
1101 | |
1102 | printf("\n%.105s\n", graph_dotted_line); | |
0e111156 NK |
1103 | printf(" %5s alloc (KB) | Hits | Order | Mig.type | %-*s | Callsite\n", |
1104 | live_page ? "Live" : "Total", gfp_len, "GFP flags"); | |
c9758cc4 NK |
1105 | printf("%.105s\n", graph_dotted_line); |
1106 | ||
1107 | while (next && n_lines--) { | |
1108 | struct page_stat *data; | |
1109 | struct symbol *sym; | |
1110 | struct map *map; | |
1111 | char buf[32]; | |
1112 | char *caller = buf; | |
1113 | ||
1114 | data = rb_entry(next, struct page_stat, node); | |
be39db9f | 1115 | sym = machine__find_kernel_function(machine, data->callsite, &map); |
a7c3899c | 1116 | if (sym) |
c9758cc4 NK |
1117 | caller = sym->name; |
1118 | else | |
1119 | scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite); | |
1120 | ||
0e111156 | 1121 | printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n", |
c9758cc4 NK |
1122 | (unsigned long long)data->alloc_bytes / 1024, |
1123 | data->nr_alloc, data->order, | |
1124 | migrate_type_str[data->migrate_type], | |
0e111156 | 1125 | gfp_len, compact_gfp_string(data->gfp_flags), caller); |
c9758cc4 NK |
1126 | |
1127 | next = rb_next(next); | |
1128 | } | |
1129 | ||
0e111156 NK |
1130 | if (n_lines == -1) { |
1131 | printf(" ... | ... | ... | ... | %-*s | ...\n", | |
1132 | gfp_len, "..."); | |
1133 | } | |
c9758cc4 NK |
1134 | |
1135 | printf("%.105s\n", graph_dotted_line); | |
0d68bc92 NK |
1136 | } |
1137 | ||
0e111156 NK |
1138 | static void print_gfp_flags(void) |
1139 | { | |
1140 | int i; | |
1141 | ||
1142 | printf("#\n"); | |
1143 | printf("# GFP flags\n"); | |
1144 | printf("# ---------\n"); | |
1145 | for (i = 0; i < nr_gfps; i++) { | |
1146 | printf("# %08x: %*s: %s\n", gfps[i].flags, | |
1147 | (int) max_gfp_len, gfps[i].compact_str, | |
1148 | gfps[i].human_readable); | |
1149 | } | |
1150 | } | |
1151 | ||
0d68bc92 | 1152 | static void print_slab_summary(void) |
ba77c9e1 | 1153 | { |
0d68bc92 NK |
1154 | printf("\nSUMMARY (SLAB allocator)"); |
1155 | printf("\n========================\n"); | |
77cfe388 NK |
1156 | printf("Total bytes requested: %'lu\n", total_requested); |
1157 | printf("Total bytes allocated: %'lu\n", total_allocated); | |
aa58e9af DA |
1158 | printf("Total bytes freed: %'lu\n", total_freed); |
1159 | if (total_allocated > total_freed) { | |
1160 | printf("Net total bytes allocated: %'lu\n", | |
1161 | total_allocated - total_freed); | |
1162 | } | |
77cfe388 | 1163 | printf("Total bytes wasted on internal fragmentation: %'lu\n", |
ba77c9e1 LZ |
1164 | total_allocated - total_requested); |
1165 | printf("Internal fragmentation: %f%%\n", | |
1166 | fragmentation(total_requested, total_allocated)); | |
77cfe388 | 1167 | printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs); |
ba77c9e1 LZ |
1168 | } |
1169 | ||
0d68bc92 NK |
1170 | static void print_page_summary(void) |
1171 | { | |
1172 | int o, m; | |
1173 | u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch; | |
1174 | u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes; | |
1175 | ||
1176 | printf("\nSUMMARY (page allocator)"); | |
1177 | printf("\n========================\n"); | |
1178 | printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total allocation requests", | |
1179 | nr_page_allocs, total_page_alloc_bytes / 1024); | |
1180 | printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total free requests", | |
1181 | nr_page_frees, total_page_free_bytes / 1024); | |
1182 | printf("\n"); | |
1183 | ||
6145c259 | 1184 | printf("%-30s: %'16"PRIu64" [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests", |
0d68bc92 | 1185 | nr_alloc_freed, (total_alloc_freed_bytes) / 1024); |
6145c259 | 1186 | printf("%-30s: %'16"PRIu64" [ %'16"PRIu64" KB ]\n", "Total alloc-only requests", |
0d68bc92 NK |
1187 | nr_page_allocs - nr_alloc_freed, |
1188 | (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024); | |
1189 | printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total free-only requests", | |
1190 | nr_page_nomatch, total_page_nomatch_bytes / 1024); | |
1191 | printf("\n"); | |
1192 | ||
1193 | printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total allocation failures", | |
1194 | nr_page_fails, total_page_fail_bytes / 1024); | |
1195 | printf("\n"); | |
1196 | ||
1197 | printf("%5s %12s %12s %12s %12s %12s\n", "Order", "Unmovable", | |
1198 | "Reclaimable", "Movable", "Reserved", "CMA/Isolated"); | |
1199 | printf("%.5s %.12s %.12s %.12s %.12s %.12s\n", graph_dotted_line, | |
1200 | graph_dotted_line, graph_dotted_line, graph_dotted_line, | |
1201 | graph_dotted_line, graph_dotted_line); | |
1202 | ||
1203 | for (o = 0; o < MAX_PAGE_ORDER; o++) { | |
1204 | printf("%5d", o); | |
1205 | for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) { | |
1206 | if (order_stats[o][m]) | |
1207 | printf(" %'12d", order_stats[o][m]); | |
1208 | else | |
1209 | printf(" %12c", '.'); | |
1210 | } | |
1211 | printf("\n"); | |
1212 | } | |
1213 | } | |
1214 | ||
1215 | static void print_slab_result(struct perf_session *session) | |
ba77c9e1 LZ |
1216 | { |
1217 | if (caller_flag) | |
0d68bc92 NK |
1218 | __print_slab_result(&root_caller_sorted, session, caller_lines, 1); |
1219 | if (alloc_flag) | |
1220 | __print_slab_result(&root_alloc_sorted, session, alloc_lines, 0); | |
1221 | print_slab_summary(); | |
1222 | } | |
1223 | ||
1224 | static void print_page_result(struct perf_session *session) | |
1225 | { | |
0e111156 NK |
1226 | if (caller_flag || alloc_flag) |
1227 | print_gfp_flags(); | |
c9758cc4 NK |
1228 | if (caller_flag) |
1229 | __print_page_caller_result(session, caller_lines); | |
ba77c9e1 | 1230 | if (alloc_flag) |
c9758cc4 | 1231 | __print_page_alloc_result(session, alloc_lines); |
0d68bc92 NK |
1232 | print_page_summary(); |
1233 | } | |
1234 | ||
1235 | static void print_result(struct perf_session *session) | |
1236 | { | |
1237 | if (kmem_slab) | |
1238 | print_slab_result(session); | |
1239 | if (kmem_page) | |
1240 | print_page_result(session); | |
ba77c9e1 LZ |
1241 | } |
1242 | ||
fb4f313d NK |
1243 | static LIST_HEAD(slab_caller_sort); |
1244 | static LIST_HEAD(slab_alloc_sort); | |
1245 | static LIST_HEAD(page_caller_sort); | |
1246 | static LIST_HEAD(page_alloc_sort); | |
29b3e152 | 1247 | |
0d68bc92 NK |
1248 | static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data, |
1249 | struct list_head *sort_list) | |
ba77c9e1 LZ |
1250 | { |
1251 | struct rb_node **new = &(root->rb_node); | |
1252 | struct rb_node *parent = NULL; | |
29b3e152 | 1253 | struct sort_dimension *sort; |
ba77c9e1 LZ |
1254 | |
1255 | while (*new) { | |
1256 | struct alloc_stat *this; | |
29b3e152 | 1257 | int cmp = 0; |
ba77c9e1 LZ |
1258 | |
1259 | this = rb_entry(*new, struct alloc_stat, node); | |
1260 | parent = *new; | |
1261 | ||
29b3e152 LZ |
1262 | list_for_each_entry(sort, sort_list, list) { |
1263 | cmp = sort->cmp(data, this); | |
1264 | if (cmp) | |
1265 | break; | |
1266 | } | |
ba77c9e1 LZ |
1267 | |
1268 | if (cmp > 0) | |
1269 | new = &((*new)->rb_left); | |
1270 | else | |
1271 | new = &((*new)->rb_right); | |
1272 | } | |
1273 | ||
1274 | rb_link_node(&data->node, parent, new); | |
1275 | rb_insert_color(&data->node, root); | |
1276 | } | |
1277 | ||
0d68bc92 NK |
1278 | static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted, |
1279 | struct list_head *sort_list) | |
ba77c9e1 LZ |
1280 | { |
1281 | struct rb_node *node; | |
1282 | struct alloc_stat *data; | |
1283 | ||
1284 | for (;;) { | |
1285 | node = rb_first(root); | |
1286 | if (!node) | |
1287 | break; | |
1288 | ||
1289 | rb_erase(node, root); | |
1290 | data = rb_entry(node, struct alloc_stat, node); | |
0d68bc92 NK |
1291 | sort_slab_insert(root_sorted, data, sort_list); |
1292 | } | |
1293 | } | |
1294 | ||
fb4f313d NK |
1295 | static void sort_page_insert(struct rb_root *root, struct page_stat *data, |
1296 | struct list_head *sort_list) | |
0d68bc92 NK |
1297 | { |
1298 | struct rb_node **new = &root->rb_node; | |
1299 | struct rb_node *parent = NULL; | |
fb4f313d | 1300 | struct sort_dimension *sort; |
0d68bc92 NK |
1301 | |
1302 | while (*new) { | |
1303 | struct page_stat *this; | |
1304 | int cmp = 0; | |
1305 | ||
1306 | this = rb_entry(*new, struct page_stat, node); | |
1307 | parent = *new; | |
1308 | ||
fb4f313d NK |
1309 | list_for_each_entry(sort, sort_list, list) { |
1310 | cmp = sort->cmp(data, this); | |
1311 | if (cmp) | |
1312 | break; | |
1313 | } | |
0d68bc92 NK |
1314 | |
1315 | if (cmp > 0) | |
1316 | new = &parent->rb_left; | |
1317 | else | |
1318 | new = &parent->rb_right; | |
1319 | } | |
1320 | ||
1321 | rb_link_node(&data->node, parent, new); | |
1322 | rb_insert_color(&data->node, root); | |
1323 | } | |
1324 | ||
fb4f313d NK |
1325 | static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted, |
1326 | struct list_head *sort_list) | |
0d68bc92 NK |
1327 | { |
1328 | struct rb_node *node; | |
1329 | struct page_stat *data; | |
1330 | ||
1331 | for (;;) { | |
1332 | node = rb_first(root); | |
1333 | if (!node) | |
1334 | break; | |
1335 | ||
1336 | rb_erase(node, root); | |
1337 | data = rb_entry(node, struct page_stat, node); | |
fb4f313d | 1338 | sort_page_insert(root_sorted, data, sort_list); |
ba77c9e1 LZ |
1339 | } |
1340 | } | |
1341 | ||
1342 | static void sort_result(void) | |
1343 | { | |
0d68bc92 NK |
1344 | if (kmem_slab) { |
1345 | __sort_slab_result(&root_alloc_stat, &root_alloc_sorted, | |
fb4f313d | 1346 | &slab_alloc_sort); |
0d68bc92 | 1347 | __sort_slab_result(&root_caller_stat, &root_caller_sorted, |
fb4f313d | 1348 | &slab_caller_sort); |
0d68bc92 NK |
1349 | } |
1350 | if (kmem_page) { | |
2a7ef02c NK |
1351 | if (live_page) |
1352 | __sort_page_result(&page_live_tree, &page_alloc_sorted, | |
1353 | &page_alloc_sort); | |
1354 | else | |
1355 | __sort_page_result(&page_alloc_tree, &page_alloc_sorted, | |
1356 | &page_alloc_sort); | |
1357 | ||
fb4f313d NK |
1358 | __sort_page_result(&page_caller_tree, &page_caller_sorted, |
1359 | &page_caller_sort); | |
0d68bc92 | 1360 | } |
ba77c9e1 LZ |
1361 | } |
1362 | ||
2b2b2c68 | 1363 | static int __cmd_kmem(struct perf_session *session) |
ba77c9e1 | 1364 | { |
d549c769 | 1365 | int err = -EINVAL; |
0d68bc92 | 1366 | struct perf_evsel *evsel; |
0f7d2f1b | 1367 | const struct perf_evsel_str_handler kmem_tracepoints[] = { |
0d68bc92 | 1368 | /* slab allocator */ |
0f7d2f1b ACM |
1369 | { "kmem:kmalloc", perf_evsel__process_alloc_event, }, |
1370 | { "kmem:kmem_cache_alloc", perf_evsel__process_alloc_event, }, | |
1371 | { "kmem:kmalloc_node", perf_evsel__process_alloc_node_event, }, | |
1372 | { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, }, | |
1373 | { "kmem:kfree", perf_evsel__process_free_event, }, | |
1374 | { "kmem:kmem_cache_free", perf_evsel__process_free_event, }, | |
0d68bc92 NK |
1375 | /* page allocator */ |
1376 | { "kmem:mm_page_alloc", perf_evsel__process_page_alloc_event, }, | |
1377 | { "kmem:mm_page_free", perf_evsel__process_page_free_event, }, | |
0f7d2f1b | 1378 | }; |
4aa65636 | 1379 | |
d549c769 | 1380 | if (!perf_session__has_traces(session, "kmem record")) |
2b2b2c68 | 1381 | goto out; |
d549c769 | 1382 | |
0f7d2f1b ACM |
1383 | if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) { |
1384 | pr_err("Initializing perf session tracepoint handlers failed\n"); | |
2b2b2c68 | 1385 | goto out; |
0f7d2f1b ACM |
1386 | } |
1387 | ||
e5cadb93 | 1388 | evlist__for_each_entry(session->evlist, evsel) { |
0d68bc92 NK |
1389 | if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") && |
1390 | perf_evsel__field(evsel, "pfn")) { | |
1391 | use_pfn = true; | |
1392 | break; | |
1393 | } | |
1394 | } | |
1395 | ||
ba77c9e1 | 1396 | setup_pager(); |
b7b61cbe | 1397 | err = perf_session__process_events(session); |
0d68bc92 NK |
1398 | if (err != 0) { |
1399 | pr_err("error during process events: %d\n", err); | |
2b2b2c68 | 1400 | goto out; |
0d68bc92 | 1401 | } |
ba77c9e1 | 1402 | sort_result(); |
4aa65636 | 1403 | print_result(session); |
2b2b2c68 | 1404 | out: |
4aa65636 | 1405 | return err; |
ba77c9e1 LZ |
1406 | } |
1407 | ||
fb4f313d NK |
1408 | /* slab sort keys */ |
1409 | static int ptr_cmp(void *a, void *b) | |
ba77c9e1 | 1410 | { |
fb4f313d NK |
1411 | struct alloc_stat *l = a; |
1412 | struct alloc_stat *r = b; | |
1413 | ||
ba77c9e1 LZ |
1414 | if (l->ptr < r->ptr) |
1415 | return -1; | |
1416 | else if (l->ptr > r->ptr) | |
1417 | return 1; | |
1418 | return 0; | |
1419 | } | |
1420 | ||
29b3e152 LZ |
1421 | static struct sort_dimension ptr_sort_dimension = { |
1422 | .name = "ptr", | |
1423 | .cmp = ptr_cmp, | |
1424 | }; | |
1425 | ||
fb4f313d | 1426 | static int slab_callsite_cmp(void *a, void *b) |
ba77c9e1 | 1427 | { |
fb4f313d NK |
1428 | struct alloc_stat *l = a; |
1429 | struct alloc_stat *r = b; | |
1430 | ||
ba77c9e1 LZ |
1431 | if (l->call_site < r->call_site) |
1432 | return -1; | |
1433 | else if (l->call_site > r->call_site) | |
1434 | return 1; | |
1435 | return 0; | |
1436 | } | |
1437 | ||
29b3e152 LZ |
1438 | static struct sort_dimension callsite_sort_dimension = { |
1439 | .name = "callsite", | |
fb4f313d | 1440 | .cmp = slab_callsite_cmp, |
29b3e152 LZ |
1441 | }; |
1442 | ||
fb4f313d | 1443 | static int hit_cmp(void *a, void *b) |
f3ced7cd | 1444 | { |
fb4f313d NK |
1445 | struct alloc_stat *l = a; |
1446 | struct alloc_stat *r = b; | |
1447 | ||
f3ced7cd PE |
1448 | if (l->hit < r->hit) |
1449 | return -1; | |
1450 | else if (l->hit > r->hit) | |
1451 | return 1; | |
1452 | return 0; | |
1453 | } | |
1454 | ||
29b3e152 LZ |
1455 | static struct sort_dimension hit_sort_dimension = { |
1456 | .name = "hit", | |
1457 | .cmp = hit_cmp, | |
1458 | }; | |
1459 | ||
fb4f313d | 1460 | static int bytes_cmp(void *a, void *b) |
ba77c9e1 | 1461 | { |
fb4f313d NK |
1462 | struct alloc_stat *l = a; |
1463 | struct alloc_stat *r = b; | |
1464 | ||
ba77c9e1 LZ |
1465 | if (l->bytes_alloc < r->bytes_alloc) |
1466 | return -1; | |
1467 | else if (l->bytes_alloc > r->bytes_alloc) | |
1468 | return 1; | |
1469 | return 0; | |
1470 | } | |
1471 | ||
29b3e152 LZ |
1472 | static struct sort_dimension bytes_sort_dimension = { |
1473 | .name = "bytes", | |
1474 | .cmp = bytes_cmp, | |
1475 | }; | |
1476 | ||
fb4f313d | 1477 | static int frag_cmp(void *a, void *b) |
f3ced7cd PE |
1478 | { |
1479 | double x, y; | |
fb4f313d NK |
1480 | struct alloc_stat *l = a; |
1481 | struct alloc_stat *r = b; | |
f3ced7cd PE |
1482 | |
1483 | x = fragmentation(l->bytes_req, l->bytes_alloc); | |
1484 | y = fragmentation(r->bytes_req, r->bytes_alloc); | |
1485 | ||
1486 | if (x < y) | |
1487 | return -1; | |
1488 | else if (x > y) | |
1489 | return 1; | |
1490 | return 0; | |
1491 | } | |
1492 | ||
29b3e152 LZ |
1493 | static struct sort_dimension frag_sort_dimension = { |
1494 | .name = "frag", | |
1495 | .cmp = frag_cmp, | |
1496 | }; | |
1497 | ||
fb4f313d | 1498 | static int pingpong_cmp(void *a, void *b) |
079d3f65 | 1499 | { |
fb4f313d NK |
1500 | struct alloc_stat *l = a; |
1501 | struct alloc_stat *r = b; | |
1502 | ||
079d3f65 LZ |
1503 | if (l->pingpong < r->pingpong) |
1504 | return -1; | |
1505 | else if (l->pingpong > r->pingpong) | |
1506 | return 1; | |
1507 | return 0; | |
1508 | } | |
1509 | ||
1510 | static struct sort_dimension pingpong_sort_dimension = { | |
1511 | .name = "pingpong", | |
1512 | .cmp = pingpong_cmp, | |
1513 | }; | |
1514 | ||
fb4f313d NK |
1515 | /* page sort keys */ |
1516 | static int page_cmp(void *a, void *b) | |
1517 | { | |
1518 | struct page_stat *l = a; | |
1519 | struct page_stat *r = b; | |
1520 | ||
1521 | if (l->page < r->page) | |
1522 | return -1; | |
1523 | else if (l->page > r->page) | |
1524 | return 1; | |
1525 | return 0; | |
1526 | } | |
1527 | ||
1528 | static struct sort_dimension page_sort_dimension = { | |
1529 | .name = "page", | |
1530 | .cmp = page_cmp, | |
1531 | }; | |
1532 | ||
1533 | static int page_callsite_cmp(void *a, void *b) | |
1534 | { | |
1535 | struct page_stat *l = a; | |
1536 | struct page_stat *r = b; | |
1537 | ||
1538 | if (l->callsite < r->callsite) | |
1539 | return -1; | |
1540 | else if (l->callsite > r->callsite) | |
1541 | return 1; | |
1542 | return 0; | |
1543 | } | |
1544 | ||
1545 | static struct sort_dimension page_callsite_sort_dimension = { | |
1546 | .name = "callsite", | |
1547 | .cmp = page_callsite_cmp, | |
1548 | }; | |
1549 | ||
1550 | static int page_hit_cmp(void *a, void *b) | |
1551 | { | |
1552 | struct page_stat *l = a; | |
1553 | struct page_stat *r = b; | |
1554 | ||
1555 | if (l->nr_alloc < r->nr_alloc) | |
1556 | return -1; | |
1557 | else if (l->nr_alloc > r->nr_alloc) | |
1558 | return 1; | |
1559 | return 0; | |
1560 | } | |
1561 | ||
1562 | static struct sort_dimension page_hit_sort_dimension = { | |
1563 | .name = "hit", | |
1564 | .cmp = page_hit_cmp, | |
1565 | }; | |
1566 | ||
1567 | static int page_bytes_cmp(void *a, void *b) | |
1568 | { | |
1569 | struct page_stat *l = a; | |
1570 | struct page_stat *r = b; | |
1571 | ||
1572 | if (l->alloc_bytes < r->alloc_bytes) | |
1573 | return -1; | |
1574 | else if (l->alloc_bytes > r->alloc_bytes) | |
1575 | return 1; | |
1576 | return 0; | |
1577 | } | |
1578 | ||
1579 | static struct sort_dimension page_bytes_sort_dimension = { | |
1580 | .name = "bytes", | |
1581 | .cmp = page_bytes_cmp, | |
1582 | }; | |
1583 | ||
1584 | static int page_order_cmp(void *a, void *b) | |
1585 | { | |
1586 | struct page_stat *l = a; | |
1587 | struct page_stat *r = b; | |
1588 | ||
1589 | if (l->order < r->order) | |
1590 | return -1; | |
1591 | else if (l->order > r->order) | |
1592 | return 1; | |
1593 | return 0; | |
1594 | } | |
1595 | ||
1596 | static struct sort_dimension page_order_sort_dimension = { | |
1597 | .name = "order", | |
1598 | .cmp = page_order_cmp, | |
1599 | }; | |
1600 | ||
1601 | static int migrate_type_cmp(void *a, void *b) | |
1602 | { | |
1603 | struct page_stat *l = a; | |
1604 | struct page_stat *r = b; | |
1605 | ||
1606 | /* for internal use to find free'd page */ | |
1607 | if (l->migrate_type == -1U) | |
1608 | return 0; | |
1609 | ||
1610 | if (l->migrate_type < r->migrate_type) | |
1611 | return -1; | |
1612 | else if (l->migrate_type > r->migrate_type) | |
1613 | return 1; | |
1614 | return 0; | |
1615 | } | |
1616 | ||
1617 | static struct sort_dimension migrate_type_sort_dimension = { | |
1618 | .name = "migtype", | |
1619 | .cmp = migrate_type_cmp, | |
1620 | }; | |
1621 | ||
1622 | static int gfp_flags_cmp(void *a, void *b) | |
1623 | { | |
1624 | struct page_stat *l = a; | |
1625 | struct page_stat *r = b; | |
1626 | ||
1627 | /* for internal use to find free'd page */ | |
1628 | if (l->gfp_flags == -1U) | |
1629 | return 0; | |
1630 | ||
1631 | if (l->gfp_flags < r->gfp_flags) | |
1632 | return -1; | |
1633 | else if (l->gfp_flags > r->gfp_flags) | |
1634 | return 1; | |
1635 | return 0; | |
1636 | } | |
1637 | ||
1638 | static struct sort_dimension gfp_flags_sort_dimension = { | |
1639 | .name = "gfp", | |
1640 | .cmp = gfp_flags_cmp, | |
1641 | }; | |
1642 | ||
1643 | static struct sort_dimension *slab_sorts[] = { | |
29b3e152 LZ |
1644 | &ptr_sort_dimension, |
1645 | &callsite_sort_dimension, | |
1646 | &hit_sort_dimension, | |
1647 | &bytes_sort_dimension, | |
1648 | &frag_sort_dimension, | |
079d3f65 | 1649 | &pingpong_sort_dimension, |
29b3e152 LZ |
1650 | }; |
1651 | ||
fb4f313d NK |
1652 | static struct sort_dimension *page_sorts[] = { |
1653 | &page_sort_dimension, | |
1654 | &page_callsite_sort_dimension, | |
1655 | &page_hit_sort_dimension, | |
1656 | &page_bytes_sort_dimension, | |
1657 | &page_order_sort_dimension, | |
1658 | &migrate_type_sort_dimension, | |
1659 | &gfp_flags_sort_dimension, | |
1660 | }; | |
29b3e152 | 1661 | |
fb4f313d | 1662 | static int slab_sort_dimension__add(const char *tok, struct list_head *list) |
29b3e152 LZ |
1663 | { |
1664 | struct sort_dimension *sort; | |
1665 | int i; | |
1666 | ||
fb4f313d NK |
1667 | for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) { |
1668 | if (!strcmp(slab_sorts[i]->name, tok)) { | |
1669 | sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i])); | |
2814eb05 | 1670 | if (!sort) { |
8d9233f2 | 1671 | pr_err("%s: memdup failed\n", __func__); |
2814eb05 ACM |
1672 | return -1; |
1673 | } | |
29b3e152 LZ |
1674 | list_add_tail(&sort->list, list); |
1675 | return 0; | |
1676 | } | |
1677 | } | |
1678 | ||
1679 | return -1; | |
1680 | } | |
1681 | ||
fb4f313d NK |
1682 | static int page_sort_dimension__add(const char *tok, struct list_head *list) |
1683 | { | |
1684 | struct sort_dimension *sort; | |
1685 | int i; | |
1686 | ||
1687 | for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) { | |
1688 | if (!strcmp(page_sorts[i]->name, tok)) { | |
1689 | sort = memdup(page_sorts[i], sizeof(*page_sorts[i])); | |
1690 | if (!sort) { | |
1691 | pr_err("%s: memdup failed\n", __func__); | |
1692 | return -1; | |
1693 | } | |
1694 | list_add_tail(&sort->list, list); | |
1695 | return 0; | |
1696 | } | |
1697 | } | |
1698 | ||
1699 | return -1; | |
1700 | } | |
1701 | ||
1702 | static int setup_slab_sorting(struct list_head *sort_list, const char *arg) | |
29b3e152 LZ |
1703 | { |
1704 | char *tok; | |
1705 | char *str = strdup(arg); | |
405f8755 | 1706 | char *pos = str; |
29b3e152 | 1707 | |
2814eb05 ACM |
1708 | if (!str) { |
1709 | pr_err("%s: strdup failed\n", __func__); | |
1710 | return -1; | |
1711 | } | |
29b3e152 LZ |
1712 | |
1713 | while (true) { | |
405f8755 | 1714 | tok = strsep(&pos, ","); |
29b3e152 LZ |
1715 | if (!tok) |
1716 | break; | |
fb4f313d | 1717 | if (slab_sort_dimension__add(tok, sort_list) < 0) { |
62d94b00 | 1718 | pr_err("Unknown slab --sort key: '%s'", tok); |
fb4f313d NK |
1719 | free(str); |
1720 | return -1; | |
1721 | } | |
1722 | } | |
1723 | ||
1724 | free(str); | |
1725 | return 0; | |
1726 | } | |
1727 | ||
1728 | static int setup_page_sorting(struct list_head *sort_list, const char *arg) | |
1729 | { | |
1730 | char *tok; | |
1731 | char *str = strdup(arg); | |
1732 | char *pos = str; | |
1733 | ||
1734 | if (!str) { | |
1735 | pr_err("%s: strdup failed\n", __func__); | |
1736 | return -1; | |
1737 | } | |
1738 | ||
1739 | while (true) { | |
1740 | tok = strsep(&pos, ","); | |
1741 | if (!tok) | |
1742 | break; | |
1743 | if (page_sort_dimension__add(tok, sort_list) < 0) { | |
62d94b00 | 1744 | pr_err("Unknown page --sort key: '%s'", tok); |
1b22859d | 1745 | free(str); |
29b3e152 LZ |
1746 | return -1; |
1747 | } | |
1748 | } | |
1749 | ||
1750 | free(str); | |
1751 | return 0; | |
1752 | } | |
1753 | ||
1d037ca1 IT |
1754 | static int parse_sort_opt(const struct option *opt __maybe_unused, |
1755 | const char *arg, int unset __maybe_unused) | |
ba77c9e1 | 1756 | { |
ba77c9e1 LZ |
1757 | if (!arg) |
1758 | return -1; | |
1759 | ||
0c160d49 NK |
1760 | if (kmem_page > kmem_slab || |
1761 | (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) { | |
fb4f313d NK |
1762 | if (caller_flag > alloc_flag) |
1763 | return setup_page_sorting(&page_caller_sort, arg); | |
1764 | else | |
1765 | return setup_page_sorting(&page_alloc_sort, arg); | |
1766 | } else { | |
1767 | if (caller_flag > alloc_flag) | |
1768 | return setup_slab_sorting(&slab_caller_sort, arg); | |
1769 | else | |
1770 | return setup_slab_sorting(&slab_alloc_sort, arg); | |
1771 | } | |
ba77c9e1 LZ |
1772 | |
1773 | return 0; | |
1774 | } | |
1775 | ||
1d037ca1 IT |
1776 | static int parse_caller_opt(const struct option *opt __maybe_unused, |
1777 | const char *arg __maybe_unused, | |
1778 | int unset __maybe_unused) | |
ba77c9e1 | 1779 | { |
90b86a9f LZ |
1780 | caller_flag = (alloc_flag + 1); |
1781 | return 0; | |
1782 | } | |
ba77c9e1 | 1783 | |
1d037ca1 IT |
1784 | static int parse_alloc_opt(const struct option *opt __maybe_unused, |
1785 | const char *arg __maybe_unused, | |
1786 | int unset __maybe_unused) | |
90b86a9f LZ |
1787 | { |
1788 | alloc_flag = (caller_flag + 1); | |
ba77c9e1 LZ |
1789 | return 0; |
1790 | } | |
1791 | ||
0d68bc92 NK |
1792 | static int parse_slab_opt(const struct option *opt __maybe_unused, |
1793 | const char *arg __maybe_unused, | |
1794 | int unset __maybe_unused) | |
1795 | { | |
1796 | kmem_slab = (kmem_page + 1); | |
1797 | return 0; | |
1798 | } | |
1799 | ||
1800 | static int parse_page_opt(const struct option *opt __maybe_unused, | |
1801 | const char *arg __maybe_unused, | |
1802 | int unset __maybe_unused) | |
1803 | { | |
1804 | kmem_page = (kmem_slab + 1); | |
1805 | return 0; | |
1806 | } | |
1807 | ||
1d037ca1 IT |
1808 | static int parse_line_opt(const struct option *opt __maybe_unused, |
1809 | const char *arg, int unset __maybe_unused) | |
ba77c9e1 LZ |
1810 | { |
1811 | int lines; | |
1812 | ||
1813 | if (!arg) | |
1814 | return -1; | |
1815 | ||
1816 | lines = strtoul(arg, NULL, 10); | |
1817 | ||
1818 | if (caller_flag > alloc_flag) | |
1819 | caller_lines = lines; | |
1820 | else | |
1821 | alloc_lines = lines; | |
1822 | ||
1823 | return 0; | |
1824 | } | |
1825 | ||
0433ffbe ACM |
1826 | static int __cmd_record(int argc, const char **argv) |
1827 | { | |
1828 | const char * const record_args[] = { | |
4a4d371a | 1829 | "record", "-a", "-R", "-c", "1", |
0d68bc92 NK |
1830 | }; |
1831 | const char * const slab_events[] = { | |
ba77c9e1 LZ |
1832 | "-e", "kmem:kmalloc", |
1833 | "-e", "kmem:kmalloc_node", | |
1834 | "-e", "kmem:kfree", | |
1835 | "-e", "kmem:kmem_cache_alloc", | |
1836 | "-e", "kmem:kmem_cache_alloc_node", | |
1837 | "-e", "kmem:kmem_cache_free", | |
0433ffbe | 1838 | }; |
0d68bc92 NK |
1839 | const char * const page_events[] = { |
1840 | "-e", "kmem:mm_page_alloc", | |
1841 | "-e", "kmem:mm_page_free", | |
1842 | }; | |
ba77c9e1 LZ |
1843 | unsigned int rec_argc, i, j; |
1844 | const char **rec_argv; | |
1845 | ||
1846 | rec_argc = ARRAY_SIZE(record_args) + argc - 1; | |
0d68bc92 NK |
1847 | if (kmem_slab) |
1848 | rec_argc += ARRAY_SIZE(slab_events); | |
1849 | if (kmem_page) | |
c9758cc4 | 1850 | rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */ |
0d68bc92 | 1851 | |
ba77c9e1 LZ |
1852 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); |
1853 | ||
ce47dc56 CS |
1854 | if (rec_argv == NULL) |
1855 | return -ENOMEM; | |
1856 | ||
ba77c9e1 LZ |
1857 | for (i = 0; i < ARRAY_SIZE(record_args); i++) |
1858 | rec_argv[i] = strdup(record_args[i]); | |
1859 | ||
0d68bc92 NK |
1860 | if (kmem_slab) { |
1861 | for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++) | |
1862 | rec_argv[i] = strdup(slab_events[j]); | |
1863 | } | |
1864 | if (kmem_page) { | |
c9758cc4 NK |
1865 | rec_argv[i++] = strdup("-g"); |
1866 | ||
0d68bc92 NK |
1867 | for (j = 0; j < ARRAY_SIZE(page_events); j++, i++) |
1868 | rec_argv[i] = strdup(page_events[j]); | |
1869 | } | |
1870 | ||
ba77c9e1 LZ |
1871 | for (j = 1; j < (unsigned int)argc; j++, i++) |
1872 | rec_argv[i] = argv[j]; | |
1873 | ||
b0ad8ea6 | 1874 | return cmd_record(i, rec_argv); |
ba77c9e1 LZ |
1875 | } |
1876 | ||
b8cbb349 | 1877 | static int kmem_config(const char *var, const char *value, void *cb __maybe_unused) |
0c160d49 NK |
1878 | { |
1879 | if (!strcmp(var, "kmem.default")) { | |
1880 | if (!strcmp(value, "slab")) | |
1881 | kmem_default = KMEM_SLAB; | |
1882 | else if (!strcmp(value, "page")) | |
1883 | kmem_default = KMEM_PAGE; | |
1884 | else | |
1885 | pr_err("invalid default value ('slab' or 'page' required): %s\n", | |
1886 | value); | |
1887 | return 0; | |
1888 | } | |
1889 | ||
b8cbb349 | 1890 | return 0; |
0c160d49 NK |
1891 | } |
1892 | ||
b0ad8ea6 | 1893 | int cmd_kmem(int argc, const char **argv) |
ba77c9e1 | 1894 | { |
fb4f313d NK |
1895 | const char * const default_slab_sort = "frag,hit,bytes"; |
1896 | const char * const default_page_sort = "bytes,hit"; | |
d1eeb77c | 1897 | struct perf_data_file file = { |
d1eeb77c YS |
1898 | .mode = PERF_DATA_MODE_READ, |
1899 | }; | |
0433ffbe ACM |
1900 | const struct option kmem_options[] = { |
1901 | OPT_STRING('i', "input", &input_name, "file", "input file name"), | |
bd72a33e NK |
1902 | OPT_INCR('v', "verbose", &verbose, |
1903 | "be more verbose (show symbol address, etc)"), | |
0433ffbe ACM |
1904 | OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL, |
1905 | "show per-callsite statistics", parse_caller_opt), | |
1906 | OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL, | |
1907 | "show per-allocation statistics", parse_alloc_opt), | |
1908 | OPT_CALLBACK('s', "sort", NULL, "key[,key2...]", | |
fb4f313d NK |
1909 | "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, " |
1910 | "page, order, migtype, gfp", parse_sort_opt), | |
0433ffbe ACM |
1911 | OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt), |
1912 | OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"), | |
d1eeb77c | 1913 | OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"), |
0d68bc92 NK |
1914 | OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator", |
1915 | parse_slab_opt), | |
1916 | OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator", | |
1917 | parse_page_opt), | |
2a7ef02c | 1918 | OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"), |
2a865bd8 DA |
1919 | OPT_STRING(0, "time", &time_str, "str", |
1920 | "Time span of interest (start,stop)"), | |
0433ffbe ACM |
1921 | OPT_END() |
1922 | }; | |
3bca2354 RR |
1923 | const char *const kmem_subcommands[] = { "record", "stat", NULL }; |
1924 | const char *kmem_usage[] = { | |
1925 | NULL, | |
0433ffbe ACM |
1926 | NULL |
1927 | }; | |
2b2b2c68 | 1928 | struct perf_session *session; |
a923e2c4 | 1929 | const char errmsg[] = "No %s allocation events found. Have you run 'perf kmem record --%s'?\n"; |
ecc4c561 ACM |
1930 | int ret = perf_config(kmem_config, NULL); |
1931 | ||
1932 | if (ret) | |
1933 | return ret; | |
2b2b2c68 | 1934 | |
3bca2354 RR |
1935 | argc = parse_options_subcommand(argc, argv, kmem_options, |
1936 | kmem_subcommands, kmem_usage, 0); | |
ba77c9e1 | 1937 | |
90b86a9f | 1938 | if (!argc) |
ba77c9e1 LZ |
1939 | usage_with_options(kmem_usage, kmem_options); |
1940 | ||
0c160d49 NK |
1941 | if (kmem_slab == 0 && kmem_page == 0) { |
1942 | if (kmem_default == KMEM_SLAB) | |
1943 | kmem_slab = 1; | |
1944 | else | |
1945 | kmem_page = 1; | |
1946 | } | |
0d68bc92 | 1947 | |
90b86a9f | 1948 | if (!strncmp(argv[0], "rec", 3)) { |
0a7e6d1b | 1949 | symbol__init(NULL); |
90b86a9f | 1950 | return __cmd_record(argc, argv); |
2b2b2c68 NK |
1951 | } |
1952 | ||
28939e1a JO |
1953 | file.path = input_name; |
1954 | ||
c9758cc4 | 1955 | kmem_session = session = perf_session__new(&file, false, &perf_kmem); |
2b2b2c68 | 1956 | if (session == NULL) |
52e02834 | 1957 | return -1; |
2b2b2c68 | 1958 | |
ecc4c561 ACM |
1959 | ret = -1; |
1960 | ||
a923e2c4 NK |
1961 | if (kmem_slab) { |
1962 | if (!perf_evlist__find_tracepoint_by_name(session->evlist, | |
1963 | "kmem:kmalloc")) { | |
1964 | pr_err(errmsg, "slab", "slab"); | |
249ca1a8 | 1965 | goto out_delete; |
a923e2c4 NK |
1966 | } |
1967 | } | |
1968 | ||
0d68bc92 | 1969 | if (kmem_page) { |
a923e2c4 | 1970 | struct perf_evsel *evsel; |
0d68bc92 | 1971 | |
a923e2c4 NK |
1972 | evsel = perf_evlist__find_tracepoint_by_name(session->evlist, |
1973 | "kmem:mm_page_alloc"); | |
1974 | if (evsel == NULL) { | |
1975 | pr_err(errmsg, "page", "page"); | |
249ca1a8 | 1976 | goto out_delete; |
0d68bc92 NK |
1977 | } |
1978 | ||
1979 | kmem_page_size = pevent_get_page_size(evsel->tp_format->pevent); | |
c9758cc4 | 1980 | symbol_conf.use_callchain = true; |
0d68bc92 NK |
1981 | } |
1982 | ||
0a7e6d1b | 1983 | symbol__init(&session->header.env); |
2b2b2c68 | 1984 | |
2a865bd8 DA |
1985 | if (perf_time__parse_str(&ptime, time_str) != 0) { |
1986 | pr_err("Invalid time string\n"); | |
1987 | return -EINVAL; | |
1988 | } | |
1989 | ||
2b2b2c68 | 1990 | if (!strcmp(argv[0], "stat")) { |
77cfe388 NK |
1991 | setlocale(LC_ALL, ""); |
1992 | ||
4b627957 | 1993 | if (cpu__setup_cpunode_map()) |
2b2b2c68 | 1994 | goto out_delete; |
90b86a9f | 1995 | |
fb4f313d NK |
1996 | if (list_empty(&slab_caller_sort)) |
1997 | setup_slab_sorting(&slab_caller_sort, default_slab_sort); | |
1998 | if (list_empty(&slab_alloc_sort)) | |
1999 | setup_slab_sorting(&slab_alloc_sort, default_slab_sort); | |
2000 | if (list_empty(&page_caller_sort)) | |
2001 | setup_page_sorting(&page_caller_sort, default_page_sort); | |
2002 | if (list_empty(&page_alloc_sort)) | |
2003 | setup_page_sorting(&page_alloc_sort, default_page_sort); | |
2004 | ||
2005 | if (kmem_page) { | |
2006 | setup_page_sorting(&page_alloc_sort_input, | |
2007 | "page,order,migtype,gfp"); | |
2008 | setup_page_sorting(&page_caller_sort_input, | |
2009 | "callsite,order,migtype,gfp"); | |
2010 | } | |
2b2b2c68 | 2011 | ret = __cmd_kmem(session); |
b00eca8c PE |
2012 | } else |
2013 | usage_with_options(kmem_usage, kmem_options); | |
7d0d3945 | 2014 | |
2b2b2c68 NK |
2015 | out_delete: |
2016 | perf_session__delete(session); | |
2017 | ||
2018 | return ret; | |
ba77c9e1 LZ |
2019 | } |
2020 |