x86/xen: resume timer irqs early
[linux/fpc-iii.git] / tools / perf / builtin-kmem.c
blob9b5f077fee5b1b65a4e51b48ef1af0727b8c134b
1 #include "builtin.h"
2 #include "perf.h"
4 #include "util/evlist.h"
5 #include "util/evsel.h"
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"
11 #include "util/session.h"
12 #include "util/tool.h"
14 #include "util/parse-options.h"
15 #include "util/trace-event.h"
17 #include "util/debug.h"
19 #include <linux/rbtree.h>
20 #include <linux/string.h>
22 struct alloc_stat;
23 typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
25 static int alloc_flag;
26 static int caller_flag;
28 static int alloc_lines = -1;
29 static int caller_lines = -1;
31 static bool raw_ip;
33 static int *cpunode_map;
34 static int max_cpu_num;
36 struct alloc_stat {
37 u64 call_site;
38 u64 ptr;
39 u64 bytes_req;
40 u64 bytes_alloc;
41 u32 hit;
42 u32 pingpong;
44 short alloc_cpu;
46 struct rb_node node;
49 static struct rb_root root_alloc_stat;
50 static struct rb_root root_alloc_sorted;
51 static struct rb_root root_caller_stat;
52 static struct rb_root root_caller_sorted;
54 static unsigned long total_requested, total_allocated;
55 static unsigned long nr_allocs, nr_cross_allocs;
57 #define PATH_SYS_NODE "/sys/devices/system/node"
59 static int init_cpunode_map(void)
61 FILE *fp;
62 int i, err = -1;
64 fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
65 if (!fp) {
66 max_cpu_num = 4096;
67 return 0;
70 if (fscanf(fp, "%d", &max_cpu_num) < 1) {
71 pr_err("Failed to read 'kernel_max' from sysfs");
72 goto out_close;
75 max_cpu_num++;
77 cpunode_map = calloc(max_cpu_num, sizeof(int));
78 if (!cpunode_map) {
79 pr_err("%s: calloc failed\n", __func__);
80 goto out_close;
83 for (i = 0; i < max_cpu_num; i++)
84 cpunode_map[i] = -1;
86 err = 0;
87 out_close:
88 fclose(fp);
89 return err;
92 static int setup_cpunode_map(void)
94 struct dirent *dent1, *dent2;
95 DIR *dir1, *dir2;
96 unsigned int cpu, mem;
97 char buf[PATH_MAX];
99 if (init_cpunode_map())
100 return -1;
102 dir1 = opendir(PATH_SYS_NODE);
103 if (!dir1)
104 return 0;
106 while ((dent1 = readdir(dir1)) != NULL) {
107 if (dent1->d_type != DT_DIR ||
108 sscanf(dent1->d_name, "node%u", &mem) < 1)
109 continue;
111 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
112 dir2 = opendir(buf);
113 if (!dir2)
114 continue;
115 while ((dent2 = readdir(dir2)) != NULL) {
116 if (dent2->d_type != DT_LNK ||
117 sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
118 continue;
119 cpunode_map[cpu] = mem;
121 closedir(dir2);
123 closedir(dir1);
124 return 0;
127 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
128 int bytes_req, int bytes_alloc, int cpu)
130 struct rb_node **node = &root_alloc_stat.rb_node;
131 struct rb_node *parent = NULL;
132 struct alloc_stat *data = NULL;
134 while (*node) {
135 parent = *node;
136 data = rb_entry(*node, struct alloc_stat, node);
138 if (ptr > data->ptr)
139 node = &(*node)->rb_right;
140 else if (ptr < data->ptr)
141 node = &(*node)->rb_left;
142 else
143 break;
146 if (data && data->ptr == ptr) {
147 data->hit++;
148 data->bytes_req += bytes_req;
149 data->bytes_alloc += bytes_alloc;
150 } else {
151 data = malloc(sizeof(*data));
152 if (!data) {
153 pr_err("%s: malloc failed\n", __func__);
154 return -1;
156 data->ptr = ptr;
157 data->pingpong = 0;
158 data->hit = 1;
159 data->bytes_req = bytes_req;
160 data->bytes_alloc = bytes_alloc;
162 rb_link_node(&data->node, parent, node);
163 rb_insert_color(&data->node, &root_alloc_stat);
165 data->call_site = call_site;
166 data->alloc_cpu = cpu;
167 return 0;
170 static int insert_caller_stat(unsigned long call_site,
171 int bytes_req, int bytes_alloc)
173 struct rb_node **node = &root_caller_stat.rb_node;
174 struct rb_node *parent = NULL;
175 struct alloc_stat *data = NULL;
177 while (*node) {
178 parent = *node;
179 data = rb_entry(*node, struct alloc_stat, node);
181 if (call_site > data->call_site)
182 node = &(*node)->rb_right;
183 else if (call_site < data->call_site)
184 node = &(*node)->rb_left;
185 else
186 break;
189 if (data && data->call_site == call_site) {
190 data->hit++;
191 data->bytes_req += bytes_req;
192 data->bytes_alloc += bytes_alloc;
193 } else {
194 data = malloc(sizeof(*data));
195 if (!data) {
196 pr_err("%s: malloc failed\n", __func__);
197 return -1;
199 data->call_site = call_site;
200 data->pingpong = 0;
201 data->hit = 1;
202 data->bytes_req = bytes_req;
203 data->bytes_alloc = bytes_alloc;
205 rb_link_node(&data->node, parent, node);
206 rb_insert_color(&data->node, &root_caller_stat);
209 return 0;
212 static int perf_evsel__process_alloc_event(struct perf_evsel *evsel,
213 struct perf_sample *sample)
215 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
216 call_site = perf_evsel__intval(evsel, sample, "call_site");
217 int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
218 bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
220 if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
221 insert_caller_stat(call_site, bytes_req, bytes_alloc))
222 return -1;
224 total_requested += bytes_req;
225 total_allocated += bytes_alloc;
227 nr_allocs++;
228 return 0;
231 static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,
232 struct perf_sample *sample)
234 int ret = perf_evsel__process_alloc_event(evsel, sample);
236 if (!ret) {
237 int node1 = cpunode_map[sample->cpu],
238 node2 = perf_evsel__intval(evsel, sample, "node");
240 if (node1 != node2)
241 nr_cross_allocs++;
244 return ret;
247 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
248 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
250 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
251 unsigned long call_site,
252 struct rb_root *root,
253 sort_fn_t sort_fn)
255 struct rb_node *node = root->rb_node;
256 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
258 while (node) {
259 struct alloc_stat *data;
260 int cmp;
262 data = rb_entry(node, struct alloc_stat, node);
264 cmp = sort_fn(&key, data);
265 if (cmp < 0)
266 node = node->rb_left;
267 else if (cmp > 0)
268 node = node->rb_right;
269 else
270 return data;
272 return NULL;
275 static int perf_evsel__process_free_event(struct perf_evsel *evsel,
276 struct perf_sample *sample)
278 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
279 struct alloc_stat *s_alloc, *s_caller;
281 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
282 if (!s_alloc)
283 return 0;
285 if ((short)sample->cpu != s_alloc->alloc_cpu) {
286 s_alloc->pingpong++;
288 s_caller = search_alloc_stat(0, s_alloc->call_site,
289 &root_caller_stat, callsite_cmp);
290 if (!s_caller)
291 return -1;
292 s_caller->pingpong++;
294 s_alloc->alloc_cpu = -1;
296 return 0;
299 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
300 struct perf_sample *sample);
302 static int process_sample_event(struct perf_tool *tool __maybe_unused,
303 union perf_event *event,
304 struct perf_sample *sample,
305 struct perf_evsel *evsel,
306 struct machine *machine)
308 struct thread *thread = machine__findnew_thread(machine, sample->pid,
309 sample->pid);
311 if (thread == NULL) {
312 pr_debug("problem processing %d event, skipping it.\n",
313 event->header.type);
314 return -1;
317 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->tid);
319 if (evsel->handler.func != NULL) {
320 tracepoint_handler f = evsel->handler.func;
321 return f(evsel, sample);
324 return 0;
327 static struct perf_tool perf_kmem = {
328 .sample = process_sample_event,
329 .comm = perf_event__process_comm,
330 .ordered_samples = true,
333 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
335 if (n_alloc == 0)
336 return 0.0;
337 else
338 return 100.0 - (100.0 * n_req / n_alloc);
341 static void __print_result(struct rb_root *root, struct perf_session *session,
342 int n_lines, int is_caller)
344 struct rb_node *next;
345 struct machine *machine = &session->machines.host;
347 printf("%.102s\n", graph_dotted_line);
348 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
349 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
350 printf("%.102s\n", graph_dotted_line);
352 next = rb_first(root);
354 while (next && n_lines--) {
355 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
356 node);
357 struct symbol *sym = NULL;
358 struct map *map;
359 char buf[BUFSIZ];
360 u64 addr;
362 if (is_caller) {
363 addr = data->call_site;
364 if (!raw_ip)
365 sym = machine__find_kernel_function(machine, addr, &map, NULL);
366 } else
367 addr = data->ptr;
369 if (sym != NULL)
370 snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
371 addr - map->unmap_ip(map, sym->start));
372 else
373 snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
374 printf(" %-34s |", buf);
376 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
377 (unsigned long long)data->bytes_alloc,
378 (unsigned long)data->bytes_alloc / data->hit,
379 (unsigned long long)data->bytes_req,
380 (unsigned long)data->bytes_req / data->hit,
381 (unsigned long)data->hit,
382 (unsigned long)data->pingpong,
383 fragmentation(data->bytes_req, data->bytes_alloc));
385 next = rb_next(next);
388 if (n_lines == -1)
389 printf(" ... | ... | ... | ... | ... | ... \n");
391 printf("%.102s\n", graph_dotted_line);
394 static void print_summary(void)
396 printf("\nSUMMARY\n=======\n");
397 printf("Total bytes requested: %lu\n", total_requested);
398 printf("Total bytes allocated: %lu\n", total_allocated);
399 printf("Total bytes wasted on internal fragmentation: %lu\n",
400 total_allocated - total_requested);
401 printf("Internal fragmentation: %f%%\n",
402 fragmentation(total_requested, total_allocated));
403 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
406 static void print_result(struct perf_session *session)
408 if (caller_flag)
409 __print_result(&root_caller_sorted, session, caller_lines, 1);
410 if (alloc_flag)
411 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
412 print_summary();
415 struct sort_dimension {
416 const char name[20];
417 sort_fn_t cmp;
418 struct list_head list;
421 static LIST_HEAD(caller_sort);
422 static LIST_HEAD(alloc_sort);
424 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
425 struct list_head *sort_list)
427 struct rb_node **new = &(root->rb_node);
428 struct rb_node *parent = NULL;
429 struct sort_dimension *sort;
431 while (*new) {
432 struct alloc_stat *this;
433 int cmp = 0;
435 this = rb_entry(*new, struct alloc_stat, node);
436 parent = *new;
438 list_for_each_entry(sort, sort_list, list) {
439 cmp = sort->cmp(data, this);
440 if (cmp)
441 break;
444 if (cmp > 0)
445 new = &((*new)->rb_left);
446 else
447 new = &((*new)->rb_right);
450 rb_link_node(&data->node, parent, new);
451 rb_insert_color(&data->node, root);
454 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
455 struct list_head *sort_list)
457 struct rb_node *node;
458 struct alloc_stat *data;
460 for (;;) {
461 node = rb_first(root);
462 if (!node)
463 break;
465 rb_erase(node, root);
466 data = rb_entry(node, struct alloc_stat, node);
467 sort_insert(root_sorted, data, sort_list);
471 static void sort_result(void)
473 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
474 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
477 static int __cmd_kmem(void)
479 int err = -EINVAL;
480 struct perf_session *session;
481 const struct perf_evsel_str_handler kmem_tracepoints[] = {
482 { "kmem:kmalloc", perf_evsel__process_alloc_event, },
483 { "kmem:kmem_cache_alloc", perf_evsel__process_alloc_event, },
484 { "kmem:kmalloc_node", perf_evsel__process_alloc_node_event, },
485 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
486 { "kmem:kfree", perf_evsel__process_free_event, },
487 { "kmem:kmem_cache_free", perf_evsel__process_free_event, },
490 session = perf_session__new(input_name, O_RDONLY, 0, false, &perf_kmem);
491 if (session == NULL)
492 return -ENOMEM;
494 if (perf_session__create_kernel_maps(session) < 0)
495 goto out_delete;
497 if (!perf_session__has_traces(session, "kmem record"))
498 goto out_delete;
500 if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
501 pr_err("Initializing perf session tracepoint handlers failed\n");
502 return -1;
505 setup_pager();
506 err = perf_session__process_events(session, &perf_kmem);
507 if (err != 0)
508 goto out_delete;
509 sort_result();
510 print_result(session);
511 out_delete:
512 perf_session__delete(session);
513 return err;
516 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
518 if (l->ptr < r->ptr)
519 return -1;
520 else if (l->ptr > r->ptr)
521 return 1;
522 return 0;
525 static struct sort_dimension ptr_sort_dimension = {
526 .name = "ptr",
527 .cmp = ptr_cmp,
530 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
532 if (l->call_site < r->call_site)
533 return -1;
534 else if (l->call_site > r->call_site)
535 return 1;
536 return 0;
539 static struct sort_dimension callsite_sort_dimension = {
540 .name = "callsite",
541 .cmp = callsite_cmp,
544 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
546 if (l->hit < r->hit)
547 return -1;
548 else if (l->hit > r->hit)
549 return 1;
550 return 0;
553 static struct sort_dimension hit_sort_dimension = {
554 .name = "hit",
555 .cmp = hit_cmp,
558 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
560 if (l->bytes_alloc < r->bytes_alloc)
561 return -1;
562 else if (l->bytes_alloc > r->bytes_alloc)
563 return 1;
564 return 0;
567 static struct sort_dimension bytes_sort_dimension = {
568 .name = "bytes",
569 .cmp = bytes_cmp,
572 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
574 double x, y;
576 x = fragmentation(l->bytes_req, l->bytes_alloc);
577 y = fragmentation(r->bytes_req, r->bytes_alloc);
579 if (x < y)
580 return -1;
581 else if (x > y)
582 return 1;
583 return 0;
586 static struct sort_dimension frag_sort_dimension = {
587 .name = "frag",
588 .cmp = frag_cmp,
591 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
593 if (l->pingpong < r->pingpong)
594 return -1;
595 else if (l->pingpong > r->pingpong)
596 return 1;
597 return 0;
600 static struct sort_dimension pingpong_sort_dimension = {
601 .name = "pingpong",
602 .cmp = pingpong_cmp,
605 static struct sort_dimension *avail_sorts[] = {
606 &ptr_sort_dimension,
607 &callsite_sort_dimension,
608 &hit_sort_dimension,
609 &bytes_sort_dimension,
610 &frag_sort_dimension,
611 &pingpong_sort_dimension,
614 #define NUM_AVAIL_SORTS ((int)ARRAY_SIZE(avail_sorts))
616 static int sort_dimension__add(const char *tok, struct list_head *list)
618 struct sort_dimension *sort;
619 int i;
621 for (i = 0; i < NUM_AVAIL_SORTS; i++) {
622 if (!strcmp(avail_sorts[i]->name, tok)) {
623 sort = memdup(avail_sorts[i], sizeof(*avail_sorts[i]));
624 if (!sort) {
625 pr_err("%s: memdup failed\n", __func__);
626 return -1;
628 list_add_tail(&sort->list, list);
629 return 0;
633 return -1;
636 static int setup_sorting(struct list_head *sort_list, const char *arg)
638 char *tok;
639 char *str = strdup(arg);
641 if (!str) {
642 pr_err("%s: strdup failed\n", __func__);
643 return -1;
646 while (true) {
647 tok = strsep(&str, ",");
648 if (!tok)
649 break;
650 if (sort_dimension__add(tok, sort_list) < 0) {
651 error("Unknown --sort key: '%s'", tok);
652 free(str);
653 return -1;
657 free(str);
658 return 0;
661 static int parse_sort_opt(const struct option *opt __maybe_unused,
662 const char *arg, int unset __maybe_unused)
664 if (!arg)
665 return -1;
667 if (caller_flag > alloc_flag)
668 return setup_sorting(&caller_sort, arg);
669 else
670 return setup_sorting(&alloc_sort, arg);
672 return 0;
675 static int parse_caller_opt(const struct option *opt __maybe_unused,
676 const char *arg __maybe_unused,
677 int unset __maybe_unused)
679 caller_flag = (alloc_flag + 1);
680 return 0;
683 static int parse_alloc_opt(const struct option *opt __maybe_unused,
684 const char *arg __maybe_unused,
685 int unset __maybe_unused)
687 alloc_flag = (caller_flag + 1);
688 return 0;
691 static int parse_line_opt(const struct option *opt __maybe_unused,
692 const char *arg, int unset __maybe_unused)
694 int lines;
696 if (!arg)
697 return -1;
699 lines = strtoul(arg, NULL, 10);
701 if (caller_flag > alloc_flag)
702 caller_lines = lines;
703 else
704 alloc_lines = lines;
706 return 0;
709 static int __cmd_record(int argc, const char **argv)
711 const char * const record_args[] = {
712 "record", "-a", "-R", "-c", "1",
713 "-e", "kmem:kmalloc",
714 "-e", "kmem:kmalloc_node",
715 "-e", "kmem:kfree",
716 "-e", "kmem:kmem_cache_alloc",
717 "-e", "kmem:kmem_cache_alloc_node",
718 "-e", "kmem:kmem_cache_free",
720 unsigned int rec_argc, i, j;
721 const char **rec_argv;
723 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
724 rec_argv = calloc(rec_argc + 1, sizeof(char *));
726 if (rec_argv == NULL)
727 return -ENOMEM;
729 for (i = 0; i < ARRAY_SIZE(record_args); i++)
730 rec_argv[i] = strdup(record_args[i]);
732 for (j = 1; j < (unsigned int)argc; j++, i++)
733 rec_argv[i] = argv[j];
735 return cmd_record(i, rec_argv, NULL);
738 int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused)
740 const char * const default_sort_order = "frag,hit,bytes";
741 const struct option kmem_options[] = {
742 OPT_STRING('i', "input", &input_name, "file", "input file name"),
743 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
744 "show per-callsite statistics", parse_caller_opt),
745 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
746 "show per-allocation statistics", parse_alloc_opt),
747 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
748 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
749 parse_sort_opt),
750 OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
751 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
752 OPT_END()
754 const char * const kmem_usage[] = {
755 "perf kmem [<options>] {record|stat}",
756 NULL
758 argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
760 if (!argc)
761 usage_with_options(kmem_usage, kmem_options);
763 symbol__init();
765 if (!strncmp(argv[0], "rec", 3)) {
766 return __cmd_record(argc, argv);
767 } else if (!strcmp(argv[0], "stat")) {
768 if (setup_cpunode_map())
769 return -1;
771 if (list_empty(&caller_sort))
772 setup_sorting(&caller_sort, default_sort_order);
773 if (list_empty(&alloc_sort))
774 setup_sorting(&alloc_sort, default_sort_order);
776 return __cmd_kmem();
777 } else
778 usage_with_options(kmem_usage, kmem_options);
780 return 0;