Linux 2.6.33.13
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / tools / perf / builtin-kmem.c
blob93c67bf53d2c96e391a649989ab6597728063b1c
1 #include "builtin.h"
2 #include "perf.h"
4 #include "util/util.h"
5 #include "util/cache.h"
6 #include "util/symbol.h"
7 #include "util/thread.h"
8 #include "util/header.h"
9 #include "util/session.h"
11 #include "util/parse-options.h"
12 #include "util/trace-event.h"
14 #include "util/debug.h"
16 #include <linux/rbtree.h>
18 struct alloc_stat;
19 typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
21 static char const *input_name = "perf.data";
23 static int alloc_flag;
24 static int caller_flag;
26 static int alloc_lines = -1;
27 static int caller_lines = -1;
29 static bool raw_ip;
31 static char default_sort_order[] = "frag,hit,bytes";
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 void init_cpunode_map(void)
61 FILE *fp;
62 int i;
64 fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
65 if (!fp) {
66 max_cpu_num = 4096;
67 return;
70 if (fscanf(fp, "%d", &max_cpu_num) < 1)
71 die("Failed to read 'kernel_max' from sysfs");
72 max_cpu_num++;
74 cpunode_map = calloc(max_cpu_num, sizeof(int));
75 if (!cpunode_map)
76 die("calloc");
77 for (i = 0; i < max_cpu_num; i++)
78 cpunode_map[i] = -1;
79 fclose(fp);
82 static void setup_cpunode_map(void)
84 struct dirent *dent1, *dent2;
85 DIR *dir1, *dir2;
86 unsigned int cpu, mem;
87 char buf[PATH_MAX];
89 init_cpunode_map();
91 dir1 = opendir(PATH_SYS_NODE);
92 if (!dir1)
93 return;
95 while (true) {
96 dent1 = readdir(dir1);
97 if (!dent1)
98 break;
100 if (sscanf(dent1->d_name, "node%u", &mem) < 1)
101 continue;
103 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
104 dir2 = opendir(buf);
105 if (!dir2)
106 continue;
107 while (true) {
108 dent2 = readdir(dir2);
109 if (!dent2)
110 break;
111 if (sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
112 continue;
113 cpunode_map[cpu] = mem;
118 static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
119 int bytes_req, int bytes_alloc, int cpu)
121 struct rb_node **node = &root_alloc_stat.rb_node;
122 struct rb_node *parent = NULL;
123 struct alloc_stat *data = NULL;
125 while (*node) {
126 parent = *node;
127 data = rb_entry(*node, struct alloc_stat, node);
129 if (ptr > data->ptr)
130 node = &(*node)->rb_right;
131 else if (ptr < data->ptr)
132 node = &(*node)->rb_left;
133 else
134 break;
137 if (data && data->ptr == ptr) {
138 data->hit++;
139 data->bytes_req += bytes_req;
140 data->bytes_alloc += bytes_alloc;
141 } else {
142 data = malloc(sizeof(*data));
143 if (!data)
144 die("malloc");
145 data->ptr = ptr;
146 data->pingpong = 0;
147 data->hit = 1;
148 data->bytes_req = bytes_req;
149 data->bytes_alloc = bytes_alloc;
151 rb_link_node(&data->node, parent, node);
152 rb_insert_color(&data->node, &root_alloc_stat);
154 data->call_site = call_site;
155 data->alloc_cpu = cpu;
158 static void insert_caller_stat(unsigned long call_site,
159 int bytes_req, int bytes_alloc)
161 struct rb_node **node = &root_caller_stat.rb_node;
162 struct rb_node *parent = NULL;
163 struct alloc_stat *data = NULL;
165 while (*node) {
166 parent = *node;
167 data = rb_entry(*node, struct alloc_stat, node);
169 if (call_site > data->call_site)
170 node = &(*node)->rb_right;
171 else if (call_site < data->call_site)
172 node = &(*node)->rb_left;
173 else
174 break;
177 if (data && data->call_site == call_site) {
178 data->hit++;
179 data->bytes_req += bytes_req;
180 data->bytes_alloc += bytes_alloc;
181 } else {
182 data = malloc(sizeof(*data));
183 if (!data)
184 die("malloc");
185 data->call_site = call_site;
186 data->pingpong = 0;
187 data->hit = 1;
188 data->bytes_req = bytes_req;
189 data->bytes_alloc = bytes_alloc;
191 rb_link_node(&data->node, parent, node);
192 rb_insert_color(&data->node, &root_caller_stat);
196 static void process_alloc_event(void *data,
197 struct event *event,
198 int cpu,
199 u64 timestamp __used,
200 struct thread *thread __used,
201 int node)
203 unsigned long call_site;
204 unsigned long ptr;
205 int bytes_req;
206 int bytes_alloc;
207 int node1, node2;
209 ptr = raw_field_value(event, "ptr", data);
210 call_site = raw_field_value(event, "call_site", data);
211 bytes_req = raw_field_value(event, "bytes_req", data);
212 bytes_alloc = raw_field_value(event, "bytes_alloc", data);
214 insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
215 insert_caller_stat(call_site, bytes_req, bytes_alloc);
217 total_requested += bytes_req;
218 total_allocated += bytes_alloc;
220 if (node) {
221 node1 = cpunode_map[cpu];
222 node2 = raw_field_value(event, "node", data);
223 if (node1 != node2)
224 nr_cross_allocs++;
226 nr_allocs++;
229 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
230 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
232 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
233 unsigned long call_site,
234 struct rb_root *root,
235 sort_fn_t sort_fn)
237 struct rb_node *node = root->rb_node;
238 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
240 while (node) {
241 struct alloc_stat *data;
242 int cmp;
244 data = rb_entry(node, struct alloc_stat, node);
246 cmp = sort_fn(&key, data);
247 if (cmp < 0)
248 node = node->rb_left;
249 else if (cmp > 0)
250 node = node->rb_right;
251 else
252 return data;
254 return NULL;
257 static void process_free_event(void *data,
258 struct event *event,
259 int cpu,
260 u64 timestamp __used,
261 struct thread *thread __used)
263 unsigned long ptr;
264 struct alloc_stat *s_alloc, *s_caller;
266 ptr = raw_field_value(event, "ptr", data);
268 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
269 if (!s_alloc)
270 return;
272 if (cpu != s_alloc->alloc_cpu) {
273 s_alloc->pingpong++;
275 s_caller = search_alloc_stat(0, s_alloc->call_site,
276 &root_caller_stat, callsite_cmp);
277 assert(s_caller);
278 s_caller->pingpong++;
280 s_alloc->alloc_cpu = -1;
283 static void
284 process_raw_event(event_t *raw_event __used, void *data,
285 int cpu, u64 timestamp, struct thread *thread)
287 struct event *event;
288 int type;
290 type = trace_parse_common_type(data);
291 event = trace_find_event(type);
293 if (!strcmp(event->name, "kmalloc") ||
294 !strcmp(event->name, "kmem_cache_alloc")) {
295 process_alloc_event(data, event, cpu, timestamp, thread, 0);
296 return;
299 if (!strcmp(event->name, "kmalloc_node") ||
300 !strcmp(event->name, "kmem_cache_alloc_node")) {
301 process_alloc_event(data, event, cpu, timestamp, thread, 1);
302 return;
305 if (!strcmp(event->name, "kfree") ||
306 !strcmp(event->name, "kmem_cache_free")) {
307 process_free_event(data, event, cpu, timestamp, thread);
308 return;
312 static int process_sample_event(event_t *event, struct perf_session *session)
314 struct sample_data data;
315 struct thread *thread;
317 memset(&data, 0, sizeof(data));
318 data.time = -1;
319 data.cpu = -1;
320 data.period = 1;
322 event__parse_sample(event, session->sample_type, &data);
324 dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
325 event->header.misc,
326 data.pid, data.tid,
327 (void *)(long)data.ip,
328 (long long)data.period);
330 thread = perf_session__findnew(session, event->ip.pid);
331 if (thread == NULL) {
332 pr_debug("problem processing %d event, skipping it.\n",
333 event->header.type);
334 return -1;
337 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
339 process_raw_event(event, data.raw_data, data.cpu,
340 data.time, thread);
342 return 0;
345 static int sample_type_check(struct perf_session *session)
347 if (!(session->sample_type & PERF_SAMPLE_RAW)) {
348 fprintf(stderr,
349 "No trace sample to read. Did you call perf record "
350 "without -R?");
351 return -1;
354 return 0;
357 static struct perf_event_ops event_ops = {
358 .process_sample_event = process_sample_event,
359 .process_comm_event = event__process_comm,
360 .sample_type_check = sample_type_check,
363 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
365 if (n_alloc == 0)
366 return 0.0;
367 else
368 return 100.0 - (100.0 * n_req / n_alloc);
371 static void __print_result(struct rb_root *root, struct perf_session *session,
372 int n_lines, int is_caller)
374 struct rb_node *next;
376 printf("%.102s\n", graph_dotted_line);
377 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
378 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
379 printf("%.102s\n", graph_dotted_line);
381 next = rb_first(root);
383 while (next && n_lines--) {
384 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
385 node);
386 struct symbol *sym = NULL;
387 char buf[BUFSIZ];
388 u64 addr;
390 if (is_caller) {
391 addr = data->call_site;
392 if (!raw_ip)
393 sym = map_groups__find_function(&session->kmaps, session, addr, NULL);
394 } else
395 addr = data->ptr;
397 if (sym != NULL)
398 snprintf(buf, sizeof(buf), "%s+%Lx", sym->name,
399 addr - sym->start);
400 else
401 snprintf(buf, sizeof(buf), "%#Lx", addr);
402 printf(" %-34s |", buf);
404 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
405 (unsigned long long)data->bytes_alloc,
406 (unsigned long)data->bytes_alloc / data->hit,
407 (unsigned long long)data->bytes_req,
408 (unsigned long)data->bytes_req / data->hit,
409 (unsigned long)data->hit,
410 (unsigned long)data->pingpong,
411 fragmentation(data->bytes_req, data->bytes_alloc));
413 next = rb_next(next);
416 if (n_lines == -1)
417 printf(" ... | ... | ... | ... | ... | ... \n");
419 printf("%.102s\n", graph_dotted_line);
422 static void print_summary(void)
424 printf("\nSUMMARY\n=======\n");
425 printf("Total bytes requested: %lu\n", total_requested);
426 printf("Total bytes allocated: %lu\n", total_allocated);
427 printf("Total bytes wasted on internal fragmentation: %lu\n",
428 total_allocated - total_requested);
429 printf("Internal fragmentation: %f%%\n",
430 fragmentation(total_requested, total_allocated));
431 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
434 static void print_result(struct perf_session *session)
436 if (caller_flag)
437 __print_result(&root_caller_sorted, session, caller_lines, 1);
438 if (alloc_flag)
439 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
440 print_summary();
443 struct sort_dimension {
444 const char name[20];
445 sort_fn_t cmp;
446 struct list_head list;
449 static LIST_HEAD(caller_sort);
450 static LIST_HEAD(alloc_sort);
452 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
453 struct list_head *sort_list)
455 struct rb_node **new = &(root->rb_node);
456 struct rb_node *parent = NULL;
457 struct sort_dimension *sort;
459 while (*new) {
460 struct alloc_stat *this;
461 int cmp = 0;
463 this = rb_entry(*new, struct alloc_stat, node);
464 parent = *new;
466 list_for_each_entry(sort, sort_list, list) {
467 cmp = sort->cmp(data, this);
468 if (cmp)
469 break;
472 if (cmp > 0)
473 new = &((*new)->rb_left);
474 else
475 new = &((*new)->rb_right);
478 rb_link_node(&data->node, parent, new);
479 rb_insert_color(&data->node, root);
482 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
483 struct list_head *sort_list)
485 struct rb_node *node;
486 struct alloc_stat *data;
488 for (;;) {
489 node = rb_first(root);
490 if (!node)
491 break;
493 rb_erase(node, root);
494 data = rb_entry(node, struct alloc_stat, node);
495 sort_insert(root_sorted, data, sort_list);
499 static void sort_result(void)
501 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
502 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
505 static int __cmd_kmem(void)
507 int err;
508 struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0);
509 if (session == NULL)
510 return -ENOMEM;
512 setup_pager();
513 err = perf_session__process_events(session, &event_ops);
514 if (err != 0)
515 goto out_delete;
516 sort_result();
517 print_result(session);
518 out_delete:
519 perf_session__delete(session);
520 return err;
523 static const char * const kmem_usage[] = {
524 "perf kmem [<options>] {record|stat}",
525 NULL
528 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
530 if (l->ptr < r->ptr)
531 return -1;
532 else if (l->ptr > r->ptr)
533 return 1;
534 return 0;
537 static struct sort_dimension ptr_sort_dimension = {
538 .name = "ptr",
539 .cmp = ptr_cmp,
542 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
544 if (l->call_site < r->call_site)
545 return -1;
546 else if (l->call_site > r->call_site)
547 return 1;
548 return 0;
551 static struct sort_dimension callsite_sort_dimension = {
552 .name = "callsite",
553 .cmp = callsite_cmp,
556 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
558 if (l->hit < r->hit)
559 return -1;
560 else if (l->hit > r->hit)
561 return 1;
562 return 0;
565 static struct sort_dimension hit_sort_dimension = {
566 .name = "hit",
567 .cmp = hit_cmp,
570 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
572 if (l->bytes_alloc < r->bytes_alloc)
573 return -1;
574 else if (l->bytes_alloc > r->bytes_alloc)
575 return 1;
576 return 0;
579 static struct sort_dimension bytes_sort_dimension = {
580 .name = "bytes",
581 .cmp = bytes_cmp,
584 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
586 double x, y;
588 x = fragmentation(l->bytes_req, l->bytes_alloc);
589 y = fragmentation(r->bytes_req, r->bytes_alloc);
591 if (x < y)
592 return -1;
593 else if (x > y)
594 return 1;
595 return 0;
598 static struct sort_dimension frag_sort_dimension = {
599 .name = "frag",
600 .cmp = frag_cmp,
603 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
605 if (l->pingpong < r->pingpong)
606 return -1;
607 else if (l->pingpong > r->pingpong)
608 return 1;
609 return 0;
612 static struct sort_dimension pingpong_sort_dimension = {
613 .name = "pingpong",
614 .cmp = pingpong_cmp,
617 static struct sort_dimension *avail_sorts[] = {
618 &ptr_sort_dimension,
619 &callsite_sort_dimension,
620 &hit_sort_dimension,
621 &bytes_sort_dimension,
622 &frag_sort_dimension,
623 &pingpong_sort_dimension,
626 #define NUM_AVAIL_SORTS \
627 (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
629 static int sort_dimension__add(const char *tok, struct list_head *list)
631 struct sort_dimension *sort;
632 int i;
634 for (i = 0; i < NUM_AVAIL_SORTS; i++) {
635 if (!strcmp(avail_sorts[i]->name, tok)) {
636 sort = malloc(sizeof(*sort));
637 if (!sort)
638 die("malloc");
639 memcpy(sort, avail_sorts[i], sizeof(*sort));
640 list_add_tail(&sort->list, list);
641 return 0;
645 return -1;
648 static int setup_sorting(struct list_head *sort_list, const char *arg)
650 char *tok;
651 char *str = strdup(arg);
653 if (!str)
654 die("strdup");
656 while (true) {
657 tok = strsep(&str, ",");
658 if (!tok)
659 break;
660 if (sort_dimension__add(tok, sort_list) < 0) {
661 error("Unknown --sort key: '%s'", tok);
662 return -1;
666 free(str);
667 return 0;
670 static int parse_sort_opt(const struct option *opt __used,
671 const char *arg, int unset __used)
673 if (!arg)
674 return -1;
676 if (caller_flag > alloc_flag)
677 return setup_sorting(&caller_sort, arg);
678 else
679 return setup_sorting(&alloc_sort, arg);
681 return 0;
684 static int parse_caller_opt(const struct option *opt __used,
685 const char *arg __used, int unset __used)
687 caller_flag = (alloc_flag + 1);
688 return 0;
691 static int parse_alloc_opt(const struct option *opt __used,
692 const char *arg __used, int unset __used)
694 alloc_flag = (caller_flag + 1);
695 return 0;
698 static int parse_line_opt(const struct option *opt __used,
699 const char *arg, int unset __used)
701 int lines;
703 if (!arg)
704 return -1;
706 lines = strtoul(arg, NULL, 10);
708 if (caller_flag > alloc_flag)
709 caller_lines = lines;
710 else
711 alloc_lines = lines;
713 return 0;
716 static const struct option kmem_options[] = {
717 OPT_STRING('i', "input", &input_name, "file",
718 "input file name"),
719 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
720 "show per-callsite statistics",
721 parse_caller_opt),
722 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
723 "show per-allocation statistics",
724 parse_alloc_opt),
725 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
726 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
727 parse_sort_opt),
728 OPT_CALLBACK('l', "line", NULL, "num",
729 "show n lines",
730 parse_line_opt),
731 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
732 OPT_END()
735 static const char *record_args[] = {
736 "record",
737 "-a",
738 "-R",
739 "-M",
740 "-f",
741 "-c", "1",
742 "-e", "kmem:kmalloc",
743 "-e", "kmem:kmalloc_node",
744 "-e", "kmem:kfree",
745 "-e", "kmem:kmem_cache_alloc",
746 "-e", "kmem:kmem_cache_alloc_node",
747 "-e", "kmem:kmem_cache_free",
750 static int __cmd_record(int argc, const char **argv)
752 unsigned int rec_argc, i, j;
753 const char **rec_argv;
755 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
756 rec_argv = calloc(rec_argc + 1, sizeof(char *));
758 for (i = 0; i < ARRAY_SIZE(record_args); i++)
759 rec_argv[i] = strdup(record_args[i]);
761 for (j = 1; j < (unsigned int)argc; j++, i++)
762 rec_argv[i] = argv[j];
764 return cmd_record(i, rec_argv, NULL);
767 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
769 argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
771 if (!argc)
772 usage_with_options(kmem_usage, kmem_options);
774 symbol__init();
776 if (!strncmp(argv[0], "rec", 3)) {
777 return __cmd_record(argc, argv);
778 } else if (!strcmp(argv[0], "stat")) {
779 setup_cpunode_map();
781 if (list_empty(&caller_sort))
782 setup_sorting(&caller_sort, default_sort_order);
783 if (list_empty(&alloc_sort))
784 setup_sorting(&alloc_sort, default_sort_order);
786 return __cmd_kmem();
787 } else
788 usage_with_options(kmem_usage, kmem_options);
790 return 0;