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>
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;
31 static char default_sort_order[] = "frag,hit,bytes";
33 static int *cpunode_map;
34 static int max_cpu_num;
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)
64 fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
70 if (fscanf(fp, "%d", &max_cpu_num) < 1)
71 die("Failed to read 'kernel_max' from sysfs");
74 cpunode_map = calloc(max_cpu_num, sizeof(int));
77 for (i = 0; i < max_cpu_num; i++)
82 static void setup_cpunode_map(void)
84 struct dirent *dent1, *dent2;
86 unsigned int cpu, mem;
91 dir1 = opendir(PATH_SYS_NODE);
96 dent1 = readdir(dir1);
100 if (sscanf(dent1->d_name, "node%u", &mem) < 1)
103 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
108 dent2 = readdir(dir2);
111 if (sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
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;
127 data = rb_entry(*node, struct alloc_stat, node);
130 node = &(*node)->rb_right;
131 else if (ptr < data->ptr)
132 node = &(*node)->rb_left;
137 if (data && data->ptr == ptr) {
139 data->bytes_req += bytes_req;
140 data->bytes_alloc += bytes_req;
142 data = malloc(sizeof(*data));
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;
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;
177 if (data && data->call_site == call_site) {
179 data->bytes_req += bytes_req;
180 data->bytes_alloc += bytes_req;
182 data = malloc(sizeof(*data));
185 data->call_site = call_site;
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,
199 u64 timestamp __used,
200 struct thread *thread __used,
203 unsigned long call_site;
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;
221 node1 = cpunode_map[cpu];
222 node2 = raw_field_value(event, "node", data);
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,
237 struct rb_node *node = root->rb_node;
238 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
241 struct alloc_stat *data;
244 data = rb_entry(node, struct alloc_stat, node);
246 cmp = sort_fn(&key, data);
248 node = node->rb_left;
250 node = node->rb_right;
257 static void process_free_event(void *data,
260 u64 timestamp __used,
261 struct thread *thread __used)
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);
272 if (cpu != s_alloc->alloc_cpu) {
275 s_caller = search_alloc_stat(0, s_alloc->call_site,
276 &root_caller_stat, callsite_cmp);
278 s_caller->pingpong++;
280 s_alloc->alloc_cpu = -1;
284 process_raw_event(event_t *raw_event __used, void *data,
285 int cpu, u64 timestamp, struct thread *thread)
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);
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);
305 if (!strcmp(event->name, "kfree") ||
306 !strcmp(event->name, "kmem_cache_free")) {
307 process_free_event(data, event, cpu, timestamp, thread);
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));
322 event__parse_sample(event, session->sample_type, &data);
324 dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
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",
337 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
339 process_raw_event(event, data.raw_data, data.cpu,
345 static struct perf_event_ops event_ops = {
346 .process_sample_event = process_sample_event,
347 .process_comm_event = event__process_comm,
348 .sample_type_check = perf_session__has_traces,
351 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
356 return 100.0 - (100.0 * n_req / n_alloc);
359 static void __print_result(struct rb_root *root, struct perf_session *session,
360 int n_lines, int is_caller)
362 struct rb_node *next;
364 printf("%.102s\n", graph_dotted_line);
365 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
366 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
367 printf("%.102s\n", graph_dotted_line);
369 next = rb_first(root);
371 while (next && n_lines--) {
372 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
374 struct symbol *sym = NULL;
379 addr = data->call_site;
381 sym = map_groups__find_function(&session->kmaps, session, addr, NULL);
386 snprintf(buf, sizeof(buf), "%s+%Lx", sym->name,
389 snprintf(buf, sizeof(buf), "%#Lx", addr);
390 printf(" %-34s |", buf);
392 printf(" %9llu/%-5lu | %9llu/%-5lu | %6lu | %8lu | %6.3f%%\n",
393 (unsigned long long)data->bytes_alloc,
394 (unsigned long)data->bytes_alloc / data->hit,
395 (unsigned long long)data->bytes_req,
396 (unsigned long)data->bytes_req / data->hit,
397 (unsigned long)data->hit,
398 (unsigned long)data->pingpong,
399 fragmentation(data->bytes_req, data->bytes_alloc));
401 next = rb_next(next);
405 printf(" ... | ... | ... | ... | ... | ... \n");
407 printf("%.102s\n", graph_dotted_line);
410 static void print_summary(void)
412 printf("\nSUMMARY\n=======\n");
413 printf("Total bytes requested: %lu\n", total_requested);
414 printf("Total bytes allocated: %lu\n", total_allocated);
415 printf("Total bytes wasted on internal fragmentation: %lu\n",
416 total_allocated - total_requested);
417 printf("Internal fragmentation: %f%%\n",
418 fragmentation(total_requested, total_allocated));
419 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
422 static void print_result(struct perf_session *session)
425 __print_result(&root_caller_sorted, session, caller_lines, 1);
427 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
431 struct sort_dimension {
434 struct list_head list;
437 static LIST_HEAD(caller_sort);
438 static LIST_HEAD(alloc_sort);
440 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
441 struct list_head *sort_list)
443 struct rb_node **new = &(root->rb_node);
444 struct rb_node *parent = NULL;
445 struct sort_dimension *sort;
448 struct alloc_stat *this;
451 this = rb_entry(*new, struct alloc_stat, node);
454 list_for_each_entry(sort, sort_list, list) {
455 cmp = sort->cmp(data, this);
461 new = &((*new)->rb_left);
463 new = &((*new)->rb_right);
466 rb_link_node(&data->node, parent, new);
467 rb_insert_color(&data->node, root);
470 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
471 struct list_head *sort_list)
473 struct rb_node *node;
474 struct alloc_stat *data;
477 node = rb_first(root);
481 rb_erase(node, root);
482 data = rb_entry(node, struct alloc_stat, node);
483 sort_insert(root_sorted, data, sort_list);
487 static void sort_result(void)
489 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
490 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
493 static int __cmd_kmem(void)
496 struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0);
501 err = perf_session__process_events(session, &event_ops);
505 print_result(session);
507 perf_session__delete(session);
511 static const char * const kmem_usage[] = {
512 "perf kmem [<options>] {record|stat}",
516 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
520 else if (l->ptr > r->ptr)
525 static struct sort_dimension ptr_sort_dimension = {
530 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
532 if (l->call_site < r->call_site)
534 else if (l->call_site > r->call_site)
539 static struct sort_dimension callsite_sort_dimension = {
544 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
548 else if (l->hit > r->hit)
553 static struct sort_dimension hit_sort_dimension = {
558 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
560 if (l->bytes_alloc < r->bytes_alloc)
562 else if (l->bytes_alloc > r->bytes_alloc)
567 static struct sort_dimension bytes_sort_dimension = {
572 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
576 x = fragmentation(l->bytes_req, l->bytes_alloc);
577 y = fragmentation(r->bytes_req, r->bytes_alloc);
586 static struct sort_dimension frag_sort_dimension = {
591 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
593 if (l->pingpong < r->pingpong)
595 else if (l->pingpong > r->pingpong)
600 static struct sort_dimension pingpong_sort_dimension = {
605 static struct sort_dimension *avail_sorts[] = {
607 &callsite_sort_dimension,
609 &bytes_sort_dimension,
610 &frag_sort_dimension,
611 &pingpong_sort_dimension,
614 #define NUM_AVAIL_SORTS \
615 (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
617 static int sort_dimension__add(const char *tok, struct list_head *list)
619 struct sort_dimension *sort;
622 for (i = 0; i < NUM_AVAIL_SORTS; i++) {
623 if (!strcmp(avail_sorts[i]->name, tok)) {
624 sort = malloc(sizeof(*sort));
627 memcpy(sort, avail_sorts[i], sizeof(*sort));
628 list_add_tail(&sort->list, list);
636 static int setup_sorting(struct list_head *sort_list, const char *arg)
639 char *str = strdup(arg);
645 tok = strsep(&str, ",");
648 if (sort_dimension__add(tok, sort_list) < 0) {
649 error("Unknown --sort key: '%s'", tok);
658 static int parse_sort_opt(const struct option *opt __used,
659 const char *arg, int unset __used)
664 if (caller_flag > alloc_flag)
665 return setup_sorting(&caller_sort, arg);
667 return setup_sorting(&alloc_sort, arg);
672 static int parse_caller_opt(const struct option *opt __used,
673 const char *arg __used, int unset __used)
675 caller_flag = (alloc_flag + 1);
679 static int parse_alloc_opt(const struct option *opt __used,
680 const char *arg __used, int unset __used)
682 alloc_flag = (caller_flag + 1);
686 static int parse_line_opt(const struct option *opt __used,
687 const char *arg, int unset __used)
694 lines = strtoul(arg, NULL, 10);
696 if (caller_flag > alloc_flag)
697 caller_lines = lines;
704 static const struct option kmem_options[] = {
705 OPT_STRING('i', "input", &input_name, "file",
707 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
708 "show per-callsite statistics",
710 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
711 "show per-allocation statistics",
713 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
714 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
716 OPT_CALLBACK('l', "line", NULL, "num",
719 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
723 static const char *record_args[] = {
730 "-e", "kmem:kmalloc",
731 "-e", "kmem:kmalloc_node",
733 "-e", "kmem:kmem_cache_alloc",
734 "-e", "kmem:kmem_cache_alloc_node",
735 "-e", "kmem:kmem_cache_free",
738 static int __cmd_record(int argc, const char **argv)
740 unsigned int rec_argc, i, j;
741 const char **rec_argv;
743 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
744 rec_argv = calloc(rec_argc + 1, sizeof(char *));
746 for (i = 0; i < ARRAY_SIZE(record_args); i++)
747 rec_argv[i] = strdup(record_args[i]);
749 for (j = 1; j < (unsigned int)argc; j++, i++)
750 rec_argv[i] = argv[j];
752 return cmd_record(i, rec_argv, NULL);
755 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
757 argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
760 usage_with_options(kmem_usage, kmem_options);
764 if (!strncmp(argv[0], "rec", 3)) {
765 return __cmd_record(argc, argv);
766 } else if (!strcmp(argv[0], "stat")) {
769 if (list_empty(&caller_sort))
770 setup_sorting(&caller_sort, default_sort_order);
771 if (list_empty(&alloc_sort))
772 setup_sorting(&alloc_sort, default_sort_order);