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);
95 while ((dent1 = readdir(dir1)) != NULL) {
96 if (dent1->d_type != DT_DIR ||
97 sscanf(dent1->d_name, "node%u", &mem) < 1)
100 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
104 while ((dent2 = readdir(dir2)) != NULL) {
105 if (dent2->d_type != DT_LNK ||
106 sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
108 cpunode_map[cpu] = mem;
113 static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
114 int bytes_req, int bytes_alloc, int cpu)
116 struct rb_node **node = &root_alloc_stat.rb_node;
117 struct rb_node *parent = NULL;
118 struct alloc_stat *data = NULL;
122 data = rb_entry(*node, struct alloc_stat, node);
125 node = &(*node)->rb_right;
126 else if (ptr < data->ptr)
127 node = &(*node)->rb_left;
132 if (data && data->ptr == ptr) {
134 data->bytes_req += bytes_req;
135 data->bytes_alloc += bytes_alloc;
137 data = malloc(sizeof(*data));
143 data->bytes_req = bytes_req;
144 data->bytes_alloc = bytes_alloc;
146 rb_link_node(&data->node, parent, node);
147 rb_insert_color(&data->node, &root_alloc_stat);
149 data->call_site = call_site;
150 data->alloc_cpu = cpu;
153 static void insert_caller_stat(unsigned long call_site,
154 int bytes_req, int bytes_alloc)
156 struct rb_node **node = &root_caller_stat.rb_node;
157 struct rb_node *parent = NULL;
158 struct alloc_stat *data = NULL;
162 data = rb_entry(*node, struct alloc_stat, node);
164 if (call_site > data->call_site)
165 node = &(*node)->rb_right;
166 else if (call_site < data->call_site)
167 node = &(*node)->rb_left;
172 if (data && data->call_site == call_site) {
174 data->bytes_req += bytes_req;
175 data->bytes_alloc += bytes_alloc;
177 data = malloc(sizeof(*data));
180 data->call_site = call_site;
183 data->bytes_req = bytes_req;
184 data->bytes_alloc = bytes_alloc;
186 rb_link_node(&data->node, parent, node);
187 rb_insert_color(&data->node, &root_caller_stat);
191 static void process_alloc_event(void *data,
194 u64 timestamp __used,
195 struct thread *thread __used,
198 unsigned long call_site;
204 ptr = raw_field_value(event, "ptr", data);
205 call_site = raw_field_value(event, "call_site", data);
206 bytes_req = raw_field_value(event, "bytes_req", data);
207 bytes_alloc = raw_field_value(event, "bytes_alloc", data);
209 insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
210 insert_caller_stat(call_site, bytes_req, bytes_alloc);
212 total_requested += bytes_req;
213 total_allocated += bytes_alloc;
216 node1 = cpunode_map[cpu];
217 node2 = raw_field_value(event, "node", data);
224 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
225 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
227 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
228 unsigned long call_site,
229 struct rb_root *root,
232 struct rb_node *node = root->rb_node;
233 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
236 struct alloc_stat *data;
239 data = rb_entry(node, struct alloc_stat, node);
241 cmp = sort_fn(&key, data);
243 node = node->rb_left;
245 node = node->rb_right;
252 static void process_free_event(void *data,
255 u64 timestamp __used,
256 struct thread *thread __used)
259 struct alloc_stat *s_alloc, *s_caller;
261 ptr = raw_field_value(event, "ptr", data);
263 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
267 if (cpu != s_alloc->alloc_cpu) {
270 s_caller = search_alloc_stat(0, s_alloc->call_site,
271 &root_caller_stat, callsite_cmp);
273 s_caller->pingpong++;
275 s_alloc->alloc_cpu = -1;
279 process_raw_event(event_t *raw_event __used, void *data,
280 int cpu, u64 timestamp, struct thread *thread)
285 type = trace_parse_common_type(data);
286 event = trace_find_event(type);
288 if (!strcmp(event->name, "kmalloc") ||
289 !strcmp(event->name, "kmem_cache_alloc")) {
290 process_alloc_event(data, event, cpu, timestamp, thread, 0);
294 if (!strcmp(event->name, "kmalloc_node") ||
295 !strcmp(event->name, "kmem_cache_alloc_node")) {
296 process_alloc_event(data, event, cpu, timestamp, thread, 1);
300 if (!strcmp(event->name, "kfree") ||
301 !strcmp(event->name, "kmem_cache_free")) {
302 process_free_event(data, event, cpu, timestamp, thread);
307 static int process_sample_event(event_t *event, struct perf_session *session)
309 struct sample_data data;
310 struct thread *thread;
312 memset(&data, 0, sizeof(data));
317 event__parse_sample(event, session->sample_type, &data);
319 dump_printf("(IP, %d): %d/%d: %#Lx period: %Ld\n", event->header.misc,
320 data.pid, data.tid, data.ip, data.period);
322 thread = perf_session__findnew(session, event->ip.pid);
323 if (thread == NULL) {
324 pr_debug("problem processing %d event, skipping it.\n",
329 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
331 process_raw_event(event, data.raw_data, data.cpu,
337 static struct perf_event_ops event_ops = {
338 .sample = process_sample_event,
339 .comm = event__process_comm,
342 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
347 return 100.0 - (100.0 * n_req / n_alloc);
350 static void __print_result(struct rb_root *root, struct perf_session *session,
351 int n_lines, int is_caller)
353 struct rb_node *next;
355 printf("%.102s\n", graph_dotted_line);
356 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
357 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
358 printf("%.102s\n", graph_dotted_line);
360 next = rb_first(root);
362 while (next && n_lines--) {
363 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
365 struct symbol *sym = NULL;
370 addr = data->call_site;
372 sym = map_groups__find_function(&session->kmaps, addr, NULL);
377 snprintf(buf, sizeof(buf), "%s+%Lx", sym->name,
380 snprintf(buf, sizeof(buf), "%#Lx", addr);
381 printf(" %-34s |", buf);
383 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
384 (unsigned long long)data->bytes_alloc,
385 (unsigned long)data->bytes_alloc / data->hit,
386 (unsigned long long)data->bytes_req,
387 (unsigned long)data->bytes_req / data->hit,
388 (unsigned long)data->hit,
389 (unsigned long)data->pingpong,
390 fragmentation(data->bytes_req, data->bytes_alloc));
392 next = rb_next(next);
396 printf(" ... | ... | ... | ... | ... | ... \n");
398 printf("%.102s\n", graph_dotted_line);
401 static void print_summary(void)
403 printf("\nSUMMARY\n=======\n");
404 printf("Total bytes requested: %lu\n", total_requested);
405 printf("Total bytes allocated: %lu\n", total_allocated);
406 printf("Total bytes wasted on internal fragmentation: %lu\n",
407 total_allocated - total_requested);
408 printf("Internal fragmentation: %f%%\n",
409 fragmentation(total_requested, total_allocated));
410 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
413 static void print_result(struct perf_session *session)
416 __print_result(&root_caller_sorted, session, caller_lines, 1);
418 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
422 struct sort_dimension {
425 struct list_head list;
428 static LIST_HEAD(caller_sort);
429 static LIST_HEAD(alloc_sort);
431 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
432 struct list_head *sort_list)
434 struct rb_node **new = &(root->rb_node);
435 struct rb_node *parent = NULL;
436 struct sort_dimension *sort;
439 struct alloc_stat *this;
442 this = rb_entry(*new, struct alloc_stat, node);
445 list_for_each_entry(sort, sort_list, list) {
446 cmp = sort->cmp(data, this);
452 new = &((*new)->rb_left);
454 new = &((*new)->rb_right);
457 rb_link_node(&data->node, parent, new);
458 rb_insert_color(&data->node, root);
461 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
462 struct list_head *sort_list)
464 struct rb_node *node;
465 struct alloc_stat *data;
468 node = rb_first(root);
472 rb_erase(node, root);
473 data = rb_entry(node, struct alloc_stat, node);
474 sort_insert(root_sorted, data, sort_list);
478 static void sort_result(void)
480 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
481 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
484 static int __cmd_kmem(void)
487 struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0);
491 if (!perf_session__has_traces(session, "kmem record"))
495 err = perf_session__process_events(session, &event_ops);
499 print_result(session);
501 perf_session__delete(session);
505 static const char * const kmem_usage[] = {
506 "perf kmem [<options>] {record|stat}",
510 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
514 else if (l->ptr > r->ptr)
519 static struct sort_dimension ptr_sort_dimension = {
524 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
526 if (l->call_site < r->call_site)
528 else if (l->call_site > r->call_site)
533 static struct sort_dimension callsite_sort_dimension = {
538 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
542 else if (l->hit > r->hit)
547 static struct sort_dimension hit_sort_dimension = {
552 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
554 if (l->bytes_alloc < r->bytes_alloc)
556 else if (l->bytes_alloc > r->bytes_alloc)
561 static struct sort_dimension bytes_sort_dimension = {
566 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
570 x = fragmentation(l->bytes_req, l->bytes_alloc);
571 y = fragmentation(r->bytes_req, r->bytes_alloc);
580 static struct sort_dimension frag_sort_dimension = {
585 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
587 if (l->pingpong < r->pingpong)
589 else if (l->pingpong > r->pingpong)
594 static struct sort_dimension pingpong_sort_dimension = {
599 static struct sort_dimension *avail_sorts[] = {
601 &callsite_sort_dimension,
603 &bytes_sort_dimension,
604 &frag_sort_dimension,
605 &pingpong_sort_dimension,
608 #define NUM_AVAIL_SORTS \
609 (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
611 static int sort_dimension__add(const char *tok, struct list_head *list)
613 struct sort_dimension *sort;
616 for (i = 0; i < NUM_AVAIL_SORTS; i++) {
617 if (!strcmp(avail_sorts[i]->name, tok)) {
618 sort = malloc(sizeof(*sort));
621 memcpy(sort, avail_sorts[i], sizeof(*sort));
622 list_add_tail(&sort->list, list);
630 static int setup_sorting(struct list_head *sort_list, const char *arg)
633 char *str = strdup(arg);
639 tok = strsep(&str, ",");
642 if (sort_dimension__add(tok, sort_list) < 0) {
643 error("Unknown --sort key: '%s'", tok);
652 static int parse_sort_opt(const struct option *opt __used,
653 const char *arg, int unset __used)
658 if (caller_flag > alloc_flag)
659 return setup_sorting(&caller_sort, arg);
661 return setup_sorting(&alloc_sort, arg);
666 static int parse_caller_opt(const struct option *opt __used,
667 const char *arg __used, int unset __used)
669 caller_flag = (alloc_flag + 1);
673 static int parse_alloc_opt(const struct option *opt __used,
674 const char *arg __used, int unset __used)
676 alloc_flag = (caller_flag + 1);
680 static int parse_line_opt(const struct option *opt __used,
681 const char *arg, int unset __used)
688 lines = strtoul(arg, NULL, 10);
690 if (caller_flag > alloc_flag)
691 caller_lines = lines;
698 static const struct option kmem_options[] = {
699 OPT_STRING('i', "input", &input_name, "file",
701 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
702 "show per-callsite statistics",
704 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
705 "show per-allocation statistics",
707 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
708 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
710 OPT_CALLBACK('l', "line", NULL, "num",
713 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
717 static const char *record_args[] = {
724 "-e", "kmem:kmalloc",
725 "-e", "kmem:kmalloc_node",
727 "-e", "kmem:kmem_cache_alloc",
728 "-e", "kmem:kmem_cache_alloc_node",
729 "-e", "kmem:kmem_cache_free",
732 static int __cmd_record(int argc, const char **argv)
734 unsigned int rec_argc, i, j;
735 const char **rec_argv;
737 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
738 rec_argv = calloc(rec_argc + 1, sizeof(char *));
740 for (i = 0; i < ARRAY_SIZE(record_args); i++)
741 rec_argv[i] = strdup(record_args[i]);
743 for (j = 1; j < (unsigned int)argc; j++, i++)
744 rec_argv[i] = argv[j];
746 return cmd_record(i, rec_argv, NULL);
749 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
751 argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
754 usage_with_options(kmem_usage, kmem_options);
758 if (!strncmp(argv[0], "rec", 3)) {
759 return __cmd_record(argc, argv);
760 } else if (!strcmp(argv[0], "stat")) {
763 if (list_empty(&caller_sort))
764 setup_sorting(&caller_sort, default_sort_order);
765 if (list_empty(&alloc_sort))
766 setup_sorting(&alloc_sort, default_sort_order);
770 usage_with_options(kmem_usage, kmem_options);