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;
278 static void process_raw_event(union perf_event *raw_event __used, void *data,
279 int cpu, u64 timestamp, struct thread *thread)
284 type = trace_parse_common_type(data);
285 event = trace_find_event(type);
287 if (!strcmp(event->name, "kmalloc") ||
288 !strcmp(event->name, "kmem_cache_alloc")) {
289 process_alloc_event(data, event, cpu, timestamp, thread, 0);
293 if (!strcmp(event->name, "kmalloc_node") ||
294 !strcmp(event->name, "kmem_cache_alloc_node")) {
295 process_alloc_event(data, event, cpu, timestamp, thread, 1);
299 if (!strcmp(event->name, "kfree") ||
300 !strcmp(event->name, "kmem_cache_free")) {
301 process_free_event(data, event, cpu, timestamp, thread);
306 static int process_sample_event(struct perf_event_ops *ops __used,
307 union perf_event *event,
308 struct perf_sample *sample,
309 struct perf_evsel *evsel __used,
310 struct machine *machine)
312 struct thread *thread = machine__findnew_thread(machine, event->ip.pid);
314 if (thread == NULL) {
315 pr_debug("problem processing %d event, skipping it.\n",
320 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
322 process_raw_event(event, sample->raw_data, sample->cpu,
323 sample->time, thread);
328 static struct perf_event_ops event_ops = {
329 .sample = process_sample_event,
330 .comm = perf_event__process_comm,
331 .ordered_samples = true,
334 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
339 return 100.0 - (100.0 * n_req / n_alloc);
342 static void __print_result(struct rb_root *root, struct perf_session *session,
343 int n_lines, int is_caller)
345 struct rb_node *next;
346 struct machine *machine;
348 printf("%.102s\n", graph_dotted_line);
349 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
350 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
351 printf("%.102s\n", graph_dotted_line);
353 next = rb_first(root);
355 machine = perf_session__find_host_machine(session);
357 pr_err("__print_result: couldn't find kernel information\n");
360 while (next && n_lines--) {
361 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
363 struct symbol *sym = NULL;
369 addr = data->call_site;
371 sym = machine__find_kernel_function(machine, addr, &map, NULL);
376 snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
377 addr - map->unmap_ip(map, sym->start));
379 snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
380 printf(" %-34s |", buf);
382 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
383 (unsigned long long)data->bytes_alloc,
384 (unsigned long)data->bytes_alloc / data->hit,
385 (unsigned long long)data->bytes_req,
386 (unsigned long)data->bytes_req / data->hit,
387 (unsigned long)data->hit,
388 (unsigned long)data->pingpong,
389 fragmentation(data->bytes_req, data->bytes_alloc));
391 next = rb_next(next);
395 printf(" ... | ... | ... | ... | ... | ... \n");
397 printf("%.102s\n", graph_dotted_line);
400 static void print_summary(void)
402 printf("\nSUMMARY\n=======\n");
403 printf("Total bytes requested: %lu\n", total_requested);
404 printf("Total bytes allocated: %lu\n", total_allocated);
405 printf("Total bytes wasted on internal fragmentation: %lu\n",
406 total_allocated - total_requested);
407 printf("Internal fragmentation: %f%%\n",
408 fragmentation(total_requested, total_allocated));
409 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
412 static void print_result(struct perf_session *session)
415 __print_result(&root_caller_sorted, session, caller_lines, 1);
417 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
421 struct sort_dimension {
424 struct list_head list;
427 static LIST_HEAD(caller_sort);
428 static LIST_HEAD(alloc_sort);
430 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
431 struct list_head *sort_list)
433 struct rb_node **new = &(root->rb_node);
434 struct rb_node *parent = NULL;
435 struct sort_dimension *sort;
438 struct alloc_stat *this;
441 this = rb_entry(*new, struct alloc_stat, node);
444 list_for_each_entry(sort, sort_list, list) {
445 cmp = sort->cmp(data, this);
451 new = &((*new)->rb_left);
453 new = &((*new)->rb_right);
456 rb_link_node(&data->node, parent, new);
457 rb_insert_color(&data->node, root);
460 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
461 struct list_head *sort_list)
463 struct rb_node *node;
464 struct alloc_stat *data;
467 node = rb_first(root);
471 rb_erase(node, root);
472 data = rb_entry(node, struct alloc_stat, node);
473 sort_insert(root_sorted, data, sort_list);
477 static void sort_result(void)
479 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
480 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
483 static int __cmd_kmem(void)
486 struct perf_session *session = perf_session__new(input_name, O_RDONLY,
487 0, false, &event_ops);
491 if (perf_session__create_kernel_maps(session) < 0)
494 if (!perf_session__has_traces(session, "kmem record"))
498 err = perf_session__process_events(session, &event_ops);
502 print_result(session);
504 perf_session__delete(session);
508 static const char * const kmem_usage[] = {
509 "perf kmem [<options>] {record|stat}",
513 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
517 else if (l->ptr > r->ptr)
522 static struct sort_dimension ptr_sort_dimension = {
527 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
529 if (l->call_site < r->call_site)
531 else if (l->call_site > r->call_site)
536 static struct sort_dimension callsite_sort_dimension = {
541 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
545 else if (l->hit > r->hit)
550 static struct sort_dimension hit_sort_dimension = {
555 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
557 if (l->bytes_alloc < r->bytes_alloc)
559 else if (l->bytes_alloc > r->bytes_alloc)
564 static struct sort_dimension bytes_sort_dimension = {
569 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
573 x = fragmentation(l->bytes_req, l->bytes_alloc);
574 y = fragmentation(r->bytes_req, r->bytes_alloc);
583 static struct sort_dimension frag_sort_dimension = {
588 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
590 if (l->pingpong < r->pingpong)
592 else if (l->pingpong > r->pingpong)
597 static struct sort_dimension pingpong_sort_dimension = {
602 static struct sort_dimension *avail_sorts[] = {
604 &callsite_sort_dimension,
606 &bytes_sort_dimension,
607 &frag_sort_dimension,
608 &pingpong_sort_dimension,
611 #define NUM_AVAIL_SORTS \
612 (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
614 static int sort_dimension__add(const char *tok, struct list_head *list)
616 struct sort_dimension *sort;
619 for (i = 0; i < NUM_AVAIL_SORTS; i++) {
620 if (!strcmp(avail_sorts[i]->name, tok)) {
621 sort = malloc(sizeof(*sort));
624 memcpy(sort, avail_sorts[i], sizeof(*sort));
625 list_add_tail(&sort->list, list);
633 static int setup_sorting(struct list_head *sort_list, const char *arg)
636 char *str = strdup(arg);
642 tok = strsep(&str, ",");
645 if (sort_dimension__add(tok, sort_list) < 0) {
646 error("Unknown --sort key: '%s'", tok);
655 static int parse_sort_opt(const struct option *opt __used,
656 const char *arg, int unset __used)
661 if (caller_flag > alloc_flag)
662 return setup_sorting(&caller_sort, arg);
664 return setup_sorting(&alloc_sort, arg);
669 static int parse_caller_opt(const struct option *opt __used,
670 const char *arg __used, int unset __used)
672 caller_flag = (alloc_flag + 1);
676 static int parse_alloc_opt(const struct option *opt __used,
677 const char *arg __used, int unset __used)
679 alloc_flag = (caller_flag + 1);
683 static int parse_line_opt(const struct option *opt __used,
684 const char *arg, int unset __used)
691 lines = strtoul(arg, NULL, 10);
693 if (caller_flag > alloc_flag)
694 caller_lines = lines;
701 static const struct option kmem_options[] = {
702 OPT_STRING('i', "input", &input_name, "file",
704 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
705 "show per-callsite statistics",
707 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
708 "show per-allocation statistics",
710 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
711 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
713 OPT_CALLBACK('l', "line", NULL, "num",
716 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
720 static const char *record_args[] = {
726 "-e", "kmem:kmalloc",
727 "-e", "kmem:kmalloc_node",
729 "-e", "kmem:kmem_cache_alloc",
730 "-e", "kmem:kmem_cache_alloc_node",
731 "-e", "kmem:kmem_cache_free",
734 static int __cmd_record(int argc, const char **argv)
736 unsigned int rec_argc, i, j;
737 const char **rec_argv;
739 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
740 rec_argv = calloc(rec_argc + 1, sizeof(char *));
742 if (rec_argv == NULL)
745 for (i = 0; i < ARRAY_SIZE(record_args); i++)
746 rec_argv[i] = strdup(record_args[i]);
748 for (j = 1; j < (unsigned int)argc; j++, i++)
749 rec_argv[i] = argv[j];
751 return cmd_record(i, rec_argv, NULL);
754 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
756 argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
759 usage_with_options(kmem_usage, kmem_options);
763 if (!strncmp(argv[0], "rec", 3)) {
764 return __cmd_record(argc, argv);
765 } else if (!strcmp(argv[0], "stat")) {
768 if (list_empty(&caller_sort))
769 setup_sorting(&caller_sort, default_sort_order);
770 if (list_empty(&alloc_sort))
771 setup_sorting(&alloc_sort, default_sort_order);
775 usage_with_options(kmem_usage, kmem_options);