4 #include "util/evlist.h"
5 #include "util/evsel.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>
22 typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
24 static int alloc_flag;
25 static int caller_flag;
27 static int alloc_lines = -1;
28 static int caller_lines = -1;
32 static int *cpunode_map;
33 static int max_cpu_num;
48 static struct rb_root root_alloc_stat;
49 static struct rb_root root_alloc_sorted;
50 static struct rb_root root_caller_stat;
51 static struct rb_root root_caller_sorted;
53 static unsigned long total_requested, total_allocated;
54 static unsigned long nr_allocs, nr_cross_allocs;
56 #define PATH_SYS_NODE "/sys/devices/system/node"
58 static int init_cpunode_map(void)
63 fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
69 if (fscanf(fp, "%d", &max_cpu_num) < 1) {
70 pr_err("Failed to read 'kernel_max' from sysfs");
76 cpunode_map = calloc(max_cpu_num, sizeof(int));
78 pr_err("%s: calloc failed\n", __func__);
82 for (i = 0; i < max_cpu_num; i++)
91 static int setup_cpunode_map(void)
93 struct dirent *dent1, *dent2;
95 unsigned int cpu, mem;
98 if (init_cpunode_map())
101 dir1 = opendir(PATH_SYS_NODE);
105 while ((dent1 = readdir(dir1)) != NULL) {
106 if (dent1->d_type != DT_DIR ||
107 sscanf(dent1->d_name, "node%u", &mem) < 1)
110 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
114 while ((dent2 = readdir(dir2)) != NULL) {
115 if (dent2->d_type != DT_LNK ||
116 sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
118 cpunode_map[cpu] = mem;
126 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
127 int bytes_req, int bytes_alloc, int cpu)
129 struct rb_node **node = &root_alloc_stat.rb_node;
130 struct rb_node *parent = NULL;
131 struct alloc_stat *data = NULL;
135 data = rb_entry(*node, struct alloc_stat, node);
138 node = &(*node)->rb_right;
139 else if (ptr < data->ptr)
140 node = &(*node)->rb_left;
145 if (data && data->ptr == ptr) {
147 data->bytes_req += bytes_req;
148 data->bytes_alloc += bytes_alloc;
150 data = malloc(sizeof(*data));
152 pr_err("%s: malloc failed\n", __func__);
158 data->bytes_req = bytes_req;
159 data->bytes_alloc = bytes_alloc;
161 rb_link_node(&data->node, parent, node);
162 rb_insert_color(&data->node, &root_alloc_stat);
164 data->call_site = call_site;
165 data->alloc_cpu = cpu;
169 static int insert_caller_stat(unsigned long call_site,
170 int bytes_req, int bytes_alloc)
172 struct rb_node **node = &root_caller_stat.rb_node;
173 struct rb_node *parent = NULL;
174 struct alloc_stat *data = NULL;
178 data = rb_entry(*node, struct alloc_stat, node);
180 if (call_site > data->call_site)
181 node = &(*node)->rb_right;
182 else if (call_site < data->call_site)
183 node = &(*node)->rb_left;
188 if (data && data->call_site == call_site) {
190 data->bytes_req += bytes_req;
191 data->bytes_alloc += bytes_alloc;
193 data = malloc(sizeof(*data));
195 pr_err("%s: malloc failed\n", __func__);
198 data->call_site = call_site;
201 data->bytes_req = bytes_req;
202 data->bytes_alloc = bytes_alloc;
204 rb_link_node(&data->node, parent, node);
205 rb_insert_color(&data->node, &root_caller_stat);
211 static int perf_evsel__process_alloc_event(struct perf_evsel *evsel,
212 struct perf_sample *sample)
214 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
215 call_site = perf_evsel__intval(evsel, sample, "call_site");
216 int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
217 bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
219 if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
220 insert_caller_stat(call_site, bytes_req, bytes_alloc))
223 total_requested += bytes_req;
224 total_allocated += bytes_alloc;
230 static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,
231 struct perf_sample *sample)
233 int ret = perf_evsel__process_alloc_event(evsel, sample);
236 int node1 = cpunode_map[sample->cpu],
237 node2 = perf_evsel__intval(evsel, sample, "node");
246 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
247 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
249 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
250 unsigned long call_site,
251 struct rb_root *root,
254 struct rb_node *node = root->rb_node;
255 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
258 struct alloc_stat *data;
261 data = rb_entry(node, struct alloc_stat, node);
263 cmp = sort_fn(&key, data);
265 node = node->rb_left;
267 node = node->rb_right;
274 static int perf_evsel__process_free_event(struct perf_evsel *evsel,
275 struct perf_sample *sample)
277 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
278 struct alloc_stat *s_alloc, *s_caller;
280 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
284 if ((short)sample->cpu != s_alloc->alloc_cpu) {
287 s_caller = search_alloc_stat(0, s_alloc->call_site,
288 &root_caller_stat, callsite_cmp);
291 s_caller->pingpong++;
293 s_alloc->alloc_cpu = -1;
298 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
299 struct perf_sample *sample);
301 static int process_sample_event(struct perf_tool *tool __maybe_unused,
302 union perf_event *event,
303 struct perf_sample *sample,
304 struct perf_evsel *evsel,
305 struct machine *machine)
307 struct thread *thread = machine__findnew_thread(machine, event->ip.pid);
309 if (thread == NULL) {
310 pr_debug("problem processing %d event, skipping it.\n",
315 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
317 if (evsel->handler.func != NULL) {
318 tracepoint_handler f = evsel->handler.func;
319 return f(evsel, sample);
325 static struct perf_tool perf_kmem = {
326 .sample = process_sample_event,
327 .comm = perf_event__process_comm,
328 .ordered_samples = true,
331 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
336 return 100.0 - (100.0 * n_req / n_alloc);
339 static void __print_result(struct rb_root *root, struct perf_session *session,
340 int n_lines, int is_caller)
342 struct rb_node *next;
343 struct machine *machine = &session->machines.host;
345 printf("%.102s\n", graph_dotted_line);
346 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
347 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
348 printf("%.102s\n", graph_dotted_line);
350 next = rb_first(root);
352 while (next && n_lines--) {
353 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
355 struct symbol *sym = NULL;
361 addr = data->call_site;
363 sym = machine__find_kernel_function(machine, addr, &map, NULL);
368 snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
369 addr - map->unmap_ip(map, sym->start));
371 snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
372 printf(" %-34s |", buf);
374 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
375 (unsigned long long)data->bytes_alloc,
376 (unsigned long)data->bytes_alloc / data->hit,
377 (unsigned long long)data->bytes_req,
378 (unsigned long)data->bytes_req / data->hit,
379 (unsigned long)data->hit,
380 (unsigned long)data->pingpong,
381 fragmentation(data->bytes_req, data->bytes_alloc));
383 next = rb_next(next);
387 printf(" ... | ... | ... | ... | ... | ... \n");
389 printf("%.102s\n", graph_dotted_line);
392 static void print_summary(void)
394 printf("\nSUMMARY\n=======\n");
395 printf("Total bytes requested: %lu\n", total_requested);
396 printf("Total bytes allocated: %lu\n", total_allocated);
397 printf("Total bytes wasted on internal fragmentation: %lu\n",
398 total_allocated - total_requested);
399 printf("Internal fragmentation: %f%%\n",
400 fragmentation(total_requested, total_allocated));
401 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
404 static void print_result(struct perf_session *session)
407 __print_result(&root_caller_sorted, session, caller_lines, 1);
409 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
413 struct sort_dimension {
416 struct list_head list;
419 static LIST_HEAD(caller_sort);
420 static LIST_HEAD(alloc_sort);
422 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
423 struct list_head *sort_list)
425 struct rb_node **new = &(root->rb_node);
426 struct rb_node *parent = NULL;
427 struct sort_dimension *sort;
430 struct alloc_stat *this;
433 this = rb_entry(*new, struct alloc_stat, node);
436 list_for_each_entry(sort, sort_list, list) {
437 cmp = sort->cmp(data, this);
443 new = &((*new)->rb_left);
445 new = &((*new)->rb_right);
448 rb_link_node(&data->node, parent, new);
449 rb_insert_color(&data->node, root);
452 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
453 struct list_head *sort_list)
455 struct rb_node *node;
456 struct alloc_stat *data;
459 node = rb_first(root);
463 rb_erase(node, root);
464 data = rb_entry(node, struct alloc_stat, node);
465 sort_insert(root_sorted, data, sort_list);
469 static void sort_result(void)
471 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
472 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
475 static int __cmd_kmem(void)
478 struct perf_session *session;
479 const struct perf_evsel_str_handler kmem_tracepoints[] = {
480 { "kmem:kmalloc", perf_evsel__process_alloc_event, },
481 { "kmem:kmem_cache_alloc", perf_evsel__process_alloc_event, },
482 { "kmem:kmalloc_node", perf_evsel__process_alloc_node_event, },
483 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
484 { "kmem:kfree", perf_evsel__process_free_event, },
485 { "kmem:kmem_cache_free", perf_evsel__process_free_event, },
488 session = perf_session__new(input_name, O_RDONLY, 0, false, &perf_kmem);
492 if (perf_session__create_kernel_maps(session) < 0)
495 if (!perf_session__has_traces(session, "kmem record"))
498 if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
499 pr_err("Initializing perf session tracepoint handlers failed\n");
504 err = perf_session__process_events(session, &perf_kmem);
508 print_result(session);
510 perf_session__delete(session);
514 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
518 else if (l->ptr > r->ptr)
523 static struct sort_dimension ptr_sort_dimension = {
528 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
530 if (l->call_site < r->call_site)
532 else if (l->call_site > r->call_site)
537 static struct sort_dimension callsite_sort_dimension = {
542 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
546 else if (l->hit > r->hit)
551 static struct sort_dimension hit_sort_dimension = {
556 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
558 if (l->bytes_alloc < r->bytes_alloc)
560 else if (l->bytes_alloc > r->bytes_alloc)
565 static struct sort_dimension bytes_sort_dimension = {
570 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
574 x = fragmentation(l->bytes_req, l->bytes_alloc);
575 y = fragmentation(r->bytes_req, r->bytes_alloc);
584 static struct sort_dimension frag_sort_dimension = {
589 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
591 if (l->pingpong < r->pingpong)
593 else if (l->pingpong > r->pingpong)
598 static struct sort_dimension pingpong_sort_dimension = {
603 static struct sort_dimension *avail_sorts[] = {
605 &callsite_sort_dimension,
607 &bytes_sort_dimension,
608 &frag_sort_dimension,
609 &pingpong_sort_dimension,
612 #define NUM_AVAIL_SORTS ((int)ARRAY_SIZE(avail_sorts))
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));
623 pr_err("%s: malloc failed\n", __func__);
626 memcpy(sort, avail_sorts[i], sizeof(*sort));
627 list_add_tail(&sort->list, list);
635 static int setup_sorting(struct list_head *sort_list, const char *arg)
638 char *str = strdup(arg);
641 pr_err("%s: strdup failed\n", __func__);
646 tok = strsep(&str, ",");
649 if (sort_dimension__add(tok, sort_list) < 0) {
650 error("Unknown --sort key: '%s'", tok);
660 static int parse_sort_opt(const struct option *opt __maybe_unused,
661 const char *arg, int unset __maybe_unused)
666 if (caller_flag > alloc_flag)
667 return setup_sorting(&caller_sort, arg);
669 return setup_sorting(&alloc_sort, arg);
674 static int parse_caller_opt(const struct option *opt __maybe_unused,
675 const char *arg __maybe_unused,
676 int unset __maybe_unused)
678 caller_flag = (alloc_flag + 1);
682 static int parse_alloc_opt(const struct option *opt __maybe_unused,
683 const char *arg __maybe_unused,
684 int unset __maybe_unused)
686 alloc_flag = (caller_flag + 1);
690 static int parse_line_opt(const struct option *opt __maybe_unused,
691 const char *arg, int unset __maybe_unused)
698 lines = strtoul(arg, NULL, 10);
700 if (caller_flag > alloc_flag)
701 caller_lines = lines;
708 static int __cmd_record(int argc, const char **argv)
710 const char * const record_args[] = {
711 "record", "-a", "-R", "-f", "-c", "1",
712 "-e", "kmem:kmalloc",
713 "-e", "kmem:kmalloc_node",
715 "-e", "kmem:kmem_cache_alloc",
716 "-e", "kmem:kmem_cache_alloc_node",
717 "-e", "kmem:kmem_cache_free",
719 unsigned int rec_argc, i, j;
720 const char **rec_argv;
722 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
723 rec_argv = calloc(rec_argc + 1, sizeof(char *));
725 if (rec_argv == NULL)
728 for (i = 0; i < ARRAY_SIZE(record_args); i++)
729 rec_argv[i] = strdup(record_args[i]);
731 for (j = 1; j < (unsigned int)argc; j++, i++)
732 rec_argv[i] = argv[j];
734 return cmd_record(i, rec_argv, NULL);
737 int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused)
739 const char * const default_sort_order = "frag,hit,bytes";
740 const struct option kmem_options[] = {
741 OPT_STRING('i', "input", &input_name, "file", "input file name"),
742 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
743 "show per-callsite statistics", parse_caller_opt),
744 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
745 "show per-allocation statistics", parse_alloc_opt),
746 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
747 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
749 OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
750 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
753 const char * const kmem_usage[] = {
754 "perf kmem [<options>] {record|stat}",
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")) {
767 if (setup_cpunode_map())
770 if (list_empty(&caller_sort))
771 setup_sorting(&caller_sort, default_sort_order);
772 if (list_empty(&alloc_sort))
773 setup_sorting(&alloc_sort, default_sort_order);
777 usage_with_options(kmem_usage, kmem_options);