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"
16 #include "util/data.h"
17 #include "util/cpumap.h"
19 #include "util/debug.h"
21 #include <linux/rbtree.h>
22 #include <linux/string.h>
25 typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
27 static int alloc_flag;
28 static int caller_flag;
30 static int alloc_lines = -1;
31 static int caller_lines = -1;
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 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
57 int bytes_req, int bytes_alloc, int cpu)
59 struct rb_node **node = &root_alloc_stat.rb_node;
60 struct rb_node *parent = NULL;
61 struct alloc_stat *data = NULL;
65 data = rb_entry(*node, struct alloc_stat, node);
68 node = &(*node)->rb_right;
69 else if (ptr < data->ptr)
70 node = &(*node)->rb_left;
75 if (data && data->ptr == ptr) {
77 data->bytes_req += bytes_req;
78 data->bytes_alloc += bytes_alloc;
80 data = malloc(sizeof(*data));
82 pr_err("%s: malloc failed\n", __func__);
88 data->bytes_req = bytes_req;
89 data->bytes_alloc = bytes_alloc;
91 rb_link_node(&data->node, parent, node);
92 rb_insert_color(&data->node, &root_alloc_stat);
94 data->call_site = call_site;
95 data->alloc_cpu = cpu;
99 static int insert_caller_stat(unsigned long call_site,
100 int bytes_req, int bytes_alloc)
102 struct rb_node **node = &root_caller_stat.rb_node;
103 struct rb_node *parent = NULL;
104 struct alloc_stat *data = NULL;
108 data = rb_entry(*node, struct alloc_stat, node);
110 if (call_site > data->call_site)
111 node = &(*node)->rb_right;
112 else if (call_site < data->call_site)
113 node = &(*node)->rb_left;
118 if (data && data->call_site == call_site) {
120 data->bytes_req += bytes_req;
121 data->bytes_alloc += bytes_alloc;
123 data = malloc(sizeof(*data));
125 pr_err("%s: malloc failed\n", __func__);
128 data->call_site = call_site;
131 data->bytes_req = bytes_req;
132 data->bytes_alloc = bytes_alloc;
134 rb_link_node(&data->node, parent, node);
135 rb_insert_color(&data->node, &root_caller_stat);
141 static int perf_evsel__process_alloc_event(struct perf_evsel *evsel,
142 struct perf_sample *sample)
144 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
145 call_site = perf_evsel__intval(evsel, sample, "call_site");
146 int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
147 bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
149 if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
150 insert_caller_stat(call_site, bytes_req, bytes_alloc))
153 total_requested += bytes_req;
154 total_allocated += bytes_alloc;
160 static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,
161 struct perf_sample *sample)
163 int ret = perf_evsel__process_alloc_event(evsel, sample);
166 int node1 = cpu__get_node(sample->cpu),
167 node2 = perf_evsel__intval(evsel, sample, "node");
176 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
177 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
179 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
180 unsigned long call_site,
181 struct rb_root *root,
184 struct rb_node *node = root->rb_node;
185 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
188 struct alloc_stat *data;
191 data = rb_entry(node, struct alloc_stat, node);
193 cmp = sort_fn(&key, data);
195 node = node->rb_left;
197 node = node->rb_right;
204 static int perf_evsel__process_free_event(struct perf_evsel *evsel,
205 struct perf_sample *sample)
207 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
208 struct alloc_stat *s_alloc, *s_caller;
210 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
214 if ((short)sample->cpu != s_alloc->alloc_cpu) {
217 s_caller = search_alloc_stat(0, s_alloc->call_site,
218 &root_caller_stat, callsite_cmp);
221 s_caller->pingpong++;
223 s_alloc->alloc_cpu = -1;
228 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
229 struct perf_sample *sample);
231 static int process_sample_event(struct perf_tool *tool __maybe_unused,
232 union perf_event *event,
233 struct perf_sample *sample,
234 struct perf_evsel *evsel,
235 struct machine *machine)
237 struct thread *thread = machine__findnew_thread(machine, sample->pid,
240 if (thread == NULL) {
241 pr_debug("problem processing %d event, skipping it.\n",
246 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
248 if (evsel->handler != NULL) {
249 tracepoint_handler f = evsel->handler;
250 return f(evsel, sample);
256 static struct perf_tool perf_kmem = {
257 .sample = process_sample_event,
258 .comm = perf_event__process_comm,
259 .ordered_samples = true,
262 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
267 return 100.0 - (100.0 * n_req / n_alloc);
270 static void __print_result(struct rb_root *root, struct perf_session *session,
271 int n_lines, int is_caller)
273 struct rb_node *next;
274 struct machine *machine = &session->machines.host;
276 printf("%.102s\n", graph_dotted_line);
277 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
278 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
279 printf("%.102s\n", graph_dotted_line);
281 next = rb_first(root);
283 while (next && n_lines--) {
284 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
286 struct symbol *sym = NULL;
292 addr = data->call_site;
294 sym = machine__find_kernel_function(machine, addr, &map, NULL);
299 snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
300 addr - map->unmap_ip(map, sym->start));
302 snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
303 printf(" %-34s |", buf);
305 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
306 (unsigned long long)data->bytes_alloc,
307 (unsigned long)data->bytes_alloc / data->hit,
308 (unsigned long long)data->bytes_req,
309 (unsigned long)data->bytes_req / data->hit,
310 (unsigned long)data->hit,
311 (unsigned long)data->pingpong,
312 fragmentation(data->bytes_req, data->bytes_alloc));
314 next = rb_next(next);
318 printf(" ... | ... | ... | ... | ... | ... \n");
320 printf("%.102s\n", graph_dotted_line);
323 static void print_summary(void)
325 printf("\nSUMMARY\n=======\n");
326 printf("Total bytes requested: %lu\n", total_requested);
327 printf("Total bytes allocated: %lu\n", total_allocated);
328 printf("Total bytes wasted on internal fragmentation: %lu\n",
329 total_allocated - total_requested);
330 printf("Internal fragmentation: %f%%\n",
331 fragmentation(total_requested, total_allocated));
332 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
335 static void print_result(struct perf_session *session)
338 __print_result(&root_caller_sorted, session, caller_lines, 1);
340 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
344 struct sort_dimension {
347 struct list_head list;
350 static LIST_HEAD(caller_sort);
351 static LIST_HEAD(alloc_sort);
353 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
354 struct list_head *sort_list)
356 struct rb_node **new = &(root->rb_node);
357 struct rb_node *parent = NULL;
358 struct sort_dimension *sort;
361 struct alloc_stat *this;
364 this = rb_entry(*new, struct alloc_stat, node);
367 list_for_each_entry(sort, sort_list, list) {
368 cmp = sort->cmp(data, this);
374 new = &((*new)->rb_left);
376 new = &((*new)->rb_right);
379 rb_link_node(&data->node, parent, new);
380 rb_insert_color(&data->node, root);
383 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
384 struct list_head *sort_list)
386 struct rb_node *node;
387 struct alloc_stat *data;
390 node = rb_first(root);
394 rb_erase(node, root);
395 data = rb_entry(node, struct alloc_stat, node);
396 sort_insert(root_sorted, data, sort_list);
400 static void sort_result(void)
402 __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
403 __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
406 static int __cmd_kmem(void)
409 struct perf_session *session;
410 const struct perf_evsel_str_handler kmem_tracepoints[] = {
411 { "kmem:kmalloc", perf_evsel__process_alloc_event, },
412 { "kmem:kmem_cache_alloc", perf_evsel__process_alloc_event, },
413 { "kmem:kmalloc_node", perf_evsel__process_alloc_node_event, },
414 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
415 { "kmem:kfree", perf_evsel__process_free_event, },
416 { "kmem:kmem_cache_free", perf_evsel__process_free_event, },
418 struct perf_data_file file = {
420 .mode = PERF_DATA_MODE_READ,
423 session = perf_session__new(&file, false, &perf_kmem);
427 if (perf_session__create_kernel_maps(session) < 0)
430 if (!perf_session__has_traces(session, "kmem record"))
433 if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
434 pr_err("Initializing perf session tracepoint handlers failed\n");
439 err = perf_session__process_events(session, &perf_kmem);
443 print_result(session);
445 perf_session__delete(session);
449 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
453 else if (l->ptr > r->ptr)
458 static struct sort_dimension ptr_sort_dimension = {
463 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
465 if (l->call_site < r->call_site)
467 else if (l->call_site > r->call_site)
472 static struct sort_dimension callsite_sort_dimension = {
477 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
481 else if (l->hit > r->hit)
486 static struct sort_dimension hit_sort_dimension = {
491 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
493 if (l->bytes_alloc < r->bytes_alloc)
495 else if (l->bytes_alloc > r->bytes_alloc)
500 static struct sort_dimension bytes_sort_dimension = {
505 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
509 x = fragmentation(l->bytes_req, l->bytes_alloc);
510 y = fragmentation(r->bytes_req, r->bytes_alloc);
519 static struct sort_dimension frag_sort_dimension = {
524 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
526 if (l->pingpong < r->pingpong)
528 else if (l->pingpong > r->pingpong)
533 static struct sort_dimension pingpong_sort_dimension = {
538 static struct sort_dimension *avail_sorts[] = {
540 &callsite_sort_dimension,
542 &bytes_sort_dimension,
543 &frag_sort_dimension,
544 &pingpong_sort_dimension,
547 #define NUM_AVAIL_SORTS ((int)ARRAY_SIZE(avail_sorts))
549 static int sort_dimension__add(const char *tok, struct list_head *list)
551 struct sort_dimension *sort;
554 for (i = 0; i < NUM_AVAIL_SORTS; i++) {
555 if (!strcmp(avail_sorts[i]->name, tok)) {
556 sort = memdup(avail_sorts[i], sizeof(*avail_sorts[i]));
558 pr_err("%s: memdup failed\n", __func__);
561 list_add_tail(&sort->list, list);
569 static int setup_sorting(struct list_head *sort_list, const char *arg)
572 char *str = strdup(arg);
575 pr_err("%s: strdup failed\n", __func__);
580 tok = strsep(&str, ",");
583 if (sort_dimension__add(tok, sort_list) < 0) {
584 error("Unknown --sort key: '%s'", tok);
594 static int parse_sort_opt(const struct option *opt __maybe_unused,
595 const char *arg, int unset __maybe_unused)
600 if (caller_flag > alloc_flag)
601 return setup_sorting(&caller_sort, arg);
603 return setup_sorting(&alloc_sort, arg);
608 static int parse_caller_opt(const struct option *opt __maybe_unused,
609 const char *arg __maybe_unused,
610 int unset __maybe_unused)
612 caller_flag = (alloc_flag + 1);
616 static int parse_alloc_opt(const struct option *opt __maybe_unused,
617 const char *arg __maybe_unused,
618 int unset __maybe_unused)
620 alloc_flag = (caller_flag + 1);
624 static int parse_line_opt(const struct option *opt __maybe_unused,
625 const char *arg, int unset __maybe_unused)
632 lines = strtoul(arg, NULL, 10);
634 if (caller_flag > alloc_flag)
635 caller_lines = lines;
642 static int __cmd_record(int argc, const char **argv)
644 const char * const record_args[] = {
645 "record", "-a", "-R", "-c", "1",
646 "-e", "kmem:kmalloc",
647 "-e", "kmem:kmalloc_node",
649 "-e", "kmem:kmem_cache_alloc",
650 "-e", "kmem:kmem_cache_alloc_node",
651 "-e", "kmem:kmem_cache_free",
653 unsigned int rec_argc, i, j;
654 const char **rec_argv;
656 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
657 rec_argv = calloc(rec_argc + 1, sizeof(char *));
659 if (rec_argv == NULL)
662 for (i = 0; i < ARRAY_SIZE(record_args); i++)
663 rec_argv[i] = strdup(record_args[i]);
665 for (j = 1; j < (unsigned int)argc; j++, i++)
666 rec_argv[i] = argv[j];
668 return cmd_record(i, rec_argv, NULL);
671 int cmd_kmem(int argc, const char **argv, const char *prefix __maybe_unused)
673 const char * const default_sort_order = "frag,hit,bytes";
674 const struct option kmem_options[] = {
675 OPT_STRING('i', "input", &input_name, "file", "input file name"),
676 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
677 "show per-callsite statistics", parse_caller_opt),
678 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
679 "show per-allocation statistics", parse_alloc_opt),
680 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
681 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
683 OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
684 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
687 const char *const kmem_subcommands[] = { "record", "stat", NULL };
688 const char *kmem_usage[] = {
692 argc = parse_options_subcommand(argc, argv, kmem_options,
693 kmem_subcommands, kmem_usage, 0);
696 usage_with_options(kmem_usage, kmem_options);
700 if (!strncmp(argv[0], "rec", 3)) {
701 return __cmd_record(argc, argv);
702 } else if (!strcmp(argv[0], "stat")) {
703 if (cpu__setup_cpunode_map())
706 if (list_empty(&caller_sort))
707 setup_sorting(&caller_sort, default_sort_order);
708 if (list_empty(&alloc_sort))
709 setup_sorting(&alloc_sort, default_sort_order);
713 usage_with_options(kmem_usage, kmem_options);