perf sched: Do not delete session object prematurely
[platform/adaptation/renesas_rcar/renesas_kernel.git] / tools / perf / builtin-kmem.c
1 #include "builtin.h"
2 #include "perf.h"
3
4 #include "util/util.h"
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"
10
11 #include "util/parse-options.h"
12 #include "util/trace-event.h"
13
14 #include "util/debug.h"
15
16 #include <linux/rbtree.h>
17
18 struct alloc_stat;
19 typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
20
21 static char const               *input_name = "perf.data";
22
23 static int                      alloc_flag;
24 static int                      caller_flag;
25
26 static int                      alloc_lines = -1;
27 static int                      caller_lines = -1;
28
29 static bool                     raw_ip;
30
31 static char                     default_sort_order[] = "frag,hit,bytes";
32
33 static int                      *cpunode_map;
34 static int                      max_cpu_num;
35
36 struct alloc_stat {
37         u64     call_site;
38         u64     ptr;
39         u64     bytes_req;
40         u64     bytes_alloc;
41         u32     hit;
42         u32     pingpong;
43
44         short   alloc_cpu;
45
46         struct rb_node node;
47 };
48
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;
53
54 static unsigned long total_requested, total_allocated;
55 static unsigned long nr_allocs, nr_cross_allocs;
56
57 #define PATH_SYS_NODE   "/sys/devices/system/node"
58
59 static void init_cpunode_map(void)
60 {
61         FILE *fp;
62         int i;
63
64         fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
65         if (!fp) {
66                 max_cpu_num = 4096;
67                 return;
68         }
69
70         if (fscanf(fp, "%d", &max_cpu_num) < 1)
71                 die("Failed to read 'kernel_max' from sysfs");
72         max_cpu_num++;
73
74         cpunode_map = calloc(max_cpu_num, sizeof(int));
75         if (!cpunode_map)
76                 die("calloc");
77         for (i = 0; i < max_cpu_num; i++)
78                 cpunode_map[i] = -1;
79         fclose(fp);
80 }
81
82 static void setup_cpunode_map(void)
83 {
84         struct dirent *dent1, *dent2;
85         DIR *dir1, *dir2;
86         unsigned int cpu, mem;
87         char buf[PATH_MAX];
88
89         init_cpunode_map();
90
91         dir1 = opendir(PATH_SYS_NODE);
92         if (!dir1)
93                 return;
94
95         while ((dent1 = readdir(dir1)) != NULL) {
96                 if (dent1->d_type != DT_DIR ||
97                     sscanf(dent1->d_name, "node%u", &mem) < 1)
98                         continue;
99
100                 snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
101                 dir2 = opendir(buf);
102                 if (!dir2)
103                         continue;
104                 while ((dent2 = readdir(dir2)) != NULL) {
105                         if (dent2->d_type != DT_LNK ||
106                             sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
107                                 continue;
108                         cpunode_map[cpu] = mem;
109                 }
110         }
111 }
112
113 static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
114                               int bytes_req, int bytes_alloc, int cpu)
115 {
116         struct rb_node **node = &root_alloc_stat.rb_node;
117         struct rb_node *parent = NULL;
118         struct alloc_stat *data = NULL;
119
120         while (*node) {
121                 parent = *node;
122                 data = rb_entry(*node, struct alloc_stat, node);
123
124                 if (ptr > data->ptr)
125                         node = &(*node)->rb_right;
126                 else if (ptr < data->ptr)
127                         node = &(*node)->rb_left;
128                 else
129                         break;
130         }
131
132         if (data && data->ptr == ptr) {
133                 data->hit++;
134                 data->bytes_req += bytes_req;
135                 data->bytes_alloc += bytes_alloc;
136         } else {
137                 data = malloc(sizeof(*data));
138                 if (!data)
139                         die("malloc");
140                 data->ptr = ptr;
141                 data->pingpong = 0;
142                 data->hit = 1;
143                 data->bytes_req = bytes_req;
144                 data->bytes_alloc = bytes_alloc;
145
146                 rb_link_node(&data->node, parent, node);
147                 rb_insert_color(&data->node, &root_alloc_stat);
148         }
149         data->call_site = call_site;
150         data->alloc_cpu = cpu;
151 }
152
153 static void insert_caller_stat(unsigned long call_site,
154                               int bytes_req, int bytes_alloc)
155 {
156         struct rb_node **node = &root_caller_stat.rb_node;
157         struct rb_node *parent = NULL;
158         struct alloc_stat *data = NULL;
159
160         while (*node) {
161                 parent = *node;
162                 data = rb_entry(*node, struct alloc_stat, node);
163
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;
168                 else
169                         break;
170         }
171
172         if (data && data->call_site == call_site) {
173                 data->hit++;
174                 data->bytes_req += bytes_req;
175                 data->bytes_alloc += bytes_alloc;
176         } else {
177                 data = malloc(sizeof(*data));
178                 if (!data)
179                         die("malloc");
180                 data->call_site = call_site;
181                 data->pingpong = 0;
182                 data->hit = 1;
183                 data->bytes_req = bytes_req;
184                 data->bytes_alloc = bytes_alloc;
185
186                 rb_link_node(&data->node, parent, node);
187                 rb_insert_color(&data->node, &root_caller_stat);
188         }
189 }
190
191 static void process_alloc_event(void *data,
192                                 struct event *event,
193                                 int cpu,
194                                 u64 timestamp __used,
195                                 struct thread *thread __used,
196                                 int node)
197 {
198         unsigned long call_site;
199         unsigned long ptr;
200         int bytes_req;
201         int bytes_alloc;
202         int node1, node2;
203
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);
208
209         insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
210         insert_caller_stat(call_site, bytes_req, bytes_alloc);
211
212         total_requested += bytes_req;
213         total_allocated += bytes_alloc;
214
215         if (node) {
216                 node1 = cpunode_map[cpu];
217                 node2 = raw_field_value(event, "node", data);
218                 if (node1 != node2)
219                         nr_cross_allocs++;
220         }
221         nr_allocs++;
222 }
223
224 static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
225 static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
226
227 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
228                                             unsigned long call_site,
229                                             struct rb_root *root,
230                                             sort_fn_t sort_fn)
231 {
232         struct rb_node *node = root->rb_node;
233         struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
234
235         while (node) {
236                 struct alloc_stat *data;
237                 int cmp;
238
239                 data = rb_entry(node, struct alloc_stat, node);
240
241                 cmp = sort_fn(&key, data);
242                 if (cmp < 0)
243                         node = node->rb_left;
244                 else if (cmp > 0)
245                         node = node->rb_right;
246                 else
247                         return data;
248         }
249         return NULL;
250 }
251
252 static void process_free_event(void *data,
253                                struct event *event,
254                                int cpu,
255                                u64 timestamp __used,
256                                struct thread *thread __used)
257 {
258         unsigned long ptr;
259         struct alloc_stat *s_alloc, *s_caller;
260
261         ptr = raw_field_value(event, "ptr", data);
262
263         s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
264         if (!s_alloc)
265                 return;
266
267         if (cpu != s_alloc->alloc_cpu) {
268                 s_alloc->pingpong++;
269
270                 s_caller = search_alloc_stat(0, s_alloc->call_site,
271                                              &root_caller_stat, callsite_cmp);
272                 assert(s_caller);
273                 s_caller->pingpong++;
274         }
275         s_alloc->alloc_cpu = -1;
276 }
277
278 static void process_raw_event(union perf_event *raw_event __used, void *data,
279                               int cpu, u64 timestamp, struct thread *thread)
280 {
281         struct event *event;
282         int type;
283
284         type = trace_parse_common_type(data);
285         event = trace_find_event(type);
286
287         if (!strcmp(event->name, "kmalloc") ||
288             !strcmp(event->name, "kmem_cache_alloc")) {
289                 process_alloc_event(data, event, cpu, timestamp, thread, 0);
290                 return;
291         }
292
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);
296                 return;
297         }
298
299         if (!strcmp(event->name, "kfree") ||
300             !strcmp(event->name, "kmem_cache_free")) {
301                 process_free_event(data, event, cpu, timestamp, thread);
302                 return;
303         }
304 }
305
306 static int process_sample_event(union perf_event *event,
307                                 struct perf_sample *sample,
308                                 struct perf_evsel *evsel __used,
309                                 struct perf_session *session)
310 {
311         struct thread *thread = perf_session__findnew(session, event->ip.pid);
312
313         if (thread == NULL) {
314                 pr_debug("problem processing %d event, skipping it.\n",
315                          event->header.type);
316                 return -1;
317         }
318
319         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
320
321         process_raw_event(event, sample->raw_data, sample->cpu,
322                           sample->time, thread);
323
324         return 0;
325 }
326
327 static struct perf_event_ops event_ops = {
328         .sample                 = process_sample_event,
329         .comm                   = perf_event__process_comm,
330         .ordered_samples        = true,
331 };
332
333 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
334 {
335         if (n_alloc == 0)
336                 return 0.0;
337         else
338                 return 100.0 - (100.0 * n_req / n_alloc);
339 }
340
341 static void __print_result(struct rb_root *root, struct perf_session *session,
342                            int n_lines, int is_caller)
343 {
344         struct rb_node *next;
345         struct machine *machine;
346
347         printf("%.102s\n", graph_dotted_line);
348         printf(" %-34s |",  is_caller ? "Callsite": "Alloc Ptr");
349         printf(" Total_alloc/Per | Total_req/Per   | Hit      | Ping-pong | Frag\n");
350         printf("%.102s\n", graph_dotted_line);
351
352         next = rb_first(root);
353
354         machine = perf_session__find_host_machine(session);
355         if (!machine) {
356                 pr_err("__print_result: couldn't find kernel information\n");
357                 return;
358         }
359         while (next && n_lines--) {
360                 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
361                                                    node);
362                 struct symbol *sym = NULL;
363                 struct map *map;
364                 char buf[BUFSIZ];
365                 u64 addr;
366
367                 if (is_caller) {
368                         addr = data->call_site;
369                         if (!raw_ip)
370                                 sym = machine__find_kernel_function(machine, addr, &map, NULL);
371                 } else
372                         addr = data->ptr;
373
374                 if (sym != NULL)
375                         snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
376                                  addr - map->unmap_ip(map, sym->start));
377                 else
378                         snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
379                 printf(" %-34s |", buf);
380
381                 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
382                        (unsigned long long)data->bytes_alloc,
383                        (unsigned long)data->bytes_alloc / data->hit,
384                        (unsigned long long)data->bytes_req,
385                        (unsigned long)data->bytes_req / data->hit,
386                        (unsigned long)data->hit,
387                        (unsigned long)data->pingpong,
388                        fragmentation(data->bytes_req, data->bytes_alloc));
389
390                 next = rb_next(next);
391         }
392
393         if (n_lines == -1)
394                 printf(" ...                                | ...             | ...             | ...    | ...      | ...   \n");
395
396         printf("%.102s\n", graph_dotted_line);
397 }
398
399 static void print_summary(void)
400 {
401         printf("\nSUMMARY\n=======\n");
402         printf("Total bytes requested: %lu\n", total_requested);
403         printf("Total bytes allocated: %lu\n", total_allocated);
404         printf("Total bytes wasted on internal fragmentation: %lu\n",
405                total_allocated - total_requested);
406         printf("Internal fragmentation: %f%%\n",
407                fragmentation(total_requested, total_allocated));
408         printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
409 }
410
411 static void print_result(struct perf_session *session)
412 {
413         if (caller_flag)
414                 __print_result(&root_caller_sorted, session, caller_lines, 1);
415         if (alloc_flag)
416                 __print_result(&root_alloc_sorted, session, alloc_lines, 0);
417         print_summary();
418 }
419
420 struct sort_dimension {
421         const char              name[20];
422         sort_fn_t               cmp;
423         struct list_head        list;
424 };
425
426 static LIST_HEAD(caller_sort);
427 static LIST_HEAD(alloc_sort);
428
429 static void sort_insert(struct rb_root *root, struct alloc_stat *data,
430                         struct list_head *sort_list)
431 {
432         struct rb_node **new = &(root->rb_node);
433         struct rb_node *parent = NULL;
434         struct sort_dimension *sort;
435
436         while (*new) {
437                 struct alloc_stat *this;
438                 int cmp = 0;
439
440                 this = rb_entry(*new, struct alloc_stat, node);
441                 parent = *new;
442
443                 list_for_each_entry(sort, sort_list, list) {
444                         cmp = sort->cmp(data, this);
445                         if (cmp)
446                                 break;
447                 }
448
449                 if (cmp > 0)
450                         new = &((*new)->rb_left);
451                 else
452                         new = &((*new)->rb_right);
453         }
454
455         rb_link_node(&data->node, parent, new);
456         rb_insert_color(&data->node, root);
457 }
458
459 static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
460                           struct list_head *sort_list)
461 {
462         struct rb_node *node;
463         struct alloc_stat *data;
464
465         for (;;) {
466                 node = rb_first(root);
467                 if (!node)
468                         break;
469
470                 rb_erase(node, root);
471                 data = rb_entry(node, struct alloc_stat, node);
472                 sort_insert(root_sorted, data, sort_list);
473         }
474 }
475
476 static void sort_result(void)
477 {
478         __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
479         __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
480 }
481
482 static int __cmd_kmem(void)
483 {
484         int err = -EINVAL;
485         struct perf_session *session = perf_session__new(input_name, O_RDONLY,
486                                                          0, false, &event_ops);
487         if (session == NULL)
488                 return -ENOMEM;
489
490         if (perf_session__create_kernel_maps(session) < 0)
491                 goto out_delete;
492
493         if (!perf_session__has_traces(session, "kmem record"))
494                 goto out_delete;
495
496         setup_pager();
497         err = perf_session__process_events(session, &event_ops);
498         if (err != 0)
499                 goto out_delete;
500         sort_result();
501         print_result(session);
502 out_delete:
503         perf_session__delete(session);
504         return err;
505 }
506
507 static const char * const kmem_usage[] = {
508         "perf kmem [<options>] {record|stat}",
509         NULL
510 };
511
512 static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
513 {
514         if (l->ptr < r->ptr)
515                 return -1;
516         else if (l->ptr > r->ptr)
517                 return 1;
518         return 0;
519 }
520
521 static struct sort_dimension ptr_sort_dimension = {
522         .name   = "ptr",
523         .cmp    = ptr_cmp,
524 };
525
526 static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
527 {
528         if (l->call_site < r->call_site)
529                 return -1;
530         else if (l->call_site > r->call_site)
531                 return 1;
532         return 0;
533 }
534
535 static struct sort_dimension callsite_sort_dimension = {
536         .name   = "callsite",
537         .cmp    = callsite_cmp,
538 };
539
540 static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
541 {
542         if (l->hit < r->hit)
543                 return -1;
544         else if (l->hit > r->hit)
545                 return 1;
546         return 0;
547 }
548
549 static struct sort_dimension hit_sort_dimension = {
550         .name   = "hit",
551         .cmp    = hit_cmp,
552 };
553
554 static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
555 {
556         if (l->bytes_alloc < r->bytes_alloc)
557                 return -1;
558         else if (l->bytes_alloc > r->bytes_alloc)
559                 return 1;
560         return 0;
561 }
562
563 static struct sort_dimension bytes_sort_dimension = {
564         .name   = "bytes",
565         .cmp    = bytes_cmp,
566 };
567
568 static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
569 {
570         double x, y;
571
572         x = fragmentation(l->bytes_req, l->bytes_alloc);
573         y = fragmentation(r->bytes_req, r->bytes_alloc);
574
575         if (x < y)
576                 return -1;
577         else if (x > y)
578                 return 1;
579         return 0;
580 }
581
582 static struct sort_dimension frag_sort_dimension = {
583         .name   = "frag",
584         .cmp    = frag_cmp,
585 };
586
587 static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
588 {
589         if (l->pingpong < r->pingpong)
590                 return -1;
591         else if (l->pingpong > r->pingpong)
592                 return 1;
593         return 0;
594 }
595
596 static struct sort_dimension pingpong_sort_dimension = {
597         .name   = "pingpong",
598         .cmp    = pingpong_cmp,
599 };
600
601 static struct sort_dimension *avail_sorts[] = {
602         &ptr_sort_dimension,
603         &callsite_sort_dimension,
604         &hit_sort_dimension,
605         &bytes_sort_dimension,
606         &frag_sort_dimension,
607         &pingpong_sort_dimension,
608 };
609
610 #define NUM_AVAIL_SORTS \
611         (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
612
613 static int sort_dimension__add(const char *tok, struct list_head *list)
614 {
615         struct sort_dimension *sort;
616         int i;
617
618         for (i = 0; i < NUM_AVAIL_SORTS; i++) {
619                 if (!strcmp(avail_sorts[i]->name, tok)) {
620                         sort = malloc(sizeof(*sort));
621                         if (!sort)
622                                 die("malloc");
623                         memcpy(sort, avail_sorts[i], sizeof(*sort));
624                         list_add_tail(&sort->list, list);
625                         return 0;
626                 }
627         }
628
629         return -1;
630 }
631
632 static int setup_sorting(struct list_head *sort_list, const char *arg)
633 {
634         char *tok;
635         char *str = strdup(arg);
636
637         if (!str)
638                 die("strdup");
639
640         while (true) {
641                 tok = strsep(&str, ",");
642                 if (!tok)
643                         break;
644                 if (sort_dimension__add(tok, sort_list) < 0) {
645                         error("Unknown --sort key: '%s'", tok);
646                         return -1;
647                 }
648         }
649
650         free(str);
651         return 0;
652 }
653
654 static int parse_sort_opt(const struct option *opt __used,
655                           const char *arg, int unset __used)
656 {
657         if (!arg)
658                 return -1;
659
660         if (caller_flag > alloc_flag)
661                 return setup_sorting(&caller_sort, arg);
662         else
663                 return setup_sorting(&alloc_sort, arg);
664
665         return 0;
666 }
667
668 static int parse_caller_opt(const struct option *opt __used,
669                           const char *arg __used, int unset __used)
670 {
671         caller_flag = (alloc_flag + 1);
672         return 0;
673 }
674
675 static int parse_alloc_opt(const struct option *opt __used,
676                           const char *arg __used, int unset __used)
677 {
678         alloc_flag = (caller_flag + 1);
679         return 0;
680 }
681
682 static int parse_line_opt(const struct option *opt __used,
683                           const char *arg, int unset __used)
684 {
685         int lines;
686
687         if (!arg)
688                 return -1;
689
690         lines = strtoul(arg, NULL, 10);
691
692         if (caller_flag > alloc_flag)
693                 caller_lines = lines;
694         else
695                 alloc_lines = lines;
696
697         return 0;
698 }
699
700 static const struct option kmem_options[] = {
701         OPT_STRING('i', "input", &input_name, "file",
702                    "input file name"),
703         OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
704                            "show per-callsite statistics",
705                            parse_caller_opt),
706         OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
707                            "show per-allocation statistics",
708                            parse_alloc_opt),
709         OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
710                      "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
711                      parse_sort_opt),
712         OPT_CALLBACK('l', "line", NULL, "num",
713                      "show n lines",
714                      parse_line_opt),
715         OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
716         OPT_END()
717 };
718
719 static const char *record_args[] = {
720         "record",
721         "-a",
722         "-R",
723         "-f",
724         "-c", "1",
725         "-e", "kmem:kmalloc",
726         "-e", "kmem:kmalloc_node",
727         "-e", "kmem:kfree",
728         "-e", "kmem:kmem_cache_alloc",
729         "-e", "kmem:kmem_cache_alloc_node",
730         "-e", "kmem:kmem_cache_free",
731 };
732
733 static int __cmd_record(int argc, const char **argv)
734 {
735         unsigned int rec_argc, i, j;
736         const char **rec_argv;
737
738         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
739         rec_argv = calloc(rec_argc + 1, sizeof(char *));
740
741         if (rec_argv == NULL)
742                 return -ENOMEM;
743
744         for (i = 0; i < ARRAY_SIZE(record_args); i++)
745                 rec_argv[i] = strdup(record_args[i]);
746
747         for (j = 1; j < (unsigned int)argc; j++, i++)
748                 rec_argv[i] = argv[j];
749
750         return cmd_record(i, rec_argv, NULL);
751 }
752
753 int cmd_kmem(int argc, const char **argv, const char *prefix __used)
754 {
755         argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
756
757         if (!argc)
758                 usage_with_options(kmem_usage, kmem_options);
759
760         symbol__init();
761
762         if (!strncmp(argv[0], "rec", 3)) {
763                 return __cmd_record(argc, argv);
764         } else if (!strcmp(argv[0], "stat")) {
765                 setup_cpunode_map();
766
767                 if (list_empty(&caller_sort))
768                         setup_sorting(&caller_sort, default_sort_order);
769                 if (list_empty(&alloc_sort))
770                         setup_sorting(&alloc_sort, default_sort_order);
771
772                 return __cmd_kmem();
773         } else
774                 usage_with_options(kmem_usage, kmem_options);
775
776         return 0;
777 }
778