4 * Builtin report command: Analyze the perf.data input file,
5 * look up and read DSOs and symbol information and display
6 * a histogram of results, along various sorting keys.
10 #include "util/util.h"
12 #include "util/list.h"
13 #include "util/cache.h"
14 #include "util/rbtree.h"
15 #include "util/symbol.h"
16 #include "util/string.h"
20 #include "util/parse-options.h"
21 #include "util/parse-events.h"
27 static char const *input_name = "perf.data";
28 static char *vmlinux = NULL;
29 static char *sort_order = "comm,dso";
31 static int show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
33 static int dump_trace = 0;
34 #define dprintf(x...) do { if (dump_trace) printf(x); } while (0)
37 static int full_paths;
39 static unsigned long page_size;
40 static unsigned long mmap_window = 32;
42 const char *perf_event_names[] = {
43 [PERF_EVENT_MMAP] = " PERF_EVENT_MMAP",
44 [PERF_EVENT_MUNMAP] = " PERF_EVENT_MUNMAP",
45 [PERF_EVENT_COMM] = " PERF_EVENT_COMM",
49 struct perf_event_header header;
55 struct perf_event_header header;
60 char filename[PATH_MAX];
64 struct perf_event_header header;
69 typedef union event_union {
70 struct perf_event_header header;
72 struct mmap_event mmap;
73 struct comm_event comm;
76 static LIST_HEAD(dsos);
77 static struct dso *kernel_dso;
79 static void dsos__add(struct dso *dso)
81 list_add_tail(&dso->node, &dsos);
84 static struct dso *dsos__find(const char *name)
88 list_for_each_entry(pos, &dsos, node)
89 if (strcmp(pos->name, name) == 0)
94 static struct dso *dsos__findnew(const char *name)
96 struct dso *dso = dsos__find(name);
102 dso = dso__new(name, 0);
106 nr = dso__load(dso, NULL);
108 fprintf(stderr, "Failed to open: %s\n", name);
111 if (!nr && verbose) {
113 "No symbols found in: %s, maybe install a debug package?\n",
126 static void dsos__fprintf(FILE *fp)
130 list_for_each_entry(pos, &dsos, node)
131 dso__fprintf(pos, fp);
134 static int load_kernel(void)
138 kernel_dso = dso__new("[kernel]", 0);
142 err = dso__load_kernel(kernel_dso, vmlinux, NULL);
144 dso__delete(kernel_dso);
147 dsos__add(kernel_dso);
152 static char __cwd[PATH_MAX];
153 static char *cwd = __cwd;
156 static int strcommon(const char *pathname)
160 while (pathname[n] == cwd[n] && n < cwdlen)
167 struct list_head node;
174 static struct map *map__new(struct mmap_event *event)
176 struct map *self = malloc(sizeof(*self));
179 const char *filename = event->filename;
180 char newfilename[PATH_MAX];
183 int n = strcommon(filename);
186 snprintf(newfilename, sizeof(newfilename),
187 ".%s", filename + n);
188 filename = newfilename;
192 self->start = event->start;
193 self->end = event->start + event->len;
194 self->pgoff = event->pgoff;
196 self->dso = dsos__findnew(filename);
197 if (self->dso == NULL)
209 struct rb_node rb_node;
210 struct list_head maps;
215 static struct thread *thread__new(pid_t pid)
217 struct thread *self = malloc(sizeof(*self));
221 self->comm = malloc(32);
223 snprintf(self->comm, 32, ":%d", self->pid);
224 INIT_LIST_HEAD(&self->maps);
230 static int thread__set_comm(struct thread *self, const char *comm)
234 self->comm = strdup(comm);
235 return self->comm ? 0 : -ENOMEM;
238 static struct rb_root threads;
239 static struct thread *last_match;
241 static struct thread *threads__findnew(pid_t pid)
243 struct rb_node **p = &threads.rb_node;
244 struct rb_node *parent = NULL;
248 * Font-end cache - PID lookups come in blocks,
249 * so most of the time we dont have to look up
252 if (last_match && last_match->pid == pid)
257 th = rb_entry(parent, struct thread, rb_node);
259 if (th->pid == pid) {
270 th = thread__new(pid);
272 rb_link_node(&th->rb_node, parent, p);
273 rb_insert_color(&th->rb_node, &threads);
280 static void thread__insert_map(struct thread *self, struct map *map)
282 list_add_tail(&map->node, &self->maps);
285 static struct map *thread__find_map(struct thread *self, uint64_t ip)
292 list_for_each_entry(pos, &self->maps, node)
293 if (ip >= pos->start && ip <= pos->end)
300 * histogram, sorted on item, collects counts
303 static struct rb_root hist;
306 struct rb_node rb_node;
308 struct thread *thread;
319 * configurable sorting bits
323 struct list_head list;
327 int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
328 int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
329 size_t (*print)(FILE *fp, struct hist_entry *);
335 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
337 return right->thread->pid - left->thread->pid;
341 sort__thread_print(FILE *fp, struct hist_entry *self)
343 return fprintf(fp, " %16s:%5d", self->thread->comm ?: "", self->thread->pid);
346 static struct sort_entry sort_thread = {
347 .header = " Command: Pid ",
348 .cmp = sort__thread_cmp,
349 .print = sort__thread_print,
355 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
357 return right->thread->pid - left->thread->pid;
361 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
363 char *comm_l = left->thread->comm;
364 char *comm_r = right->thread->comm;
366 if (!comm_l || !comm_r) {
367 if (!comm_l && !comm_r)
375 return strcmp(comm_l, comm_r);
379 sort__comm_print(FILE *fp, struct hist_entry *self)
381 return fprintf(fp, " %16s", self->thread->comm);
384 static struct sort_entry sort_comm = {
385 .header = " Command",
386 .cmp = sort__comm_cmp,
387 .collapse = sort__comm_collapse,
388 .print = sort__comm_print,
394 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
396 struct dso *dso_l = left->dso;
397 struct dso *dso_r = right->dso;
399 if (!dso_l || !dso_r) {
400 if (!dso_l && !dso_r)
408 return strcmp(dso_l->name, dso_r->name);
412 sort__dso_print(FILE *fp, struct hist_entry *self)
415 return fprintf(fp, " %-25s", self->dso->name);
417 return fprintf(fp, " %016llx", (__u64)self->ip);
420 static struct sort_entry sort_dso = {
421 .header = " Shared Object ",
422 .cmp = sort__dso_cmp,
423 .print = sort__dso_print,
429 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
433 if (left->sym == right->sym)
436 ip_l = left->sym ? left->sym->start : left->ip;
437 ip_r = right->sym ? right->sym->start : right->ip;
439 return (int64_t)(ip_r - ip_l);
443 sort__sym_print(FILE *fp, struct hist_entry *self)
448 ret += fprintf(fp, " %#018llx", (__u64)self->ip);
451 ret += fprintf(fp, " %s: ", self->dso->name);
453 ret += fprintf(fp, " %#016llx: ", (__u64)self->ip);
456 ret += fprintf(fp, "%s", self->sym->name);
458 ret += fprintf(fp, "%#016llx", (__u64)self->ip);
463 static struct sort_entry sort_sym = {
464 .header = " Shared Object: Symbol",
465 .cmp = sort__sym_cmp,
466 .print = sort__sym_print,
469 static int sort__need_collapse = 0;
471 struct sort_dimension {
473 struct sort_entry *entry;
477 static struct sort_dimension sort_dimensions[] = {
478 { .name = "pid", .entry = &sort_thread, },
479 { .name = "comm", .entry = &sort_comm, },
480 { .name = "dso", .entry = &sort_dso, },
481 { .name = "symbol", .entry = &sort_sym, },
484 static LIST_HEAD(hist_entry__sort_list);
486 static int sort_dimension__add(char *tok)
490 for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
491 struct sort_dimension *sd = &sort_dimensions[i];
496 if (strncasecmp(tok, sd->name, strlen(tok)))
499 if (sd->entry->collapse)
500 sort__need_collapse = 1;
502 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
512 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
514 struct sort_entry *se;
517 list_for_each_entry(se, &hist_entry__sort_list, list) {
518 cmp = se->cmp(left, right);
527 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
529 struct sort_entry *se;
532 list_for_each_entry(se, &hist_entry__sort_list, list) {
533 int64_t (*f)(struct hist_entry *, struct hist_entry *);
535 f = se->collapse ?: se->cmp;
537 cmp = f(left, right);
546 hist_entry__fprintf(FILE *fp, struct hist_entry *self, uint64_t total_samples)
548 struct sort_entry *se;
552 ret = fprintf(fp, " %6.2f%%",
553 (self->count * 100.0) / total_samples);
555 ret = fprintf(fp, "%12d ", self->count);
557 list_for_each_entry(se, &hist_entry__sort_list, list)
558 ret += se->print(fp, self);
560 ret += fprintf(fp, "\n");
566 * collect histogram counts
570 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
571 struct symbol *sym, uint64_t ip, char level)
573 struct rb_node **p = &hist.rb_node;
574 struct rb_node *parent = NULL;
575 struct hist_entry *he;
576 struct hist_entry entry = {
589 he = rb_entry(parent, struct hist_entry, rb_node);
591 cmp = hist_entry__cmp(&entry, he);
604 he = malloc(sizeof(*he));
608 rb_link_node(&he->rb_node, parent, p);
609 rb_insert_color(&he->rb_node, &hist);
614 static void hist_entry__free(struct hist_entry *he)
620 * collapse the histogram
623 static struct rb_root collapse_hists;
625 static void collapse__insert_entry(struct hist_entry *he)
627 struct rb_node **p = &collapse_hists.rb_node;
628 struct rb_node *parent = NULL;
629 struct hist_entry *iter;
634 iter = rb_entry(parent, struct hist_entry, rb_node);
636 cmp = hist_entry__collapse(iter, he);
639 iter->count += he->count;
640 hist_entry__free(he);
650 rb_link_node(&he->rb_node, parent, p);
651 rb_insert_color(&he->rb_node, &collapse_hists);
654 static void collapse__resort(void)
656 struct rb_node *next;
657 struct hist_entry *n;
659 if (!sort__need_collapse)
662 next = rb_first(&hist);
664 n = rb_entry(next, struct hist_entry, rb_node);
665 next = rb_next(&n->rb_node);
667 rb_erase(&n->rb_node, &hist);
668 collapse__insert_entry(n);
673 * reverse the map, sort on count.
676 static struct rb_root output_hists;
678 static void output__insert_entry(struct hist_entry *he)
680 struct rb_node **p = &output_hists.rb_node;
681 struct rb_node *parent = NULL;
682 struct hist_entry *iter;
686 iter = rb_entry(parent, struct hist_entry, rb_node);
688 if (he->count > iter->count)
694 rb_link_node(&he->rb_node, parent, p);
695 rb_insert_color(&he->rb_node, &output_hists);
698 static void output__resort(void)
700 struct rb_node *next;
701 struct hist_entry *n;
703 if (sort__need_collapse)
704 next = rb_first(&collapse_hists);
706 next = rb_first(&hist);
709 n = rb_entry(next, struct hist_entry, rb_node);
710 next = rb_next(&n->rb_node);
712 rb_erase(&n->rb_node, &hist);
713 output__insert_entry(n);
717 static size_t output__fprintf(FILE *fp, uint64_t total_samples)
719 struct hist_entry *pos;
720 struct sort_entry *se;
726 fprintf(fp, "# Overhead");
727 list_for_each_entry(se, &hist_entry__sort_list, list)
728 fprintf(fp, " %s", se->header);
731 fprintf(fp, "# ........");
732 list_for_each_entry(se, &hist_entry__sort_list, list) {
736 for (i = 0; i < strlen(se->header)-1; i++)
743 for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
744 pos = rb_entry(nd, struct hist_entry, rb_node);
745 ret += hist_entry__fprintf(fp, pos, total_samples);
751 static void register_idle_thread(void)
753 struct thread *thread = threads__findnew(0);
755 if (thread == NULL ||
756 thread__set_comm(thread, "[idle]")) {
757 fprintf(stderr, "problem inserting idle task.\n");
762 static unsigned long total = 0, total_mmap = 0, total_comm = 0, total_unknown = 0;
765 process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
769 struct dso *dso = NULL;
770 struct thread *thread = threads__findnew(event->ip.pid);
771 uint64_t ip = event->ip.ip;
772 struct map *map = NULL;
774 dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
775 (void *)(offset + head),
776 (void *)(long)(event->header.size),
781 dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
783 if (thread == NULL) {
784 fprintf(stderr, "problem processing %d event, skipping it.\n",
789 if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
795 dprintf(" ...... dso: %s\n", dso->name);
797 } else if (event->header.misc & PERF_EVENT_MISC_USER) {
802 map = thread__find_map(thread, ip);
805 ip -= map->start + map->pgoff;
808 * If this is outside of all known maps,
809 * and is a negative address, try to look it
810 * up in the kernel dso, as it might be a
811 * vsyscall (which executes in user-mode):
813 if ((long long)ip < 0)
816 dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
821 dprintf(" ...... dso: [hypervisor]\n");
824 if (show & show_mask) {
825 struct symbol *sym = dso__find_symbol(dso, ip);
827 if (hist_entry__add(thread, map, dso, sym, ip, level)) {
829 "problem incrementing symbol count, skipping event\n");
839 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
841 struct thread *thread = threads__findnew(event->mmap.pid);
842 struct map *map = map__new(&event->mmap);
844 dprintf("%p [%p]: PERF_EVENT_MMAP: [%p(%p) @ %p]: %s\n",
845 (void *)(offset + head),
846 (void *)(long)(event->header.size),
847 (void *)(long)event->mmap.start,
848 (void *)(long)event->mmap.len,
849 (void *)(long)event->mmap.pgoff,
850 event->mmap.filename);
852 if (thread == NULL || map == NULL) {
853 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
857 thread__insert_map(thread, map);
864 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
866 struct thread *thread = threads__findnew(event->comm.pid);
868 dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
869 (void *)(offset + head),
870 (void *)(long)(event->header.size),
871 event->comm.comm, event->comm.pid);
873 if (thread == NULL ||
874 thread__set_comm(thread, event->comm.comm)) {
875 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
884 process_event(event_t *event, unsigned long offset, unsigned long head)
886 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW)
887 return process_overflow_event(event, offset, head);
889 switch (event->header.type) {
890 case PERF_EVENT_MMAP:
891 return process_mmap_event(event, offset, head);
893 case PERF_EVENT_COMM:
894 return process_comm_event(event, offset, head);
897 * We dont process them right now but they are fine:
899 case PERF_EVENT_MUNMAP:
900 case PERF_EVENT_PERIOD:
901 case PERF_EVENT_THROTTLE:
902 case PERF_EVENT_UNTHROTTLE:
912 static int __cmd_report(void)
914 int ret, rc = EXIT_FAILURE;
915 unsigned long offset = 0;
916 unsigned long head = 0;
922 register_idle_thread();
924 input = open(input_name, O_RDONLY);
926 perror("failed to open file");
930 ret = fstat(input, &stat);
932 perror("failed to stat file");
937 fprintf(stderr, "zero-sized file, nothing to do!\n");
941 if (load_kernel() < 0) {
942 perror("failed to load kernel symbols");
947 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
948 perror("failed to get the current directory");
951 cwdlen = strlen(cwd);
957 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
958 MAP_SHARED, input, offset);
959 if (buf == MAP_FAILED) {
960 perror("failed to mmap file");
965 event = (event_t *)(buf + head);
967 size = event->header.size;
971 if (head + event->header.size >= page_size * mmap_window) {
972 unsigned long shift = page_size * (head / page_size);
975 ret = munmap(buf, page_size * mmap_window);
983 size = event->header.size;
985 if (!size || process_event(event, offset, head) < 0) {
987 dprintf("%p [%p]: skipping unknown header type: %d\n",
988 (void *)(offset + head),
989 (void *)(long)(event->header.size),
995 * assume we lost track of the stream, check alignment, and
996 * increment a single u64 in the hope to catch on again 'soon'.
999 if (unlikely(head & 7))
1007 if (offset + head < stat.st_size)
1013 dprintf(" IP events: %10ld\n", total);
1014 dprintf(" mmap events: %10ld\n", total_mmap);
1015 dprintf(" comm events: %10ld\n", total_comm);
1016 dprintf(" unknown events: %10ld\n", total_unknown);
1022 dsos__fprintf(stdout);
1026 output__fprintf(stdout, total);
1031 static const char * const report_usage[] = {
1032 "perf report [<options>] <command>",
1036 static const struct option options[] = {
1037 OPT_STRING('i', "input", &input_name, "file",
1039 OPT_BOOLEAN('v', "verbose", &verbose,
1040 "be more verbose (show symbol address, etc)"),
1041 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1042 "dump raw trace in ASCII"),
1043 OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1044 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1045 "sort by key(s): pid, comm, dso, symbol. Default: pid,symbol"),
1046 OPT_BOOLEAN('P', "full-paths", &full_paths,
1047 "Don't shorten the pathnames taking into account the cwd"),
1051 static void setup_sorting(void)
1053 char *tmp, *tok, *str = strdup(sort_order);
1055 for (tok = strtok_r(str, ", ", &tmp);
1056 tok; tok = strtok_r(NULL, ", ", &tmp)) {
1057 if (sort_dimension__add(tok) < 0) {
1058 error("Unknown --sort key: `%s'", tok);
1059 usage_with_options(report_usage, options);
1066 int cmd_report(int argc, const char **argv, const char *prefix)
1070 page_size = getpagesize();
1072 parse_options(argc, argv, options, report_usage, 0);
1078 return __cmd_report();