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/color.h"
13 #include "util/list.h"
14 #include "util/cache.h"
15 #include "util/rbtree.h"
16 #include "util/symbol.h"
17 #include "util/string.h"
21 #include "util/parse-options.h"
22 #include "util/parse-events.h"
28 static char const *input_name = "perf.data";
29 static char *vmlinux = NULL;
31 static char default_sort_order[] = "comm,dso";
32 static char *sort_order = default_sort_order;
35 static int show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
37 static int dump_trace = 0;
38 #define dprintf(x...) do { if (dump_trace) printf(x); } while (0)
41 static int full_paths;
43 static unsigned long page_size;
44 static unsigned long mmap_window = 32;
47 struct perf_event_header header;
54 struct perf_event_header header;
59 char filename[PATH_MAX];
63 struct perf_event_header header;
69 struct perf_event_header header;
74 struct perf_event_header header;
80 typedef union event_union {
81 struct perf_event_header header;
83 struct mmap_event mmap;
84 struct comm_event comm;
85 struct fork_event fork;
86 struct period_event period;
89 static LIST_HEAD(dsos);
90 static struct dso *kernel_dso;
91 static struct dso *vdso;
93 static void dsos__add(struct dso *dso)
95 list_add_tail(&dso->node, &dsos);
98 static struct dso *dsos__find(const char *name)
102 list_for_each_entry(pos, &dsos, node)
103 if (strcmp(pos->name, name) == 0)
108 static struct dso *dsos__findnew(const char *name)
110 struct dso *dso = dsos__find(name);
116 dso = dso__new(name, 0);
120 nr = dso__load(dso, NULL, verbose);
123 fprintf(stderr, "Failed to open: %s\n", name);
126 if (!nr && verbose) {
128 "No symbols found in: %s, maybe install a debug package?\n",
141 static void dsos__fprintf(FILE *fp)
145 list_for_each_entry(pos, &dsos, node)
146 dso__fprintf(pos, fp);
149 static struct symbol *vdso__find_symbol(struct dso *dso, __u64 ip)
151 return dso__find_symbol(kernel_dso, ip);
154 static int load_kernel(void)
158 kernel_dso = dso__new("[kernel]", 0);
162 err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose);
164 dso__delete(kernel_dso);
167 dsos__add(kernel_dso);
169 vdso = dso__new("[vdso]", 0);
173 vdso->find_symbol = vdso__find_symbol;
180 static char __cwd[PATH_MAX];
181 static char *cwd = __cwd;
184 static int strcommon(const char *pathname)
188 while (pathname[n] == cwd[n] && n < cwdlen)
195 struct list_head node;
199 __u64 (*map_ip)(struct map *, __u64);
203 static __u64 map__map_ip(struct map *map, __u64 ip)
205 return ip - map->start + map->pgoff;
208 static __u64 vdso__map_ip(struct map *map, __u64 ip)
213 static inline int is_anon_memory(const char *filename)
215 return strcmp(filename, "//anon") == 0;
218 static struct map *map__new(struct mmap_event *event)
220 struct map *self = malloc(sizeof(*self));
223 const char *filename = event->filename;
224 char newfilename[PATH_MAX];
228 int n = strcommon(filename);
231 snprintf(newfilename, sizeof(newfilename),
232 ".%s", filename + n);
233 filename = newfilename;
237 anon = is_anon_memory(filename);
240 snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", event->pid);
241 filename = newfilename;
244 self->start = event->start;
245 self->end = event->start + event->len;
246 self->pgoff = event->pgoff;
248 self->dso = dsos__findnew(filename);
249 if (self->dso == NULL)
252 if (self->dso == vdso || anon)
253 self->map_ip = vdso__map_ip;
255 self->map_ip = map__map_ip;
263 static struct map *map__clone(struct map *self)
265 struct map *map = malloc(sizeof(*self));
270 memcpy(map, self, sizeof(*self));
275 static int map__overlap(struct map *l, struct map *r)
277 if (l->start > r->start) {
283 if (l->end > r->start)
289 static size_t map__fprintf(struct map *self, FILE *fp)
291 return fprintf(fp, " %Lx-%Lx %Lx %s\n",
292 self->start, self->end, self->pgoff, self->dso->name);
297 struct rb_node rb_node;
298 struct list_head maps;
303 static struct thread *thread__new(pid_t pid)
305 struct thread *self = malloc(sizeof(*self));
309 self->comm = malloc(32);
311 snprintf(self->comm, 32, ":%d", self->pid);
312 INIT_LIST_HEAD(&self->maps);
318 static int thread__set_comm(struct thread *self, const char *comm)
322 self->comm = strdup(comm);
323 return self->comm ? 0 : -ENOMEM;
326 static size_t thread__fprintf(struct thread *self, FILE *fp)
329 size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
331 list_for_each_entry(pos, &self->maps, node)
332 ret += map__fprintf(pos, fp);
338 static struct rb_root threads;
339 static struct thread *last_match;
341 static struct thread *threads__findnew(pid_t pid)
343 struct rb_node **p = &threads.rb_node;
344 struct rb_node *parent = NULL;
348 * Font-end cache - PID lookups come in blocks,
349 * so most of the time we dont have to look up
352 if (last_match && last_match->pid == pid)
357 th = rb_entry(parent, struct thread, rb_node);
359 if (th->pid == pid) {
370 th = thread__new(pid);
372 rb_link_node(&th->rb_node, parent, p);
373 rb_insert_color(&th->rb_node, &threads);
380 static void thread__insert_map(struct thread *self, struct map *map)
382 struct map *pos, *tmp;
384 list_for_each_entry_safe(pos, tmp, &self->maps, node) {
385 if (map__overlap(pos, map)) {
386 list_del_init(&pos->node);
392 list_add_tail(&map->node, &self->maps);
395 static int thread__fork(struct thread *self, struct thread *parent)
401 self->comm = strdup(parent->comm);
405 list_for_each_entry(map, &parent->maps, node) {
406 struct map *new = map__clone(map);
409 thread__insert_map(self, new);
415 static struct map *thread__find_map(struct thread *self, __u64 ip)
422 list_for_each_entry(pos, &self->maps, node)
423 if (ip >= pos->start && ip <= pos->end)
429 static size_t threads__fprintf(FILE *fp)
434 for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
435 struct thread *pos = rb_entry(nd, struct thread, rb_node);
437 ret += thread__fprintf(pos, fp);
444 * histogram, sorted on item, collects counts
447 static struct rb_root hist;
450 struct rb_node rb_node;
452 struct thread *thread;
463 * configurable sorting bits
467 struct list_head list;
471 int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
472 int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
473 size_t (*print)(FILE *fp, struct hist_entry *);
479 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
481 return right->thread->pid - left->thread->pid;
485 sort__thread_print(FILE *fp, struct hist_entry *self)
487 return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
490 static struct sort_entry sort_thread = {
491 .header = " Command: Pid",
492 .cmp = sort__thread_cmp,
493 .print = sort__thread_print,
499 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
501 return right->thread->pid - left->thread->pid;
505 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
507 char *comm_l = left->thread->comm;
508 char *comm_r = right->thread->comm;
510 if (!comm_l || !comm_r) {
511 if (!comm_l && !comm_r)
519 return strcmp(comm_l, comm_r);
523 sort__comm_print(FILE *fp, struct hist_entry *self)
525 return fprintf(fp, "%16s", self->thread->comm);
528 static struct sort_entry sort_comm = {
529 .header = " Command",
530 .cmp = sort__comm_cmp,
531 .collapse = sort__comm_collapse,
532 .print = sort__comm_print,
538 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
540 struct dso *dso_l = left->dso;
541 struct dso *dso_r = right->dso;
543 if (!dso_l || !dso_r) {
544 if (!dso_l && !dso_r)
552 return strcmp(dso_l->name, dso_r->name);
556 sort__dso_print(FILE *fp, struct hist_entry *self)
559 return fprintf(fp, "%-25s", self->dso->name);
561 return fprintf(fp, "%016llx ", (__u64)self->ip);
564 static struct sort_entry sort_dso = {
565 .header = "Shared Object ",
566 .cmp = sort__dso_cmp,
567 .print = sort__dso_print,
573 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
577 if (left->sym == right->sym)
580 ip_l = left->sym ? left->sym->start : left->ip;
581 ip_r = right->sym ? right->sym->start : right->ip;
583 return (int64_t)(ip_r - ip_l);
587 sort__sym_print(FILE *fp, struct hist_entry *self)
592 ret += fprintf(fp, "%#018llx ", (__u64)self->ip);
595 ret += fprintf(fp, "[%c] %s",
596 self->dso == kernel_dso ? 'k' : '.', self->sym->name);
598 ret += fprintf(fp, "%#016llx", (__u64)self->ip);
604 static struct sort_entry sort_sym = {
606 .cmp = sort__sym_cmp,
607 .print = sort__sym_print,
610 static int sort__need_collapse = 0;
612 struct sort_dimension {
614 struct sort_entry *entry;
618 static struct sort_dimension sort_dimensions[] = {
619 { .name = "pid", .entry = &sort_thread, },
620 { .name = "comm", .entry = &sort_comm, },
621 { .name = "dso", .entry = &sort_dso, },
622 { .name = "symbol", .entry = &sort_sym, },
625 static LIST_HEAD(hist_entry__sort_list);
627 static int sort_dimension__add(char *tok)
631 for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
632 struct sort_dimension *sd = &sort_dimensions[i];
637 if (strncasecmp(tok, sd->name, strlen(tok)))
640 if (sd->entry->collapse)
641 sort__need_collapse = 1;
643 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
653 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
655 struct sort_entry *se;
658 list_for_each_entry(se, &hist_entry__sort_list, list) {
659 cmp = se->cmp(left, right);
668 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
670 struct sort_entry *se;
673 list_for_each_entry(se, &hist_entry__sort_list, list) {
674 int64_t (*f)(struct hist_entry *, struct hist_entry *);
676 f = se->collapse ?: se->cmp;
678 cmp = f(left, right);
687 hist_entry__fprintf(FILE *fp, struct hist_entry *self, __u64 total_samples)
689 struct sort_entry *se;
693 double percent = self->count * 100.0 / total_samples;
694 char *color = PERF_COLOR_NORMAL;
697 * We color high-overhead entries in red, mid-overhead
698 * entries in green - and keep the low overhead places
701 if (percent >= 5.0) {
702 color = PERF_COLOR_RED;
705 color = PERF_COLOR_GREEN;
708 ret = color_fprintf(fp, color, " %6.2f%%",
709 (self->count * 100.0) / total_samples);
711 ret = fprintf(fp, "%12Ld ", self->count);
713 list_for_each_entry(se, &hist_entry__sort_list, list) {
715 ret += se->print(fp, self);
718 ret += fprintf(fp, "\n");
724 * collect histogram counts
728 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
729 struct symbol *sym, __u64 ip, char level, __u64 count)
731 struct rb_node **p = &hist.rb_node;
732 struct rb_node *parent = NULL;
733 struct hist_entry *he;
734 struct hist_entry entry = {
747 he = rb_entry(parent, struct hist_entry, rb_node);
749 cmp = hist_entry__cmp(&entry, he);
762 he = malloc(sizeof(*he));
766 rb_link_node(&he->rb_node, parent, p);
767 rb_insert_color(&he->rb_node, &hist);
772 static void hist_entry__free(struct hist_entry *he)
778 * collapse the histogram
781 static struct rb_root collapse_hists;
783 static void collapse__insert_entry(struct hist_entry *he)
785 struct rb_node **p = &collapse_hists.rb_node;
786 struct rb_node *parent = NULL;
787 struct hist_entry *iter;
792 iter = rb_entry(parent, struct hist_entry, rb_node);
794 cmp = hist_entry__collapse(iter, he);
797 iter->count += he->count;
798 hist_entry__free(he);
808 rb_link_node(&he->rb_node, parent, p);
809 rb_insert_color(&he->rb_node, &collapse_hists);
812 static void collapse__resort(void)
814 struct rb_node *next;
815 struct hist_entry *n;
817 if (!sort__need_collapse)
820 next = rb_first(&hist);
822 n = rb_entry(next, struct hist_entry, rb_node);
823 next = rb_next(&n->rb_node);
825 rb_erase(&n->rb_node, &hist);
826 collapse__insert_entry(n);
831 * reverse the map, sort on count.
834 static struct rb_root output_hists;
836 static void output__insert_entry(struct hist_entry *he)
838 struct rb_node **p = &output_hists.rb_node;
839 struct rb_node *parent = NULL;
840 struct hist_entry *iter;
844 iter = rb_entry(parent, struct hist_entry, rb_node);
846 if (he->count > iter->count)
852 rb_link_node(&he->rb_node, parent, p);
853 rb_insert_color(&he->rb_node, &output_hists);
856 static void output__resort(void)
858 struct rb_node *next;
859 struct hist_entry *n;
860 struct rb_root *tree = &hist;
862 if (sort__need_collapse)
863 tree = &collapse_hists;
865 next = rb_first(tree);
868 n = rb_entry(next, struct hist_entry, rb_node);
869 next = rb_next(&n->rb_node);
871 rb_erase(&n->rb_node, tree);
872 output__insert_entry(n);
876 static size_t output__fprintf(FILE *fp, __u64 total_samples)
878 struct hist_entry *pos;
879 struct sort_entry *se;
885 fprintf(fp, "# (%Ld samples)\n", (__u64)total_samples);
888 fprintf(fp, "# Overhead");
889 list_for_each_entry(se, &hist_entry__sort_list, list)
890 fprintf(fp, " %s", se->header);
893 fprintf(fp, "# ........");
894 list_for_each_entry(se, &hist_entry__sort_list, list) {
898 for (i = 0; i < strlen(se->header); i++)
905 for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
906 pos = rb_entry(nd, struct hist_entry, rb_node);
907 ret += hist_entry__fprintf(fp, pos, total_samples);
910 if (!strcmp(sort_order, default_sort_order)) {
912 fprintf(fp, "# (For more details, try: perf report --sort comm,dso,symbol)\n");
920 static void register_idle_thread(void)
922 struct thread *thread = threads__findnew(0);
924 if (thread == NULL ||
925 thread__set_comm(thread, "[idle]")) {
926 fprintf(stderr, "problem inserting idle task.\n");
931 static unsigned long total = 0,
938 process_overflow_event(event_t *event, unsigned long offset, unsigned long head)
942 struct dso *dso = NULL;
943 struct thread *thread = threads__findnew(event->ip.pid);
944 __u64 ip = event->ip.ip;
946 struct map *map = NULL;
948 if (event->header.type & PERF_SAMPLE_PERIOD)
949 period = event->ip.period;
951 dprintf("%p [%p]: PERF_EVENT (IP, %d): %d: %p period: %Ld\n",
952 (void *)(offset + head),
953 (void *)(long)(event->header.size),
959 dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
961 if (thread == NULL) {
962 fprintf(stderr, "problem processing %d event, skipping it.\n",
967 if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
973 dprintf(" ...... dso: %s\n", dso->name);
975 } else if (event->header.misc & PERF_EVENT_MISC_USER) {
980 map = thread__find_map(thread, ip);
982 ip = map->map_ip(map, ip);
986 * If this is outside of all known maps,
987 * and is a negative address, try to look it
988 * up in the kernel dso, as it might be a
989 * vsyscall (which executes in user-mode):
991 if ((long long)ip < 0)
994 dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
999 dprintf(" ...... dso: [hypervisor]\n");
1002 if (show & show_mask) {
1003 struct symbol *sym = NULL;
1006 sym = dso->find_symbol(dso, ip);
1008 if (hist_entry__add(thread, map, dso, sym, ip, level, period)) {
1010 "problem incrementing symbol count, skipping event\n");
1020 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
1022 struct thread *thread = threads__findnew(event->mmap.pid);
1023 struct map *map = map__new(&event->mmap);
1025 dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
1026 (void *)(offset + head),
1027 (void *)(long)(event->header.size),
1029 (void *)(long)event->mmap.start,
1030 (void *)(long)event->mmap.len,
1031 (void *)(long)event->mmap.pgoff,
1032 event->mmap.filename);
1034 if (thread == NULL || map == NULL) {
1035 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
1039 thread__insert_map(thread, map);
1046 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
1048 struct thread *thread = threads__findnew(event->comm.pid);
1050 dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
1051 (void *)(offset + head),
1052 (void *)(long)(event->header.size),
1053 event->comm.comm, event->comm.pid);
1055 if (thread == NULL ||
1056 thread__set_comm(thread, event->comm.comm)) {
1057 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
1066 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
1068 struct thread *thread = threads__findnew(event->fork.pid);
1069 struct thread *parent = threads__findnew(event->fork.ppid);
1071 dprintf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
1072 (void *)(offset + head),
1073 (void *)(long)(event->header.size),
1074 event->fork.pid, event->fork.ppid);
1076 if (!thread || !parent || thread__fork(thread, parent)) {
1077 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
1086 process_period_event(event_t *event, unsigned long offset, unsigned long head)
1088 dprintf("%p [%p]: PERF_EVENT_PERIOD: time:%Ld, id:%Ld: period:%Ld\n",
1089 (void *)(offset + head),
1090 (void *)(long)(event->header.size),
1093 event->period.sample_period);
1099 process_event(event_t *event, unsigned long offset, unsigned long head)
1101 if (event->header.misc & PERF_EVENT_MISC_OVERFLOW)
1102 return process_overflow_event(event, offset, head);
1104 switch (event->header.type) {
1105 case PERF_EVENT_MMAP:
1106 return process_mmap_event(event, offset, head);
1108 case PERF_EVENT_COMM:
1109 return process_comm_event(event, offset, head);
1111 case PERF_EVENT_FORK:
1112 return process_fork_event(event, offset, head);
1114 case PERF_EVENT_PERIOD:
1115 return process_period_event(event, offset, head);
1117 * We dont process them right now but they are fine:
1120 case PERF_EVENT_THROTTLE:
1121 case PERF_EVENT_UNTHROTTLE:
1131 static int __cmd_report(void)
1133 int ret, rc = EXIT_FAILURE;
1134 unsigned long offset = 0;
1135 unsigned long head = 0;
1141 register_idle_thread();
1143 input = open(input_name, O_RDONLY);
1145 fprintf(stderr, " failed to open file: %s", input_name);
1146 if (!strcmp(input_name, "perf.data"))
1147 fprintf(stderr, " (try 'perf record' first)");
1148 fprintf(stderr, "\n");
1152 ret = fstat(input, &stat);
1154 perror("failed to stat file");
1158 if (!stat.st_size) {
1159 fprintf(stderr, "zero-sized file, nothing to do!\n");
1163 if (load_kernel() < 0) {
1164 perror("failed to load kernel symbols");
1165 return EXIT_FAILURE;
1169 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1170 perror("failed to get the current directory");
1171 return EXIT_FAILURE;
1173 cwdlen = strlen(cwd);
1179 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1180 MAP_SHARED, input, offset);
1181 if (buf == MAP_FAILED) {
1182 perror("failed to mmap file");
1187 event = (event_t *)(buf + head);
1189 size = event->header.size;
1193 if (head + event->header.size >= page_size * mmap_window) {
1194 unsigned long shift = page_size * (head / page_size);
1197 ret = munmap(buf, page_size * mmap_window);
1205 size = event->header.size;
1207 dprintf("%p [%p]: event: %d\n",
1208 (void *)(offset + head),
1209 (void *)(long)event->header.size,
1210 event->header.type);
1212 if (!size || process_event(event, offset, head) < 0) {
1214 dprintf("%p [%p]: skipping unknown header type: %d\n",
1215 (void *)(offset + head),
1216 (void *)(long)(event->header.size),
1217 event->header.type);
1222 * assume we lost track of the stream, check alignment, and
1223 * increment a single u64 in the hope to catch on again 'soon'.
1226 if (unlikely(head & 7))
1234 if (offset + head < stat.st_size)
1240 dprintf(" IP events: %10ld\n", total);
1241 dprintf(" mmap events: %10ld\n", total_mmap);
1242 dprintf(" comm events: %10ld\n", total_comm);
1243 dprintf(" fork events: %10ld\n", total_fork);
1244 dprintf(" unknown events: %10ld\n", total_unknown);
1250 threads__fprintf(stdout);
1253 dsos__fprintf(stdout);
1257 output__fprintf(stdout, total);
1262 static const char * const report_usage[] = {
1263 "perf report [<options>] <command>",
1267 static const struct option options[] = {
1268 OPT_STRING('i', "input", &input_name, "file",
1270 OPT_BOOLEAN('v', "verbose", &verbose,
1271 "be more verbose (show symbol address, etc)"),
1272 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1273 "dump raw trace in ASCII"),
1274 OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
1275 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1276 "sort by key(s): pid, comm, dso, symbol. Default: pid,symbol"),
1277 OPT_BOOLEAN('P', "full-paths", &full_paths,
1278 "Don't shorten the pathnames taking into account the cwd"),
1282 static void setup_sorting(void)
1284 char *tmp, *tok, *str = strdup(sort_order);
1286 for (tok = strtok_r(str, ", ", &tmp);
1287 tok; tok = strtok_r(NULL, ", ", &tmp)) {
1288 if (sort_dimension__add(tok) < 0) {
1289 error("Unknown --sort key: `%s'", tok);
1290 usage_with_options(report_usage, options);
1297 int cmd_report(int argc, const char **argv, const char *prefix)
1301 page_size = getpagesize();
1303 argc = parse_options(argc, argv, options, report_usage, 0);
1308 * Any (unrecognized) arguments left?
1311 usage_with_options(report_usage, options);
1315 return __cmd_report();