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 <linux/list.h>
14 #include "util/cache.h"
15 #include <linux/rbtree.h>
16 #include "util/symbol.h"
17 #include "util/string.h"
18 #include "util/callchain.h"
19 #include "util/strlist.h"
20 #include "util/values.h"
23 #include "util/header.h"
25 #include "util/parse-options.h"
26 #include "util/parse-events.h"
32 static char const *input_name = "perf.data";
33 static char *vmlinux = NULL;
35 static char default_sort_order[] = "comm,dso,symbol";
36 static char *sort_order = default_sort_order;
37 static char *dso_list_str, *comm_list_str, *sym_list_str,
39 static struct strlist *dso_list, *comm_list, *sym_list;
40 static char *field_sep;
43 static int show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
45 static int dump_trace = 0;
46 #define dprintf(x...) do { if (dump_trace) printf(x); } while (0)
47 #define cdprintf(x...) do { if (dump_trace) color_fprintf(stdout, color, x); } while (0)
50 #define eprintf(x...) do { if (verbose) fprintf(stderr, x); } while (0)
54 static int full_paths;
55 static int show_nr_samples;
57 static int show_threads;
58 static struct perf_read_values show_threads_values;
60 static unsigned long page_size;
61 static unsigned long mmap_window = 32;
63 static char default_parent_pattern[] = "^sys_|^do_page_fault";
64 static char *parent_pattern = default_parent_pattern;
65 static regex_t parent_regex;
67 static int exclude_other = 1;
69 static char callchain_default_opt[] = "fractal,0.5";
74 struct callchain_param callchain_param = {
75 .mode = CHAIN_GRAPH_REL,
79 static u64 sample_type;
82 struct perf_event_header header;
85 unsigned char __more_data[];
89 struct perf_event_header header;
94 char filename[PATH_MAX];
98 struct perf_event_header header;
104 struct perf_event_header header;
110 struct perf_event_header header;
116 struct perf_event_header header;
124 typedef union event_union {
125 struct perf_event_header header;
127 struct mmap_event mmap;
128 struct comm_event comm;
129 struct fork_event fork;
130 struct lost_event lost;
131 struct read_event read;
134 static int repsep_fprintf(FILE *fp, const char *fmt, ...)
141 n = vfprintf(fp, fmt, ap);
144 n = vasprintf(&bf, fmt, ap);
148 sep = strchr(sep, *field_sep);
161 static LIST_HEAD(dsos);
162 static struct dso *kernel_dso;
163 static struct dso *vdso;
164 static struct dso *hypervisor_dso;
166 static void dsos__add(struct dso *dso)
168 list_add_tail(&dso->node, &dsos);
171 static struct dso *dsos__find(const char *name)
175 list_for_each_entry(pos, &dsos, node)
176 if (strcmp(pos->name, name) == 0)
181 static struct dso *dsos__findnew(const char *name)
183 struct dso *dso = dsos__find(name);
189 dso = dso__new(name, 0);
193 nr = dso__load(dso, NULL, verbose);
195 eprintf("Failed to open: %s\n", name);
199 eprintf("No symbols found in: %s, maybe install a debug package?\n", name);
210 static void dsos__fprintf(FILE *fp)
214 list_for_each_entry(pos, &dsos, node)
215 dso__fprintf(pos, fp);
218 static struct symbol *vdso__find_symbol(struct dso *dso, u64 ip)
220 return dso__find_symbol(dso, ip);
223 static int load_kernel(void)
227 kernel_dso = dso__new("[kernel]", 0);
231 err = dso__load_kernel(kernel_dso, vmlinux, NULL, verbose, modules);
233 dso__delete(kernel_dso);
236 dsos__add(kernel_dso);
238 vdso = dso__new("[vdso]", 0);
242 vdso->find_symbol = vdso__find_symbol;
246 hypervisor_dso = dso__new("[hypervisor]", 0);
249 dsos__add(hypervisor_dso);
254 static char __cwd[PATH_MAX];
255 static char *cwd = __cwd;
258 static int strcommon(const char *pathname)
262 while (n < cwdlen && pathname[n] == cwd[n])
269 struct list_head node;
273 u64 (*map_ip)(struct map *, u64);
277 static u64 map__map_ip(struct map *map, u64 ip)
279 return ip - map->start + map->pgoff;
282 static u64 vdso__map_ip(struct map *map __used, u64 ip)
287 static inline int is_anon_memory(const char *filename)
289 return strcmp(filename, "//anon") == 0;
292 static struct map *map__new(struct mmap_event *event)
294 struct map *self = malloc(sizeof(*self));
297 const char *filename = event->filename;
298 char newfilename[PATH_MAX];
302 int n = strcommon(filename);
305 snprintf(newfilename, sizeof(newfilename),
306 ".%s", filename + n);
307 filename = newfilename;
311 anon = is_anon_memory(filename);
314 snprintf(newfilename, sizeof(newfilename), "/tmp/perf-%d.map", event->pid);
315 filename = newfilename;
318 self->start = event->start;
319 self->end = event->start + event->len;
320 self->pgoff = event->pgoff;
322 self->dso = dsos__findnew(filename);
323 if (self->dso == NULL)
326 if (self->dso == vdso || anon)
327 self->map_ip = vdso__map_ip;
329 self->map_ip = map__map_ip;
337 static struct map *map__clone(struct map *self)
339 struct map *map = malloc(sizeof(*self));
344 memcpy(map, self, sizeof(*self));
349 static int map__overlap(struct map *l, struct map *r)
351 if (l->start > r->start) {
357 if (l->end > r->start)
363 static size_t map__fprintf(struct map *self, FILE *fp)
365 return fprintf(fp, " %Lx-%Lx %Lx %s\n",
366 self->start, self->end, self->pgoff, self->dso->name);
371 struct rb_node rb_node;
372 struct list_head maps;
377 static struct thread *thread__new(pid_t pid)
379 struct thread *self = malloc(sizeof(*self));
383 self->comm = malloc(32);
385 snprintf(self->comm, 32, ":%d", self->pid);
386 INIT_LIST_HEAD(&self->maps);
392 static unsigned int dsos__col_width,
396 static int thread__set_comm(struct thread *self, const char *comm)
400 self->comm = strdup(comm);
404 if (!col_width_list_str && !field_sep &&
405 (!comm_list || strlist__has_entry(comm_list, comm))) {
406 unsigned int slen = strlen(comm);
407 if (slen > comms__col_width) {
408 comms__col_width = slen;
409 threads__col_width = slen + 6;
416 static size_t thread__fprintf(struct thread *self, FILE *fp)
419 size_t ret = fprintf(fp, "Thread %d %s\n", self->pid, self->comm);
421 list_for_each_entry(pos, &self->maps, node)
422 ret += map__fprintf(pos, fp);
428 static struct rb_root threads;
429 static struct thread *last_match;
431 static struct thread *threads__findnew(pid_t pid)
433 struct rb_node **p = &threads.rb_node;
434 struct rb_node *parent = NULL;
438 * Font-end cache - PID lookups come in blocks,
439 * so most of the time we dont have to look up
442 if (last_match && last_match->pid == pid)
447 th = rb_entry(parent, struct thread, rb_node);
449 if (th->pid == pid) {
460 th = thread__new(pid);
462 rb_link_node(&th->rb_node, parent, p);
463 rb_insert_color(&th->rb_node, &threads);
470 static void thread__insert_map(struct thread *self, struct map *map)
472 struct map *pos, *tmp;
474 list_for_each_entry_safe(pos, tmp, &self->maps, node) {
475 if (map__overlap(pos, map)) {
477 printf("overlapping maps:\n");
478 map__fprintf(map, stdout);
479 map__fprintf(pos, stdout);
482 if (map->start <= pos->start && map->end > pos->start)
483 pos->start = map->end;
485 if (map->end >= pos->end && map->start < pos->end)
486 pos->end = map->start;
489 printf("after collision:\n");
490 map__fprintf(pos, stdout);
493 if (pos->start >= pos->end) {
494 list_del_init(&pos->node);
500 list_add_tail(&map->node, &self->maps);
503 static int thread__fork(struct thread *self, struct thread *parent)
509 self->comm = strdup(parent->comm);
513 list_for_each_entry(map, &parent->maps, node) {
514 struct map *new = map__clone(map);
517 thread__insert_map(self, new);
523 static struct map *thread__find_map(struct thread *self, u64 ip)
530 list_for_each_entry(pos, &self->maps, node)
531 if (ip >= pos->start && ip <= pos->end)
537 static size_t threads__fprintf(FILE *fp)
542 for (nd = rb_first(&threads); nd; nd = rb_next(nd)) {
543 struct thread *pos = rb_entry(nd, struct thread, rb_node);
545 ret += thread__fprintf(pos, fp);
552 * histogram, sorted on item, collects counts
555 static struct rb_root hist;
558 struct rb_node rb_node;
560 struct thread *thread;
564 struct symbol *parent;
567 struct callchain_node callchain;
568 struct rb_root sorted_chain;
574 * configurable sorting bits
578 struct list_head list;
582 int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
583 int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
584 size_t (*print)(FILE *fp, struct hist_entry *, unsigned int width);
589 static int64_t cmp_null(void *l, void *r)
602 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
604 return right->thread->pid - left->thread->pid;
608 sort__thread_print(FILE *fp, struct hist_entry *self, unsigned int width)
610 return repsep_fprintf(fp, "%*s:%5d", width - 6,
611 self->thread->comm ?: "", self->thread->pid);
614 static struct sort_entry sort_thread = {
615 .header = "Command: Pid",
616 .cmp = sort__thread_cmp,
617 .print = sort__thread_print,
618 .width = &threads__col_width,
624 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
626 return right->thread->pid - left->thread->pid;
630 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
632 char *comm_l = left->thread->comm;
633 char *comm_r = right->thread->comm;
635 if (!comm_l || !comm_r)
636 return cmp_null(comm_l, comm_r);
638 return strcmp(comm_l, comm_r);
642 sort__comm_print(FILE *fp, struct hist_entry *self, unsigned int width)
644 return repsep_fprintf(fp, "%*s", width, self->thread->comm);
647 static struct sort_entry sort_comm = {
649 .cmp = sort__comm_cmp,
650 .collapse = sort__comm_collapse,
651 .print = sort__comm_print,
652 .width = &comms__col_width,
658 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
660 struct dso *dso_l = left->dso;
661 struct dso *dso_r = right->dso;
663 if (!dso_l || !dso_r)
664 return cmp_null(dso_l, dso_r);
666 return strcmp(dso_l->name, dso_r->name);
670 sort__dso_print(FILE *fp, struct hist_entry *self, unsigned int width)
673 return repsep_fprintf(fp, "%-*s", width, self->dso->name);
675 return repsep_fprintf(fp, "%*llx", width, (u64)self->ip);
678 static struct sort_entry sort_dso = {
679 .header = "Shared Object",
680 .cmp = sort__dso_cmp,
681 .print = sort__dso_print,
682 .width = &dsos__col_width,
688 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
692 if (left->sym == right->sym)
695 ip_l = left->sym ? left->sym->start : left->ip;
696 ip_r = right->sym ? right->sym->start : right->ip;
698 return (int64_t)(ip_r - ip_l);
702 sort__sym_print(FILE *fp, struct hist_entry *self, unsigned int width __used)
707 ret += repsep_fprintf(fp, "%#018llx %c ", (u64)self->ip,
708 dso__symtab_origin(self->dso));
710 ret += repsep_fprintf(fp, "[%c] ", self->level);
712 ret += repsep_fprintf(fp, "%s", self->sym->name);
714 if (self->sym->module)
715 ret += repsep_fprintf(fp, "\t[%s]",
716 self->sym->module->name);
718 ret += repsep_fprintf(fp, "%#016llx", (u64)self->ip);
724 static struct sort_entry sort_sym = {
726 .cmp = sort__sym_cmp,
727 .print = sort__sym_print,
733 sort__parent_cmp(struct hist_entry *left, struct hist_entry *right)
735 struct symbol *sym_l = left->parent;
736 struct symbol *sym_r = right->parent;
738 if (!sym_l || !sym_r)
739 return cmp_null(sym_l, sym_r);
741 return strcmp(sym_l->name, sym_r->name);
745 sort__parent_print(FILE *fp, struct hist_entry *self, unsigned int width)
747 return repsep_fprintf(fp, "%-*s", width,
748 self->parent ? self->parent->name : "[other]");
751 static unsigned int parent_symbol__col_width;
753 static struct sort_entry sort_parent = {
754 .header = "Parent symbol",
755 .cmp = sort__parent_cmp,
756 .print = sort__parent_print,
757 .width = &parent_symbol__col_width,
760 static int sort__need_collapse = 0;
761 static int sort__has_parent = 0;
763 struct sort_dimension {
765 struct sort_entry *entry;
769 static struct sort_dimension sort_dimensions[] = {
770 { .name = "pid", .entry = &sort_thread, },
771 { .name = "comm", .entry = &sort_comm, },
772 { .name = "dso", .entry = &sort_dso, },
773 { .name = "symbol", .entry = &sort_sym, },
774 { .name = "parent", .entry = &sort_parent, },
777 static LIST_HEAD(hist_entry__sort_list);
779 static int sort_dimension__add(char *tok)
783 for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
784 struct sort_dimension *sd = &sort_dimensions[i];
789 if (strncasecmp(tok, sd->name, strlen(tok)))
792 if (sd->entry->collapse)
793 sort__need_collapse = 1;
795 if (sd->entry == &sort_parent) {
796 int ret = regcomp(&parent_regex, parent_pattern, REG_EXTENDED);
800 regerror(ret, &parent_regex, err, sizeof(err));
801 fprintf(stderr, "Invalid regex: %s\n%s",
802 parent_pattern, err);
805 sort__has_parent = 1;
808 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
818 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
820 struct sort_entry *se;
823 list_for_each_entry(se, &hist_entry__sort_list, list) {
824 cmp = se->cmp(left, right);
833 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
835 struct sort_entry *se;
838 list_for_each_entry(se, &hist_entry__sort_list, list) {
839 int64_t (*f)(struct hist_entry *, struct hist_entry *);
841 f = se->collapse ?: se->cmp;
843 cmp = f(left, right);
851 static size_t ipchain__fprintf_graph_line(FILE *fp, int depth, int depth_mask)
856 ret += fprintf(fp, "%s", " ");
858 for (i = 0; i < depth; i++)
859 if (depth_mask & (1 << i))
860 ret += fprintf(fp, "| ");
862 ret += fprintf(fp, " ");
864 ret += fprintf(fp, "\n");
869 ipchain__fprintf_graph(FILE *fp, struct callchain_list *chain, int depth,
870 int depth_mask, int count, u64 total_samples,
876 ret += fprintf(fp, "%s", " ");
877 for (i = 0; i < depth; i++) {
878 if (depth_mask & (1 << i))
879 ret += fprintf(fp, "|");
881 ret += fprintf(fp, " ");
882 if (!count && i == depth - 1) {
885 percent = hits * 100.0 / total_samples;
886 ret += percent_color_fprintf(fp, "--%2.2f%%-- ", percent);
888 ret += fprintf(fp, "%s", " ");
891 ret += fprintf(fp, "%s\n", chain->sym->name);
893 ret += fprintf(fp, "%p\n", (void *)(long)chain->ip);
898 static struct symbol *rem_sq_bracket;
899 static struct callchain_list rem_hits;
901 static void init_rem_hits(void)
903 rem_sq_bracket = malloc(sizeof(*rem_sq_bracket) + 6);
904 if (!rem_sq_bracket) {
905 fprintf(stderr, "Not enough memory to display remaining hits\n");
909 strcpy(rem_sq_bracket->name, "[...]");
910 rem_hits.sym = rem_sq_bracket;
914 callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
915 u64 total_samples, int depth, int depth_mask)
917 struct rb_node *node, *next;
918 struct callchain_node *child;
919 struct callchain_list *chain;
920 int new_depth_mask = depth_mask;
926 if (callchain_param.mode == CHAIN_GRAPH_REL)
927 new_total = self->children_hit;
929 new_total = total_samples;
931 remaining = new_total;
933 node = rb_first(&self->rb_root);
937 child = rb_entry(node, struct callchain_node, rb_node);
938 cumul = cumul_hits(child);
942 * The depth mask manages the output of pipes that show
943 * the depth. We don't want to keep the pipes of the current
944 * level for the last child of this depth.
945 * Except if we have remaining filtered hits. They will
946 * supersede the last child
948 next = rb_next(node);
949 if (!next && (callchain_param.mode != CHAIN_GRAPH_REL || !remaining))
950 new_depth_mask &= ~(1 << (depth - 1));
953 * But we keep the older depth mask for the line seperator
954 * to keep the level link until we reach the last child
956 ret += ipchain__fprintf_graph_line(fp, depth, depth_mask);
958 list_for_each_entry(chain, &child->val, list) {
959 if (chain->ip >= PERF_CONTEXT_MAX)
961 ret += ipchain__fprintf_graph(fp, chain, depth,
966 ret += callchain__fprintf_graph(fp, child, new_total,
968 new_depth_mask | (1 << depth));
972 if (callchain_param.mode == CHAIN_GRAPH_REL &&
973 remaining && remaining != new_total) {
978 new_depth_mask &= ~(1 << (depth - 1));
980 ret += ipchain__fprintf_graph(fp, &rem_hits, depth,
981 new_depth_mask, 0, new_total,
989 callchain__fprintf_flat(FILE *fp, struct callchain_node *self,
992 struct callchain_list *chain;
998 ret += callchain__fprintf_flat(fp, self->parent, total_samples);
1001 list_for_each_entry(chain, &self->val, list) {
1002 if (chain->ip >= PERF_CONTEXT_MAX)
1005 ret += fprintf(fp, " %s\n", chain->sym->name);
1007 ret += fprintf(fp, " %p\n",
1008 (void *)(long)chain->ip);
1015 hist_entry_callchain__fprintf(FILE *fp, struct hist_entry *self,
1018 struct rb_node *rb_node;
1019 struct callchain_node *chain;
1022 rb_node = rb_first(&self->sorted_chain);
1026 chain = rb_entry(rb_node, struct callchain_node, rb_node);
1027 percent = chain->hit * 100.0 / total_samples;
1028 switch (callchain_param.mode) {
1030 ret += percent_color_fprintf(fp, " %6.2f%%\n",
1032 ret += callchain__fprintf_flat(fp, chain, total_samples);
1034 case CHAIN_GRAPH_ABS: /* Falldown */
1035 case CHAIN_GRAPH_REL:
1036 ret += callchain__fprintf_graph(fp, chain,
1037 total_samples, 1, 1);
1041 ret += fprintf(fp, "\n");
1042 rb_node = rb_next(rb_node);
1050 hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
1052 struct sort_entry *se;
1055 if (exclude_other && !self->parent)
1059 ret = percent_color_fprintf(fp,
1060 field_sep ? "%.2f" : " %6.2f%%",
1061 (self->count * 100.0) / total_samples);
1063 ret = fprintf(fp, field_sep ? "%lld" : "%12lld ", self->count);
1065 if (show_nr_samples) {
1067 fprintf(fp, "%c%lld", *field_sep, self->count);
1069 fprintf(fp, "%11lld", self->count);
1072 list_for_each_entry(se, &hist_entry__sort_list, list) {
1076 fprintf(fp, "%s", field_sep ?: " ");
1077 ret += se->print(fp, self, se->width ? *se->width : 0);
1080 ret += fprintf(fp, "\n");
1083 hist_entry_callchain__fprintf(fp, self, total_samples);
1092 static void dso__calc_col_width(struct dso *self)
1094 if (!col_width_list_str && !field_sep &&
1095 (!dso_list || strlist__has_entry(dso_list, self->name))) {
1096 unsigned int slen = strlen(self->name);
1097 if (slen > dsos__col_width)
1098 dsos__col_width = slen;
1101 self->slen_calculated = 1;
1104 static struct symbol *
1105 resolve_symbol(struct thread *thread, struct map **mapp,
1106 struct dso **dsop, u64 *ipp)
1108 struct dso *dso = dsop ? *dsop : NULL;
1109 struct map *map = mapp ? *mapp : NULL;
1121 map = thread__find_map(thread, ip);
1124 * We have to do this here as we may have a dso
1125 * with no symbol hit that has a name longer than
1126 * the ones with symbols sampled.
1128 if (!sort_dso.elide && !map->dso->slen_calculated)
1129 dso__calc_col_width(map->dso);
1134 ip = map->map_ip(map, ip);
1139 * If this is outside of all known maps,
1140 * and is a negative address, try to look it
1141 * up in the kernel dso, as it might be a
1142 * vsyscall (which executes in user-mode):
1144 if ((long long)ip < 0)
1147 dprintf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
1148 dprintf(" ...... map: %Lx -> %Lx\n", *ipp, ip);
1157 return dso->find_symbol(dso, ip);
1160 static int call__match(struct symbol *sym)
1162 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
1168 static struct symbol **
1169 resolve_callchain(struct thread *thread, struct map *map __used,
1170 struct ip_callchain *chain, struct hist_entry *entry)
1172 u64 context = PERF_CONTEXT_MAX;
1173 struct symbol **syms = NULL;
1177 syms = calloc(chain->nr, sizeof(*syms));
1179 fprintf(stderr, "Can't allocate memory for symbols\n");
1184 for (i = 0; i < chain->nr; i++) {
1185 u64 ip = chain->ips[i];
1186 struct dso *dso = NULL;
1189 if (ip >= PERF_CONTEXT_MAX) {
1195 case PERF_CONTEXT_HV:
1196 dso = hypervisor_dso;
1198 case PERF_CONTEXT_KERNEL:
1205 sym = resolve_symbol(thread, NULL, &dso, &ip);
1208 if (sort__has_parent && call__match(sym) &&
1210 entry->parent = sym;
1221 * collect histogram counts
1225 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
1226 struct symbol *sym, u64 ip, struct ip_callchain *chain,
1227 char level, u64 count)
1229 struct rb_node **p = &hist.rb_node;
1230 struct rb_node *parent = NULL;
1231 struct hist_entry *he;
1232 struct symbol **syms = NULL;
1233 struct hist_entry entry = {
1242 .sorted_chain = RB_ROOT
1246 if ((sort__has_parent || callchain) && chain)
1247 syms = resolve_callchain(thread, map, chain, &entry);
1249 while (*p != NULL) {
1251 he = rb_entry(parent, struct hist_entry, rb_node);
1253 cmp = hist_entry__cmp(&entry, he);
1258 append_chain(&he->callchain, chain, syms);
1267 p = &(*p)->rb_right;
1270 he = malloc(sizeof(*he));
1275 callchain_init(&he->callchain);
1276 append_chain(&he->callchain, chain, syms);
1279 rb_link_node(&he->rb_node, parent, p);
1280 rb_insert_color(&he->rb_node, &hist);
1285 static void hist_entry__free(struct hist_entry *he)
1291 * collapse the histogram
1294 static struct rb_root collapse_hists;
1296 static void collapse__insert_entry(struct hist_entry *he)
1298 struct rb_node **p = &collapse_hists.rb_node;
1299 struct rb_node *parent = NULL;
1300 struct hist_entry *iter;
1303 while (*p != NULL) {
1305 iter = rb_entry(parent, struct hist_entry, rb_node);
1307 cmp = hist_entry__collapse(iter, he);
1310 iter->count += he->count;
1311 hist_entry__free(he);
1318 p = &(*p)->rb_right;
1321 rb_link_node(&he->rb_node, parent, p);
1322 rb_insert_color(&he->rb_node, &collapse_hists);
1325 static void collapse__resort(void)
1327 struct rb_node *next;
1328 struct hist_entry *n;
1330 if (!sort__need_collapse)
1333 next = rb_first(&hist);
1335 n = rb_entry(next, struct hist_entry, rb_node);
1336 next = rb_next(&n->rb_node);
1338 rb_erase(&n->rb_node, &hist);
1339 collapse__insert_entry(n);
1344 * reverse the map, sort on count.
1347 static struct rb_root output_hists;
1349 static void output__insert_entry(struct hist_entry *he, u64 min_callchain_hits)
1351 struct rb_node **p = &output_hists.rb_node;
1352 struct rb_node *parent = NULL;
1353 struct hist_entry *iter;
1356 callchain_param.sort(&he->sorted_chain, &he->callchain,
1357 min_callchain_hits, &callchain_param);
1359 while (*p != NULL) {
1361 iter = rb_entry(parent, struct hist_entry, rb_node);
1363 if (he->count > iter->count)
1366 p = &(*p)->rb_right;
1369 rb_link_node(&he->rb_node, parent, p);
1370 rb_insert_color(&he->rb_node, &output_hists);
1373 static void output__resort(u64 total_samples)
1375 struct rb_node *next;
1376 struct hist_entry *n;
1377 struct rb_root *tree = &hist;
1378 u64 min_callchain_hits;
1380 min_callchain_hits = total_samples * (callchain_param.min_percent / 100);
1382 if (sort__need_collapse)
1383 tree = &collapse_hists;
1385 next = rb_first(tree);
1388 n = rb_entry(next, struct hist_entry, rb_node);
1389 next = rb_next(&n->rb_node);
1391 rb_erase(&n->rb_node, tree);
1392 output__insert_entry(n, min_callchain_hits);
1396 static size_t output__fprintf(FILE *fp, u64 total_samples)
1398 struct hist_entry *pos;
1399 struct sort_entry *se;
1403 char *col_width = col_width_list_str;
1407 fprintf(fp, "# Samples: %Ld\n", (u64)total_samples);
1410 fprintf(fp, "# Overhead");
1411 if (show_nr_samples) {
1413 fprintf(fp, "%cSamples", *field_sep);
1415 fputs(" Samples ", fp);
1417 list_for_each_entry(se, &hist_entry__sort_list, list) {
1421 fprintf(fp, "%c%s", *field_sep, se->header);
1424 width = strlen(se->header);
1426 if (col_width_list_str) {
1428 *se->width = atoi(col_width);
1429 col_width = strchr(col_width, ',');
1434 width = *se->width = max(*se->width, width);
1436 fprintf(fp, " %*s", width, se->header);
1443 fprintf(fp, "# ........");
1444 if (show_nr_samples)
1445 fprintf(fp, " ..........");
1446 list_for_each_entry(se, &hist_entry__sort_list, list) {
1456 width = strlen(se->header);
1457 for (i = 0; i < width; i++)
1465 for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
1466 pos = rb_entry(nd, struct hist_entry, rb_node);
1467 ret += hist_entry__fprintf(fp, pos, total_samples);
1470 if (sort_order == default_sort_order &&
1471 parent_pattern == default_parent_pattern) {
1473 fprintf(fp, "# (For a higher level overview, try: perf report --sort comm,dso)\n");
1478 free(rem_sq_bracket);
1481 perf_read_values_display(fp, &show_threads_values);
1486 static void register_idle_thread(void)
1488 struct thread *thread = threads__findnew(0);
1490 if (thread == NULL ||
1491 thread__set_comm(thread, "[idle]")) {
1492 fprintf(stderr, "problem inserting idle task.\n");
1497 static unsigned long total = 0,
1504 static int validate_chain(struct ip_callchain *chain, event_t *event)
1506 unsigned int chain_size;
1508 chain_size = event->header.size;
1509 chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
1511 if (chain->nr*sizeof(u64) > chain_size)
1518 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1522 struct dso *dso = NULL;
1523 struct thread *thread = threads__findnew(event->ip.pid);
1524 u64 ip = event->ip.ip;
1526 struct map *map = NULL;
1527 void *more_data = event->ip.__more_data;
1528 struct ip_callchain *chain = NULL;
1531 if (sample_type & PERF_SAMPLE_PERIOD) {
1532 period = *(u64 *)more_data;
1533 more_data += sizeof(u64);
1536 dprintf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d: %p period: %Ld\n",
1537 (void *)(offset + head),
1538 (void *)(long)(event->header.size),
1544 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
1547 chain = (void *)more_data;
1549 dprintf("... chain: nr:%Lu\n", chain->nr);
1551 if (validate_chain(chain, event) < 0) {
1552 eprintf("call-chain problem with event, skipping it.\n");
1557 for (i = 0; i < chain->nr; i++)
1558 dprintf("..... %2d: %016Lx\n", i, chain->ips[i]);
1562 dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1564 if (thread == NULL) {
1565 eprintf("problem processing %d event, skipping it.\n",
1566 event->header.type);
1570 if (comm_list && !strlist__has_entry(comm_list, thread->comm))
1573 cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
1575 if (cpumode == PERF_EVENT_MISC_KERNEL) {
1581 dprintf(" ...... dso: %s\n", dso->name);
1583 } else if (cpumode == PERF_EVENT_MISC_USER) {
1592 dso = hypervisor_dso;
1594 dprintf(" ...... dso: [hypervisor]\n");
1597 if (show & show_mask) {
1598 struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip);
1600 if (dso_list && dso && dso->name && !strlist__has_entry(dso_list, dso->name))
1603 if (sym_list && sym && !strlist__has_entry(sym_list, sym->name))
1606 if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) {
1607 eprintf("problem incrementing symbol count, skipping event\n");
1617 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
1619 struct thread *thread = threads__findnew(event->mmap.pid);
1620 struct map *map = map__new(&event->mmap);
1622 dprintf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
1623 (void *)(offset + head),
1624 (void *)(long)(event->header.size),
1626 (void *)(long)event->mmap.start,
1627 (void *)(long)event->mmap.len,
1628 (void *)(long)event->mmap.pgoff,
1629 event->mmap.filename);
1631 if (thread == NULL || map == NULL) {
1632 dprintf("problem processing PERF_EVENT_MMAP, skipping event.\n");
1636 thread__insert_map(thread, map);
1643 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
1645 struct thread *thread = threads__findnew(event->comm.pid);
1647 dprintf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
1648 (void *)(offset + head),
1649 (void *)(long)(event->header.size),
1650 event->comm.comm, event->comm.pid);
1652 if (thread == NULL ||
1653 thread__set_comm(thread, event->comm.comm)) {
1654 dprintf("problem processing PERF_EVENT_COMM, skipping event.\n");
1663 process_task_event(event_t *event, unsigned long offset, unsigned long head)
1665 struct thread *thread = threads__findnew(event->fork.pid);
1666 struct thread *parent = threads__findnew(event->fork.ppid);
1668 dprintf("%p [%p]: PERF_EVENT_%s: (%d:%d):(%d:%d)\n",
1669 (void *)(offset + head),
1670 (void *)(long)(event->header.size),
1671 event->header.type == PERF_EVENT_FORK ? "FORK" : "EXIT",
1672 event->fork.pid, event->fork.tid,
1673 event->fork.ppid, event->fork.ptid);
1676 * A thread clone will have the same PID for both
1679 if (thread == parent)
1682 if (event->header.type == PERF_EVENT_EXIT)
1685 if (!thread || !parent || thread__fork(thread, parent)) {
1686 dprintf("problem processing PERF_EVENT_FORK, skipping event.\n");
1695 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
1697 dprintf("%p [%p]: PERF_EVENT_LOST: id:%Ld: lost:%Ld\n",
1698 (void *)(offset + head),
1699 (void *)(long)(event->header.size),
1703 total_lost += event->lost.lost;
1708 static void trace_event(event_t *event)
1710 unsigned char *raw_event = (void *)event;
1711 char *color = PERF_COLOR_BLUE;
1718 cdprintf("\n. ... raw event: size %d bytes\n", event->header.size);
1720 for (i = 0; i < event->header.size; i++) {
1721 if ((i & 15) == 0) {
1723 cdprintf(" %04x: ", i);
1726 cdprintf(" %02x", raw_event[i]);
1728 if (((i & 15) == 15) || i == event->header.size-1) {
1730 for (j = 0; j < 15-(i & 15); j++)
1732 for (j = 0; j < (i & 15); j++) {
1733 if (isprint(raw_event[i-15+j]))
1734 cdprintf("%c", raw_event[i-15+j]);
1744 static struct perf_header *header;
1746 static struct perf_counter_attr *perf_header__find_attr(u64 id)
1750 for (i = 0; i < header->attrs; i++) {
1751 struct perf_header_attr *attr = header->attr[i];
1754 for (j = 0; j < attr->ids; j++) {
1755 if (attr->id[j] == id)
1764 process_read_event(event_t *event, unsigned long offset, unsigned long head)
1766 struct perf_counter_attr *attr = perf_header__find_attr(event->read.id);
1769 char *name = attr ? __event_name(attr->type, attr->config)
1771 perf_read_values_add_value(&show_threads_values,
1772 event->read.pid, event->read.tid,
1778 dprintf("%p [%p]: PERF_EVENT_READ: %d %d %s %Lu\n",
1779 (void *)(offset + head),
1780 (void *)(long)(event->header.size),
1783 attr ? __event_name(attr->type, attr->config)
1791 process_event(event_t *event, unsigned long offset, unsigned long head)
1795 switch (event->header.type) {
1796 case PERF_EVENT_SAMPLE:
1797 return process_sample_event(event, offset, head);
1799 case PERF_EVENT_MMAP:
1800 return process_mmap_event(event, offset, head);
1802 case PERF_EVENT_COMM:
1803 return process_comm_event(event, offset, head);
1805 case PERF_EVENT_FORK:
1806 case PERF_EVENT_EXIT:
1807 return process_task_event(event, offset, head);
1809 case PERF_EVENT_LOST:
1810 return process_lost_event(event, offset, head);
1812 case PERF_EVENT_READ:
1813 return process_read_event(event, offset, head);
1816 * We dont process them right now but they are fine:
1819 case PERF_EVENT_THROTTLE:
1820 case PERF_EVENT_UNTHROTTLE:
1830 static u64 perf_header__sample_type(void)
1832 u64 sample_type = 0;
1835 for (i = 0; i < header->attrs; i++) {
1836 struct perf_header_attr *attr = header->attr[i];
1839 sample_type = attr->attr.sample_type;
1840 else if (sample_type != attr->attr.sample_type)
1841 die("non matching sample_type");
1847 static int __cmd_report(void)
1849 int ret, rc = EXIT_FAILURE;
1850 unsigned long offset = 0;
1851 unsigned long head, shift;
1857 register_idle_thread();
1860 perf_read_values_init(&show_threads_values);
1862 input = open(input_name, O_RDONLY);
1864 fprintf(stderr, " failed to open file: %s", input_name);
1865 if (!strcmp(input_name, "perf.data"))
1866 fprintf(stderr, " (try 'perf record' first)");
1867 fprintf(stderr, "\n");
1871 ret = fstat(input, &stat);
1873 perror("failed to stat file");
1877 if (!stat.st_size) {
1878 fprintf(stderr, "zero-sized file, nothing to do!\n");
1882 header = perf_header__read(input);
1883 head = header->data_offset;
1885 sample_type = perf_header__sample_type();
1887 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1888 if (sort__has_parent) {
1889 fprintf(stderr, "selected --sort parent, but no"
1890 " callchain data. Did you call"
1891 " perf record without -g?\n");
1895 fprintf(stderr, "selected -c but no callchain data."
1896 " Did you call perf record without"
1900 } else if (callchain_param.mode != CHAIN_NONE && !callchain) {
1902 if (register_callchain_param(&callchain_param) < 0) {
1903 fprintf(stderr, "Can't register callchain"
1909 if (load_kernel() < 0) {
1910 perror("failed to load kernel symbols");
1911 return EXIT_FAILURE;
1915 if (getcwd(__cwd, sizeof(__cwd)) == NULL) {
1916 perror("failed to get the current directory");
1917 return EXIT_FAILURE;
1919 cwdlen = strlen(cwd);
1925 shift = page_size * (head / page_size);
1930 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1931 MAP_SHARED, input, offset);
1932 if (buf == MAP_FAILED) {
1933 perror("failed to mmap file");
1938 event = (event_t *)(buf + head);
1940 size = event->header.size;
1944 if (head + event->header.size >= page_size * mmap_window) {
1947 shift = page_size * (head / page_size);
1949 ret = munmap(buf, page_size * mmap_window);
1957 size = event->header.size;
1959 dprintf("\n%p [%p]: event: %d\n",
1960 (void *)(offset + head),
1961 (void *)(long)event->header.size,
1962 event->header.type);
1964 if (!size || process_event(event, offset, head) < 0) {
1966 dprintf("%p [%p]: skipping unknown header type: %d\n",
1967 (void *)(offset + head),
1968 (void *)(long)(event->header.size),
1969 event->header.type);
1974 * assume we lost track of the stream, check alignment, and
1975 * increment a single u64 in the hope to catch on again 'soon'.
1978 if (unlikely(head & 7))
1986 if (offset + head >= header->data_offset + header->data_size)
1989 if (offset + head < (unsigned long)stat.st_size)
1996 dprintf(" IP events: %10ld\n", total);
1997 dprintf(" mmap events: %10ld\n", total_mmap);
1998 dprintf(" comm events: %10ld\n", total_comm);
1999 dprintf(" fork events: %10ld\n", total_fork);
2000 dprintf(" lost events: %10ld\n", total_lost);
2001 dprintf(" unknown events: %10ld\n", total_unknown);
2007 threads__fprintf(stdout);
2010 dsos__fprintf(stdout);
2013 output__resort(total);
2014 output__fprintf(stdout, total);
2017 perf_read_values_destroy(&show_threads_values);
2023 parse_callchain_opt(const struct option *opt __used, const char *arg,
2034 tok = strtok((char *)arg, ",");
2038 /* get the output mode */
2039 if (!strncmp(tok, "graph", strlen(arg)))
2040 callchain_param.mode = CHAIN_GRAPH_ABS;
2042 else if (!strncmp(tok, "flat", strlen(arg)))
2043 callchain_param.mode = CHAIN_FLAT;
2045 else if (!strncmp(tok, "fractal", strlen(arg)))
2046 callchain_param.mode = CHAIN_GRAPH_REL;
2048 else if (!strncmp(tok, "none", strlen(arg))) {
2049 callchain_param.mode = CHAIN_NONE;
2058 /* get the min percentage */
2059 tok = strtok(NULL, ",");
2063 callchain_param.min_percent = strtod(tok, &endptr);
2068 if (register_callchain_param(&callchain_param) < 0) {
2069 fprintf(stderr, "Can't register callchain params\n");
2075 static const char * const report_usage[] = {
2076 "perf report [<options>] <command>",
2080 static const struct option options[] = {
2081 OPT_STRING('i', "input", &input_name, "file",
2083 OPT_BOOLEAN('v', "verbose", &verbose,
2084 "be more verbose (show symbol address, etc)"),
2085 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
2086 "dump raw trace in ASCII"),
2087 OPT_STRING('k', "vmlinux", &vmlinux, "file", "vmlinux pathname"),
2088 OPT_BOOLEAN('m', "modules", &modules,
2089 "load module symbols - WARNING: use only with -k and LIVE kernel"),
2090 OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples,
2091 "Show a column with the number of samples"),
2092 OPT_BOOLEAN('T', "threads", &show_threads,
2093 "Show per-thread event counters"),
2094 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
2095 "sort by key(s): pid, comm, dso, symbol, parent"),
2096 OPT_BOOLEAN('P', "full-paths", &full_paths,
2097 "Don't shorten the pathnames taking into account the cwd"),
2098 OPT_STRING('p', "parent", &parent_pattern, "regex",
2099 "regex filter to identify parent, see: '--sort parent'"),
2100 OPT_BOOLEAN('x', "exclude-other", &exclude_other,
2101 "Only display entries with parent-match"),
2102 OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
2103 "Display callchains using output_type and min percent threshold. "
2104 "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
2105 OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]",
2106 "only consider symbols in these dsos"),
2107 OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]",
2108 "only consider symbols in these comms"),
2109 OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]",
2110 "only consider these symbols"),
2111 OPT_STRING('w', "column-widths", &col_width_list_str,
2113 "don't try to adjust column width, use these fixed values"),
2114 OPT_STRING('t', "field-separator", &field_sep, "separator",
2115 "separator for columns, no spaces will be added between "
2116 "columns '.' is reserved."),
2120 static void setup_sorting(void)
2122 char *tmp, *tok, *str = strdup(sort_order);
2124 for (tok = strtok_r(str, ", ", &tmp);
2125 tok; tok = strtok_r(NULL, ", ", &tmp)) {
2126 if (sort_dimension__add(tok) < 0) {
2127 error("Unknown --sort key: `%s'", tok);
2128 usage_with_options(report_usage, options);
2135 static void setup_list(struct strlist **list, const char *list_str,
2136 struct sort_entry *se, const char *list_name,
2140 *list = strlist__new(true, list_str);
2142 fprintf(stderr, "problems parsing %s list\n",
2146 if (strlist__nr_entries(*list) == 1) {
2147 fprintf(fp, "# %s: %s\n", list_name,
2148 strlist__entry(*list, 0)->s);
2154 int cmd_report(int argc, const char **argv, const char *prefix __used)
2158 page_size = getpagesize();
2160 argc = parse_options(argc, argv, options, report_usage, 0);
2164 if (parent_pattern != default_parent_pattern) {
2165 sort_dimension__add("parent");
2166 sort_parent.elide = 1;
2171 * Any (unrecognized) arguments left?
2174 usage_with_options(report_usage, options);
2178 setup_list(&dso_list, dso_list_str, &sort_dso, "dso", stdout);
2179 setup_list(&comm_list, comm_list_str, &sort_comm, "comm", stdout);
2180 setup_list(&sym_list, sym_list_str, &sort_sym, "symbol", stdout);
2182 if (field_sep && *field_sep == '.') {
2183 fputs("'.' is the only non valid --field-separator argument\n",
2188 return __cmd_report();