1 // SPDX-License-Identifier: GPL-2.0
19 #include <sys/types.h>
23 #include "linux/hash.h"
25 #include "bpf-event.h"
27 #include "sane_ctype.h"
28 #include <symbol/kallsyms.h>
29 #include <linux/mman.h>
31 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
33 static void dsos__init(struct dsos *dsos)
35 INIT_LIST_HEAD(&dsos->head);
37 init_rwsem(&dsos->lock);
40 static void machine__threads_init(struct machine *machine)
44 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
45 struct threads *threads = &machine->threads[i];
46 threads->entries = RB_ROOT_CACHED;
47 init_rwsem(&threads->lock);
49 INIT_LIST_HEAD(&threads->dead);
50 threads->last_match = NULL;
54 static int machine__set_mmap_name(struct machine *machine)
56 if (machine__is_host(machine))
57 machine->mmap_name = strdup("[kernel.kallsyms]");
58 else if (machine__is_default_guest(machine))
59 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
60 else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
62 machine->mmap_name = NULL;
64 return machine->mmap_name ? 0 : -ENOMEM;
67 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
71 memset(machine, 0, sizeof(*machine));
72 map_groups__init(&machine->kmaps, machine);
73 RB_CLEAR_NODE(&machine->rb_node);
74 dsos__init(&machine->dsos);
76 machine__threads_init(machine);
78 machine->vdso_info = NULL;
83 machine->id_hdr_size = 0;
84 machine->kptr_restrict_warned = false;
85 machine->comm_exec = false;
86 machine->kernel_start = 0;
87 machine->vmlinux_map = NULL;
89 machine->root_dir = strdup(root_dir);
90 if (machine->root_dir == NULL)
93 if (machine__set_mmap_name(machine))
96 if (pid != HOST_KERNEL_ID) {
97 struct thread *thread = machine__findnew_thread(machine, -1,
104 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
105 thread__set_comm(thread, comm, 0);
109 machine->current_tid = NULL;
114 zfree(&machine->root_dir);
115 zfree(&machine->mmap_name);
120 struct machine *machine__new_host(void)
122 struct machine *machine = malloc(sizeof(*machine));
124 if (machine != NULL) {
125 machine__init(machine, "", HOST_KERNEL_ID);
127 if (machine__create_kernel_maps(machine) < 0)
137 struct machine *machine__new_kallsyms(void)
139 struct machine *machine = machine__new_host();
142 * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
143 * ask for not using the kcore parsing code, once this one is fixed
144 * to create a map per module.
146 if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
147 machine__delete(machine);
154 static void dsos__purge(struct dsos *dsos)
158 down_write(&dsos->lock);
160 list_for_each_entry_safe(pos, n, &dsos->head, node) {
161 RB_CLEAR_NODE(&pos->rb_node);
163 list_del_init(&pos->node);
167 up_write(&dsos->lock);
170 static void dsos__exit(struct dsos *dsos)
173 exit_rwsem(&dsos->lock);
176 void machine__delete_threads(struct machine *machine)
181 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
182 struct threads *threads = &machine->threads[i];
183 down_write(&threads->lock);
184 nd = rb_first_cached(&threads->entries);
186 struct thread *t = rb_entry(nd, struct thread, rb_node);
189 __machine__remove_thread(machine, t, false);
191 up_write(&threads->lock);
195 void machine__exit(struct machine *machine)
202 machine__destroy_kernel_maps(machine);
203 map_groups__exit(&machine->kmaps);
204 dsos__exit(&machine->dsos);
205 machine__exit_vdso(machine);
206 zfree(&machine->root_dir);
207 zfree(&machine->mmap_name);
208 zfree(&machine->current_tid);
210 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
211 struct threads *threads = &machine->threads[i];
212 exit_rwsem(&threads->lock);
216 void machine__delete(struct machine *machine)
219 machine__exit(machine);
224 void machines__init(struct machines *machines)
226 machine__init(&machines->host, "", HOST_KERNEL_ID);
227 machines->guests = RB_ROOT_CACHED;
230 void machines__exit(struct machines *machines)
232 machine__exit(&machines->host);
236 struct machine *machines__add(struct machines *machines, pid_t pid,
237 const char *root_dir)
239 struct rb_node **p = &machines->guests.rb_root.rb_node;
240 struct rb_node *parent = NULL;
241 struct machine *pos, *machine = malloc(sizeof(*machine));
242 bool leftmost = true;
247 if (machine__init(machine, root_dir, pid) != 0) {
254 pos = rb_entry(parent, struct machine, rb_node);
263 rb_link_node(&machine->rb_node, parent, p);
264 rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
269 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
273 machines->host.comm_exec = comm_exec;
275 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
276 struct machine *machine = rb_entry(nd, struct machine, rb_node);
278 machine->comm_exec = comm_exec;
282 struct machine *machines__find(struct machines *machines, pid_t pid)
284 struct rb_node **p = &machines->guests.rb_root.rb_node;
285 struct rb_node *parent = NULL;
286 struct machine *machine;
287 struct machine *default_machine = NULL;
289 if (pid == HOST_KERNEL_ID)
290 return &machines->host;
294 machine = rb_entry(parent, struct machine, rb_node);
295 if (pid < machine->pid)
297 else if (pid > machine->pid)
302 default_machine = machine;
305 return default_machine;
308 struct machine *machines__findnew(struct machines *machines, pid_t pid)
311 const char *root_dir = "";
312 struct machine *machine = machines__find(machines, pid);
314 if (machine && (machine->pid == pid))
317 if ((pid != HOST_KERNEL_ID) &&
318 (pid != DEFAULT_GUEST_KERNEL_ID) &&
319 (symbol_conf.guestmount)) {
320 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
321 if (access(path, R_OK)) {
322 static struct strlist *seen;
325 seen = strlist__new(NULL, NULL);
327 if (!strlist__has_entry(seen, path)) {
328 pr_err("Can't access file %s\n", path);
329 strlist__add(seen, path);
337 machine = machines__add(machines, pid, root_dir);
342 void machines__process_guests(struct machines *machines,
343 machine__process_t process, void *data)
347 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
348 struct machine *pos = rb_entry(nd, struct machine, rb_node);
353 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
355 struct rb_node *node;
356 struct machine *machine;
358 machines->host.id_hdr_size = id_hdr_size;
360 for (node = rb_first_cached(&machines->guests); node;
361 node = rb_next(node)) {
362 machine = rb_entry(node, struct machine, rb_node);
363 machine->id_hdr_size = id_hdr_size;
369 static void machine__update_thread_pid(struct machine *machine,
370 struct thread *th, pid_t pid)
372 struct thread *leader;
374 if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
379 if (th->pid_ == th->tid)
382 leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
387 leader->mg = map_groups__new(machine);
392 if (th->mg == leader->mg)
397 * Maps are created from MMAP events which provide the pid and
398 * tid. Consequently there never should be any maps on a thread
399 * with an unknown pid. Just print an error if there are.
401 if (!map_groups__empty(th->mg))
402 pr_err("Discarding thread maps for %d:%d\n",
404 map_groups__put(th->mg);
407 th->mg = map_groups__get(leader->mg);
412 pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
417 * Front-end cache - TID lookups come in blocks,
418 * so most of the time we dont have to look up
421 static struct thread*
422 __threads__get_last_match(struct threads *threads, struct machine *machine,
427 th = threads->last_match;
429 if (th->tid == tid) {
430 machine__update_thread_pid(machine, th, pid);
431 return thread__get(th);
434 threads->last_match = NULL;
440 static struct thread*
441 threads__get_last_match(struct threads *threads, struct machine *machine,
444 struct thread *th = NULL;
446 if (perf_singlethreaded)
447 th = __threads__get_last_match(threads, machine, pid, tid);
453 __threads__set_last_match(struct threads *threads, struct thread *th)
455 threads->last_match = th;
459 threads__set_last_match(struct threads *threads, struct thread *th)
461 if (perf_singlethreaded)
462 __threads__set_last_match(threads, th);
466 * Caller must eventually drop thread->refcnt returned with a successful
467 * lookup/new thread inserted.
469 static struct thread *____machine__findnew_thread(struct machine *machine,
470 struct threads *threads,
471 pid_t pid, pid_t tid,
474 struct rb_node **p = &threads->entries.rb_root.rb_node;
475 struct rb_node *parent = NULL;
477 bool leftmost = true;
479 th = threads__get_last_match(threads, machine, pid, tid);
485 th = rb_entry(parent, struct thread, rb_node);
487 if (th->tid == tid) {
488 threads__set_last_match(threads, th);
489 machine__update_thread_pid(machine, th, pid);
490 return thread__get(th);
504 th = thread__new(pid, tid);
506 rb_link_node(&th->rb_node, parent, p);
507 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
510 * We have to initialize map_groups separately
511 * after rb tree is updated.
513 * The reason is that we call machine__findnew_thread
514 * within thread__init_map_groups to find the thread
515 * leader and that would screwed the rb tree.
517 if (thread__init_map_groups(th, machine)) {
518 rb_erase_cached(&th->rb_node, &threads->entries);
519 RB_CLEAR_NODE(&th->rb_node);
524 * It is now in the rbtree, get a ref
527 threads__set_last_match(threads, th);
534 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
536 return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
539 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
542 struct threads *threads = machine__threads(machine, tid);
545 down_write(&threads->lock);
546 th = __machine__findnew_thread(machine, pid, tid);
547 up_write(&threads->lock);
551 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
554 struct threads *threads = machine__threads(machine, tid);
557 down_read(&threads->lock);
558 th = ____machine__findnew_thread(machine, threads, pid, tid, false);
559 up_read(&threads->lock);
563 struct comm *machine__thread_exec_comm(struct machine *machine,
564 struct thread *thread)
566 if (machine->comm_exec)
567 return thread__exec_comm(thread);
569 return thread__comm(thread);
572 int machine__process_comm_event(struct machine *machine, union perf_event *event,
573 struct perf_sample *sample)
575 struct thread *thread = machine__findnew_thread(machine,
578 bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
582 machine->comm_exec = true;
585 perf_event__fprintf_comm(event, stdout);
587 if (thread == NULL ||
588 __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
589 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
598 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
599 union perf_event *event,
600 struct perf_sample *sample __maybe_unused)
602 struct thread *thread = machine__findnew_thread(machine,
603 event->namespaces.pid,
604 event->namespaces.tid);
607 WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
608 "\nWARNING: kernel seems to support more namespaces than perf"
609 " tool.\nTry updating the perf tool..\n\n");
611 WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
612 "\nWARNING: perf tool seems to support more namespaces than"
613 " the kernel.\nTry updating the kernel..\n\n");
616 perf_event__fprintf_namespaces(event, stdout);
618 if (thread == NULL ||
619 thread__set_namespaces(thread, sample->time, &event->namespaces)) {
620 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
629 int machine__process_lost_event(struct machine *machine __maybe_unused,
630 union perf_event *event, struct perf_sample *sample __maybe_unused)
632 dump_printf(": id:%" PRIu64 ": lost:%" PRIu64 "\n",
633 event->lost.id, event->lost.lost);
637 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
638 union perf_event *event, struct perf_sample *sample)
640 dump_printf(": id:%" PRIu64 ": lost samples :%" PRIu64 "\n",
641 sample->id, event->lost_samples.lost);
645 static struct dso *machine__findnew_module_dso(struct machine *machine,
647 const char *filename)
651 down_write(&machine->dsos.lock);
653 dso = __dsos__find(&machine->dsos, m->name, true);
655 dso = __dsos__addnew(&machine->dsos, m->name);
659 dso__set_module_info(dso, m, machine);
660 dso__set_long_name(dso, strdup(filename), true);
665 up_write(&machine->dsos.lock);
669 int machine__process_aux_event(struct machine *machine __maybe_unused,
670 union perf_event *event)
673 perf_event__fprintf_aux(event, stdout);
677 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
678 union perf_event *event)
681 perf_event__fprintf_itrace_start(event, stdout);
685 int machine__process_switch_event(struct machine *machine __maybe_unused,
686 union perf_event *event)
689 perf_event__fprintf_switch(event, stdout);
693 static int machine__process_ksymbol_register(struct machine *machine,
694 union perf_event *event,
695 struct perf_sample *sample __maybe_unused)
700 map = map_groups__find(&machine->kmaps, event->ksymbol_event.addr);
702 map = dso__new_map(event->ksymbol_event.name);
706 map->start = event->ksymbol_event.addr;
707 map->pgoff = map->start;
708 map->end = map->start + event->ksymbol_event.len;
709 map_groups__insert(&machine->kmaps, map);
712 sym = symbol__new(event->ksymbol_event.addr, event->ksymbol_event.len,
713 0, 0, event->ksymbol_event.name);
716 dso__insert_symbol(map->dso, sym);
720 static int machine__process_ksymbol_unregister(struct machine *machine,
721 union perf_event *event,
722 struct perf_sample *sample __maybe_unused)
726 map = map_groups__find(&machine->kmaps, event->ksymbol_event.addr);
728 map_groups__remove(&machine->kmaps, map);
733 int machine__process_ksymbol(struct machine *machine __maybe_unused,
734 union perf_event *event,
735 struct perf_sample *sample)
738 perf_event__fprintf_ksymbol(event, stdout);
740 if (event->ksymbol_event.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
741 return machine__process_ksymbol_unregister(machine, event,
743 return machine__process_ksymbol_register(machine, event, sample);
746 static void dso__adjust_kmod_long_name(struct dso *dso, const char *filename)
748 const char *dup_filename;
750 if (!filename || !dso || !dso->long_name)
752 if (dso->long_name[0] != '[')
754 if (!strchr(filename, '/'))
757 dup_filename = strdup(filename);
761 dso__set_long_name(dso, dup_filename, true);
764 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
765 const char *filename)
767 struct map *map = NULL;
768 struct dso *dso = NULL;
771 if (kmod_path__parse_name(&m, filename))
774 map = map_groups__find_by_name(&machine->kmaps, m.name);
777 * If the map's dso is an offline module, give dso__load()
778 * a chance to find the file path of that module by fixing
781 dso__adjust_kmod_long_name(map->dso, filename);
785 dso = machine__findnew_module_dso(machine, &m, filename);
789 map = map__new2(start, dso);
793 map_groups__insert(&machine->kmaps, map);
795 /* Put the map here because map_groups__insert alread got it */
798 /* put the dso here, corresponding to machine__findnew_module_dso */
804 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
807 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
809 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
810 struct machine *pos = rb_entry(nd, struct machine, rb_node);
811 ret += __dsos__fprintf(&pos->dsos.head, fp);
817 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
818 bool (skip)(struct dso *dso, int parm), int parm)
820 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
823 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
824 bool (skip)(struct dso *dso, int parm), int parm)
827 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
829 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
830 struct machine *pos = rb_entry(nd, struct machine, rb_node);
831 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
836 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
840 struct dso *kdso = machine__kernel_map(machine)->dso;
842 if (kdso->has_build_id) {
843 char filename[PATH_MAX];
844 if (dso__build_id_filename(kdso, filename, sizeof(filename),
846 printed += fprintf(fp, "[0] %s\n", filename);
849 for (i = 0; i < vmlinux_path__nr_entries; ++i)
850 printed += fprintf(fp, "[%d] %s\n",
851 i + kdso->has_build_id, vmlinux_path[i]);
856 size_t machine__fprintf(struct machine *machine, FILE *fp)
862 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
863 struct threads *threads = &machine->threads[i];
865 down_read(&threads->lock);
867 ret = fprintf(fp, "Threads: %u\n", threads->nr);
869 for (nd = rb_first_cached(&threads->entries); nd;
871 struct thread *pos = rb_entry(nd, struct thread, rb_node);
873 ret += thread__fprintf(pos, fp);
876 up_read(&threads->lock);
881 static struct dso *machine__get_kernel(struct machine *machine)
883 const char *vmlinux_name = machine->mmap_name;
886 if (machine__is_host(machine)) {
887 if (symbol_conf.vmlinux_name)
888 vmlinux_name = symbol_conf.vmlinux_name;
890 kernel = machine__findnew_kernel(machine, vmlinux_name,
891 "[kernel]", DSO_TYPE_KERNEL);
893 if (symbol_conf.default_guest_vmlinux_name)
894 vmlinux_name = symbol_conf.default_guest_vmlinux_name;
896 kernel = machine__findnew_kernel(machine, vmlinux_name,
898 DSO_TYPE_GUEST_KERNEL);
901 if (kernel != NULL && (!kernel->has_build_id))
902 dso__read_running_kernel_build_id(kernel, machine);
907 struct process_args {
911 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
914 if (machine__is_default_guest(machine))
915 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
917 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
920 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
922 /* Figure out the start address of kernel map from /proc/kallsyms.
923 * Returns the name of the start symbol in *symbol_name. Pass in NULL as
924 * symbol_name if it's not that important.
926 static int machine__get_running_kernel_start(struct machine *machine,
927 const char **symbol_name,
928 u64 *start, u64 *end)
930 char filename[PATH_MAX];
935 machine__get_kallsyms_filename(machine, filename, PATH_MAX);
937 if (symbol__restricted_filename(filename, "/proc/kallsyms"))
940 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
941 err = kallsyms__get_function_start(filename, name, &addr);
954 err = kallsyms__get_function_start(filename, "_etext", &addr);
961 int machine__create_extra_kernel_map(struct machine *machine,
963 struct extra_kernel_map *xm)
968 map = map__new2(xm->start, kernel);
973 map->pgoff = xm->pgoff;
975 kmap = map__kmap(map);
977 kmap->kmaps = &machine->kmaps;
978 strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
980 map_groups__insert(&machine->kmaps, map);
982 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
983 kmap->name, map->start, map->end);
990 static u64 find_entry_trampoline(struct dso *dso)
992 /* Duplicates are removed so lookup all aliases */
993 const char *syms[] = {
995 "__entry_trampoline_start",
996 "entry_SYSCALL_64_trampoline",
998 struct symbol *sym = dso__first_symbol(dso);
1001 for (; sym; sym = dso__next_symbol(sym)) {
1002 if (sym->binding != STB_GLOBAL)
1004 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1005 if (!strcmp(sym->name, syms[i]))
1014 * These values can be used for kernels that do not have symbols for the entry
1015 * trampolines in kallsyms.
1017 #define X86_64_CPU_ENTRY_AREA_PER_CPU 0xfffffe0000000000ULL
1018 #define X86_64_CPU_ENTRY_AREA_SIZE 0x2c000
1019 #define X86_64_ENTRY_TRAMPOLINE 0x6000
1021 /* Map x86_64 PTI entry trampolines */
1022 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1025 struct map_groups *kmaps = &machine->kmaps;
1026 struct maps *maps = &kmaps->maps;
1027 int nr_cpus_avail, cpu;
1033 * In the vmlinux case, pgoff is a virtual address which must now be
1034 * mapped to a vmlinux offset.
1036 for (map = maps__first(maps); map; map = map__next(map)) {
1037 struct kmap *kmap = __map__kmap(map);
1038 struct map *dest_map;
1040 if (!kmap || !is_entry_trampoline(kmap->name))
1043 dest_map = map_groups__find(kmaps, map->pgoff);
1044 if (dest_map != map)
1045 map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1048 if (found || machine->trampolines_mapped)
1051 pgoff = find_entry_trampoline(kernel);
1055 nr_cpus_avail = machine__nr_cpus_avail(machine);
1057 /* Add a 1 page map for each CPU's entry trampoline */
1058 for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1059 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1060 cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1061 X86_64_ENTRY_TRAMPOLINE;
1062 struct extra_kernel_map xm = {
1064 .end = va + page_size,
1068 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1070 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1074 machine->trampolines_mapped = nr_cpus_avail;
1079 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1080 struct dso *kernel __maybe_unused)
1086 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1091 /* In case of renewal the kernel map, destroy previous one */
1092 machine__destroy_kernel_maps(machine);
1094 machine->vmlinux_map = map__new2(0, kernel);
1095 if (machine->vmlinux_map == NULL)
1098 machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1099 map = machine__kernel_map(machine);
1100 kmap = map__kmap(map);
1104 kmap->kmaps = &machine->kmaps;
1105 map_groups__insert(&machine->kmaps, map);
1110 void machine__destroy_kernel_maps(struct machine *machine)
1113 struct map *map = machine__kernel_map(machine);
1118 kmap = map__kmap(map);
1119 map_groups__remove(&machine->kmaps, map);
1120 if (kmap && kmap->ref_reloc_sym) {
1121 zfree((char **)&kmap->ref_reloc_sym->name);
1122 zfree(&kmap->ref_reloc_sym);
1125 map__zput(machine->vmlinux_map);
1128 int machines__create_guest_kernel_maps(struct machines *machines)
1131 struct dirent **namelist = NULL;
1133 char path[PATH_MAX];
1137 if (symbol_conf.default_guest_vmlinux_name ||
1138 symbol_conf.default_guest_modules ||
1139 symbol_conf.default_guest_kallsyms) {
1140 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1143 if (symbol_conf.guestmount) {
1144 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1147 for (i = 0; i < items; i++) {
1148 if (!isdigit(namelist[i]->d_name[0])) {
1149 /* Filter out . and .. */
1152 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1153 if ((*endp != '\0') ||
1154 (endp == namelist[i]->d_name) ||
1155 (errno == ERANGE)) {
1156 pr_debug("invalid directory (%s). Skipping.\n",
1157 namelist[i]->d_name);
1160 sprintf(path, "%s/%s/proc/kallsyms",
1161 symbol_conf.guestmount,
1162 namelist[i]->d_name);
1163 ret = access(path, R_OK);
1165 pr_debug("Can't access file %s\n", path);
1168 machines__create_kernel_maps(machines, pid);
1177 void machines__destroy_kernel_maps(struct machines *machines)
1179 struct rb_node *next = rb_first_cached(&machines->guests);
1181 machine__destroy_kernel_maps(&machines->host);
1184 struct machine *pos = rb_entry(next, struct machine, rb_node);
1186 next = rb_next(&pos->rb_node);
1187 rb_erase_cached(&pos->rb_node, &machines->guests);
1188 machine__delete(pos);
1192 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1194 struct machine *machine = machines__findnew(machines, pid);
1196 if (machine == NULL)
1199 return machine__create_kernel_maps(machine);
1202 int machine__load_kallsyms(struct machine *machine, const char *filename)
1204 struct map *map = machine__kernel_map(machine);
1205 int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1208 dso__set_loaded(map->dso);
1210 * Since /proc/kallsyms will have multiple sessions for the
1211 * kernel, with modules between them, fixup the end of all
1214 map_groups__fixup_end(&machine->kmaps);
1220 int machine__load_vmlinux_path(struct machine *machine)
1222 struct map *map = machine__kernel_map(machine);
1223 int ret = dso__load_vmlinux_path(map->dso, map);
1226 dso__set_loaded(map->dso);
1231 static char *get_kernel_version(const char *root_dir)
1233 char version[PATH_MAX];
1236 const char *prefix = "Linux version ";
1238 sprintf(version, "%s/proc/version", root_dir);
1239 file = fopen(version, "r");
1243 tmp = fgets(version, sizeof(version), file);
1248 name = strstr(version, prefix);
1251 name += strlen(prefix);
1252 tmp = strchr(name, ' ');
1256 return strdup(name);
1259 static bool is_kmod_dso(struct dso *dso)
1261 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1262 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1265 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
1266 struct kmod_path *m)
1269 struct map *map = map_groups__find_by_name(mg, m->name);
1274 long_name = strdup(path);
1275 if (long_name == NULL)
1278 dso__set_long_name(map->dso, long_name, true);
1279 dso__kernel_module_get_build_id(map->dso, "");
1282 * Full name could reveal us kmod compression, so
1283 * we need to update the symtab_type if needed.
1285 if (m->comp && is_kmod_dso(map->dso)) {
1286 map->dso->symtab_type++;
1287 map->dso->comp = m->comp;
1293 static int map_groups__set_modules_path_dir(struct map_groups *mg,
1294 const char *dir_name, int depth)
1296 struct dirent *dent;
1297 DIR *dir = opendir(dir_name);
1301 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1305 while ((dent = readdir(dir)) != NULL) {
1306 char path[PATH_MAX];
1309 /*sshfs might return bad dent->d_type, so we have to stat*/
1310 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1311 if (stat(path, &st))
1314 if (S_ISDIR(st.st_mode)) {
1315 if (!strcmp(dent->d_name, ".") ||
1316 !strcmp(dent->d_name, ".."))
1319 /* Do not follow top-level source and build symlinks */
1321 if (!strcmp(dent->d_name, "source") ||
1322 !strcmp(dent->d_name, "build"))
1326 ret = map_groups__set_modules_path_dir(mg, path,
1333 ret = kmod_path__parse_name(&m, dent->d_name);
1338 ret = map_groups__set_module_path(mg, path, &m);
1352 static int machine__set_modules_path(struct machine *machine)
1355 char modules_path[PATH_MAX];
1357 version = get_kernel_version(machine->root_dir);
1361 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1362 machine->root_dir, version);
1365 return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1367 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1368 const char *name __maybe_unused)
1373 static int machine__create_module(void *arg, const char *name, u64 start,
1376 struct machine *machine = arg;
1379 if (arch__fix_module_text_start(&start, name) < 0)
1382 map = machine__findnew_module_map(machine, start, name);
1385 map->end = start + size;
1387 dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1392 static int machine__create_modules(struct machine *machine)
1394 const char *modules;
1395 char path[PATH_MAX];
1397 if (machine__is_default_guest(machine)) {
1398 modules = symbol_conf.default_guest_modules;
1400 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1404 if (symbol__restricted_filename(modules, "/proc/modules"))
1407 if (modules__parse(modules, machine, machine__create_module))
1410 if (!machine__set_modules_path(machine))
1413 pr_debug("Problems setting modules path maps, continuing anyway...\n");
1418 static void machine__set_kernel_mmap(struct machine *machine,
1421 machine->vmlinux_map->start = start;
1422 machine->vmlinux_map->end = end;
1424 * Be a bit paranoid here, some perf.data file came with
1425 * a zero sized synthesized MMAP event for the kernel.
1427 if (start == 0 && end == 0)
1428 machine->vmlinux_map->end = ~0ULL;
1431 static void machine__update_kernel_mmap(struct machine *machine,
1434 struct map *map = machine__kernel_map(machine);
1437 map_groups__remove(&machine->kmaps, map);
1439 machine__set_kernel_mmap(machine, start, end);
1441 map_groups__insert(&machine->kmaps, map);
1445 int machine__create_kernel_maps(struct machine *machine)
1447 struct dso *kernel = machine__get_kernel(machine);
1448 const char *name = NULL;
1450 u64 start = 0, end = ~0ULL;
1456 ret = __machine__create_kernel_maps(machine, kernel);
1460 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1461 if (machine__is_host(machine))
1462 pr_debug("Problems creating module maps, "
1463 "continuing anyway...\n");
1465 pr_debug("Problems creating module maps for guest %d, "
1466 "continuing anyway...\n", machine->pid);
1469 if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1471 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1472 machine__destroy_kernel_maps(machine);
1478 * we have a real start address now, so re-order the kmaps
1479 * assume it's the last in the kmaps
1481 machine__update_kernel_mmap(machine, start, end);
1484 if (machine__create_extra_kernel_maps(machine, kernel))
1485 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1488 /* update end address of the kernel map using adjacent module address */
1489 map = map__next(machine__kernel_map(machine));
1491 machine__set_kernel_mmap(machine, start, map->start);
1499 static bool machine__uses_kcore(struct machine *machine)
1503 list_for_each_entry(dso, &machine->dsos.head, node) {
1504 if (dso__is_kcore(dso))
1511 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1512 union perf_event *event)
1514 return machine__is(machine, "x86_64") &&
1515 is_entry_trampoline(event->mmap.filename);
1518 static int machine__process_extra_kernel_map(struct machine *machine,
1519 union perf_event *event)
1521 struct map *kernel_map = machine__kernel_map(machine);
1522 struct dso *kernel = kernel_map ? kernel_map->dso : NULL;
1523 struct extra_kernel_map xm = {
1524 .start = event->mmap.start,
1525 .end = event->mmap.start + event->mmap.len,
1526 .pgoff = event->mmap.pgoff,
1532 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1534 return machine__create_extra_kernel_map(machine, kernel, &xm);
1537 static int machine__process_kernel_mmap_event(struct machine *machine,
1538 union perf_event *event)
1541 enum dso_kernel_type kernel_type;
1542 bool is_kernel_mmap;
1544 /* If we have maps from kcore then we do not need or want any others */
1545 if (machine__uses_kcore(machine))
1548 if (machine__is_host(machine))
1549 kernel_type = DSO_TYPE_KERNEL;
1551 kernel_type = DSO_TYPE_GUEST_KERNEL;
1553 is_kernel_mmap = memcmp(event->mmap.filename,
1555 strlen(machine->mmap_name) - 1) == 0;
1556 if (event->mmap.filename[0] == '/' ||
1557 (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1558 map = machine__findnew_module_map(machine, event->mmap.start,
1559 event->mmap.filename);
1563 map->end = map->start + event->mmap.len;
1564 } else if (is_kernel_mmap) {
1565 const char *symbol_name = (event->mmap.filename +
1566 strlen(machine->mmap_name));
1568 * Should be there already, from the build-id table in
1571 struct dso *kernel = NULL;
1574 down_read(&machine->dsos.lock);
1576 list_for_each_entry(dso, &machine->dsos.head, node) {
1579 * The cpumode passed to is_kernel_module is not the
1580 * cpumode of *this* event. If we insist on passing
1581 * correct cpumode to is_kernel_module, we should
1582 * record the cpumode when we adding this dso to the
1585 * However we don't really need passing correct
1586 * cpumode. We know the correct cpumode must be kernel
1587 * mode (if not, we should not link it onto kernel_dsos
1590 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1591 * is_kernel_module() treats it as a kernel cpumode.
1595 is_kernel_module(dso->long_name,
1596 PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1604 up_read(&machine->dsos.lock);
1607 kernel = machine__findnew_dso(machine, machine->mmap_name);
1611 kernel->kernel = kernel_type;
1612 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1617 if (strstr(kernel->long_name, "vmlinux"))
1618 dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1620 machine__update_kernel_mmap(machine, event->mmap.start,
1621 event->mmap.start + event->mmap.len);
1624 * Avoid using a zero address (kptr_restrict) for the ref reloc
1625 * symbol. Effectively having zero here means that at record
1626 * time /proc/sys/kernel/kptr_restrict was non zero.
1628 if (event->mmap.pgoff != 0) {
1629 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1634 if (machine__is_default_guest(machine)) {
1636 * preload dso of guest kernel and modules
1638 dso__load(kernel, machine__kernel_map(machine));
1640 } else if (perf_event__is_extra_kernel_mmap(machine, event)) {
1641 return machine__process_extra_kernel_map(machine, event);
1648 int machine__process_mmap2_event(struct machine *machine,
1649 union perf_event *event,
1650 struct perf_sample *sample)
1652 struct thread *thread;
1657 perf_event__fprintf_mmap2(event, stdout);
1659 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1660 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1661 ret = machine__process_kernel_mmap_event(machine, event);
1667 thread = machine__findnew_thread(machine, event->mmap2.pid,
1672 map = map__new(machine, event->mmap2.start,
1673 event->mmap2.len, event->mmap2.pgoff,
1675 event->mmap2.min, event->mmap2.ino,
1676 event->mmap2.ino_generation,
1679 event->mmap2.filename, thread);
1682 goto out_problem_map;
1684 ret = thread__insert_map(thread, map);
1686 goto out_problem_insert;
1688 thread__put(thread);
1695 thread__put(thread);
1697 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1701 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1702 struct perf_sample *sample)
1704 struct thread *thread;
1710 perf_event__fprintf_mmap(event, stdout);
1712 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1713 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1714 ret = machine__process_kernel_mmap_event(machine, event);
1720 thread = machine__findnew_thread(machine, event->mmap.pid,
1725 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1728 map = map__new(machine, event->mmap.start,
1729 event->mmap.len, event->mmap.pgoff,
1730 0, 0, 0, 0, prot, 0,
1731 event->mmap.filename,
1735 goto out_problem_map;
1737 ret = thread__insert_map(thread, map);
1739 goto out_problem_insert;
1741 thread__put(thread);
1748 thread__put(thread);
1750 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1754 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1756 struct threads *threads = machine__threads(machine, th->tid);
1758 if (threads->last_match == th)
1759 threads__set_last_match(threads, NULL);
1761 BUG_ON(refcount_read(&th->refcnt) == 0);
1763 down_write(&threads->lock);
1764 rb_erase_cached(&th->rb_node, &threads->entries);
1765 RB_CLEAR_NODE(&th->rb_node);
1768 * Move it first to the dead_threads list, then drop the reference,
1769 * if this is the last reference, then the thread__delete destructor
1770 * will be called and we will remove it from the dead_threads list.
1772 list_add_tail(&th->node, &threads->dead);
1774 up_write(&threads->lock);
1778 void machine__remove_thread(struct machine *machine, struct thread *th)
1780 return __machine__remove_thread(machine, th, true);
1783 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1784 struct perf_sample *sample)
1786 struct thread *thread = machine__find_thread(machine,
1789 struct thread *parent = machine__findnew_thread(machine,
1792 bool do_maps_clone = true;
1796 perf_event__fprintf_task(event, stdout);
1799 * There may be an existing thread that is not actually the parent,
1800 * either because we are processing events out of order, or because the
1801 * (fork) event that would have removed the thread was lost. Assume the
1802 * latter case and continue on as best we can.
1804 if (parent->pid_ != (pid_t)event->fork.ppid) {
1805 dump_printf("removing erroneous parent thread %d/%d\n",
1806 parent->pid_, parent->tid);
1807 machine__remove_thread(machine, parent);
1808 thread__put(parent);
1809 parent = machine__findnew_thread(machine, event->fork.ppid,
1813 /* if a thread currently exists for the thread id remove it */
1814 if (thread != NULL) {
1815 machine__remove_thread(machine, thread);
1816 thread__put(thread);
1819 thread = machine__findnew_thread(machine, event->fork.pid,
1822 * When synthesizing FORK events, we are trying to create thread
1823 * objects for the already running tasks on the machine.
1825 * Normally, for a kernel FORK event, we want to clone the parent's
1826 * maps because that is what the kernel just did.
1828 * But when synthesizing, this should not be done. If we do, we end up
1829 * with overlapping maps as we process the sythesized MMAP2 events that
1830 * get delivered shortly thereafter.
1832 * Use the FORK event misc flags in an internal way to signal this
1833 * situation, so we can elide the map clone when appropriate.
1835 if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1836 do_maps_clone = false;
1838 if (thread == NULL || parent == NULL ||
1839 thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1840 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1843 thread__put(thread);
1844 thread__put(parent);
1849 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1850 struct perf_sample *sample __maybe_unused)
1852 struct thread *thread = machine__find_thread(machine,
1857 perf_event__fprintf_task(event, stdout);
1859 if (thread != NULL) {
1860 thread__exited(thread);
1861 thread__put(thread);
1867 int machine__process_event(struct machine *machine, union perf_event *event,
1868 struct perf_sample *sample)
1872 switch (event->header.type) {
1873 case PERF_RECORD_COMM:
1874 ret = machine__process_comm_event(machine, event, sample); break;
1875 case PERF_RECORD_MMAP:
1876 ret = machine__process_mmap_event(machine, event, sample); break;
1877 case PERF_RECORD_NAMESPACES:
1878 ret = machine__process_namespaces_event(machine, event, sample); break;
1879 case PERF_RECORD_MMAP2:
1880 ret = machine__process_mmap2_event(machine, event, sample); break;
1881 case PERF_RECORD_FORK:
1882 ret = machine__process_fork_event(machine, event, sample); break;
1883 case PERF_RECORD_EXIT:
1884 ret = machine__process_exit_event(machine, event, sample); break;
1885 case PERF_RECORD_LOST:
1886 ret = machine__process_lost_event(machine, event, sample); break;
1887 case PERF_RECORD_AUX:
1888 ret = machine__process_aux_event(machine, event); break;
1889 case PERF_RECORD_ITRACE_START:
1890 ret = machine__process_itrace_start_event(machine, event); break;
1891 case PERF_RECORD_LOST_SAMPLES:
1892 ret = machine__process_lost_samples_event(machine, event, sample); break;
1893 case PERF_RECORD_SWITCH:
1894 case PERF_RECORD_SWITCH_CPU_WIDE:
1895 ret = machine__process_switch_event(machine, event); break;
1896 case PERF_RECORD_KSYMBOL:
1897 ret = machine__process_ksymbol(machine, event, sample); break;
1898 case PERF_RECORD_BPF_EVENT:
1899 ret = machine__process_bpf_event(machine, event, sample); break;
1908 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1910 if (!regexec(regex, sym->name, 0, NULL, 0))
1915 static void ip__resolve_ams(struct thread *thread,
1916 struct addr_map_symbol *ams,
1919 struct addr_location al;
1921 memset(&al, 0, sizeof(al));
1923 * We cannot use the header.misc hint to determine whether a
1924 * branch stack address is user, kernel, guest, hypervisor.
1925 * Branches may straddle the kernel/user/hypervisor boundaries.
1926 * Thus, we have to try consecutively until we find a match
1927 * or else, the symbol is unknown
1929 thread__find_cpumode_addr_location(thread, ip, &al);
1932 ams->al_addr = al.addr;
1938 static void ip__resolve_data(struct thread *thread,
1939 u8 m, struct addr_map_symbol *ams,
1940 u64 addr, u64 phys_addr)
1942 struct addr_location al;
1944 memset(&al, 0, sizeof(al));
1946 thread__find_symbol(thread, m, addr, &al);
1949 ams->al_addr = al.addr;
1952 ams->phys_addr = phys_addr;
1955 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1956 struct addr_location *al)
1958 struct mem_info *mi = mem_info__new();
1963 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1964 ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
1965 sample->addr, sample->phys_addr);
1966 mi->data_src.val = sample->data_src;
1971 static char *callchain_srcline(struct map *map, struct symbol *sym, u64 ip)
1973 char *srcline = NULL;
1975 if (!map || callchain_param.key == CCKEY_FUNCTION)
1978 srcline = srcline__tree_find(&map->dso->srclines, ip);
1980 bool show_sym = false;
1981 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
1983 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
1984 sym, show_sym, show_addr, ip);
1985 srcline__tree_insert(&map->dso->srclines, ip, srcline);
1996 static int add_callchain_ip(struct thread *thread,
1997 struct callchain_cursor *cursor,
1998 struct symbol **parent,
1999 struct addr_location *root_al,
2003 struct branch_flags *flags,
2004 struct iterations *iter,
2007 struct addr_location al;
2008 int nr_loop_iter = 0;
2009 u64 iter_cycles = 0;
2010 const char *srcline = NULL;
2015 thread__find_cpumode_addr_location(thread, ip, &al);
2017 if (ip >= PERF_CONTEXT_MAX) {
2019 case PERF_CONTEXT_HV:
2020 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2022 case PERF_CONTEXT_KERNEL:
2023 *cpumode = PERF_RECORD_MISC_KERNEL;
2025 case PERF_CONTEXT_USER:
2026 *cpumode = PERF_RECORD_MISC_USER;
2029 pr_debug("invalid callchain context: "
2030 "%"PRId64"\n", (s64) ip);
2032 * It seems the callchain is corrupted.
2035 callchain_cursor_reset(cursor);
2040 thread__find_symbol(thread, *cpumode, ip, &al);
2043 if (al.sym != NULL) {
2044 if (perf_hpp_list.parent && !*parent &&
2045 symbol__match_regex(al.sym, &parent_regex))
2047 else if (have_ignore_callees && root_al &&
2048 symbol__match_regex(al.sym, &ignore_callees_regex)) {
2049 /* Treat this symbol as the root,
2050 forgetting its callees. */
2052 callchain_cursor_reset(cursor);
2056 if (symbol_conf.hide_unresolved && al.sym == NULL)
2060 nr_loop_iter = iter->nr_loop_iter;
2061 iter_cycles = iter->cycles;
2064 srcline = callchain_srcline(al.map, al.sym, al.addr);
2065 return callchain_cursor_append(cursor, ip, al.map, al.sym,
2066 branch, flags, nr_loop_iter,
2067 iter_cycles, branch_from, srcline);
2070 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2071 struct addr_location *al)
2074 const struct branch_stack *bs = sample->branch_stack;
2075 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2080 for (i = 0; i < bs->nr; i++) {
2081 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
2082 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
2083 bi[i].flags = bs->entries[i].flags;
2088 static void save_iterations(struct iterations *iter,
2089 struct branch_entry *be, int nr)
2093 iter->nr_loop_iter++;
2096 for (i = 0; i < nr; i++)
2097 iter->cycles += be[i].flags.cycles;
2102 #define NO_ENTRY 0xff
2104 #define PERF_MAX_BRANCH_DEPTH 127
2107 static int remove_loops(struct branch_entry *l, int nr,
2108 struct iterations *iter)
2111 unsigned char chash[CHASHSZ];
2113 memset(chash, NO_ENTRY, sizeof(chash));
2115 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2117 for (i = 0; i < nr; i++) {
2118 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2120 /* no collision handling for now */
2121 if (chash[h] == NO_ENTRY) {
2123 } else if (l[chash[h]].from == l[i].from) {
2124 bool is_loop = true;
2125 /* check if it is a real loop */
2127 for (j = chash[h]; j < i && i + off < nr; j++, off++)
2128 if (l[j].from != l[i + off].from) {
2135 save_iterations(iter + i + off,
2138 memmove(iter + i, iter + i + off,
2141 memmove(l + i, l + i + off,
2153 * Recolve LBR callstack chain sample
2155 * 1 on success get LBR callchain information
2156 * 0 no available LBR callchain information, should try fp
2157 * negative error code on other errors.
2159 static int resolve_lbr_callchain_sample(struct thread *thread,
2160 struct callchain_cursor *cursor,
2161 struct perf_sample *sample,
2162 struct symbol **parent,
2163 struct addr_location *root_al,
2166 struct ip_callchain *chain = sample->callchain;
2167 int chain_nr = min(max_stack, (int)chain->nr), i;
2168 u8 cpumode = PERF_RECORD_MISC_USER;
2169 u64 ip, branch_from = 0;
2171 for (i = 0; i < chain_nr; i++) {
2172 if (chain->ips[i] == PERF_CONTEXT_USER)
2176 /* LBR only affects the user callchain */
2177 if (i != chain_nr) {
2178 struct branch_stack *lbr_stack = sample->branch_stack;
2179 int lbr_nr = lbr_stack->nr, j, k;
2181 struct branch_flags *flags;
2183 * LBR callstack can only get user call chain.
2184 * The mix_chain_nr is kernel call chain
2185 * number plus LBR user call chain number.
2186 * i is kernel call chain number,
2187 * 1 is PERF_CONTEXT_USER,
2188 * lbr_nr + 1 is the user call chain number.
2189 * For details, please refer to the comments
2190 * in callchain__printf
2192 int mix_chain_nr = i + 1 + lbr_nr + 1;
2194 for (j = 0; j < mix_chain_nr; j++) {
2199 if (callchain_param.order == ORDER_CALLEE) {
2202 else if (j > i + 1) {
2204 ip = lbr_stack->entries[k].from;
2206 flags = &lbr_stack->entries[k].flags;
2208 ip = lbr_stack->entries[0].to;
2210 flags = &lbr_stack->entries[0].flags;
2212 lbr_stack->entries[0].from;
2217 ip = lbr_stack->entries[k].from;
2219 flags = &lbr_stack->entries[k].flags;
2221 else if (j > lbr_nr)
2222 ip = chain->ips[i + 1 - (j - lbr_nr)];
2224 ip = lbr_stack->entries[0].to;
2226 flags = &lbr_stack->entries[0].flags;
2228 lbr_stack->entries[0].from;
2232 err = add_callchain_ip(thread, cursor, parent,
2233 root_al, &cpumode, ip,
2234 branch, flags, NULL,
2237 return (err < 0) ? err : 0;
2245 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2246 struct callchain_cursor *cursor,
2247 struct symbol **parent,
2248 struct addr_location *root_al,
2249 u8 *cpumode, int ent)
2253 while (--ent >= 0) {
2254 u64 ip = chain->ips[ent];
2256 if (ip >= PERF_CONTEXT_MAX) {
2257 err = add_callchain_ip(thread, cursor, parent,
2258 root_al, cpumode, ip,
2259 false, NULL, NULL, 0);
2266 static int thread__resolve_callchain_sample(struct thread *thread,
2267 struct callchain_cursor *cursor,
2268 struct perf_evsel *evsel,
2269 struct perf_sample *sample,
2270 struct symbol **parent,
2271 struct addr_location *root_al,
2274 struct branch_stack *branch = sample->branch_stack;
2275 struct ip_callchain *chain = sample->callchain;
2277 u8 cpumode = PERF_RECORD_MISC_USER;
2278 int i, j, err, nr_entries;
2283 chain_nr = chain->nr;
2285 if (perf_evsel__has_branch_callstack(evsel)) {
2286 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2287 root_al, max_stack);
2289 return (err < 0) ? err : 0;
2293 * Based on DWARF debug information, some architectures skip
2294 * a callchain entry saved by the kernel.
2296 skip_idx = arch_skip_callchain_idx(thread, chain);
2299 * Add branches to call stack for easier browsing. This gives
2300 * more context for a sample than just the callers.
2302 * This uses individual histograms of paths compared to the
2303 * aggregated histograms the normal LBR mode uses.
2305 * Limitations for now:
2306 * - No extra filters
2307 * - No annotations (should annotate somehow)
2310 if (branch && callchain_param.branch_callstack) {
2311 int nr = min(max_stack, (int)branch->nr);
2312 struct branch_entry be[nr];
2313 struct iterations iter[nr];
2315 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2316 pr_warning("corrupted branch chain. skipping...\n");
2320 for (i = 0; i < nr; i++) {
2321 if (callchain_param.order == ORDER_CALLEE) {
2322 be[i] = branch->entries[i];
2328 * Check for overlap into the callchain.
2329 * The return address is one off compared to
2330 * the branch entry. To adjust for this
2331 * assume the calling instruction is not longer
2334 if (i == skip_idx ||
2335 chain->ips[first_call] >= PERF_CONTEXT_MAX)
2337 else if (be[i].from < chain->ips[first_call] &&
2338 be[i].from >= chain->ips[first_call] - 8)
2341 be[i] = branch->entries[branch->nr - i - 1];
2344 memset(iter, 0, sizeof(struct iterations) * nr);
2345 nr = remove_loops(be, nr, iter);
2347 for (i = 0; i < nr; i++) {
2348 err = add_callchain_ip(thread, cursor, parent,
2355 err = add_callchain_ip(thread, cursor, parent, root_al,
2372 if (callchain_param.order != ORDER_CALLEE) {
2373 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2374 &cpumode, chain->nr - first_call);
2376 return (err < 0) ? err : 0;
2378 for (i = first_call, nr_entries = 0;
2379 i < chain_nr && nr_entries < max_stack; i++) {
2382 if (callchain_param.order == ORDER_CALLEE)
2385 j = chain->nr - i - 1;
2387 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2392 if (ip < PERF_CONTEXT_MAX)
2394 else if (callchain_param.order != ORDER_CALLEE) {
2395 err = find_prev_cpumode(chain, thread, cursor, parent,
2396 root_al, &cpumode, j);
2398 return (err < 0) ? err : 0;
2402 err = add_callchain_ip(thread, cursor, parent,
2403 root_al, &cpumode, ip,
2404 false, NULL, NULL, 0);
2407 return (err < 0) ? err : 0;
2413 static int append_inlines(struct callchain_cursor *cursor,
2414 struct map *map, struct symbol *sym, u64 ip)
2416 struct inline_node *inline_node;
2417 struct inline_list *ilist;
2421 if (!symbol_conf.inline_name || !map || !sym)
2424 addr = map__map_ip(map, ip);
2425 addr = map__rip_2objdump(map, addr);
2427 inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2429 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2432 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2435 list_for_each_entry(ilist, &inline_node->val, list) {
2436 ret = callchain_cursor_append(cursor, ip, map,
2437 ilist->symbol, false,
2438 NULL, 0, 0, 0, ilist->srcline);
2447 static int unwind_entry(struct unwind_entry *entry, void *arg)
2449 struct callchain_cursor *cursor = arg;
2450 const char *srcline = NULL;
2451 u64 addr = entry->ip;
2453 if (symbol_conf.hide_unresolved && entry->sym == NULL)
2456 if (append_inlines(cursor, entry->map, entry->sym, entry->ip) == 0)
2460 * Convert entry->ip from a virtual address to an offset in
2461 * its corresponding binary.
2464 addr = map__map_ip(entry->map, entry->ip);
2466 srcline = callchain_srcline(entry->map, entry->sym, addr);
2467 return callchain_cursor_append(cursor, entry->ip,
2468 entry->map, entry->sym,
2469 false, NULL, 0, 0, 0, srcline);
2472 static int thread__resolve_callchain_unwind(struct thread *thread,
2473 struct callchain_cursor *cursor,
2474 struct perf_evsel *evsel,
2475 struct perf_sample *sample,
2478 /* Can we do dwarf post unwind? */
2479 if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2480 (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
2483 /* Bail out if nothing was captured. */
2484 if ((!sample->user_regs.regs) ||
2485 (!sample->user_stack.size))
2488 return unwind__get_entries(unwind_entry, cursor,
2489 thread, sample, max_stack);
2492 int thread__resolve_callchain(struct thread *thread,
2493 struct callchain_cursor *cursor,
2494 struct perf_evsel *evsel,
2495 struct perf_sample *sample,
2496 struct symbol **parent,
2497 struct addr_location *root_al,
2502 callchain_cursor_reset(cursor);
2504 if (callchain_param.order == ORDER_CALLEE) {
2505 ret = thread__resolve_callchain_sample(thread, cursor,
2511 ret = thread__resolve_callchain_unwind(thread, cursor,
2515 ret = thread__resolve_callchain_unwind(thread, cursor,
2520 ret = thread__resolve_callchain_sample(thread, cursor,
2529 int machine__for_each_thread(struct machine *machine,
2530 int (*fn)(struct thread *thread, void *p),
2533 struct threads *threads;
2535 struct thread *thread;
2539 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2540 threads = &machine->threads[i];
2541 for (nd = rb_first_cached(&threads->entries); nd;
2543 thread = rb_entry(nd, struct thread, rb_node);
2544 rc = fn(thread, priv);
2549 list_for_each_entry(thread, &threads->dead, node) {
2550 rc = fn(thread, priv);
2558 int machines__for_each_thread(struct machines *machines,
2559 int (*fn)(struct thread *thread, void *p),
2565 rc = machine__for_each_thread(&machines->host, fn, priv);
2569 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2570 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2572 rc = machine__for_each_thread(machine, fn, priv);
2579 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
2580 struct target *target, struct thread_map *threads,
2581 perf_event__handler_t process, bool data_mmap,
2582 unsigned int nr_threads_synthesize)
2584 if (target__has_task(target))
2585 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap);
2586 else if (target__has_cpu(target))
2587 return perf_event__synthesize_threads(tool, process,
2589 nr_threads_synthesize);
2590 /* command specified */
2594 pid_t machine__get_current_tid(struct machine *machine, int cpu)
2596 if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid)
2599 return machine->current_tid[cpu];
2602 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
2605 struct thread *thread;
2610 if (!machine->current_tid) {
2613 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t));
2614 if (!machine->current_tid)
2616 for (i = 0; i < MAX_NR_CPUS; i++)
2617 machine->current_tid[i] = -1;
2620 if (cpu >= MAX_NR_CPUS) {
2621 pr_err("Requested CPU %d too large. ", cpu);
2622 pr_err("Consider raising MAX_NR_CPUS\n");
2626 machine->current_tid[cpu] = tid;
2628 thread = machine__findnew_thread(machine, pid, tid);
2633 thread__put(thread);
2639 * Compares the raw arch string. N.B. see instead perf_env__arch() if a
2640 * normalized arch is needed.
2642 bool machine__is(struct machine *machine, const char *arch)
2644 return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
2647 int machine__nr_cpus_avail(struct machine *machine)
2649 return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
2652 int machine__get_kernel_start(struct machine *machine)
2654 struct map *map = machine__kernel_map(machine);
2658 * The only addresses above 2^63 are kernel addresses of a 64-bit
2659 * kernel. Note that addresses are unsigned so that on a 32-bit system
2660 * all addresses including kernel addresses are less than 2^32. In
2661 * that case (32-bit system), if the kernel mapping is unknown, all
2662 * addresses will be assumed to be in user space - see
2663 * machine__kernel_ip().
2665 machine->kernel_start = 1ULL << 63;
2667 err = map__load(map);
2669 * On x86_64, PTI entry trampolines are less than the
2670 * start of kernel text, but still above 2^63. So leave
2671 * kernel_start = 1ULL << 63 for x86_64.
2673 if (!err && !machine__is(machine, "x86_64"))
2674 machine->kernel_start = map->start;
2679 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
2681 u8 addr_cpumode = cpumode;
2684 if (!machine->single_address_space)
2687 kernel_ip = machine__kernel_ip(machine, addr);
2689 case PERF_RECORD_MISC_KERNEL:
2690 case PERF_RECORD_MISC_USER:
2691 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
2692 PERF_RECORD_MISC_USER;
2694 case PERF_RECORD_MISC_GUEST_KERNEL:
2695 case PERF_RECORD_MISC_GUEST_USER:
2696 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
2697 PERF_RECORD_MISC_GUEST_USER;
2703 return addr_cpumode;
2706 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2708 return dsos__findnew(&machine->dsos, filename);
2711 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2713 struct machine *machine = vmachine;
2715 struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
2720 *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2721 *addrp = map->unmap_ip(map, sym->start);