1 // SPDX-License-Identifier: GPL-2.0
19 #include <sys/types.h>
23 #include "linux/hash.h"
25 #include "bpf-event.h"
27 #include <linux/ctype.h>
28 #include <symbol/kallsyms.h>
29 #include <linux/mman.h>
30 #include <linux/zalloc.h>
32 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
34 static void dsos__init(struct dsos *dsos)
36 INIT_LIST_HEAD(&dsos->head);
38 init_rwsem(&dsos->lock);
41 static void machine__threads_init(struct machine *machine)
45 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
46 struct threads *threads = &machine->threads[i];
47 threads->entries = RB_ROOT_CACHED;
48 init_rwsem(&threads->lock);
50 INIT_LIST_HEAD(&threads->dead);
51 threads->last_match = NULL;
55 static int machine__set_mmap_name(struct machine *machine)
57 if (machine__is_host(machine))
58 machine->mmap_name = strdup("[kernel.kallsyms]");
59 else if (machine__is_default_guest(machine))
60 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
61 else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
63 machine->mmap_name = NULL;
65 return machine->mmap_name ? 0 : -ENOMEM;
68 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
72 memset(machine, 0, sizeof(*machine));
73 map_groups__init(&machine->kmaps, machine);
74 RB_CLEAR_NODE(&machine->rb_node);
75 dsos__init(&machine->dsos);
77 machine__threads_init(machine);
79 machine->vdso_info = NULL;
84 machine->id_hdr_size = 0;
85 machine->kptr_restrict_warned = false;
86 machine->comm_exec = false;
87 machine->kernel_start = 0;
88 machine->vmlinux_map = NULL;
90 machine->root_dir = strdup(root_dir);
91 if (machine->root_dir == NULL)
94 if (machine__set_mmap_name(machine))
97 if (pid != HOST_KERNEL_ID) {
98 struct thread *thread = machine__findnew_thread(machine, -1,
105 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
106 thread__set_comm(thread, comm, 0);
110 machine->current_tid = NULL;
115 zfree(&machine->root_dir);
116 zfree(&machine->mmap_name);
121 struct machine *machine__new_host(void)
123 struct machine *machine = malloc(sizeof(*machine));
125 if (machine != NULL) {
126 machine__init(machine, "", HOST_KERNEL_ID);
128 if (machine__create_kernel_maps(machine) < 0)
138 struct machine *machine__new_kallsyms(void)
140 struct machine *machine = machine__new_host();
143 * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
144 * ask for not using the kcore parsing code, once this one is fixed
145 * to create a map per module.
147 if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
148 machine__delete(machine);
155 static void dsos__purge(struct dsos *dsos)
159 down_write(&dsos->lock);
161 list_for_each_entry_safe(pos, n, &dsos->head, node) {
162 RB_CLEAR_NODE(&pos->rb_node);
164 list_del_init(&pos->node);
168 up_write(&dsos->lock);
171 static void dsos__exit(struct dsos *dsos)
174 exit_rwsem(&dsos->lock);
177 void machine__delete_threads(struct machine *machine)
182 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
183 struct threads *threads = &machine->threads[i];
184 down_write(&threads->lock);
185 nd = rb_first_cached(&threads->entries);
187 struct thread *t = rb_entry(nd, struct thread, rb_node);
190 __machine__remove_thread(machine, t, false);
192 up_write(&threads->lock);
196 void machine__exit(struct machine *machine)
203 machine__destroy_kernel_maps(machine);
204 map_groups__exit(&machine->kmaps);
205 dsos__exit(&machine->dsos);
206 machine__exit_vdso(machine);
207 zfree(&machine->root_dir);
208 zfree(&machine->mmap_name);
209 zfree(&machine->current_tid);
211 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
212 struct threads *threads = &machine->threads[i];
213 struct thread *thread, *n;
215 * Forget about the dead, at this point whatever threads were
216 * left in the dead lists better have a reference count taken
217 * by who is using them, and then, when they drop those references
218 * and it finally hits zero, thread__put() will check and see that
219 * its not in the dead threads list and will not try to remove it
220 * from there, just calling thread__delete() straight away.
222 list_for_each_entry_safe(thread, n, &threads->dead, node)
223 list_del_init(&thread->node);
225 exit_rwsem(&threads->lock);
229 void machine__delete(struct machine *machine)
232 machine__exit(machine);
237 void machines__init(struct machines *machines)
239 machine__init(&machines->host, "", HOST_KERNEL_ID);
240 machines->guests = RB_ROOT_CACHED;
243 void machines__exit(struct machines *machines)
245 machine__exit(&machines->host);
249 struct machine *machines__add(struct machines *machines, pid_t pid,
250 const char *root_dir)
252 struct rb_node **p = &machines->guests.rb_root.rb_node;
253 struct rb_node *parent = NULL;
254 struct machine *pos, *machine = malloc(sizeof(*machine));
255 bool leftmost = true;
260 if (machine__init(machine, root_dir, pid) != 0) {
267 pos = rb_entry(parent, struct machine, rb_node);
276 rb_link_node(&machine->rb_node, parent, p);
277 rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
282 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
286 machines->host.comm_exec = comm_exec;
288 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
289 struct machine *machine = rb_entry(nd, struct machine, rb_node);
291 machine->comm_exec = comm_exec;
295 struct machine *machines__find(struct machines *machines, pid_t pid)
297 struct rb_node **p = &machines->guests.rb_root.rb_node;
298 struct rb_node *parent = NULL;
299 struct machine *machine;
300 struct machine *default_machine = NULL;
302 if (pid == HOST_KERNEL_ID)
303 return &machines->host;
307 machine = rb_entry(parent, struct machine, rb_node);
308 if (pid < machine->pid)
310 else if (pid > machine->pid)
315 default_machine = machine;
318 return default_machine;
321 struct machine *machines__findnew(struct machines *machines, pid_t pid)
324 const char *root_dir = "";
325 struct machine *machine = machines__find(machines, pid);
327 if (machine && (machine->pid == pid))
330 if ((pid != HOST_KERNEL_ID) &&
331 (pid != DEFAULT_GUEST_KERNEL_ID) &&
332 (symbol_conf.guestmount)) {
333 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
334 if (access(path, R_OK)) {
335 static struct strlist *seen;
338 seen = strlist__new(NULL, NULL);
340 if (!strlist__has_entry(seen, path)) {
341 pr_err("Can't access file %s\n", path);
342 strlist__add(seen, path);
350 machine = machines__add(machines, pid, root_dir);
355 void machines__process_guests(struct machines *machines,
356 machine__process_t process, void *data)
360 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
361 struct machine *pos = rb_entry(nd, struct machine, rb_node);
366 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
368 struct rb_node *node;
369 struct machine *machine;
371 machines->host.id_hdr_size = id_hdr_size;
373 for (node = rb_first_cached(&machines->guests); node;
374 node = rb_next(node)) {
375 machine = rb_entry(node, struct machine, rb_node);
376 machine->id_hdr_size = id_hdr_size;
382 static void machine__update_thread_pid(struct machine *machine,
383 struct thread *th, pid_t pid)
385 struct thread *leader;
387 if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
392 if (th->pid_ == th->tid)
395 leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
400 leader->mg = map_groups__new(machine);
405 if (th->mg == leader->mg)
410 * Maps are created from MMAP events which provide the pid and
411 * tid. Consequently there never should be any maps on a thread
412 * with an unknown pid. Just print an error if there are.
414 if (!map_groups__empty(th->mg))
415 pr_err("Discarding thread maps for %d:%d\n",
417 map_groups__put(th->mg);
420 th->mg = map_groups__get(leader->mg);
425 pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
430 * Front-end cache - TID lookups come in blocks,
431 * so most of the time we dont have to look up
434 static struct thread*
435 __threads__get_last_match(struct threads *threads, struct machine *machine,
440 th = threads->last_match;
442 if (th->tid == tid) {
443 machine__update_thread_pid(machine, th, pid);
444 return thread__get(th);
447 threads->last_match = NULL;
453 static struct thread*
454 threads__get_last_match(struct threads *threads, struct machine *machine,
457 struct thread *th = NULL;
459 if (perf_singlethreaded)
460 th = __threads__get_last_match(threads, machine, pid, tid);
466 __threads__set_last_match(struct threads *threads, struct thread *th)
468 threads->last_match = th;
472 threads__set_last_match(struct threads *threads, struct thread *th)
474 if (perf_singlethreaded)
475 __threads__set_last_match(threads, th);
479 * Caller must eventually drop thread->refcnt returned with a successful
480 * lookup/new thread inserted.
482 static struct thread *____machine__findnew_thread(struct machine *machine,
483 struct threads *threads,
484 pid_t pid, pid_t tid,
487 struct rb_node **p = &threads->entries.rb_root.rb_node;
488 struct rb_node *parent = NULL;
490 bool leftmost = true;
492 th = threads__get_last_match(threads, machine, pid, tid);
498 th = rb_entry(parent, struct thread, rb_node);
500 if (th->tid == tid) {
501 threads__set_last_match(threads, th);
502 machine__update_thread_pid(machine, th, pid);
503 return thread__get(th);
517 th = thread__new(pid, tid);
519 rb_link_node(&th->rb_node, parent, p);
520 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
523 * We have to initialize map_groups separately
524 * after rb tree is updated.
526 * The reason is that we call machine__findnew_thread
527 * within thread__init_map_groups to find the thread
528 * leader and that would screwed the rb tree.
530 if (thread__init_map_groups(th, machine)) {
531 rb_erase_cached(&th->rb_node, &threads->entries);
532 RB_CLEAR_NODE(&th->rb_node);
537 * It is now in the rbtree, get a ref
540 threads__set_last_match(threads, th);
547 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
549 return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
552 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
555 struct threads *threads = machine__threads(machine, tid);
558 down_write(&threads->lock);
559 th = __machine__findnew_thread(machine, pid, tid);
560 up_write(&threads->lock);
564 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
567 struct threads *threads = machine__threads(machine, tid);
570 down_read(&threads->lock);
571 th = ____machine__findnew_thread(machine, threads, pid, tid, false);
572 up_read(&threads->lock);
576 struct comm *machine__thread_exec_comm(struct machine *machine,
577 struct thread *thread)
579 if (machine->comm_exec)
580 return thread__exec_comm(thread);
582 return thread__comm(thread);
585 int machine__process_comm_event(struct machine *machine, union perf_event *event,
586 struct perf_sample *sample)
588 struct thread *thread = machine__findnew_thread(machine,
591 bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
595 machine->comm_exec = true;
598 perf_event__fprintf_comm(event, stdout);
600 if (thread == NULL ||
601 __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
602 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
611 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
612 union perf_event *event,
613 struct perf_sample *sample __maybe_unused)
615 struct thread *thread = machine__findnew_thread(machine,
616 event->namespaces.pid,
617 event->namespaces.tid);
620 WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
621 "\nWARNING: kernel seems to support more namespaces than perf"
622 " tool.\nTry updating the perf tool..\n\n");
624 WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
625 "\nWARNING: perf tool seems to support more namespaces than"
626 " the kernel.\nTry updating the kernel..\n\n");
629 perf_event__fprintf_namespaces(event, stdout);
631 if (thread == NULL ||
632 thread__set_namespaces(thread, sample->time, &event->namespaces)) {
633 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
642 int machine__process_lost_event(struct machine *machine __maybe_unused,
643 union perf_event *event, struct perf_sample *sample __maybe_unused)
645 dump_printf(": id:%" PRIu64 ": lost:%" PRIu64 "\n",
646 event->lost.id, event->lost.lost);
650 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
651 union perf_event *event, struct perf_sample *sample)
653 dump_printf(": id:%" PRIu64 ": lost samples :%" PRIu64 "\n",
654 sample->id, event->lost_samples.lost);
658 static struct dso *machine__findnew_module_dso(struct machine *machine,
660 const char *filename)
664 down_write(&machine->dsos.lock);
666 dso = __dsos__find(&machine->dsos, m->name, true);
668 dso = __dsos__addnew(&machine->dsos, m->name);
672 dso__set_module_info(dso, m, machine);
673 dso__set_long_name(dso, strdup(filename), true);
678 up_write(&machine->dsos.lock);
682 int machine__process_aux_event(struct machine *machine __maybe_unused,
683 union perf_event *event)
686 perf_event__fprintf_aux(event, stdout);
690 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
691 union perf_event *event)
694 perf_event__fprintf_itrace_start(event, stdout);
698 int machine__process_switch_event(struct machine *machine __maybe_unused,
699 union perf_event *event)
702 perf_event__fprintf_switch(event, stdout);
706 static int machine__process_ksymbol_register(struct machine *machine,
707 union perf_event *event,
708 struct perf_sample *sample __maybe_unused)
713 map = map_groups__find(&machine->kmaps, event->ksymbol_event.addr);
715 map = dso__new_map(event->ksymbol_event.name);
719 map->start = event->ksymbol_event.addr;
720 map->end = map->start + event->ksymbol_event.len;
721 map_groups__insert(&machine->kmaps, map);
724 sym = symbol__new(map->map_ip(map, map->start),
725 event->ksymbol_event.len,
726 0, 0, event->ksymbol_event.name);
729 dso__insert_symbol(map->dso, sym);
733 static int machine__process_ksymbol_unregister(struct machine *machine,
734 union perf_event *event,
735 struct perf_sample *sample __maybe_unused)
739 map = map_groups__find(&machine->kmaps, event->ksymbol_event.addr);
741 map_groups__remove(&machine->kmaps, map);
746 int machine__process_ksymbol(struct machine *machine __maybe_unused,
747 union perf_event *event,
748 struct perf_sample *sample)
751 perf_event__fprintf_ksymbol(event, stdout);
753 if (event->ksymbol_event.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
754 return machine__process_ksymbol_unregister(machine, event,
756 return machine__process_ksymbol_register(machine, event, sample);
759 static void dso__adjust_kmod_long_name(struct dso *dso, const char *filename)
761 const char *dup_filename;
763 if (!filename || !dso || !dso->long_name)
765 if (dso->long_name[0] != '[')
767 if (!strchr(filename, '/'))
770 dup_filename = strdup(filename);
774 dso__set_long_name(dso, dup_filename, true);
777 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
778 const char *filename)
780 struct map *map = NULL;
781 struct dso *dso = NULL;
784 if (kmod_path__parse_name(&m, filename))
787 map = map_groups__find_by_name(&machine->kmaps, m.name);
790 * If the map's dso is an offline module, give dso__load()
791 * a chance to find the file path of that module by fixing
794 dso__adjust_kmod_long_name(map->dso, filename);
798 dso = machine__findnew_module_dso(machine, &m, filename);
802 map = map__new2(start, dso);
806 map_groups__insert(&machine->kmaps, map);
808 /* Put the map here because map_groups__insert alread got it */
811 /* put the dso here, corresponding to machine__findnew_module_dso */
817 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
820 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
822 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
823 struct machine *pos = rb_entry(nd, struct machine, rb_node);
824 ret += __dsos__fprintf(&pos->dsos.head, fp);
830 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
831 bool (skip)(struct dso *dso, int parm), int parm)
833 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
836 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
837 bool (skip)(struct dso *dso, int parm), int parm)
840 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
842 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
843 struct machine *pos = rb_entry(nd, struct machine, rb_node);
844 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
849 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
853 struct dso *kdso = machine__kernel_map(machine)->dso;
855 if (kdso->has_build_id) {
856 char filename[PATH_MAX];
857 if (dso__build_id_filename(kdso, filename, sizeof(filename),
859 printed += fprintf(fp, "[0] %s\n", filename);
862 for (i = 0; i < vmlinux_path__nr_entries; ++i)
863 printed += fprintf(fp, "[%d] %s\n",
864 i + kdso->has_build_id, vmlinux_path[i]);
869 size_t machine__fprintf(struct machine *machine, FILE *fp)
875 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
876 struct threads *threads = &machine->threads[i];
878 down_read(&threads->lock);
880 ret = fprintf(fp, "Threads: %u\n", threads->nr);
882 for (nd = rb_first_cached(&threads->entries); nd;
884 struct thread *pos = rb_entry(nd, struct thread, rb_node);
886 ret += thread__fprintf(pos, fp);
889 up_read(&threads->lock);
894 static struct dso *machine__get_kernel(struct machine *machine)
896 const char *vmlinux_name = machine->mmap_name;
899 if (machine__is_host(machine)) {
900 if (symbol_conf.vmlinux_name)
901 vmlinux_name = symbol_conf.vmlinux_name;
903 kernel = machine__findnew_kernel(machine, vmlinux_name,
904 "[kernel]", DSO_TYPE_KERNEL);
906 if (symbol_conf.default_guest_vmlinux_name)
907 vmlinux_name = symbol_conf.default_guest_vmlinux_name;
909 kernel = machine__findnew_kernel(machine, vmlinux_name,
911 DSO_TYPE_GUEST_KERNEL);
914 if (kernel != NULL && (!kernel->has_build_id))
915 dso__read_running_kernel_build_id(kernel, machine);
920 struct process_args {
924 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
927 if (machine__is_default_guest(machine))
928 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
930 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
933 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
935 /* Figure out the start address of kernel map from /proc/kallsyms.
936 * Returns the name of the start symbol in *symbol_name. Pass in NULL as
937 * symbol_name if it's not that important.
939 static int machine__get_running_kernel_start(struct machine *machine,
940 const char **symbol_name,
941 u64 *start, u64 *end)
943 char filename[PATH_MAX];
948 machine__get_kallsyms_filename(machine, filename, PATH_MAX);
950 if (symbol__restricted_filename(filename, "/proc/kallsyms"))
953 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
954 err = kallsyms__get_function_start(filename, name, &addr);
967 err = kallsyms__get_function_start(filename, "_etext", &addr);
974 int machine__create_extra_kernel_map(struct machine *machine,
976 struct extra_kernel_map *xm)
981 map = map__new2(xm->start, kernel);
986 map->pgoff = xm->pgoff;
988 kmap = map__kmap(map);
990 kmap->kmaps = &machine->kmaps;
991 strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
993 map_groups__insert(&machine->kmaps, map);
995 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
996 kmap->name, map->start, map->end);
1003 static u64 find_entry_trampoline(struct dso *dso)
1005 /* Duplicates are removed so lookup all aliases */
1006 const char *syms[] = {
1007 "_entry_trampoline",
1008 "__entry_trampoline_start",
1009 "entry_SYSCALL_64_trampoline",
1011 struct symbol *sym = dso__first_symbol(dso);
1014 for (; sym; sym = dso__next_symbol(sym)) {
1015 if (sym->binding != STB_GLOBAL)
1017 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1018 if (!strcmp(sym->name, syms[i]))
1027 * These values can be used for kernels that do not have symbols for the entry
1028 * trampolines in kallsyms.
1030 #define X86_64_CPU_ENTRY_AREA_PER_CPU 0xfffffe0000000000ULL
1031 #define X86_64_CPU_ENTRY_AREA_SIZE 0x2c000
1032 #define X86_64_ENTRY_TRAMPOLINE 0x6000
1034 /* Map x86_64 PTI entry trampolines */
1035 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1038 struct map_groups *kmaps = &machine->kmaps;
1039 struct maps *maps = &kmaps->maps;
1040 int nr_cpus_avail, cpu;
1046 * In the vmlinux case, pgoff is a virtual address which must now be
1047 * mapped to a vmlinux offset.
1049 for (map = maps__first(maps); map; map = map__next(map)) {
1050 struct kmap *kmap = __map__kmap(map);
1051 struct map *dest_map;
1053 if (!kmap || !is_entry_trampoline(kmap->name))
1056 dest_map = map_groups__find(kmaps, map->pgoff);
1057 if (dest_map != map)
1058 map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1061 if (found || machine->trampolines_mapped)
1064 pgoff = find_entry_trampoline(kernel);
1068 nr_cpus_avail = machine__nr_cpus_avail(machine);
1070 /* Add a 1 page map for each CPU's entry trampoline */
1071 for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1072 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1073 cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1074 X86_64_ENTRY_TRAMPOLINE;
1075 struct extra_kernel_map xm = {
1077 .end = va + page_size,
1081 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1083 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1087 machine->trampolines_mapped = nr_cpus_avail;
1092 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1093 struct dso *kernel __maybe_unused)
1099 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1104 /* In case of renewal the kernel map, destroy previous one */
1105 machine__destroy_kernel_maps(machine);
1107 machine->vmlinux_map = map__new2(0, kernel);
1108 if (machine->vmlinux_map == NULL)
1111 machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1112 map = machine__kernel_map(machine);
1113 kmap = map__kmap(map);
1117 kmap->kmaps = &machine->kmaps;
1118 map_groups__insert(&machine->kmaps, map);
1123 void machine__destroy_kernel_maps(struct machine *machine)
1126 struct map *map = machine__kernel_map(machine);
1131 kmap = map__kmap(map);
1132 map_groups__remove(&machine->kmaps, map);
1133 if (kmap && kmap->ref_reloc_sym) {
1134 zfree((char **)&kmap->ref_reloc_sym->name);
1135 zfree(&kmap->ref_reloc_sym);
1138 map__zput(machine->vmlinux_map);
1141 int machines__create_guest_kernel_maps(struct machines *machines)
1144 struct dirent **namelist = NULL;
1146 char path[PATH_MAX];
1150 if (symbol_conf.default_guest_vmlinux_name ||
1151 symbol_conf.default_guest_modules ||
1152 symbol_conf.default_guest_kallsyms) {
1153 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1156 if (symbol_conf.guestmount) {
1157 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1160 for (i = 0; i < items; i++) {
1161 if (!isdigit(namelist[i]->d_name[0])) {
1162 /* Filter out . and .. */
1165 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1166 if ((*endp != '\0') ||
1167 (endp == namelist[i]->d_name) ||
1168 (errno == ERANGE)) {
1169 pr_debug("invalid directory (%s). Skipping.\n",
1170 namelist[i]->d_name);
1173 sprintf(path, "%s/%s/proc/kallsyms",
1174 symbol_conf.guestmount,
1175 namelist[i]->d_name);
1176 ret = access(path, R_OK);
1178 pr_debug("Can't access file %s\n", path);
1181 machines__create_kernel_maps(machines, pid);
1190 void machines__destroy_kernel_maps(struct machines *machines)
1192 struct rb_node *next = rb_first_cached(&machines->guests);
1194 machine__destroy_kernel_maps(&machines->host);
1197 struct machine *pos = rb_entry(next, struct machine, rb_node);
1199 next = rb_next(&pos->rb_node);
1200 rb_erase_cached(&pos->rb_node, &machines->guests);
1201 machine__delete(pos);
1205 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1207 struct machine *machine = machines__findnew(machines, pid);
1209 if (machine == NULL)
1212 return machine__create_kernel_maps(machine);
1215 int machine__load_kallsyms(struct machine *machine, const char *filename)
1217 struct map *map = machine__kernel_map(machine);
1218 int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1221 dso__set_loaded(map->dso);
1223 * Since /proc/kallsyms will have multiple sessions for the
1224 * kernel, with modules between them, fixup the end of all
1227 map_groups__fixup_end(&machine->kmaps);
1233 int machine__load_vmlinux_path(struct machine *machine)
1235 struct map *map = machine__kernel_map(machine);
1236 int ret = dso__load_vmlinux_path(map->dso, map);
1239 dso__set_loaded(map->dso);
1244 static char *get_kernel_version(const char *root_dir)
1246 char version[PATH_MAX];
1249 const char *prefix = "Linux version ";
1251 sprintf(version, "%s/proc/version", root_dir);
1252 file = fopen(version, "r");
1256 tmp = fgets(version, sizeof(version), file);
1261 name = strstr(version, prefix);
1264 name += strlen(prefix);
1265 tmp = strchr(name, ' ');
1269 return strdup(name);
1272 static bool is_kmod_dso(struct dso *dso)
1274 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1275 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1278 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
1279 struct kmod_path *m)
1282 struct map *map = map_groups__find_by_name(mg, m->name);
1287 long_name = strdup(path);
1288 if (long_name == NULL)
1291 dso__set_long_name(map->dso, long_name, true);
1292 dso__kernel_module_get_build_id(map->dso, "");
1295 * Full name could reveal us kmod compression, so
1296 * we need to update the symtab_type if needed.
1298 if (m->comp && is_kmod_dso(map->dso)) {
1299 map->dso->symtab_type++;
1300 map->dso->comp = m->comp;
1306 static int map_groups__set_modules_path_dir(struct map_groups *mg,
1307 const char *dir_name, int depth)
1309 struct dirent *dent;
1310 DIR *dir = opendir(dir_name);
1314 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1318 while ((dent = readdir(dir)) != NULL) {
1319 char path[PATH_MAX];
1322 /*sshfs might return bad dent->d_type, so we have to stat*/
1323 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1324 if (stat(path, &st))
1327 if (S_ISDIR(st.st_mode)) {
1328 if (!strcmp(dent->d_name, ".") ||
1329 !strcmp(dent->d_name, ".."))
1332 /* Do not follow top-level source and build symlinks */
1334 if (!strcmp(dent->d_name, "source") ||
1335 !strcmp(dent->d_name, "build"))
1339 ret = map_groups__set_modules_path_dir(mg, path,
1346 ret = kmod_path__parse_name(&m, dent->d_name);
1351 ret = map_groups__set_module_path(mg, path, &m);
1365 static int machine__set_modules_path(struct machine *machine)
1368 char modules_path[PATH_MAX];
1370 version = get_kernel_version(machine->root_dir);
1374 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1375 machine->root_dir, version);
1378 return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1380 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1381 const char *name __maybe_unused)
1386 static int machine__create_module(void *arg, const char *name, u64 start,
1389 struct machine *machine = arg;
1392 if (arch__fix_module_text_start(&start, name) < 0)
1395 map = machine__findnew_module_map(machine, start, name);
1398 map->end = start + size;
1400 dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1405 static int machine__create_modules(struct machine *machine)
1407 const char *modules;
1408 char path[PATH_MAX];
1410 if (machine__is_default_guest(machine)) {
1411 modules = symbol_conf.default_guest_modules;
1413 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1417 if (symbol__restricted_filename(modules, "/proc/modules"))
1420 if (modules__parse(modules, machine, machine__create_module))
1423 if (!machine__set_modules_path(machine))
1426 pr_debug("Problems setting modules path maps, continuing anyway...\n");
1431 static void machine__set_kernel_mmap(struct machine *machine,
1434 machine->vmlinux_map->start = start;
1435 machine->vmlinux_map->end = end;
1437 * Be a bit paranoid here, some perf.data file came with
1438 * a zero sized synthesized MMAP event for the kernel.
1440 if (start == 0 && end == 0)
1441 machine->vmlinux_map->end = ~0ULL;
1444 static void machine__update_kernel_mmap(struct machine *machine,
1447 struct map *map = machine__kernel_map(machine);
1450 map_groups__remove(&machine->kmaps, map);
1452 machine__set_kernel_mmap(machine, start, end);
1454 map_groups__insert(&machine->kmaps, map);
1458 int machine__create_kernel_maps(struct machine *machine)
1460 struct dso *kernel = machine__get_kernel(machine);
1461 const char *name = NULL;
1463 u64 start = 0, end = ~0ULL;
1469 ret = __machine__create_kernel_maps(machine, kernel);
1473 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1474 if (machine__is_host(machine))
1475 pr_debug("Problems creating module maps, "
1476 "continuing anyway...\n");
1478 pr_debug("Problems creating module maps for guest %d, "
1479 "continuing anyway...\n", machine->pid);
1482 if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1484 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1485 machine__destroy_kernel_maps(machine);
1491 * we have a real start address now, so re-order the kmaps
1492 * assume it's the last in the kmaps
1494 machine__update_kernel_mmap(machine, start, end);
1497 if (machine__create_extra_kernel_maps(machine, kernel))
1498 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1501 /* update end address of the kernel map using adjacent module address */
1502 map = map__next(machine__kernel_map(machine));
1504 machine__set_kernel_mmap(machine, start, map->start);
1512 static bool machine__uses_kcore(struct machine *machine)
1516 list_for_each_entry(dso, &machine->dsos.head, node) {
1517 if (dso__is_kcore(dso))
1524 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1525 union perf_event *event)
1527 return machine__is(machine, "x86_64") &&
1528 is_entry_trampoline(event->mmap.filename);
1531 static int machine__process_extra_kernel_map(struct machine *machine,
1532 union perf_event *event)
1534 struct map *kernel_map = machine__kernel_map(machine);
1535 struct dso *kernel = kernel_map ? kernel_map->dso : NULL;
1536 struct extra_kernel_map xm = {
1537 .start = event->mmap.start,
1538 .end = event->mmap.start + event->mmap.len,
1539 .pgoff = event->mmap.pgoff,
1545 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1547 return machine__create_extra_kernel_map(machine, kernel, &xm);
1550 static int machine__process_kernel_mmap_event(struct machine *machine,
1551 union perf_event *event)
1554 enum dso_kernel_type kernel_type;
1555 bool is_kernel_mmap;
1557 /* If we have maps from kcore then we do not need or want any others */
1558 if (machine__uses_kcore(machine))
1561 if (machine__is_host(machine))
1562 kernel_type = DSO_TYPE_KERNEL;
1564 kernel_type = DSO_TYPE_GUEST_KERNEL;
1566 is_kernel_mmap = memcmp(event->mmap.filename,
1568 strlen(machine->mmap_name) - 1) == 0;
1569 if (event->mmap.filename[0] == '/' ||
1570 (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1571 map = machine__findnew_module_map(machine, event->mmap.start,
1572 event->mmap.filename);
1576 map->end = map->start + event->mmap.len;
1577 } else if (is_kernel_mmap) {
1578 const char *symbol_name = (event->mmap.filename +
1579 strlen(machine->mmap_name));
1581 * Should be there already, from the build-id table in
1584 struct dso *kernel = NULL;
1587 down_read(&machine->dsos.lock);
1589 list_for_each_entry(dso, &machine->dsos.head, node) {
1592 * The cpumode passed to is_kernel_module is not the
1593 * cpumode of *this* event. If we insist on passing
1594 * correct cpumode to is_kernel_module, we should
1595 * record the cpumode when we adding this dso to the
1598 * However we don't really need passing correct
1599 * cpumode. We know the correct cpumode must be kernel
1600 * mode (if not, we should not link it onto kernel_dsos
1603 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1604 * is_kernel_module() treats it as a kernel cpumode.
1608 is_kernel_module(dso->long_name,
1609 PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1617 up_read(&machine->dsos.lock);
1620 kernel = machine__findnew_dso(machine, machine->mmap_name);
1624 kernel->kernel = kernel_type;
1625 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1630 if (strstr(kernel->long_name, "vmlinux"))
1631 dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1633 machine__update_kernel_mmap(machine, event->mmap.start,
1634 event->mmap.start + event->mmap.len);
1637 * Avoid using a zero address (kptr_restrict) for the ref reloc
1638 * symbol. Effectively having zero here means that at record
1639 * time /proc/sys/kernel/kptr_restrict was non zero.
1641 if (event->mmap.pgoff != 0) {
1642 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1647 if (machine__is_default_guest(machine)) {
1649 * preload dso of guest kernel and modules
1651 dso__load(kernel, machine__kernel_map(machine));
1653 } else if (perf_event__is_extra_kernel_mmap(machine, event)) {
1654 return machine__process_extra_kernel_map(machine, event);
1661 int machine__process_mmap2_event(struct machine *machine,
1662 union perf_event *event,
1663 struct perf_sample *sample)
1665 struct thread *thread;
1670 perf_event__fprintf_mmap2(event, stdout);
1672 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1673 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1674 ret = machine__process_kernel_mmap_event(machine, event);
1680 thread = machine__findnew_thread(machine, event->mmap2.pid,
1685 map = map__new(machine, event->mmap2.start,
1686 event->mmap2.len, event->mmap2.pgoff,
1688 event->mmap2.min, event->mmap2.ino,
1689 event->mmap2.ino_generation,
1692 event->mmap2.filename, thread);
1695 goto out_problem_map;
1697 ret = thread__insert_map(thread, map);
1699 goto out_problem_insert;
1701 thread__put(thread);
1708 thread__put(thread);
1710 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1714 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1715 struct perf_sample *sample)
1717 struct thread *thread;
1723 perf_event__fprintf_mmap(event, stdout);
1725 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1726 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1727 ret = machine__process_kernel_mmap_event(machine, event);
1733 thread = machine__findnew_thread(machine, event->mmap.pid,
1738 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1741 map = map__new(machine, event->mmap.start,
1742 event->mmap.len, event->mmap.pgoff,
1743 0, 0, 0, 0, prot, 0,
1744 event->mmap.filename,
1748 goto out_problem_map;
1750 ret = thread__insert_map(thread, map);
1752 goto out_problem_insert;
1754 thread__put(thread);
1761 thread__put(thread);
1763 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1767 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1769 struct threads *threads = machine__threads(machine, th->tid);
1771 if (threads->last_match == th)
1772 threads__set_last_match(threads, NULL);
1775 down_write(&threads->lock);
1777 BUG_ON(refcount_read(&th->refcnt) == 0);
1779 rb_erase_cached(&th->rb_node, &threads->entries);
1780 RB_CLEAR_NODE(&th->rb_node);
1783 * Move it first to the dead_threads list, then drop the reference,
1784 * if this is the last reference, then the thread__delete destructor
1785 * will be called and we will remove it from the dead_threads list.
1787 list_add_tail(&th->node, &threads->dead);
1790 * We need to do the put here because if this is the last refcount,
1791 * then we will be touching the threads->dead head when removing the
1797 up_write(&threads->lock);
1800 void machine__remove_thread(struct machine *machine, struct thread *th)
1802 return __machine__remove_thread(machine, th, true);
1805 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1806 struct perf_sample *sample)
1808 struct thread *thread = machine__find_thread(machine,
1811 struct thread *parent = machine__findnew_thread(machine,
1814 bool do_maps_clone = true;
1818 perf_event__fprintf_task(event, stdout);
1821 * There may be an existing thread that is not actually the parent,
1822 * either because we are processing events out of order, or because the
1823 * (fork) event that would have removed the thread was lost. Assume the
1824 * latter case and continue on as best we can.
1826 if (parent->pid_ != (pid_t)event->fork.ppid) {
1827 dump_printf("removing erroneous parent thread %d/%d\n",
1828 parent->pid_, parent->tid);
1829 machine__remove_thread(machine, parent);
1830 thread__put(parent);
1831 parent = machine__findnew_thread(machine, event->fork.ppid,
1835 /* if a thread currently exists for the thread id remove it */
1836 if (thread != NULL) {
1837 machine__remove_thread(machine, thread);
1838 thread__put(thread);
1841 thread = machine__findnew_thread(machine, event->fork.pid,
1844 * When synthesizing FORK events, we are trying to create thread
1845 * objects for the already running tasks on the machine.
1847 * Normally, for a kernel FORK event, we want to clone the parent's
1848 * maps because that is what the kernel just did.
1850 * But when synthesizing, this should not be done. If we do, we end up
1851 * with overlapping maps as we process the sythesized MMAP2 events that
1852 * get delivered shortly thereafter.
1854 * Use the FORK event misc flags in an internal way to signal this
1855 * situation, so we can elide the map clone when appropriate.
1857 if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1858 do_maps_clone = false;
1860 if (thread == NULL || parent == NULL ||
1861 thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1862 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1865 thread__put(thread);
1866 thread__put(parent);
1871 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1872 struct perf_sample *sample __maybe_unused)
1874 struct thread *thread = machine__find_thread(machine,
1879 perf_event__fprintf_task(event, stdout);
1881 if (thread != NULL) {
1882 thread__exited(thread);
1883 thread__put(thread);
1889 int machine__process_event(struct machine *machine, union perf_event *event,
1890 struct perf_sample *sample)
1894 switch (event->header.type) {
1895 case PERF_RECORD_COMM:
1896 ret = machine__process_comm_event(machine, event, sample); break;
1897 case PERF_RECORD_MMAP:
1898 ret = machine__process_mmap_event(machine, event, sample); break;
1899 case PERF_RECORD_NAMESPACES:
1900 ret = machine__process_namespaces_event(machine, event, sample); break;
1901 case PERF_RECORD_MMAP2:
1902 ret = machine__process_mmap2_event(machine, event, sample); break;
1903 case PERF_RECORD_FORK:
1904 ret = machine__process_fork_event(machine, event, sample); break;
1905 case PERF_RECORD_EXIT:
1906 ret = machine__process_exit_event(machine, event, sample); break;
1907 case PERF_RECORD_LOST:
1908 ret = machine__process_lost_event(machine, event, sample); break;
1909 case PERF_RECORD_AUX:
1910 ret = machine__process_aux_event(machine, event); break;
1911 case PERF_RECORD_ITRACE_START:
1912 ret = machine__process_itrace_start_event(machine, event); break;
1913 case PERF_RECORD_LOST_SAMPLES:
1914 ret = machine__process_lost_samples_event(machine, event, sample); break;
1915 case PERF_RECORD_SWITCH:
1916 case PERF_RECORD_SWITCH_CPU_WIDE:
1917 ret = machine__process_switch_event(machine, event); break;
1918 case PERF_RECORD_KSYMBOL:
1919 ret = machine__process_ksymbol(machine, event, sample); break;
1920 case PERF_RECORD_BPF_EVENT:
1921 ret = machine__process_bpf_event(machine, event, sample); break;
1930 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1932 if (!regexec(regex, sym->name, 0, NULL, 0))
1937 static void ip__resolve_ams(struct thread *thread,
1938 struct addr_map_symbol *ams,
1941 struct addr_location al;
1943 memset(&al, 0, sizeof(al));
1945 * We cannot use the header.misc hint to determine whether a
1946 * branch stack address is user, kernel, guest, hypervisor.
1947 * Branches may straddle the kernel/user/hypervisor boundaries.
1948 * Thus, we have to try consecutively until we find a match
1949 * or else, the symbol is unknown
1951 thread__find_cpumode_addr_location(thread, ip, &al);
1954 ams->al_addr = al.addr;
1960 static void ip__resolve_data(struct thread *thread,
1961 u8 m, struct addr_map_symbol *ams,
1962 u64 addr, u64 phys_addr)
1964 struct addr_location al;
1966 memset(&al, 0, sizeof(al));
1968 thread__find_symbol(thread, m, addr, &al);
1971 ams->al_addr = al.addr;
1974 ams->phys_addr = phys_addr;
1977 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1978 struct addr_location *al)
1980 struct mem_info *mi = mem_info__new();
1985 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1986 ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
1987 sample->addr, sample->phys_addr);
1988 mi->data_src.val = sample->data_src;
1993 static char *callchain_srcline(struct map *map, struct symbol *sym, u64 ip)
1995 char *srcline = NULL;
1997 if (!map || callchain_param.key == CCKEY_FUNCTION)
2000 srcline = srcline__tree_find(&map->dso->srclines, ip);
2002 bool show_sym = false;
2003 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2005 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
2006 sym, show_sym, show_addr, ip);
2007 srcline__tree_insert(&map->dso->srclines, ip, srcline);
2018 static int add_callchain_ip(struct thread *thread,
2019 struct callchain_cursor *cursor,
2020 struct symbol **parent,
2021 struct addr_location *root_al,
2025 struct branch_flags *flags,
2026 struct iterations *iter,
2029 struct addr_location al;
2030 int nr_loop_iter = 0;
2031 u64 iter_cycles = 0;
2032 const char *srcline = NULL;
2037 thread__find_cpumode_addr_location(thread, ip, &al);
2039 if (ip >= PERF_CONTEXT_MAX) {
2041 case PERF_CONTEXT_HV:
2042 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2044 case PERF_CONTEXT_KERNEL:
2045 *cpumode = PERF_RECORD_MISC_KERNEL;
2047 case PERF_CONTEXT_USER:
2048 *cpumode = PERF_RECORD_MISC_USER;
2051 pr_debug("invalid callchain context: "
2052 "%"PRId64"\n", (s64) ip);
2054 * It seems the callchain is corrupted.
2057 callchain_cursor_reset(cursor);
2062 thread__find_symbol(thread, *cpumode, ip, &al);
2065 if (al.sym != NULL) {
2066 if (perf_hpp_list.parent && !*parent &&
2067 symbol__match_regex(al.sym, &parent_regex))
2069 else if (have_ignore_callees && root_al &&
2070 symbol__match_regex(al.sym, &ignore_callees_regex)) {
2071 /* Treat this symbol as the root,
2072 forgetting its callees. */
2074 callchain_cursor_reset(cursor);
2078 if (symbol_conf.hide_unresolved && al.sym == NULL)
2082 nr_loop_iter = iter->nr_loop_iter;
2083 iter_cycles = iter->cycles;
2086 srcline = callchain_srcline(al.map, al.sym, al.addr);
2087 return callchain_cursor_append(cursor, ip, al.map, al.sym,
2088 branch, flags, nr_loop_iter,
2089 iter_cycles, branch_from, srcline);
2092 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2093 struct addr_location *al)
2096 const struct branch_stack *bs = sample->branch_stack;
2097 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2102 for (i = 0; i < bs->nr; i++) {
2103 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
2104 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
2105 bi[i].flags = bs->entries[i].flags;
2110 static void save_iterations(struct iterations *iter,
2111 struct branch_entry *be, int nr)
2115 iter->nr_loop_iter++;
2118 for (i = 0; i < nr; i++)
2119 iter->cycles += be[i].flags.cycles;
2124 #define NO_ENTRY 0xff
2126 #define PERF_MAX_BRANCH_DEPTH 127
2129 static int remove_loops(struct branch_entry *l, int nr,
2130 struct iterations *iter)
2133 unsigned char chash[CHASHSZ];
2135 memset(chash, NO_ENTRY, sizeof(chash));
2137 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2139 for (i = 0; i < nr; i++) {
2140 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2142 /* no collision handling for now */
2143 if (chash[h] == NO_ENTRY) {
2145 } else if (l[chash[h]].from == l[i].from) {
2146 bool is_loop = true;
2147 /* check if it is a real loop */
2149 for (j = chash[h]; j < i && i + off < nr; j++, off++)
2150 if (l[j].from != l[i + off].from) {
2157 save_iterations(iter + i + off,
2160 memmove(iter + i, iter + i + off,
2163 memmove(l + i, l + i + off,
2175 * Recolve LBR callstack chain sample
2177 * 1 on success get LBR callchain information
2178 * 0 no available LBR callchain information, should try fp
2179 * negative error code on other errors.
2181 static int resolve_lbr_callchain_sample(struct thread *thread,
2182 struct callchain_cursor *cursor,
2183 struct perf_sample *sample,
2184 struct symbol **parent,
2185 struct addr_location *root_al,
2188 struct ip_callchain *chain = sample->callchain;
2189 int chain_nr = min(max_stack, (int)chain->nr), i;
2190 u8 cpumode = PERF_RECORD_MISC_USER;
2191 u64 ip, branch_from = 0;
2193 for (i = 0; i < chain_nr; i++) {
2194 if (chain->ips[i] == PERF_CONTEXT_USER)
2198 /* LBR only affects the user callchain */
2199 if (i != chain_nr) {
2200 struct branch_stack *lbr_stack = sample->branch_stack;
2201 int lbr_nr = lbr_stack->nr, j, k;
2203 struct branch_flags *flags;
2205 * LBR callstack can only get user call chain.
2206 * The mix_chain_nr is kernel call chain
2207 * number plus LBR user call chain number.
2208 * i is kernel call chain number,
2209 * 1 is PERF_CONTEXT_USER,
2210 * lbr_nr + 1 is the user call chain number.
2211 * For details, please refer to the comments
2212 * in callchain__printf
2214 int mix_chain_nr = i + 1 + lbr_nr + 1;
2216 for (j = 0; j < mix_chain_nr; j++) {
2221 if (callchain_param.order == ORDER_CALLEE) {
2224 else if (j > i + 1) {
2226 ip = lbr_stack->entries[k].from;
2228 flags = &lbr_stack->entries[k].flags;
2230 ip = lbr_stack->entries[0].to;
2232 flags = &lbr_stack->entries[0].flags;
2234 lbr_stack->entries[0].from;
2239 ip = lbr_stack->entries[k].from;
2241 flags = &lbr_stack->entries[k].flags;
2243 else if (j > lbr_nr)
2244 ip = chain->ips[i + 1 - (j - lbr_nr)];
2246 ip = lbr_stack->entries[0].to;
2248 flags = &lbr_stack->entries[0].flags;
2250 lbr_stack->entries[0].from;
2254 err = add_callchain_ip(thread, cursor, parent,
2255 root_al, &cpumode, ip,
2256 branch, flags, NULL,
2259 return (err < 0) ? err : 0;
2267 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2268 struct callchain_cursor *cursor,
2269 struct symbol **parent,
2270 struct addr_location *root_al,
2271 u8 *cpumode, int ent)
2275 while (--ent >= 0) {
2276 u64 ip = chain->ips[ent];
2278 if (ip >= PERF_CONTEXT_MAX) {
2279 err = add_callchain_ip(thread, cursor, parent,
2280 root_al, cpumode, ip,
2281 false, NULL, NULL, 0);
2288 static int thread__resolve_callchain_sample(struct thread *thread,
2289 struct callchain_cursor *cursor,
2290 struct evsel *evsel,
2291 struct perf_sample *sample,
2292 struct symbol **parent,
2293 struct addr_location *root_al,
2296 struct branch_stack *branch = sample->branch_stack;
2297 struct ip_callchain *chain = sample->callchain;
2299 u8 cpumode = PERF_RECORD_MISC_USER;
2300 int i, j, err, nr_entries;
2305 chain_nr = chain->nr;
2307 if (perf_evsel__has_branch_callstack(evsel)) {
2308 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2309 root_al, max_stack);
2311 return (err < 0) ? err : 0;
2315 * Based on DWARF debug information, some architectures skip
2316 * a callchain entry saved by the kernel.
2318 skip_idx = arch_skip_callchain_idx(thread, chain);
2321 * Add branches to call stack for easier browsing. This gives
2322 * more context for a sample than just the callers.
2324 * This uses individual histograms of paths compared to the
2325 * aggregated histograms the normal LBR mode uses.
2327 * Limitations for now:
2328 * - No extra filters
2329 * - No annotations (should annotate somehow)
2332 if (branch && callchain_param.branch_callstack) {
2333 int nr = min(max_stack, (int)branch->nr);
2334 struct branch_entry be[nr];
2335 struct iterations iter[nr];
2337 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2338 pr_warning("corrupted branch chain. skipping...\n");
2342 for (i = 0; i < nr; i++) {
2343 if (callchain_param.order == ORDER_CALLEE) {
2344 be[i] = branch->entries[i];
2350 * Check for overlap into the callchain.
2351 * The return address is one off compared to
2352 * the branch entry. To adjust for this
2353 * assume the calling instruction is not longer
2356 if (i == skip_idx ||
2357 chain->ips[first_call] >= PERF_CONTEXT_MAX)
2359 else if (be[i].from < chain->ips[first_call] &&
2360 be[i].from >= chain->ips[first_call] - 8)
2363 be[i] = branch->entries[branch->nr - i - 1];
2366 memset(iter, 0, sizeof(struct iterations) * nr);
2367 nr = remove_loops(be, nr, iter);
2369 for (i = 0; i < nr; i++) {
2370 err = add_callchain_ip(thread, cursor, parent,
2377 err = add_callchain_ip(thread, cursor, parent, root_al,
2394 if (callchain_param.order != ORDER_CALLEE) {
2395 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2396 &cpumode, chain->nr - first_call);
2398 return (err < 0) ? err : 0;
2400 for (i = first_call, nr_entries = 0;
2401 i < chain_nr && nr_entries < max_stack; i++) {
2404 if (callchain_param.order == ORDER_CALLEE)
2407 j = chain->nr - i - 1;
2409 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2414 if (ip < PERF_CONTEXT_MAX)
2416 else if (callchain_param.order != ORDER_CALLEE) {
2417 err = find_prev_cpumode(chain, thread, cursor, parent,
2418 root_al, &cpumode, j);
2420 return (err < 0) ? err : 0;
2424 err = add_callchain_ip(thread, cursor, parent,
2425 root_al, &cpumode, ip,
2426 false, NULL, NULL, 0);
2429 return (err < 0) ? err : 0;
2435 static int append_inlines(struct callchain_cursor *cursor,
2436 struct map *map, struct symbol *sym, u64 ip)
2438 struct inline_node *inline_node;
2439 struct inline_list *ilist;
2443 if (!symbol_conf.inline_name || !map || !sym)
2446 addr = map__map_ip(map, ip);
2447 addr = map__rip_2objdump(map, addr);
2449 inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2451 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2454 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2457 list_for_each_entry(ilist, &inline_node->val, list) {
2458 ret = callchain_cursor_append(cursor, ip, map,
2459 ilist->symbol, false,
2460 NULL, 0, 0, 0, ilist->srcline);
2469 static int unwind_entry(struct unwind_entry *entry, void *arg)
2471 struct callchain_cursor *cursor = arg;
2472 const char *srcline = NULL;
2473 u64 addr = entry->ip;
2475 if (symbol_conf.hide_unresolved && entry->sym == NULL)
2478 if (append_inlines(cursor, entry->map, entry->sym, entry->ip) == 0)
2482 * Convert entry->ip from a virtual address to an offset in
2483 * its corresponding binary.
2486 addr = map__map_ip(entry->map, entry->ip);
2488 srcline = callchain_srcline(entry->map, entry->sym, addr);
2489 return callchain_cursor_append(cursor, entry->ip,
2490 entry->map, entry->sym,
2491 false, NULL, 0, 0, 0, srcline);
2494 static int thread__resolve_callchain_unwind(struct thread *thread,
2495 struct callchain_cursor *cursor,
2496 struct evsel *evsel,
2497 struct perf_sample *sample,
2500 /* Can we do dwarf post unwind? */
2501 if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2502 (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
2505 /* Bail out if nothing was captured. */
2506 if ((!sample->user_regs.regs) ||
2507 (!sample->user_stack.size))
2510 return unwind__get_entries(unwind_entry, cursor,
2511 thread, sample, max_stack);
2514 int thread__resolve_callchain(struct thread *thread,
2515 struct callchain_cursor *cursor,
2516 struct evsel *evsel,
2517 struct perf_sample *sample,
2518 struct symbol **parent,
2519 struct addr_location *root_al,
2524 callchain_cursor_reset(cursor);
2526 if (callchain_param.order == ORDER_CALLEE) {
2527 ret = thread__resolve_callchain_sample(thread, cursor,
2533 ret = thread__resolve_callchain_unwind(thread, cursor,
2537 ret = thread__resolve_callchain_unwind(thread, cursor,
2542 ret = thread__resolve_callchain_sample(thread, cursor,
2551 int machine__for_each_thread(struct machine *machine,
2552 int (*fn)(struct thread *thread, void *p),
2555 struct threads *threads;
2557 struct thread *thread;
2561 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2562 threads = &machine->threads[i];
2563 for (nd = rb_first_cached(&threads->entries); nd;
2565 thread = rb_entry(nd, struct thread, rb_node);
2566 rc = fn(thread, priv);
2571 list_for_each_entry(thread, &threads->dead, node) {
2572 rc = fn(thread, priv);
2580 int machines__for_each_thread(struct machines *machines,
2581 int (*fn)(struct thread *thread, void *p),
2587 rc = machine__for_each_thread(&machines->host, fn, priv);
2591 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2592 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2594 rc = machine__for_each_thread(machine, fn, priv);
2601 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
2602 struct target *target, struct perf_thread_map *threads,
2603 perf_event__handler_t process, bool data_mmap,
2604 unsigned int nr_threads_synthesize)
2606 if (target__has_task(target))
2607 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap);
2608 else if (target__has_cpu(target))
2609 return perf_event__synthesize_threads(tool, process,
2611 nr_threads_synthesize);
2612 /* command specified */
2616 pid_t machine__get_current_tid(struct machine *machine, int cpu)
2618 if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid)
2621 return machine->current_tid[cpu];
2624 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
2627 struct thread *thread;
2632 if (!machine->current_tid) {
2635 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t));
2636 if (!machine->current_tid)
2638 for (i = 0; i < MAX_NR_CPUS; i++)
2639 machine->current_tid[i] = -1;
2642 if (cpu >= MAX_NR_CPUS) {
2643 pr_err("Requested CPU %d too large. ", cpu);
2644 pr_err("Consider raising MAX_NR_CPUS\n");
2648 machine->current_tid[cpu] = tid;
2650 thread = machine__findnew_thread(machine, pid, tid);
2655 thread__put(thread);
2661 * Compares the raw arch string. N.B. see instead perf_env__arch() if a
2662 * normalized arch is needed.
2664 bool machine__is(struct machine *machine, const char *arch)
2666 return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
2669 int machine__nr_cpus_avail(struct machine *machine)
2671 return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
2674 int machine__get_kernel_start(struct machine *machine)
2676 struct map *map = machine__kernel_map(machine);
2680 * The only addresses above 2^63 are kernel addresses of a 64-bit
2681 * kernel. Note that addresses are unsigned so that on a 32-bit system
2682 * all addresses including kernel addresses are less than 2^32. In
2683 * that case (32-bit system), if the kernel mapping is unknown, all
2684 * addresses will be assumed to be in user space - see
2685 * machine__kernel_ip().
2687 machine->kernel_start = 1ULL << 63;
2689 err = map__load(map);
2691 * On x86_64, PTI entry trampolines are less than the
2692 * start of kernel text, but still above 2^63. So leave
2693 * kernel_start = 1ULL << 63 for x86_64.
2695 if (!err && !machine__is(machine, "x86_64"))
2696 machine->kernel_start = map->start;
2701 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
2703 u8 addr_cpumode = cpumode;
2706 if (!machine->single_address_space)
2709 kernel_ip = machine__kernel_ip(machine, addr);
2711 case PERF_RECORD_MISC_KERNEL:
2712 case PERF_RECORD_MISC_USER:
2713 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
2714 PERF_RECORD_MISC_USER;
2716 case PERF_RECORD_MISC_GUEST_KERNEL:
2717 case PERF_RECORD_MISC_GUEST_USER:
2718 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
2719 PERF_RECORD_MISC_GUEST_USER;
2725 return addr_cpumode;
2728 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2730 return dsos__findnew(&machine->dsos, filename);
2733 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2735 struct machine *machine = vmachine;
2737 struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
2742 *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2743 *addrp = map->unmap_ip(map, sym->start);