1 // SPDX-License-Identifier: GPL-2.0
16 #include "map_symbol.h"
18 #include "mem-events.h"
28 #include <sys/types.h>
32 #include "linux/hash.h"
34 #include "bpf-event.h"
35 #include <internal/lib.h> // page_size
38 #include <linux/ctype.h>
39 #include <symbol/kallsyms.h>
40 #include <linux/mman.h>
41 #include <linux/string.h>
42 #include <linux/zalloc.h>
44 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
46 static struct dso *machine__kernel_dso(struct machine *machine)
48 return machine->vmlinux_map->dso;
51 static void dsos__init(struct dsos *dsos)
53 INIT_LIST_HEAD(&dsos->head);
55 init_rwsem(&dsos->lock);
58 static void machine__threads_init(struct machine *machine)
62 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
63 struct threads *threads = &machine->threads[i];
64 threads->entries = RB_ROOT_CACHED;
65 init_rwsem(&threads->lock);
67 INIT_LIST_HEAD(&threads->dead);
68 threads->last_match = NULL;
72 static int machine__set_mmap_name(struct machine *machine)
74 if (machine__is_host(machine))
75 machine->mmap_name = strdup("[kernel.kallsyms]");
76 else if (machine__is_default_guest(machine))
77 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
78 else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
80 machine->mmap_name = NULL;
82 return machine->mmap_name ? 0 : -ENOMEM;
85 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
89 memset(machine, 0, sizeof(*machine));
90 maps__init(&machine->kmaps, machine);
91 RB_CLEAR_NODE(&machine->rb_node);
92 dsos__init(&machine->dsos);
94 machine__threads_init(machine);
96 machine->vdso_info = NULL;
101 machine->id_hdr_size = 0;
102 machine->kptr_restrict_warned = false;
103 machine->comm_exec = false;
104 machine->kernel_start = 0;
105 machine->vmlinux_map = NULL;
107 machine->root_dir = strdup(root_dir);
108 if (machine->root_dir == NULL)
111 if (machine__set_mmap_name(machine))
114 if (pid != HOST_KERNEL_ID) {
115 struct thread *thread = machine__findnew_thread(machine, -1,
122 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
123 thread__set_comm(thread, comm, 0);
127 machine->current_tid = NULL;
132 zfree(&machine->root_dir);
133 zfree(&machine->mmap_name);
138 struct machine *machine__new_host(void)
140 struct machine *machine = malloc(sizeof(*machine));
142 if (machine != NULL) {
143 machine__init(machine, "", HOST_KERNEL_ID);
145 if (machine__create_kernel_maps(machine) < 0)
155 struct machine *machine__new_kallsyms(void)
157 struct machine *machine = machine__new_host();
160 * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
161 * ask for not using the kcore parsing code, once this one is fixed
162 * to create a map per module.
164 if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
165 machine__delete(machine);
172 static void dsos__purge(struct dsos *dsos)
176 down_write(&dsos->lock);
178 list_for_each_entry_safe(pos, n, &dsos->head, node) {
179 RB_CLEAR_NODE(&pos->rb_node);
181 list_del_init(&pos->node);
185 up_write(&dsos->lock);
188 static void dsos__exit(struct dsos *dsos)
191 exit_rwsem(&dsos->lock);
194 void machine__delete_threads(struct machine *machine)
199 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
200 struct threads *threads = &machine->threads[i];
201 down_write(&threads->lock);
202 nd = rb_first_cached(&threads->entries);
204 struct thread *t = rb_entry(nd, struct thread, rb_node);
207 __machine__remove_thread(machine, t, false);
209 up_write(&threads->lock);
213 void machine__exit(struct machine *machine)
220 machine__destroy_kernel_maps(machine);
221 maps__exit(&machine->kmaps);
222 dsos__exit(&machine->dsos);
223 machine__exit_vdso(machine);
224 zfree(&machine->root_dir);
225 zfree(&machine->mmap_name);
226 zfree(&machine->current_tid);
228 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
229 struct threads *threads = &machine->threads[i];
230 struct thread *thread, *n;
232 * Forget about the dead, at this point whatever threads were
233 * left in the dead lists better have a reference count taken
234 * by who is using them, and then, when they drop those references
235 * and it finally hits zero, thread__put() will check and see that
236 * its not in the dead threads list and will not try to remove it
237 * from there, just calling thread__delete() straight away.
239 list_for_each_entry_safe(thread, n, &threads->dead, node)
240 list_del_init(&thread->node);
242 exit_rwsem(&threads->lock);
246 void machine__delete(struct machine *machine)
249 machine__exit(machine);
254 void machines__init(struct machines *machines)
256 machine__init(&machines->host, "", HOST_KERNEL_ID);
257 machines->guests = RB_ROOT_CACHED;
260 void machines__exit(struct machines *machines)
262 machine__exit(&machines->host);
266 struct machine *machines__add(struct machines *machines, pid_t pid,
267 const char *root_dir)
269 struct rb_node **p = &machines->guests.rb_root.rb_node;
270 struct rb_node *parent = NULL;
271 struct machine *pos, *machine = malloc(sizeof(*machine));
272 bool leftmost = true;
277 if (machine__init(machine, root_dir, pid) != 0) {
284 pos = rb_entry(parent, struct machine, rb_node);
293 rb_link_node(&machine->rb_node, parent, p);
294 rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
299 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
303 machines->host.comm_exec = comm_exec;
305 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
306 struct machine *machine = rb_entry(nd, struct machine, rb_node);
308 machine->comm_exec = comm_exec;
312 struct machine *machines__find(struct machines *machines, pid_t pid)
314 struct rb_node **p = &machines->guests.rb_root.rb_node;
315 struct rb_node *parent = NULL;
316 struct machine *machine;
317 struct machine *default_machine = NULL;
319 if (pid == HOST_KERNEL_ID)
320 return &machines->host;
324 machine = rb_entry(parent, struct machine, rb_node);
325 if (pid < machine->pid)
327 else if (pid > machine->pid)
332 default_machine = machine;
335 return default_machine;
338 struct machine *machines__findnew(struct machines *machines, pid_t pid)
341 const char *root_dir = "";
342 struct machine *machine = machines__find(machines, pid);
344 if (machine && (machine->pid == pid))
347 if ((pid != HOST_KERNEL_ID) &&
348 (pid != DEFAULT_GUEST_KERNEL_ID) &&
349 (symbol_conf.guestmount)) {
350 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
351 if (access(path, R_OK)) {
352 static struct strlist *seen;
355 seen = strlist__new(NULL, NULL);
357 if (!strlist__has_entry(seen, path)) {
358 pr_err("Can't access file %s\n", path);
359 strlist__add(seen, path);
367 machine = machines__add(machines, pid, root_dir);
372 void machines__process_guests(struct machines *machines,
373 machine__process_t process, void *data)
377 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
378 struct machine *pos = rb_entry(nd, struct machine, rb_node);
383 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
385 struct rb_node *node;
386 struct machine *machine;
388 machines->host.id_hdr_size = id_hdr_size;
390 for (node = rb_first_cached(&machines->guests); node;
391 node = rb_next(node)) {
392 machine = rb_entry(node, struct machine, rb_node);
393 machine->id_hdr_size = id_hdr_size;
399 static void machine__update_thread_pid(struct machine *machine,
400 struct thread *th, pid_t pid)
402 struct thread *leader;
404 if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
409 if (th->pid_ == th->tid)
412 leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
417 leader->maps = maps__new(machine);
422 if (th->maps == leader->maps)
427 * Maps are created from MMAP events which provide the pid and
428 * tid. Consequently there never should be any maps on a thread
429 * with an unknown pid. Just print an error if there are.
431 if (!maps__empty(th->maps))
432 pr_err("Discarding thread maps for %d:%d\n",
437 th->maps = maps__get(leader->maps);
442 pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
447 * Front-end cache - TID lookups come in blocks,
448 * so most of the time we dont have to look up
451 static struct thread*
452 __threads__get_last_match(struct threads *threads, struct machine *machine,
457 th = threads->last_match;
459 if (th->tid == tid) {
460 machine__update_thread_pid(machine, th, pid);
461 return thread__get(th);
464 threads->last_match = NULL;
470 static struct thread*
471 threads__get_last_match(struct threads *threads, struct machine *machine,
474 struct thread *th = NULL;
476 if (perf_singlethreaded)
477 th = __threads__get_last_match(threads, machine, pid, tid);
483 __threads__set_last_match(struct threads *threads, struct thread *th)
485 threads->last_match = th;
489 threads__set_last_match(struct threads *threads, struct thread *th)
491 if (perf_singlethreaded)
492 __threads__set_last_match(threads, th);
496 * Caller must eventually drop thread->refcnt returned with a successful
497 * lookup/new thread inserted.
499 static struct thread *____machine__findnew_thread(struct machine *machine,
500 struct threads *threads,
501 pid_t pid, pid_t tid,
504 struct rb_node **p = &threads->entries.rb_root.rb_node;
505 struct rb_node *parent = NULL;
507 bool leftmost = true;
509 th = threads__get_last_match(threads, machine, pid, tid);
515 th = rb_entry(parent, struct thread, rb_node);
517 if (th->tid == tid) {
518 threads__set_last_match(threads, th);
519 machine__update_thread_pid(machine, th, pid);
520 return thread__get(th);
534 th = thread__new(pid, tid);
536 rb_link_node(&th->rb_node, parent, p);
537 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
540 * We have to initialize maps separately after rb tree is updated.
542 * The reason is that we call machine__findnew_thread
543 * within thread__init_maps to find the thread
544 * leader and that would screwed the rb tree.
546 if (thread__init_maps(th, machine)) {
547 rb_erase_cached(&th->rb_node, &threads->entries);
548 RB_CLEAR_NODE(&th->rb_node);
553 * It is now in the rbtree, get a ref
556 threads__set_last_match(threads, th);
563 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
565 return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
568 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
571 struct threads *threads = machine__threads(machine, tid);
574 down_write(&threads->lock);
575 th = __machine__findnew_thread(machine, pid, tid);
576 up_write(&threads->lock);
580 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
583 struct threads *threads = machine__threads(machine, tid);
586 down_read(&threads->lock);
587 th = ____machine__findnew_thread(machine, threads, pid, tid, false);
588 up_read(&threads->lock);
592 struct comm *machine__thread_exec_comm(struct machine *machine,
593 struct thread *thread)
595 if (machine->comm_exec)
596 return thread__exec_comm(thread);
598 return thread__comm(thread);
601 int machine__process_comm_event(struct machine *machine, union perf_event *event,
602 struct perf_sample *sample)
604 struct thread *thread = machine__findnew_thread(machine,
607 bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
611 machine->comm_exec = true;
614 perf_event__fprintf_comm(event, stdout);
616 if (thread == NULL ||
617 __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
618 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
627 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
628 union perf_event *event,
629 struct perf_sample *sample __maybe_unused)
631 struct thread *thread = machine__findnew_thread(machine,
632 event->namespaces.pid,
633 event->namespaces.tid);
636 WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
637 "\nWARNING: kernel seems to support more namespaces than perf"
638 " tool.\nTry updating the perf tool..\n\n");
640 WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
641 "\nWARNING: perf tool seems to support more namespaces than"
642 " the kernel.\nTry updating the kernel..\n\n");
645 perf_event__fprintf_namespaces(event, stdout);
647 if (thread == NULL ||
648 thread__set_namespaces(thread, sample->time, &event->namespaces)) {
649 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
658 int machine__process_cgroup_event(struct machine *machine,
659 union perf_event *event,
660 struct perf_sample *sample __maybe_unused)
665 perf_event__fprintf_cgroup(event, stdout);
667 cgrp = cgroup__findnew(machine->env, event->cgroup.id, event->cgroup.path);
674 int machine__process_lost_event(struct machine *machine __maybe_unused,
675 union perf_event *event, struct perf_sample *sample __maybe_unused)
677 dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
678 event->lost.id, event->lost.lost);
682 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
683 union perf_event *event, struct perf_sample *sample)
685 dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
686 sample->id, event->lost_samples.lost);
690 static struct dso *machine__findnew_module_dso(struct machine *machine,
692 const char *filename)
696 down_write(&machine->dsos.lock);
698 dso = __dsos__find(&machine->dsos, m->name, true);
700 dso = __dsos__addnew(&machine->dsos, m->name);
704 dso__set_module_info(dso, m, machine);
705 dso__set_long_name(dso, strdup(filename), true);
706 dso->kernel = DSO_TYPE_KERNEL;
711 up_write(&machine->dsos.lock);
715 int machine__process_aux_event(struct machine *machine __maybe_unused,
716 union perf_event *event)
719 perf_event__fprintf_aux(event, stdout);
723 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
724 union perf_event *event)
727 perf_event__fprintf_itrace_start(event, stdout);
731 int machine__process_switch_event(struct machine *machine __maybe_unused,
732 union perf_event *event)
735 perf_event__fprintf_switch(event, stdout);
739 static int is_bpf_image(const char *name)
741 return strncmp(name, "bpf_trampoline_", sizeof("bpf_trampoline_") - 1) == 0 ||
742 strncmp(name, "bpf_dispatcher_", sizeof("bpf_dispatcher_") - 1) == 0;
745 static int machine__process_ksymbol_register(struct machine *machine,
746 union perf_event *event,
747 struct perf_sample *sample __maybe_unused)
750 struct map *map = maps__find(&machine->kmaps, event->ksymbol.addr);
753 struct dso *dso = dso__new(event->ksymbol.name);
756 dso->kernel = DSO_TYPE_KERNEL;
757 map = map__new2(0, dso);
765 map->start = event->ksymbol.addr;
766 map->end = map->start + event->ksymbol.len;
767 maps__insert(&machine->kmaps, map);
768 dso__set_loaded(dso);
770 if (is_bpf_image(event->ksymbol.name)) {
771 dso->binary_type = DSO_BINARY_TYPE__BPF_IMAGE;
772 dso__set_long_name(dso, "", false);
776 sym = symbol__new(map->map_ip(map, map->start),
778 0, 0, event->ksymbol.name);
781 dso__insert_symbol(map->dso, sym);
785 static int machine__process_ksymbol_unregister(struct machine *machine,
786 union perf_event *event,
787 struct perf_sample *sample __maybe_unused)
791 map = maps__find(&machine->kmaps, event->ksymbol.addr);
793 maps__remove(&machine->kmaps, map);
798 int machine__process_ksymbol(struct machine *machine __maybe_unused,
799 union perf_event *event,
800 struct perf_sample *sample)
803 perf_event__fprintf_ksymbol(event, stdout);
805 if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
806 return machine__process_ksymbol_unregister(machine, event,
808 return machine__process_ksymbol_register(machine, event, sample);
811 static struct map *machine__addnew_module_map(struct machine *machine, u64 start,
812 const char *filename)
814 struct map *map = NULL;
818 if (kmod_path__parse_name(&m, filename))
821 dso = machine__findnew_module_dso(machine, &m, filename);
825 map = map__new2(start, dso);
829 maps__insert(&machine->kmaps, map);
831 /* Put the map here because maps__insert alread got it */
834 /* put the dso here, corresponding to machine__findnew_module_dso */
840 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
843 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
845 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
846 struct machine *pos = rb_entry(nd, struct machine, rb_node);
847 ret += __dsos__fprintf(&pos->dsos.head, fp);
853 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
854 bool (skip)(struct dso *dso, int parm), int parm)
856 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
859 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
860 bool (skip)(struct dso *dso, int parm), int parm)
863 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
865 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
866 struct machine *pos = rb_entry(nd, struct machine, rb_node);
867 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
872 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
876 struct dso *kdso = machine__kernel_dso(machine);
878 if (kdso->has_build_id) {
879 char filename[PATH_MAX];
880 if (dso__build_id_filename(kdso, filename, sizeof(filename),
882 printed += fprintf(fp, "[0] %s\n", filename);
885 for (i = 0; i < vmlinux_path__nr_entries; ++i)
886 printed += fprintf(fp, "[%d] %s\n",
887 i + kdso->has_build_id, vmlinux_path[i]);
892 size_t machine__fprintf(struct machine *machine, FILE *fp)
898 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
899 struct threads *threads = &machine->threads[i];
901 down_read(&threads->lock);
903 ret = fprintf(fp, "Threads: %u\n", threads->nr);
905 for (nd = rb_first_cached(&threads->entries); nd;
907 struct thread *pos = rb_entry(nd, struct thread, rb_node);
909 ret += thread__fprintf(pos, fp);
912 up_read(&threads->lock);
917 static struct dso *machine__get_kernel(struct machine *machine)
919 const char *vmlinux_name = machine->mmap_name;
922 if (machine__is_host(machine)) {
923 if (symbol_conf.vmlinux_name)
924 vmlinux_name = symbol_conf.vmlinux_name;
926 kernel = machine__findnew_kernel(machine, vmlinux_name,
927 "[kernel]", DSO_TYPE_KERNEL);
929 if (symbol_conf.default_guest_vmlinux_name)
930 vmlinux_name = symbol_conf.default_guest_vmlinux_name;
932 kernel = machine__findnew_kernel(machine, vmlinux_name,
934 DSO_TYPE_GUEST_KERNEL);
937 if (kernel != NULL && (!kernel->has_build_id))
938 dso__read_running_kernel_build_id(kernel, machine);
943 struct process_args {
947 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
950 if (machine__is_default_guest(machine))
951 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
953 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
956 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
958 /* Figure out the start address of kernel map from /proc/kallsyms.
959 * Returns the name of the start symbol in *symbol_name. Pass in NULL as
960 * symbol_name if it's not that important.
962 static int machine__get_running_kernel_start(struct machine *machine,
963 const char **symbol_name,
964 u64 *start, u64 *end)
966 char filename[PATH_MAX];
971 machine__get_kallsyms_filename(machine, filename, PATH_MAX);
973 if (symbol__restricted_filename(filename, "/proc/kallsyms"))
976 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
977 err = kallsyms__get_function_start(filename, name, &addr);
990 err = kallsyms__get_function_start(filename, "_etext", &addr);
997 int machine__create_extra_kernel_map(struct machine *machine,
999 struct extra_kernel_map *xm)
1004 map = map__new2(xm->start, kernel);
1009 map->pgoff = xm->pgoff;
1011 kmap = map__kmap(map);
1013 strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1015 maps__insert(&machine->kmaps, map);
1017 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1018 kmap->name, map->start, map->end);
1025 static u64 find_entry_trampoline(struct dso *dso)
1027 /* Duplicates are removed so lookup all aliases */
1028 const char *syms[] = {
1029 "_entry_trampoline",
1030 "__entry_trampoline_start",
1031 "entry_SYSCALL_64_trampoline",
1033 struct symbol *sym = dso__first_symbol(dso);
1036 for (; sym; sym = dso__next_symbol(sym)) {
1037 if (sym->binding != STB_GLOBAL)
1039 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1040 if (!strcmp(sym->name, syms[i]))
1049 * These values can be used for kernels that do not have symbols for the entry
1050 * trampolines in kallsyms.
1052 #define X86_64_CPU_ENTRY_AREA_PER_CPU 0xfffffe0000000000ULL
1053 #define X86_64_CPU_ENTRY_AREA_SIZE 0x2c000
1054 #define X86_64_ENTRY_TRAMPOLINE 0x6000
1056 /* Map x86_64 PTI entry trampolines */
1057 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1060 struct maps *kmaps = &machine->kmaps;
1061 int nr_cpus_avail, cpu;
1067 * In the vmlinux case, pgoff is a virtual address which must now be
1068 * mapped to a vmlinux offset.
1070 maps__for_each_entry(kmaps, map) {
1071 struct kmap *kmap = __map__kmap(map);
1072 struct map *dest_map;
1074 if (!kmap || !is_entry_trampoline(kmap->name))
1077 dest_map = maps__find(kmaps, map->pgoff);
1078 if (dest_map != map)
1079 map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1082 if (found || machine->trampolines_mapped)
1085 pgoff = find_entry_trampoline(kernel);
1089 nr_cpus_avail = machine__nr_cpus_avail(machine);
1091 /* Add a 1 page map for each CPU's entry trampoline */
1092 for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1093 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1094 cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1095 X86_64_ENTRY_TRAMPOLINE;
1096 struct extra_kernel_map xm = {
1098 .end = va + page_size,
1102 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1104 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1108 machine->trampolines_mapped = nr_cpus_avail;
1113 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1114 struct dso *kernel __maybe_unused)
1120 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1122 /* In case of renewal the kernel map, destroy previous one */
1123 machine__destroy_kernel_maps(machine);
1125 machine->vmlinux_map = map__new2(0, kernel);
1126 if (machine->vmlinux_map == NULL)
1129 machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1130 maps__insert(&machine->kmaps, machine->vmlinux_map);
1134 void machine__destroy_kernel_maps(struct machine *machine)
1137 struct map *map = machine__kernel_map(machine);
1142 kmap = map__kmap(map);
1143 maps__remove(&machine->kmaps, map);
1144 if (kmap && kmap->ref_reloc_sym) {
1145 zfree((char **)&kmap->ref_reloc_sym->name);
1146 zfree(&kmap->ref_reloc_sym);
1149 map__zput(machine->vmlinux_map);
1152 int machines__create_guest_kernel_maps(struct machines *machines)
1155 struct dirent **namelist = NULL;
1157 char path[PATH_MAX];
1161 if (symbol_conf.default_guest_vmlinux_name ||
1162 symbol_conf.default_guest_modules ||
1163 symbol_conf.default_guest_kallsyms) {
1164 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1167 if (symbol_conf.guestmount) {
1168 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1171 for (i = 0; i < items; i++) {
1172 if (!isdigit(namelist[i]->d_name[0])) {
1173 /* Filter out . and .. */
1176 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1177 if ((*endp != '\0') ||
1178 (endp == namelist[i]->d_name) ||
1179 (errno == ERANGE)) {
1180 pr_debug("invalid directory (%s). Skipping.\n",
1181 namelist[i]->d_name);
1184 sprintf(path, "%s/%s/proc/kallsyms",
1185 symbol_conf.guestmount,
1186 namelist[i]->d_name);
1187 ret = access(path, R_OK);
1189 pr_debug("Can't access file %s\n", path);
1192 machines__create_kernel_maps(machines, pid);
1201 void machines__destroy_kernel_maps(struct machines *machines)
1203 struct rb_node *next = rb_first_cached(&machines->guests);
1205 machine__destroy_kernel_maps(&machines->host);
1208 struct machine *pos = rb_entry(next, struct machine, rb_node);
1210 next = rb_next(&pos->rb_node);
1211 rb_erase_cached(&pos->rb_node, &machines->guests);
1212 machine__delete(pos);
1216 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1218 struct machine *machine = machines__findnew(machines, pid);
1220 if (machine == NULL)
1223 return machine__create_kernel_maps(machine);
1226 int machine__load_kallsyms(struct machine *machine, const char *filename)
1228 struct map *map = machine__kernel_map(machine);
1229 int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1232 dso__set_loaded(map->dso);
1234 * Since /proc/kallsyms will have multiple sessions for the
1235 * kernel, with modules between them, fixup the end of all
1238 maps__fixup_end(&machine->kmaps);
1244 int machine__load_vmlinux_path(struct machine *machine)
1246 struct map *map = machine__kernel_map(machine);
1247 int ret = dso__load_vmlinux_path(map->dso, map);
1250 dso__set_loaded(map->dso);
1255 static char *get_kernel_version(const char *root_dir)
1257 char version[PATH_MAX];
1260 const char *prefix = "Linux version ";
1262 sprintf(version, "%s/proc/version", root_dir);
1263 file = fopen(version, "r");
1267 tmp = fgets(version, sizeof(version), file);
1272 name = strstr(version, prefix);
1275 name += strlen(prefix);
1276 tmp = strchr(name, ' ');
1280 return strdup(name);
1283 static bool is_kmod_dso(struct dso *dso)
1285 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1286 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1289 static int maps__set_module_path(struct maps *maps, const char *path, struct kmod_path *m)
1292 struct map *map = maps__find_by_name(maps, m->name);
1297 long_name = strdup(path);
1298 if (long_name == NULL)
1301 dso__set_long_name(map->dso, long_name, true);
1302 dso__kernel_module_get_build_id(map->dso, "");
1305 * Full name could reveal us kmod compression, so
1306 * we need to update the symtab_type if needed.
1308 if (m->comp && is_kmod_dso(map->dso)) {
1309 map->dso->symtab_type++;
1310 map->dso->comp = m->comp;
1316 static int maps__set_modules_path_dir(struct maps *maps, const char *dir_name, int depth)
1318 struct dirent *dent;
1319 DIR *dir = opendir(dir_name);
1323 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1327 while ((dent = readdir(dir)) != NULL) {
1328 char path[PATH_MAX];
1331 /*sshfs might return bad dent->d_type, so we have to stat*/
1332 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1333 if (stat(path, &st))
1336 if (S_ISDIR(st.st_mode)) {
1337 if (!strcmp(dent->d_name, ".") ||
1338 !strcmp(dent->d_name, ".."))
1341 /* Do not follow top-level source and build symlinks */
1343 if (!strcmp(dent->d_name, "source") ||
1344 !strcmp(dent->d_name, "build"))
1348 ret = maps__set_modules_path_dir(maps, path, depth + 1);
1354 ret = kmod_path__parse_name(&m, dent->d_name);
1359 ret = maps__set_module_path(maps, path, &m);
1373 static int machine__set_modules_path(struct machine *machine)
1376 char modules_path[PATH_MAX];
1378 version = get_kernel_version(machine->root_dir);
1382 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1383 machine->root_dir, version);
1386 return maps__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1388 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1389 u64 *size __maybe_unused,
1390 const char *name __maybe_unused)
1395 static int machine__create_module(void *arg, const char *name, u64 start,
1398 struct machine *machine = arg;
1401 if (arch__fix_module_text_start(&start, &size, name) < 0)
1404 map = machine__addnew_module_map(machine, start, name);
1407 map->end = start + size;
1409 dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1414 static int machine__create_modules(struct machine *machine)
1416 const char *modules;
1417 char path[PATH_MAX];
1419 if (machine__is_default_guest(machine)) {
1420 modules = symbol_conf.default_guest_modules;
1422 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1426 if (symbol__restricted_filename(modules, "/proc/modules"))
1429 if (modules__parse(modules, machine, machine__create_module))
1432 if (!machine__set_modules_path(machine))
1435 pr_debug("Problems setting modules path maps, continuing anyway...\n");
1440 static void machine__set_kernel_mmap(struct machine *machine,
1443 machine->vmlinux_map->start = start;
1444 machine->vmlinux_map->end = end;
1446 * Be a bit paranoid here, some perf.data file came with
1447 * a zero sized synthesized MMAP event for the kernel.
1449 if (start == 0 && end == 0)
1450 machine->vmlinux_map->end = ~0ULL;
1453 static void machine__update_kernel_mmap(struct machine *machine,
1456 struct map *map = machine__kernel_map(machine);
1459 maps__remove(&machine->kmaps, map);
1461 machine__set_kernel_mmap(machine, start, end);
1463 maps__insert(&machine->kmaps, map);
1467 int machine__create_kernel_maps(struct machine *machine)
1469 struct dso *kernel = machine__get_kernel(machine);
1470 const char *name = NULL;
1472 u64 start = 0, end = ~0ULL;
1478 ret = __machine__create_kernel_maps(machine, kernel);
1482 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1483 if (machine__is_host(machine))
1484 pr_debug("Problems creating module maps, "
1485 "continuing anyway...\n");
1487 pr_debug("Problems creating module maps for guest %d, "
1488 "continuing anyway...\n", machine->pid);
1491 if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1493 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1494 machine__destroy_kernel_maps(machine);
1500 * we have a real start address now, so re-order the kmaps
1501 * assume it's the last in the kmaps
1503 machine__update_kernel_mmap(machine, start, end);
1506 if (machine__create_extra_kernel_maps(machine, kernel))
1507 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1510 /* update end address of the kernel map using adjacent module address */
1511 map = map__next(machine__kernel_map(machine));
1513 machine__set_kernel_mmap(machine, start, map->start);
1521 static bool machine__uses_kcore(struct machine *machine)
1525 list_for_each_entry(dso, &machine->dsos.head, node) {
1526 if (dso__is_kcore(dso))
1533 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1534 union perf_event *event)
1536 return machine__is(machine, "x86_64") &&
1537 is_entry_trampoline(event->mmap.filename);
1540 static int machine__process_extra_kernel_map(struct machine *machine,
1541 union perf_event *event)
1543 struct dso *kernel = machine__kernel_dso(machine);
1544 struct extra_kernel_map xm = {
1545 .start = event->mmap.start,
1546 .end = event->mmap.start + event->mmap.len,
1547 .pgoff = event->mmap.pgoff,
1553 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1555 return machine__create_extra_kernel_map(machine, kernel, &xm);
1558 static int machine__process_kernel_mmap_event(struct machine *machine,
1559 union perf_event *event)
1562 enum dso_kernel_type kernel_type;
1563 bool is_kernel_mmap;
1565 /* If we have maps from kcore then we do not need or want any others */
1566 if (machine__uses_kcore(machine))
1569 if (machine__is_host(machine))
1570 kernel_type = DSO_TYPE_KERNEL;
1572 kernel_type = DSO_TYPE_GUEST_KERNEL;
1574 is_kernel_mmap = memcmp(event->mmap.filename,
1576 strlen(machine->mmap_name) - 1) == 0;
1577 if (event->mmap.filename[0] == '/' ||
1578 (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1579 map = machine__addnew_module_map(machine, event->mmap.start,
1580 event->mmap.filename);
1584 map->end = map->start + event->mmap.len;
1585 } else if (is_kernel_mmap) {
1586 const char *symbol_name = (event->mmap.filename +
1587 strlen(machine->mmap_name));
1589 * Should be there already, from the build-id table in
1592 struct dso *kernel = NULL;
1595 down_read(&machine->dsos.lock);
1597 list_for_each_entry(dso, &machine->dsos.head, node) {
1600 * The cpumode passed to is_kernel_module is not the
1601 * cpumode of *this* event. If we insist on passing
1602 * correct cpumode to is_kernel_module, we should
1603 * record the cpumode when we adding this dso to the
1606 * However we don't really need passing correct
1607 * cpumode. We know the correct cpumode must be kernel
1608 * mode (if not, we should not link it onto kernel_dsos
1611 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1612 * is_kernel_module() treats it as a kernel cpumode.
1616 is_kernel_module(dso->long_name,
1617 PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1625 up_read(&machine->dsos.lock);
1628 kernel = machine__findnew_dso(machine, machine->mmap_name);
1632 kernel->kernel = kernel_type;
1633 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1638 if (strstr(kernel->long_name, "vmlinux"))
1639 dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1641 machine__update_kernel_mmap(machine, event->mmap.start,
1642 event->mmap.start + event->mmap.len);
1645 * Avoid using a zero address (kptr_restrict) for the ref reloc
1646 * symbol. Effectively having zero here means that at record
1647 * time /proc/sys/kernel/kptr_restrict was non zero.
1649 if (event->mmap.pgoff != 0) {
1650 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1655 if (machine__is_default_guest(machine)) {
1657 * preload dso of guest kernel and modules
1659 dso__load(kernel, machine__kernel_map(machine));
1661 } else if (perf_event__is_extra_kernel_mmap(machine, event)) {
1662 return machine__process_extra_kernel_map(machine, event);
1669 int machine__process_mmap2_event(struct machine *machine,
1670 union perf_event *event,
1671 struct perf_sample *sample)
1673 struct thread *thread;
1675 struct dso_id dso_id = {
1676 .maj = event->mmap2.maj,
1677 .min = event->mmap2.min,
1678 .ino = event->mmap2.ino,
1679 .ino_generation = event->mmap2.ino_generation,
1684 perf_event__fprintf_mmap2(event, stdout);
1686 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1687 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1688 ret = machine__process_kernel_mmap_event(machine, event);
1694 thread = machine__findnew_thread(machine, event->mmap2.pid,
1699 map = map__new(machine, event->mmap2.start,
1700 event->mmap2.len, event->mmap2.pgoff,
1701 &dso_id, event->mmap2.prot,
1703 event->mmap2.filename, thread);
1706 goto out_problem_map;
1708 ret = thread__insert_map(thread, map);
1710 goto out_problem_insert;
1712 thread__put(thread);
1719 thread__put(thread);
1721 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1725 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1726 struct perf_sample *sample)
1728 struct thread *thread;
1734 perf_event__fprintf_mmap(event, stdout);
1736 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1737 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1738 ret = machine__process_kernel_mmap_event(machine, event);
1744 thread = machine__findnew_thread(machine, event->mmap.pid,
1749 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1752 map = map__new(machine, event->mmap.start,
1753 event->mmap.len, event->mmap.pgoff,
1754 NULL, prot, 0, event->mmap.filename, thread);
1757 goto out_problem_map;
1759 ret = thread__insert_map(thread, map);
1761 goto out_problem_insert;
1763 thread__put(thread);
1770 thread__put(thread);
1772 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1776 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1778 struct threads *threads = machine__threads(machine, th->tid);
1780 if (threads->last_match == th)
1781 threads__set_last_match(threads, NULL);
1784 down_write(&threads->lock);
1786 BUG_ON(refcount_read(&th->refcnt) == 0);
1788 rb_erase_cached(&th->rb_node, &threads->entries);
1789 RB_CLEAR_NODE(&th->rb_node);
1792 * Move it first to the dead_threads list, then drop the reference,
1793 * if this is the last reference, then the thread__delete destructor
1794 * will be called and we will remove it from the dead_threads list.
1796 list_add_tail(&th->node, &threads->dead);
1799 * We need to do the put here because if this is the last refcount,
1800 * then we will be touching the threads->dead head when removing the
1806 up_write(&threads->lock);
1809 void machine__remove_thread(struct machine *machine, struct thread *th)
1811 return __machine__remove_thread(machine, th, true);
1814 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1815 struct perf_sample *sample)
1817 struct thread *thread = machine__find_thread(machine,
1820 struct thread *parent = machine__findnew_thread(machine,
1823 bool do_maps_clone = true;
1827 perf_event__fprintf_task(event, stdout);
1830 * There may be an existing thread that is not actually the parent,
1831 * either because we are processing events out of order, or because the
1832 * (fork) event that would have removed the thread was lost. Assume the
1833 * latter case and continue on as best we can.
1835 if (parent->pid_ != (pid_t)event->fork.ppid) {
1836 dump_printf("removing erroneous parent thread %d/%d\n",
1837 parent->pid_, parent->tid);
1838 machine__remove_thread(machine, parent);
1839 thread__put(parent);
1840 parent = machine__findnew_thread(machine, event->fork.ppid,
1844 /* if a thread currently exists for the thread id remove it */
1845 if (thread != NULL) {
1846 machine__remove_thread(machine, thread);
1847 thread__put(thread);
1850 thread = machine__findnew_thread(machine, event->fork.pid,
1853 * When synthesizing FORK events, we are trying to create thread
1854 * objects for the already running tasks on the machine.
1856 * Normally, for a kernel FORK event, we want to clone the parent's
1857 * maps because that is what the kernel just did.
1859 * But when synthesizing, this should not be done. If we do, we end up
1860 * with overlapping maps as we process the sythesized MMAP2 events that
1861 * get delivered shortly thereafter.
1863 * Use the FORK event misc flags in an internal way to signal this
1864 * situation, so we can elide the map clone when appropriate.
1866 if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1867 do_maps_clone = false;
1869 if (thread == NULL || parent == NULL ||
1870 thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1871 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1874 thread__put(thread);
1875 thread__put(parent);
1880 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1881 struct perf_sample *sample __maybe_unused)
1883 struct thread *thread = machine__find_thread(machine,
1888 perf_event__fprintf_task(event, stdout);
1890 if (thread != NULL) {
1891 thread__exited(thread);
1892 thread__put(thread);
1898 int machine__process_event(struct machine *machine, union perf_event *event,
1899 struct perf_sample *sample)
1903 switch (event->header.type) {
1904 case PERF_RECORD_COMM:
1905 ret = machine__process_comm_event(machine, event, sample); break;
1906 case PERF_RECORD_MMAP:
1907 ret = machine__process_mmap_event(machine, event, sample); break;
1908 case PERF_RECORD_NAMESPACES:
1909 ret = machine__process_namespaces_event(machine, event, sample); break;
1910 case PERF_RECORD_CGROUP:
1911 ret = machine__process_cgroup_event(machine, event, sample); break;
1912 case PERF_RECORD_MMAP2:
1913 ret = machine__process_mmap2_event(machine, event, sample); break;
1914 case PERF_RECORD_FORK:
1915 ret = machine__process_fork_event(machine, event, sample); break;
1916 case PERF_RECORD_EXIT:
1917 ret = machine__process_exit_event(machine, event, sample); break;
1918 case PERF_RECORD_LOST:
1919 ret = machine__process_lost_event(machine, event, sample); break;
1920 case PERF_RECORD_AUX:
1921 ret = machine__process_aux_event(machine, event); break;
1922 case PERF_RECORD_ITRACE_START:
1923 ret = machine__process_itrace_start_event(machine, event); break;
1924 case PERF_RECORD_LOST_SAMPLES:
1925 ret = machine__process_lost_samples_event(machine, event, sample); break;
1926 case PERF_RECORD_SWITCH:
1927 case PERF_RECORD_SWITCH_CPU_WIDE:
1928 ret = machine__process_switch_event(machine, event); break;
1929 case PERF_RECORD_KSYMBOL:
1930 ret = machine__process_ksymbol(machine, event, sample); break;
1931 case PERF_RECORD_BPF_EVENT:
1932 ret = machine__process_bpf(machine, event, sample); break;
1941 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1943 if (!regexec(regex, sym->name, 0, NULL, 0))
1948 static void ip__resolve_ams(struct thread *thread,
1949 struct addr_map_symbol *ams,
1952 struct addr_location al;
1954 memset(&al, 0, sizeof(al));
1956 * We cannot use the header.misc hint to determine whether a
1957 * branch stack address is user, kernel, guest, hypervisor.
1958 * Branches may straddle the kernel/user/hypervisor boundaries.
1959 * Thus, we have to try consecutively until we find a match
1960 * or else, the symbol is unknown
1962 thread__find_cpumode_addr_location(thread, ip, &al);
1965 ams->al_addr = al.addr;
1966 ams->ms.maps = al.maps;
1967 ams->ms.sym = al.sym;
1968 ams->ms.map = al.map;
1972 static void ip__resolve_data(struct thread *thread,
1973 u8 m, struct addr_map_symbol *ams,
1974 u64 addr, u64 phys_addr)
1976 struct addr_location al;
1978 memset(&al, 0, sizeof(al));
1980 thread__find_symbol(thread, m, addr, &al);
1983 ams->al_addr = al.addr;
1984 ams->ms.maps = al.maps;
1985 ams->ms.sym = al.sym;
1986 ams->ms.map = al.map;
1987 ams->phys_addr = phys_addr;
1990 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1991 struct addr_location *al)
1993 struct mem_info *mi = mem_info__new();
1998 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1999 ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
2000 sample->addr, sample->phys_addr);
2001 mi->data_src.val = sample->data_src;
2006 static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2008 struct map *map = ms->map;
2009 char *srcline = NULL;
2011 if (!map || callchain_param.key == CCKEY_FUNCTION)
2014 srcline = srcline__tree_find(&map->dso->srclines, ip);
2016 bool show_sym = false;
2017 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2019 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
2020 ms->sym, show_sym, show_addr, ip);
2021 srcline__tree_insert(&map->dso->srclines, ip, srcline);
2032 static int add_callchain_ip(struct thread *thread,
2033 struct callchain_cursor *cursor,
2034 struct symbol **parent,
2035 struct addr_location *root_al,
2039 struct branch_flags *flags,
2040 struct iterations *iter,
2043 struct map_symbol ms;
2044 struct addr_location al;
2045 int nr_loop_iter = 0;
2046 u64 iter_cycles = 0;
2047 const char *srcline = NULL;
2052 thread__find_cpumode_addr_location(thread, ip, &al);
2054 if (ip >= PERF_CONTEXT_MAX) {
2056 case PERF_CONTEXT_HV:
2057 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2059 case PERF_CONTEXT_KERNEL:
2060 *cpumode = PERF_RECORD_MISC_KERNEL;
2062 case PERF_CONTEXT_USER:
2063 *cpumode = PERF_RECORD_MISC_USER;
2066 pr_debug("invalid callchain context: "
2067 "%"PRId64"\n", (s64) ip);
2069 * It seems the callchain is corrupted.
2072 callchain_cursor_reset(cursor);
2077 thread__find_symbol(thread, *cpumode, ip, &al);
2080 if (al.sym != NULL) {
2081 if (perf_hpp_list.parent && !*parent &&
2082 symbol__match_regex(al.sym, &parent_regex))
2084 else if (have_ignore_callees && root_al &&
2085 symbol__match_regex(al.sym, &ignore_callees_regex)) {
2086 /* Treat this symbol as the root,
2087 forgetting its callees. */
2089 callchain_cursor_reset(cursor);
2093 if (symbol_conf.hide_unresolved && al.sym == NULL)
2097 nr_loop_iter = iter->nr_loop_iter;
2098 iter_cycles = iter->cycles;
2104 srcline = callchain_srcline(&ms, al.addr);
2105 return callchain_cursor_append(cursor, ip, &ms,
2106 branch, flags, nr_loop_iter,
2107 iter_cycles, branch_from, srcline);
2110 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2111 struct addr_location *al)
2114 const struct branch_stack *bs = sample->branch_stack;
2115 struct branch_entry *entries = perf_sample__branch_entries(sample);
2116 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2121 for (i = 0; i < bs->nr; i++) {
2122 ip__resolve_ams(al->thread, &bi[i].to, entries[i].to);
2123 ip__resolve_ams(al->thread, &bi[i].from, entries[i].from);
2124 bi[i].flags = entries[i].flags;
2129 static void save_iterations(struct iterations *iter,
2130 struct branch_entry *be, int nr)
2134 iter->nr_loop_iter++;
2137 for (i = 0; i < nr; i++)
2138 iter->cycles += be[i].flags.cycles;
2143 #define NO_ENTRY 0xff
2145 #define PERF_MAX_BRANCH_DEPTH 127
2148 static int remove_loops(struct branch_entry *l, int nr,
2149 struct iterations *iter)
2152 unsigned char chash[CHASHSZ];
2154 memset(chash, NO_ENTRY, sizeof(chash));
2156 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2158 for (i = 0; i < nr; i++) {
2159 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2161 /* no collision handling for now */
2162 if (chash[h] == NO_ENTRY) {
2164 } else if (l[chash[h]].from == l[i].from) {
2165 bool is_loop = true;
2166 /* check if it is a real loop */
2168 for (j = chash[h]; j < i && i + off < nr; j++, off++)
2169 if (l[j].from != l[i + off].from) {
2176 save_iterations(iter + i + off,
2179 memmove(iter + i, iter + i + off,
2182 memmove(l + i, l + i + off,
2193 static int lbr_callchain_add_kernel_ip(struct thread *thread,
2194 struct callchain_cursor *cursor,
2195 struct perf_sample *sample,
2196 struct symbol **parent,
2197 struct addr_location *root_al,
2199 bool callee, int end)
2201 struct ip_callchain *chain = sample->callchain;
2202 u8 cpumode = PERF_RECORD_MISC_USER;
2206 for (i = 0; i < end + 1; i++) {
2207 err = add_callchain_ip(thread, cursor, parent,
2208 root_al, &cpumode, chain->ips[i],
2209 false, NULL, NULL, branch_from);
2216 for (i = end; i >= 0; i--) {
2217 err = add_callchain_ip(thread, cursor, parent,
2218 root_al, &cpumode, chain->ips[i],
2219 false, NULL, NULL, branch_from);
2227 static void save_lbr_cursor_node(struct thread *thread,
2228 struct callchain_cursor *cursor,
2231 struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2236 if (cursor->pos == cursor->nr) {
2237 lbr_stitch->prev_lbr_cursor[idx].valid = false;
2242 cursor->curr = cursor->first;
2244 cursor->curr = cursor->curr->next;
2245 memcpy(&lbr_stitch->prev_lbr_cursor[idx], cursor->curr,
2246 sizeof(struct callchain_cursor_node));
2248 lbr_stitch->prev_lbr_cursor[idx].valid = true;
2252 static int lbr_callchain_add_lbr_ip(struct thread *thread,
2253 struct callchain_cursor *cursor,
2254 struct perf_sample *sample,
2255 struct symbol **parent,
2256 struct addr_location *root_al,
2260 struct branch_stack *lbr_stack = sample->branch_stack;
2261 struct branch_entry *entries = perf_sample__branch_entries(sample);
2262 u8 cpumode = PERF_RECORD_MISC_USER;
2263 int lbr_nr = lbr_stack->nr;
2264 struct branch_flags *flags;
2269 * The curr and pos are not used in writing session. They are cleared
2270 * in callchain_cursor_commit() when the writing session is closed.
2271 * Using curr and pos to track the current cursor node.
2273 if (thread->lbr_stitch) {
2274 cursor->curr = NULL;
2275 cursor->pos = cursor->nr;
2277 cursor->curr = cursor->first;
2278 for (i = 0; i < (int)(cursor->nr - 1); i++)
2279 cursor->curr = cursor->curr->next;
2284 /* Add LBR ip from first entries.to */
2286 flags = &entries[0].flags;
2287 *branch_from = entries[0].from;
2288 err = add_callchain_ip(thread, cursor, parent,
2289 root_al, &cpumode, ip,
2296 * The number of cursor node increases.
2297 * Move the current cursor node.
2298 * But does not need to save current cursor node for entry 0.
2299 * It's impossible to stitch the whole LBRs of previous sample.
2301 if (thread->lbr_stitch && (cursor->pos != cursor->nr)) {
2303 cursor->curr = cursor->first;
2305 cursor->curr = cursor->curr->next;
2309 /* Add LBR ip from entries.from one by one. */
2310 for (i = 0; i < lbr_nr; i++) {
2311 ip = entries[i].from;
2312 flags = &entries[i].flags;
2313 err = add_callchain_ip(thread, cursor, parent,
2314 root_al, &cpumode, ip,
2319 save_lbr_cursor_node(thread, cursor, i);
2324 /* Add LBR ip from entries.from one by one. */
2325 for (i = lbr_nr - 1; i >= 0; i--) {
2326 ip = entries[i].from;
2327 flags = &entries[i].flags;
2328 err = add_callchain_ip(thread, cursor, parent,
2329 root_al, &cpumode, ip,
2334 save_lbr_cursor_node(thread, cursor, i);
2337 /* Add LBR ip from first entries.to */
2339 flags = &entries[0].flags;
2340 *branch_from = entries[0].from;
2341 err = add_callchain_ip(thread, cursor, parent,
2342 root_al, &cpumode, ip,
2351 static int lbr_callchain_add_stitched_lbr_ip(struct thread *thread,
2352 struct callchain_cursor *cursor)
2354 struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2355 struct callchain_cursor_node *cnode;
2356 struct stitch_list *stitch_node;
2359 list_for_each_entry(stitch_node, &lbr_stitch->lists, node) {
2360 cnode = &stitch_node->cursor;
2362 err = callchain_cursor_append(cursor, cnode->ip,
2365 &cnode->branch_flags,
2366 cnode->nr_loop_iter,
2376 static struct stitch_list *get_stitch_node(struct thread *thread)
2378 struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2379 struct stitch_list *stitch_node;
2381 if (!list_empty(&lbr_stitch->free_lists)) {
2382 stitch_node = list_first_entry(&lbr_stitch->free_lists,
2383 struct stitch_list, node);
2384 list_del(&stitch_node->node);
2389 return malloc(sizeof(struct stitch_list));
2392 static bool has_stitched_lbr(struct thread *thread,
2393 struct perf_sample *cur,
2394 struct perf_sample *prev,
2395 unsigned int max_lbr,
2398 struct branch_stack *cur_stack = cur->branch_stack;
2399 struct branch_entry *cur_entries = perf_sample__branch_entries(cur);
2400 struct branch_stack *prev_stack = prev->branch_stack;
2401 struct branch_entry *prev_entries = perf_sample__branch_entries(prev);
2402 struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2403 int i, j, nr_identical_branches = 0;
2404 struct stitch_list *stitch_node;
2405 u64 cur_base, distance;
2407 if (!cur_stack || !prev_stack)
2410 /* Find the physical index of the base-of-stack for current sample. */
2411 cur_base = max_lbr - cur_stack->nr + cur_stack->hw_idx + 1;
2413 distance = (prev_stack->hw_idx > cur_base) ? (prev_stack->hw_idx - cur_base) :
2414 (max_lbr + prev_stack->hw_idx - cur_base);
2415 /* Previous sample has shorter stack. Nothing can be stitched. */
2416 if (distance + 1 > prev_stack->nr)
2420 * Check if there are identical LBRs between two samples.
2421 * Identicall LBRs must have same from, to and flags values. Also,
2422 * they have to be saved in the same LBR registers (same physical
2425 * Starts from the base-of-stack of current sample.
2427 for (i = distance, j = cur_stack->nr - 1; (i >= 0) && (j >= 0); i--, j--) {
2428 if ((prev_entries[i].from != cur_entries[j].from) ||
2429 (prev_entries[i].to != cur_entries[j].to) ||
2430 (prev_entries[i].flags.value != cur_entries[j].flags.value))
2432 nr_identical_branches++;
2435 if (!nr_identical_branches)
2439 * Save the LBRs between the base-of-stack of previous sample
2440 * and the base-of-stack of current sample into lbr_stitch->lists.
2441 * These LBRs will be stitched later.
2443 for (i = prev_stack->nr - 1; i > (int)distance; i--) {
2445 if (!lbr_stitch->prev_lbr_cursor[i].valid)
2448 stitch_node = get_stitch_node(thread);
2452 memcpy(&stitch_node->cursor, &lbr_stitch->prev_lbr_cursor[i],
2453 sizeof(struct callchain_cursor_node));
2456 list_add(&stitch_node->node, &lbr_stitch->lists);
2458 list_add_tail(&stitch_node->node, &lbr_stitch->lists);
2464 static bool alloc_lbr_stitch(struct thread *thread, unsigned int max_lbr)
2466 if (thread->lbr_stitch)
2469 thread->lbr_stitch = zalloc(sizeof(*thread->lbr_stitch));
2470 if (!thread->lbr_stitch)
2473 thread->lbr_stitch->prev_lbr_cursor = calloc(max_lbr + 1, sizeof(struct callchain_cursor_node));
2474 if (!thread->lbr_stitch->prev_lbr_cursor)
2475 goto free_lbr_stitch;
2477 INIT_LIST_HEAD(&thread->lbr_stitch->lists);
2478 INIT_LIST_HEAD(&thread->lbr_stitch->free_lists);
2483 zfree(&thread->lbr_stitch);
2485 pr_warning("Failed to allocate space for stitched LBRs. Disable LBR stitch\n");
2486 thread->lbr_stitch_enable = false;
2491 * Recolve LBR callstack chain sample
2493 * 1 on success get LBR callchain information
2494 * 0 no available LBR callchain information, should try fp
2495 * negative error code on other errors.
2497 static int resolve_lbr_callchain_sample(struct thread *thread,
2498 struct callchain_cursor *cursor,
2499 struct perf_sample *sample,
2500 struct symbol **parent,
2501 struct addr_location *root_al,
2503 unsigned int max_lbr)
2505 bool callee = (callchain_param.order == ORDER_CALLEE);
2506 struct ip_callchain *chain = sample->callchain;
2507 int chain_nr = min(max_stack, (int)chain->nr), i;
2508 struct lbr_stitch *lbr_stitch;
2509 bool stitched_lbr = false;
2510 u64 branch_from = 0;
2513 for (i = 0; i < chain_nr; i++) {
2514 if (chain->ips[i] == PERF_CONTEXT_USER)
2518 /* LBR only affects the user callchain */
2522 if (thread->lbr_stitch_enable && !sample->no_hw_idx &&
2523 (max_lbr > 0) && alloc_lbr_stitch(thread, max_lbr)) {
2524 lbr_stitch = thread->lbr_stitch;
2526 stitched_lbr = has_stitched_lbr(thread, sample,
2527 &lbr_stitch->prev_sample,
2530 if (!stitched_lbr && !list_empty(&lbr_stitch->lists)) {
2531 list_replace_init(&lbr_stitch->lists,
2532 &lbr_stitch->free_lists);
2534 memcpy(&lbr_stitch->prev_sample, sample, sizeof(*sample));
2539 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2540 parent, root_al, branch_from,
2545 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2546 root_al, &branch_from, true);
2551 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2558 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2562 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2563 root_al, &branch_from, false);
2568 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2569 parent, root_al, branch_from,
2577 return (err < 0) ? err : 0;
2580 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2581 struct callchain_cursor *cursor,
2582 struct symbol **parent,
2583 struct addr_location *root_al,
2584 u8 *cpumode, int ent)
2588 while (--ent >= 0) {
2589 u64 ip = chain->ips[ent];
2591 if (ip >= PERF_CONTEXT_MAX) {
2592 err = add_callchain_ip(thread, cursor, parent,
2593 root_al, cpumode, ip,
2594 false, NULL, NULL, 0);
2601 static int thread__resolve_callchain_sample(struct thread *thread,
2602 struct callchain_cursor *cursor,
2603 struct evsel *evsel,
2604 struct perf_sample *sample,
2605 struct symbol **parent,
2606 struct addr_location *root_al,
2609 struct branch_stack *branch = sample->branch_stack;
2610 struct branch_entry *entries = perf_sample__branch_entries(sample);
2611 struct ip_callchain *chain = sample->callchain;
2613 u8 cpumode = PERF_RECORD_MISC_USER;
2614 int i, j, err, nr_entries;
2619 chain_nr = chain->nr;
2621 if (evsel__has_branch_callstack(evsel)) {
2622 struct perf_env *env = evsel__env(evsel);
2624 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2626 !env ? 0 : env->max_branches);
2628 return (err < 0) ? err : 0;
2632 * Based on DWARF debug information, some architectures skip
2633 * a callchain entry saved by the kernel.
2635 skip_idx = arch_skip_callchain_idx(thread, chain);
2638 * Add branches to call stack for easier browsing. This gives
2639 * more context for a sample than just the callers.
2641 * This uses individual histograms of paths compared to the
2642 * aggregated histograms the normal LBR mode uses.
2644 * Limitations for now:
2645 * - No extra filters
2646 * - No annotations (should annotate somehow)
2649 if (branch && callchain_param.branch_callstack) {
2650 int nr = min(max_stack, (int)branch->nr);
2651 struct branch_entry be[nr];
2652 struct iterations iter[nr];
2654 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2655 pr_warning("corrupted branch chain. skipping...\n");
2659 for (i = 0; i < nr; i++) {
2660 if (callchain_param.order == ORDER_CALLEE) {
2667 * Check for overlap into the callchain.
2668 * The return address is one off compared to
2669 * the branch entry. To adjust for this
2670 * assume the calling instruction is not longer
2673 if (i == skip_idx ||
2674 chain->ips[first_call] >= PERF_CONTEXT_MAX)
2676 else if (be[i].from < chain->ips[first_call] &&
2677 be[i].from >= chain->ips[first_call] - 8)
2680 be[i] = entries[branch->nr - i - 1];
2683 memset(iter, 0, sizeof(struct iterations) * nr);
2684 nr = remove_loops(be, nr, iter);
2686 for (i = 0; i < nr; i++) {
2687 err = add_callchain_ip(thread, cursor, parent,
2694 err = add_callchain_ip(thread, cursor, parent, root_al,
2711 if (chain && callchain_param.order != ORDER_CALLEE) {
2712 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2713 &cpumode, chain->nr - first_call);
2715 return (err < 0) ? err : 0;
2717 for (i = first_call, nr_entries = 0;
2718 i < chain_nr && nr_entries < max_stack; i++) {
2721 if (callchain_param.order == ORDER_CALLEE)
2724 j = chain->nr - i - 1;
2726 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2731 if (ip < PERF_CONTEXT_MAX)
2733 else if (callchain_param.order != ORDER_CALLEE) {
2734 err = find_prev_cpumode(chain, thread, cursor, parent,
2735 root_al, &cpumode, j);
2737 return (err < 0) ? err : 0;
2741 err = add_callchain_ip(thread, cursor, parent,
2742 root_al, &cpumode, ip,
2743 false, NULL, NULL, 0);
2746 return (err < 0) ? err : 0;
2752 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
2754 struct symbol *sym = ms->sym;
2755 struct map *map = ms->map;
2756 struct inline_node *inline_node;
2757 struct inline_list *ilist;
2761 if (!symbol_conf.inline_name || !map || !sym)
2764 addr = map__map_ip(map, ip);
2765 addr = map__rip_2objdump(map, addr);
2767 inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2769 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2772 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2775 list_for_each_entry(ilist, &inline_node->val, list) {
2776 struct map_symbol ilist_ms = {
2779 .sym = ilist->symbol,
2781 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
2782 NULL, 0, 0, 0, ilist->srcline);
2791 static int unwind_entry(struct unwind_entry *entry, void *arg)
2793 struct callchain_cursor *cursor = arg;
2794 const char *srcline = NULL;
2795 u64 addr = entry->ip;
2797 if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
2800 if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
2804 * Convert entry->ip from a virtual address to an offset in
2805 * its corresponding binary.
2808 addr = map__map_ip(entry->ms.map, entry->ip);
2810 srcline = callchain_srcline(&entry->ms, addr);
2811 return callchain_cursor_append(cursor, entry->ip, &entry->ms,
2812 false, NULL, 0, 0, 0, srcline);
2815 static int thread__resolve_callchain_unwind(struct thread *thread,
2816 struct callchain_cursor *cursor,
2817 struct evsel *evsel,
2818 struct perf_sample *sample,
2821 /* Can we do dwarf post unwind? */
2822 if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2823 (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
2826 /* Bail out if nothing was captured. */
2827 if ((!sample->user_regs.regs) ||
2828 (!sample->user_stack.size))
2831 return unwind__get_entries(unwind_entry, cursor,
2832 thread, sample, max_stack);
2835 int thread__resolve_callchain(struct thread *thread,
2836 struct callchain_cursor *cursor,
2837 struct evsel *evsel,
2838 struct perf_sample *sample,
2839 struct symbol **parent,
2840 struct addr_location *root_al,
2845 callchain_cursor_reset(cursor);
2847 if (callchain_param.order == ORDER_CALLEE) {
2848 ret = thread__resolve_callchain_sample(thread, cursor,
2854 ret = thread__resolve_callchain_unwind(thread, cursor,
2858 ret = thread__resolve_callchain_unwind(thread, cursor,
2863 ret = thread__resolve_callchain_sample(thread, cursor,
2872 int machine__for_each_thread(struct machine *machine,
2873 int (*fn)(struct thread *thread, void *p),
2876 struct threads *threads;
2878 struct thread *thread;
2882 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2883 threads = &machine->threads[i];
2884 for (nd = rb_first_cached(&threads->entries); nd;
2886 thread = rb_entry(nd, struct thread, rb_node);
2887 rc = fn(thread, priv);
2892 list_for_each_entry(thread, &threads->dead, node) {
2893 rc = fn(thread, priv);
2901 int machines__for_each_thread(struct machines *machines,
2902 int (*fn)(struct thread *thread, void *p),
2908 rc = machine__for_each_thread(&machines->host, fn, priv);
2912 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2913 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2915 rc = machine__for_each_thread(machine, fn, priv);
2922 pid_t machine__get_current_tid(struct machine *machine, int cpu)
2924 int nr_cpus = min(machine->env->nr_cpus_online, MAX_NR_CPUS);
2926 if (cpu < 0 || cpu >= nr_cpus || !machine->current_tid)
2929 return machine->current_tid[cpu];
2932 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
2935 struct thread *thread;
2936 int nr_cpus = min(machine->env->nr_cpus_online, MAX_NR_CPUS);
2941 if (!machine->current_tid) {
2944 machine->current_tid = calloc(nr_cpus, sizeof(pid_t));
2945 if (!machine->current_tid)
2947 for (i = 0; i < nr_cpus; i++)
2948 machine->current_tid[i] = -1;
2951 if (cpu >= nr_cpus) {
2952 pr_err("Requested CPU %d too large. ", cpu);
2953 pr_err("Consider raising MAX_NR_CPUS\n");
2957 machine->current_tid[cpu] = tid;
2959 thread = machine__findnew_thread(machine, pid, tid);
2964 thread__put(thread);
2970 * Compares the raw arch string. N.B. see instead perf_env__arch() if a
2971 * normalized arch is needed.
2973 bool machine__is(struct machine *machine, const char *arch)
2975 return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
2978 int machine__nr_cpus_avail(struct machine *machine)
2980 return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
2983 int machine__get_kernel_start(struct machine *machine)
2985 struct map *map = machine__kernel_map(machine);
2989 * The only addresses above 2^63 are kernel addresses of a 64-bit
2990 * kernel. Note that addresses are unsigned so that on a 32-bit system
2991 * all addresses including kernel addresses are less than 2^32. In
2992 * that case (32-bit system), if the kernel mapping is unknown, all
2993 * addresses will be assumed to be in user space - see
2994 * machine__kernel_ip().
2996 machine->kernel_start = 1ULL << 63;
2998 err = map__load(map);
3000 * On x86_64, PTI entry trampolines are less than the
3001 * start of kernel text, but still above 2^63. So leave
3002 * kernel_start = 1ULL << 63 for x86_64.
3004 if (!err && !machine__is(machine, "x86_64"))
3005 machine->kernel_start = map->start;
3010 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
3012 u8 addr_cpumode = cpumode;
3015 if (!machine->single_address_space)
3018 kernel_ip = machine__kernel_ip(machine, addr);
3020 case PERF_RECORD_MISC_KERNEL:
3021 case PERF_RECORD_MISC_USER:
3022 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
3023 PERF_RECORD_MISC_USER;
3025 case PERF_RECORD_MISC_GUEST_KERNEL:
3026 case PERF_RECORD_MISC_GUEST_USER:
3027 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
3028 PERF_RECORD_MISC_GUEST_USER;
3034 return addr_cpumode;
3037 struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id)
3039 return dsos__findnew_id(&machine->dsos, filename, id);
3042 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
3044 return machine__findnew_dso_id(machine, filename, NULL);
3047 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
3049 struct machine *machine = vmachine;
3051 struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
3056 *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
3057 *addrp = map->unmap_ip(map, sym->start);