1 // SPDX-License-Identifier: GPL-2.0
16 #include "map_symbol.h"
18 #include "mem-events.h"
29 #include <sys/types.h>
33 #include "linux/hash.h"
35 #include "bpf-event.h"
36 #include <internal/lib.h> // page_size
38 #include "arm64-frame-pointer-unwind-support.h"
40 #include <linux/ctype.h>
41 #include <symbol/kallsyms.h>
42 #include <linux/mman.h>
43 #include <linux/string.h>
44 #include <linux/zalloc.h>
46 static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
47 struct thread *th, bool lock);
48 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip);
50 static struct dso *machine__kernel_dso(struct machine *machine)
52 return map__dso(machine->vmlinux_map);
55 static void dsos__init(struct dsos *dsos)
57 INIT_LIST_HEAD(&dsos->head);
59 init_rwsem(&dsos->lock);
62 static void machine__threads_init(struct machine *machine)
66 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
67 struct threads *threads = &machine->threads[i];
68 threads->entries = RB_ROOT_CACHED;
69 init_rwsem(&threads->lock);
71 INIT_LIST_HEAD(&threads->dead);
72 threads->last_match = NULL;
76 static int thread_rb_node__cmp_tid(const void *key, const struct rb_node *nd)
78 int to_find = (int) *((pid_t *)key);
80 return to_find - (int)thread__tid(rb_entry(nd, struct thread_rb_node, rb_node)->thread);
83 static struct thread_rb_node *thread_rb_node__find(const struct thread *th,
86 pid_t to_find = thread__tid(th);
87 struct rb_node *nd = rb_find(&to_find, tree, thread_rb_node__cmp_tid);
89 return rb_entry(nd, struct thread_rb_node, rb_node);
92 static int machine__set_mmap_name(struct machine *machine)
94 if (machine__is_host(machine))
95 machine->mmap_name = strdup("[kernel.kallsyms]");
96 else if (machine__is_default_guest(machine))
97 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
98 else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
100 machine->mmap_name = NULL;
102 return machine->mmap_name ? 0 : -ENOMEM;
105 static void thread__set_guest_comm(struct thread *thread, pid_t pid)
109 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
110 thread__set_comm(thread, comm, 0);
113 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
117 memset(machine, 0, sizeof(*machine));
118 machine->kmaps = maps__new(machine);
119 if (machine->kmaps == NULL)
122 RB_CLEAR_NODE(&machine->rb_node);
123 dsos__init(&machine->dsos);
125 machine__threads_init(machine);
127 machine->vdso_info = NULL;
132 machine->id_hdr_size = 0;
133 machine->kptr_restrict_warned = false;
134 machine->comm_exec = false;
135 machine->kernel_start = 0;
136 machine->vmlinux_map = NULL;
138 machine->root_dir = strdup(root_dir);
139 if (machine->root_dir == NULL)
142 if (machine__set_mmap_name(machine))
145 if (pid != HOST_KERNEL_ID) {
146 struct thread *thread = machine__findnew_thread(machine, -1,
152 thread__set_guest_comm(thread, pid);
156 machine->current_tid = NULL;
161 zfree(&machine->kmaps);
162 zfree(&machine->root_dir);
163 zfree(&machine->mmap_name);
168 struct machine *machine__new_host(void)
170 struct machine *machine = malloc(sizeof(*machine));
172 if (machine != NULL) {
173 machine__init(machine, "", HOST_KERNEL_ID);
175 if (machine__create_kernel_maps(machine) < 0)
185 struct machine *machine__new_kallsyms(void)
187 struct machine *machine = machine__new_host();
190 * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
191 * ask for not using the kcore parsing code, once this one is fixed
192 * to create a map per module.
194 if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
195 machine__delete(machine);
202 static void dsos__purge(struct dsos *dsos)
206 down_write(&dsos->lock);
208 list_for_each_entry_safe(pos, n, &dsos->head, node) {
209 RB_CLEAR_NODE(&pos->rb_node);
211 list_del_init(&pos->node);
215 up_write(&dsos->lock);
218 static void dsos__exit(struct dsos *dsos)
221 exit_rwsem(&dsos->lock);
224 void machine__delete_threads(struct machine *machine)
229 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
230 struct threads *threads = &machine->threads[i];
231 down_write(&threads->lock);
232 nd = rb_first_cached(&threads->entries);
234 struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
237 __machine__remove_thread(machine, trb, trb->thread, false);
239 up_write(&threads->lock);
243 void machine__exit(struct machine *machine)
250 machine__destroy_kernel_maps(machine);
251 maps__zput(machine->kmaps);
252 dsos__exit(&machine->dsos);
253 machine__exit_vdso(machine);
254 zfree(&machine->root_dir);
255 zfree(&machine->mmap_name);
256 zfree(&machine->current_tid);
257 zfree(&machine->kallsyms_filename);
259 machine__delete_threads(machine);
260 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
261 struct threads *threads = &machine->threads[i];
263 exit_rwsem(&threads->lock);
267 void machine__delete(struct machine *machine)
270 machine__exit(machine);
275 void machines__init(struct machines *machines)
277 machine__init(&machines->host, "", HOST_KERNEL_ID);
278 machines->guests = RB_ROOT_CACHED;
281 void machines__exit(struct machines *machines)
283 machine__exit(&machines->host);
287 struct machine *machines__add(struct machines *machines, pid_t pid,
288 const char *root_dir)
290 struct rb_node **p = &machines->guests.rb_root.rb_node;
291 struct rb_node *parent = NULL;
292 struct machine *pos, *machine = malloc(sizeof(*machine));
293 bool leftmost = true;
298 if (machine__init(machine, root_dir, pid) != 0) {
305 pos = rb_entry(parent, struct machine, rb_node);
314 rb_link_node(&machine->rb_node, parent, p);
315 rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
317 machine->machines = machines;
322 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
326 machines->host.comm_exec = comm_exec;
328 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
329 struct machine *machine = rb_entry(nd, struct machine, rb_node);
331 machine->comm_exec = comm_exec;
335 struct machine *machines__find(struct machines *machines, pid_t pid)
337 struct rb_node **p = &machines->guests.rb_root.rb_node;
338 struct rb_node *parent = NULL;
339 struct machine *machine;
340 struct machine *default_machine = NULL;
342 if (pid == HOST_KERNEL_ID)
343 return &machines->host;
347 machine = rb_entry(parent, struct machine, rb_node);
348 if (pid < machine->pid)
350 else if (pid > machine->pid)
355 default_machine = machine;
358 return default_machine;
361 struct machine *machines__findnew(struct machines *machines, pid_t pid)
364 const char *root_dir = "";
365 struct machine *machine = machines__find(machines, pid);
367 if (machine && (machine->pid == pid))
370 if ((pid != HOST_KERNEL_ID) &&
371 (pid != DEFAULT_GUEST_KERNEL_ID) &&
372 (symbol_conf.guestmount)) {
373 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
374 if (access(path, R_OK)) {
375 static struct strlist *seen;
378 seen = strlist__new(NULL, NULL);
380 if (!strlist__has_entry(seen, path)) {
381 pr_err("Can't access file %s\n", path);
382 strlist__add(seen, path);
390 machine = machines__add(machines, pid, root_dir);
395 struct machine *machines__find_guest(struct machines *machines, pid_t pid)
397 struct machine *machine = machines__find(machines, pid);
400 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID);
405 * A common case for KVM test programs is that the test program acts as the
406 * hypervisor, creating, running and destroying the virtual machine, and
407 * providing the guest object code from its own object code. In this case,
408 * the VM is not running an OS, but only the functions loaded into it by the
409 * hypervisor test program, and conveniently, loaded at the same virtual
412 * Normally to resolve addresses, MMAP events are needed to map addresses
413 * back to the object code and debug symbols for that object code.
415 * Currently, there is no way to get such mapping information from guests
416 * but, in the scenario described above, the guest has the same mappings
417 * as the hypervisor, so support for that scenario can be achieved.
419 * To support that, copy the host thread's maps to the guest thread's maps.
420 * Note, we do not discover the guest until we encounter a guest event,
421 * which works well because it is not until then that we know that the host
422 * thread's maps have been set up.
424 * This function returns the guest thread. Apart from keeping the data
425 * structures sane, using a thread belonging to the guest machine, instead
426 * of the host thread, allows it to have its own comm (refer
427 * thread__set_guest_comm()).
429 static struct thread *findnew_guest_code(struct machine *machine,
430 struct machine *host_machine,
433 struct thread *host_thread;
434 struct thread *thread;
440 thread = machine__findnew_thread(machine, -1, pid);
444 /* Assume maps are set up if there are any */
445 if (maps__nr_maps(thread__maps(thread)))
448 host_thread = machine__find_thread(host_machine, -1, pid);
452 thread__set_guest_comm(thread, pid);
455 * Guest code can be found in hypervisor process at the same address
458 err = maps__clone(thread, thread__maps(host_thread));
459 thread__put(host_thread);
466 thread__zput(thread);
470 struct thread *machines__findnew_guest_code(struct machines *machines, pid_t pid)
472 struct machine *host_machine = machines__find(machines, HOST_KERNEL_ID);
473 struct machine *machine = machines__findnew(machines, pid);
475 return findnew_guest_code(machine, host_machine, pid);
478 struct thread *machine__findnew_guest_code(struct machine *machine, pid_t pid)
480 struct machines *machines = machine->machines;
481 struct machine *host_machine;
486 host_machine = machines__find(machines, HOST_KERNEL_ID);
488 return findnew_guest_code(machine, host_machine, pid);
491 void machines__process_guests(struct machines *machines,
492 machine__process_t process, void *data)
496 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
497 struct machine *pos = rb_entry(nd, struct machine, rb_node);
502 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
504 struct rb_node *node;
505 struct machine *machine;
507 machines->host.id_hdr_size = id_hdr_size;
509 for (node = rb_first_cached(&machines->guests); node;
510 node = rb_next(node)) {
511 machine = rb_entry(node, struct machine, rb_node);
512 machine->id_hdr_size = id_hdr_size;
518 static void machine__update_thread_pid(struct machine *machine,
519 struct thread *th, pid_t pid)
521 struct thread *leader;
523 if (pid == thread__pid(th) || pid == -1 || thread__pid(th) != -1)
526 thread__set_pid(th, pid);
528 if (thread__pid(th) == thread__tid(th))
531 leader = __machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
535 if (!thread__maps(leader))
536 thread__set_maps(leader, maps__new(machine));
538 if (!thread__maps(leader))
541 if (thread__maps(th) == thread__maps(leader))
544 if (thread__maps(th)) {
546 * Maps are created from MMAP events which provide the pid and
547 * tid. Consequently there never should be any maps on a thread
548 * with an unknown pid. Just print an error if there are.
550 if (!maps__empty(thread__maps(th)))
551 pr_err("Discarding thread maps for %d:%d\n",
552 thread__pid(th), thread__tid(th));
553 maps__put(thread__maps(th));
556 thread__set_maps(th, maps__get(thread__maps(leader)));
561 pr_err("Failed to join map groups for %d:%d\n", thread__pid(th), thread__tid(th));
566 * Front-end cache - TID lookups come in blocks,
567 * so most of the time we dont have to look up
570 static struct thread*
571 __threads__get_last_match(struct threads *threads, struct machine *machine,
576 th = threads->last_match;
578 if (thread__tid(th) == tid) {
579 machine__update_thread_pid(machine, th, pid);
580 return thread__get(th);
582 thread__put(threads->last_match);
583 threads->last_match = NULL;
589 static struct thread*
590 threads__get_last_match(struct threads *threads, struct machine *machine,
593 struct thread *th = NULL;
595 if (perf_singlethreaded)
596 th = __threads__get_last_match(threads, machine, pid, tid);
602 __threads__set_last_match(struct threads *threads, struct thread *th)
604 thread__put(threads->last_match);
605 threads->last_match = thread__get(th);
609 threads__set_last_match(struct threads *threads, struct thread *th)
611 if (perf_singlethreaded)
612 __threads__set_last_match(threads, th);
616 * Caller must eventually drop thread->refcnt returned with a successful
617 * lookup/new thread inserted.
619 static struct thread *____machine__findnew_thread(struct machine *machine,
620 struct threads *threads,
621 pid_t pid, pid_t tid,
624 struct rb_node **p = &threads->entries.rb_root.rb_node;
625 struct rb_node *parent = NULL;
627 struct thread_rb_node *nd;
628 bool leftmost = true;
630 th = threads__get_last_match(threads, machine, pid, tid);
636 th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
638 if (thread__tid(th) == tid) {
639 threads__set_last_match(threads, th);
640 machine__update_thread_pid(machine, th, pid);
641 return thread__get(th);
644 if (tid < thread__tid(th))
655 th = thread__new(pid, tid);
659 nd = malloc(sizeof(*nd));
666 rb_link_node(&nd->rb_node, parent, p);
667 rb_insert_color_cached(&nd->rb_node, &threads->entries, leftmost);
669 * We have to initialize maps separately after rb tree is updated.
671 * The reason is that we call machine__findnew_thread within
672 * thread__init_maps to find the thread leader and that would screwed
675 if (thread__init_maps(th, machine)) {
676 pr_err("Thread init failed thread %d\n", pid);
677 rb_erase_cached(&nd->rb_node, &threads->entries);
678 RB_CLEAR_NODE(&nd->rb_node);
684 * It is now in the rbtree, get a ref
686 threads__set_last_match(threads, th);
689 return thread__get(th);
692 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
694 return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
697 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
700 struct threads *threads = machine__threads(machine, tid);
703 down_write(&threads->lock);
704 th = __machine__findnew_thread(machine, pid, tid);
705 up_write(&threads->lock);
709 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
712 struct threads *threads = machine__threads(machine, tid);
715 down_read(&threads->lock);
716 th = ____machine__findnew_thread(machine, threads, pid, tid, false);
717 up_read(&threads->lock);
722 * Threads are identified by pid and tid, and the idle task has pid == tid == 0.
723 * So here a single thread is created for that, but actually there is a separate
724 * idle task per cpu, so there should be one 'struct thread' per cpu, but there
725 * is only 1. That causes problems for some tools, requiring workarounds. For
726 * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu().
728 struct thread *machine__idle_thread(struct machine *machine)
730 struct thread *thread = machine__findnew_thread(machine, 0, 0);
732 if (!thread || thread__set_comm(thread, "swapper", 0) ||
733 thread__set_namespaces(thread, 0, NULL))
734 pr_err("problem inserting idle task for machine pid %d\n", machine->pid);
739 struct comm *machine__thread_exec_comm(struct machine *machine,
740 struct thread *thread)
742 if (machine->comm_exec)
743 return thread__exec_comm(thread);
745 return thread__comm(thread);
748 int machine__process_comm_event(struct machine *machine, union perf_event *event,
749 struct perf_sample *sample)
751 struct thread *thread = machine__findnew_thread(machine,
754 bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
758 machine->comm_exec = true;
761 perf_event__fprintf_comm(event, stdout);
763 if (thread == NULL ||
764 __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
765 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
774 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
775 union perf_event *event,
776 struct perf_sample *sample __maybe_unused)
778 struct thread *thread = machine__findnew_thread(machine,
779 event->namespaces.pid,
780 event->namespaces.tid);
783 WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
784 "\nWARNING: kernel seems to support more namespaces than perf"
785 " tool.\nTry updating the perf tool..\n\n");
787 WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
788 "\nWARNING: perf tool seems to support more namespaces than"
789 " the kernel.\nTry updating the kernel..\n\n");
792 perf_event__fprintf_namespaces(event, stdout);
794 if (thread == NULL ||
795 thread__set_namespaces(thread, sample->time, &event->namespaces)) {
796 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
805 int machine__process_cgroup_event(struct machine *machine,
806 union perf_event *event,
807 struct perf_sample *sample __maybe_unused)
812 perf_event__fprintf_cgroup(event, stdout);
814 cgrp = cgroup__findnew(machine->env, event->cgroup.id, event->cgroup.path);
821 int machine__process_lost_event(struct machine *machine __maybe_unused,
822 union perf_event *event, struct perf_sample *sample __maybe_unused)
824 dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
825 event->lost.id, event->lost.lost);
829 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
830 union perf_event *event, struct perf_sample *sample)
832 dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
833 sample->id, event->lost_samples.lost);
837 static struct dso *machine__findnew_module_dso(struct machine *machine,
839 const char *filename)
843 down_write(&machine->dsos.lock);
845 dso = __dsos__find(&machine->dsos, m->name, true);
847 dso = __dsos__addnew(&machine->dsos, m->name);
851 dso__set_module_info(dso, m, machine);
852 dso__set_long_name(dso, strdup(filename), true);
853 dso->kernel = DSO_SPACE__KERNEL;
858 up_write(&machine->dsos.lock);
862 int machine__process_aux_event(struct machine *machine __maybe_unused,
863 union perf_event *event)
866 perf_event__fprintf_aux(event, stdout);
870 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
871 union perf_event *event)
874 perf_event__fprintf_itrace_start(event, stdout);
878 int machine__process_aux_output_hw_id_event(struct machine *machine __maybe_unused,
879 union perf_event *event)
882 perf_event__fprintf_aux_output_hw_id(event, stdout);
886 int machine__process_switch_event(struct machine *machine __maybe_unused,
887 union perf_event *event)
890 perf_event__fprintf_switch(event, stdout);
894 static int machine__process_ksymbol_register(struct machine *machine,
895 union perf_event *event,
896 struct perf_sample *sample __maybe_unused)
900 struct map *map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
901 bool put_map = false;
905 dso = dso__new(event->ksymbol.name);
911 dso->kernel = DSO_SPACE__KERNEL;
912 map = map__new2(0, dso);
919 * The inserted map has a get on it, we need to put to release
920 * the reference count here, but do it after all accesses are
924 if (event->ksymbol.ksym_type == PERF_RECORD_KSYMBOL_TYPE_OOL) {
925 dso->binary_type = DSO_BINARY_TYPE__OOL;
926 dso->data.file_size = event->ksymbol.len;
927 dso__set_loaded(dso);
930 map__set_start(map, event->ksymbol.addr);
931 map__set_end(map, map__start(map) + event->ksymbol.len);
932 err = maps__insert(machine__kernel_maps(machine), map);
938 dso__set_loaded(dso);
940 if (is_bpf_image(event->ksymbol.name)) {
941 dso->binary_type = DSO_BINARY_TYPE__BPF_IMAGE;
942 dso__set_long_name(dso, "", false);
948 sym = symbol__new(map__map_ip(map, map__start(map)),
950 0, 0, event->ksymbol.name);
955 dso__insert_symbol(dso, sym);
962 static int machine__process_ksymbol_unregister(struct machine *machine,
963 union perf_event *event,
964 struct perf_sample *sample __maybe_unused)
969 map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
973 if (RC_CHK_ACCESS(map) != RC_CHK_ACCESS(machine->vmlinux_map))
974 maps__remove(machine__kernel_maps(machine), map);
976 struct dso *dso = map__dso(map);
978 sym = dso__find_symbol(dso, map__map_ip(map, map__start(map)));
980 dso__delete_symbol(dso, sym);
986 int machine__process_ksymbol(struct machine *machine __maybe_unused,
987 union perf_event *event,
988 struct perf_sample *sample)
991 perf_event__fprintf_ksymbol(event, stdout);
993 if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
994 return machine__process_ksymbol_unregister(machine, event,
996 return machine__process_ksymbol_register(machine, event, sample);
999 int machine__process_text_poke(struct machine *machine, union perf_event *event,
1000 struct perf_sample *sample __maybe_unused)
1002 struct map *map = maps__find(machine__kernel_maps(machine), event->text_poke.addr);
1003 u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1004 struct dso *dso = map ? map__dso(map) : NULL;
1007 perf_event__fprintf_text_poke(event, machine, stdout);
1009 if (!event->text_poke.new_len)
1012 if (cpumode != PERF_RECORD_MISC_KERNEL) {
1013 pr_debug("%s: unsupported cpumode - ignoring\n", __func__);
1018 u8 *new_bytes = event->text_poke.bytes + event->text_poke.old_len;
1022 * Kernel maps might be changed when loading symbols so loading
1023 * must be done prior to using kernel maps.
1026 ret = dso__data_write_cache_addr(dso, map, machine,
1027 event->text_poke.addr,
1029 event->text_poke.new_len);
1030 if (ret != event->text_poke.new_len)
1031 pr_debug("Failed to write kernel text poke at %#" PRI_lx64 "\n",
1032 event->text_poke.addr);
1034 pr_debug("Failed to find kernel text poke address map for %#" PRI_lx64 "\n",
1035 event->text_poke.addr);
1041 static struct map *machine__addnew_module_map(struct machine *machine, u64 start,
1042 const char *filename)
1044 struct map *map = NULL;
1049 if (kmod_path__parse_name(&m, filename))
1052 dso = machine__findnew_module_dso(machine, &m, filename);
1056 map = map__new2(start, dso);
1060 err = maps__insert(machine__kernel_maps(machine), map);
1061 /* If maps__insert failed, return NULL. */
1067 /* put the dso here, corresponding to machine__findnew_module_dso */
1073 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
1076 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
1078 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
1079 struct machine *pos = rb_entry(nd, struct machine, rb_node);
1080 ret += __dsos__fprintf(&pos->dsos.head, fp);
1086 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
1087 bool (skip)(struct dso *dso, int parm), int parm)
1089 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
1092 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
1093 bool (skip)(struct dso *dso, int parm), int parm)
1096 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
1098 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
1099 struct machine *pos = rb_entry(nd, struct machine, rb_node);
1100 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
1105 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
1109 struct dso *kdso = machine__kernel_dso(machine);
1111 if (kdso->has_build_id) {
1112 char filename[PATH_MAX];
1113 if (dso__build_id_filename(kdso, filename, sizeof(filename),
1115 printed += fprintf(fp, "[0] %s\n", filename);
1118 for (i = 0; i < vmlinux_path__nr_entries; ++i)
1119 printed += fprintf(fp, "[%d] %s\n",
1120 i + kdso->has_build_id, vmlinux_path[i]);
1125 size_t machine__fprintf(struct machine *machine, FILE *fp)
1131 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
1132 struct threads *threads = &machine->threads[i];
1134 down_read(&threads->lock);
1136 ret = fprintf(fp, "Threads: %u\n", threads->nr);
1138 for (nd = rb_first_cached(&threads->entries); nd;
1140 struct thread *pos = rb_entry(nd, struct thread_rb_node, rb_node)->thread;
1142 ret += thread__fprintf(pos, fp);
1145 up_read(&threads->lock);
1150 static struct dso *machine__get_kernel(struct machine *machine)
1152 const char *vmlinux_name = machine->mmap_name;
1155 if (machine__is_host(machine)) {
1156 if (symbol_conf.vmlinux_name)
1157 vmlinux_name = symbol_conf.vmlinux_name;
1159 kernel = machine__findnew_kernel(machine, vmlinux_name,
1160 "[kernel]", DSO_SPACE__KERNEL);
1162 if (symbol_conf.default_guest_vmlinux_name)
1163 vmlinux_name = symbol_conf.default_guest_vmlinux_name;
1165 kernel = machine__findnew_kernel(machine, vmlinux_name,
1167 DSO_SPACE__KERNEL_GUEST);
1170 if (kernel != NULL && (!kernel->has_build_id))
1171 dso__read_running_kernel_build_id(kernel, machine);
1176 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
1179 if (machine__is_default_guest(machine))
1180 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
1182 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
1185 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
1187 /* Figure out the start address of kernel map from /proc/kallsyms.
1188 * Returns the name of the start symbol in *symbol_name. Pass in NULL as
1189 * symbol_name if it's not that important.
1191 static int machine__get_running_kernel_start(struct machine *machine,
1192 const char **symbol_name,
1193 u64 *start, u64 *end)
1195 char filename[PATH_MAX];
1200 machine__get_kallsyms_filename(machine, filename, PATH_MAX);
1202 if (symbol__restricted_filename(filename, "/proc/kallsyms"))
1205 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
1206 err = kallsyms__get_function_start(filename, name, &addr);
1215 *symbol_name = name;
1219 err = kallsyms__get_symbol_start(filename, "_edata", &addr);
1221 err = kallsyms__get_function_start(filename, "_etext", &addr);
1228 int machine__create_extra_kernel_map(struct machine *machine,
1230 struct extra_kernel_map *xm)
1236 map = map__new2(xm->start, kernel);
1240 map__set_end(map, xm->end);
1241 map__set_pgoff(map, xm->pgoff);
1243 kmap = map__kmap(map);
1245 strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1247 err = maps__insert(machine__kernel_maps(machine), map);
1250 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1251 kmap->name, map__start(map), map__end(map));
1259 static u64 find_entry_trampoline(struct dso *dso)
1261 /* Duplicates are removed so lookup all aliases */
1262 const char *syms[] = {
1263 "_entry_trampoline",
1264 "__entry_trampoline_start",
1265 "entry_SYSCALL_64_trampoline",
1267 struct symbol *sym = dso__first_symbol(dso);
1270 for (; sym; sym = dso__next_symbol(sym)) {
1271 if (sym->binding != STB_GLOBAL)
1273 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1274 if (!strcmp(sym->name, syms[i]))
1283 * These values can be used for kernels that do not have symbols for the entry
1284 * trampolines in kallsyms.
1286 #define X86_64_CPU_ENTRY_AREA_PER_CPU 0xfffffe0000000000ULL
1287 #define X86_64_CPU_ENTRY_AREA_SIZE 0x2c000
1288 #define X86_64_ENTRY_TRAMPOLINE 0x6000
1290 /* Map x86_64 PTI entry trampolines */
1291 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1294 struct maps *kmaps = machine__kernel_maps(machine);
1295 int nr_cpus_avail, cpu;
1297 struct map_rb_node *rb_node;
1301 * In the vmlinux case, pgoff is a virtual address which must now be
1302 * mapped to a vmlinux offset.
1304 maps__for_each_entry(kmaps, rb_node) {
1305 struct map *dest_map, *map = rb_node->map;
1306 struct kmap *kmap = __map__kmap(map);
1308 if (!kmap || !is_entry_trampoline(kmap->name))
1311 dest_map = maps__find(kmaps, map__pgoff(map));
1312 if (dest_map != map)
1313 map__set_pgoff(map, map__map_ip(dest_map, map__pgoff(map)));
1316 if (found || machine->trampolines_mapped)
1319 pgoff = find_entry_trampoline(kernel);
1323 nr_cpus_avail = machine__nr_cpus_avail(machine);
1325 /* Add a 1 page map for each CPU's entry trampoline */
1326 for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1327 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1328 cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1329 X86_64_ENTRY_TRAMPOLINE;
1330 struct extra_kernel_map xm = {
1332 .end = va + page_size,
1336 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1338 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1342 machine->trampolines_mapped = nr_cpus_avail;
1347 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1348 struct dso *kernel __maybe_unused)
1354 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1356 /* In case of renewal the kernel map, destroy previous one */
1357 machine__destroy_kernel_maps(machine);
1359 map__put(machine->vmlinux_map);
1360 machine->vmlinux_map = map__new2(0, kernel);
1361 if (machine->vmlinux_map == NULL)
1364 map__set_map_ip(machine->vmlinux_map, identity__map_ip);
1365 map__set_unmap_ip(machine->vmlinux_map, identity__map_ip);
1366 return maps__insert(machine__kernel_maps(machine), machine->vmlinux_map);
1369 void machine__destroy_kernel_maps(struct machine *machine)
1372 struct map *map = machine__kernel_map(machine);
1377 kmap = map__kmap(map);
1378 maps__remove(machine__kernel_maps(machine), map);
1379 if (kmap && kmap->ref_reloc_sym) {
1380 zfree((char **)&kmap->ref_reloc_sym->name);
1381 zfree(&kmap->ref_reloc_sym);
1384 map__zput(machine->vmlinux_map);
1387 int machines__create_guest_kernel_maps(struct machines *machines)
1390 struct dirent **namelist = NULL;
1392 char path[PATH_MAX];
1396 if (symbol_conf.default_guest_vmlinux_name ||
1397 symbol_conf.default_guest_modules ||
1398 symbol_conf.default_guest_kallsyms) {
1399 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1402 if (symbol_conf.guestmount) {
1403 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1406 for (i = 0; i < items; i++) {
1407 if (!isdigit(namelist[i]->d_name[0])) {
1408 /* Filter out . and .. */
1411 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1412 if ((*endp != '\0') ||
1413 (endp == namelist[i]->d_name) ||
1414 (errno == ERANGE)) {
1415 pr_debug("invalid directory (%s). Skipping.\n",
1416 namelist[i]->d_name);
1419 sprintf(path, "%s/%s/proc/kallsyms",
1420 symbol_conf.guestmount,
1421 namelist[i]->d_name);
1422 ret = access(path, R_OK);
1424 pr_debug("Can't access file %s\n", path);
1427 machines__create_kernel_maps(machines, pid);
1436 void machines__destroy_kernel_maps(struct machines *machines)
1438 struct rb_node *next = rb_first_cached(&machines->guests);
1440 machine__destroy_kernel_maps(&machines->host);
1443 struct machine *pos = rb_entry(next, struct machine, rb_node);
1445 next = rb_next(&pos->rb_node);
1446 rb_erase_cached(&pos->rb_node, &machines->guests);
1447 machine__delete(pos);
1451 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1453 struct machine *machine = machines__findnew(machines, pid);
1455 if (machine == NULL)
1458 return machine__create_kernel_maps(machine);
1461 int machine__load_kallsyms(struct machine *machine, const char *filename)
1463 struct map *map = machine__kernel_map(machine);
1464 struct dso *dso = map__dso(map);
1465 int ret = __dso__load_kallsyms(dso, filename, map, true);
1468 dso__set_loaded(dso);
1470 * Since /proc/kallsyms will have multiple sessions for the
1471 * kernel, with modules between them, fixup the end of all
1474 maps__fixup_end(machine__kernel_maps(machine));
1480 int machine__load_vmlinux_path(struct machine *machine)
1482 struct map *map = machine__kernel_map(machine);
1483 struct dso *dso = map__dso(map);
1484 int ret = dso__load_vmlinux_path(dso, map);
1487 dso__set_loaded(dso);
1492 static char *get_kernel_version(const char *root_dir)
1494 char version[PATH_MAX];
1497 const char *prefix = "Linux version ";
1499 sprintf(version, "%s/proc/version", root_dir);
1500 file = fopen(version, "r");
1504 tmp = fgets(version, sizeof(version), file);
1509 name = strstr(version, prefix);
1512 name += strlen(prefix);
1513 tmp = strchr(name, ' ');
1517 return strdup(name);
1520 static bool is_kmod_dso(struct dso *dso)
1522 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1523 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1526 static int maps__set_module_path(struct maps *maps, const char *path, struct kmod_path *m)
1530 struct map *map = maps__find_by_name(maps, m->name);
1535 long_name = strdup(path);
1536 if (long_name == NULL)
1539 dso = map__dso(map);
1540 dso__set_long_name(dso, long_name, true);
1541 dso__kernel_module_get_build_id(dso, "");
1544 * Full name could reveal us kmod compression, so
1545 * we need to update the symtab_type if needed.
1547 if (m->comp && is_kmod_dso(dso)) {
1549 dso->comp = m->comp;
1555 static int maps__set_modules_path_dir(struct maps *maps, const char *dir_name, int depth)
1557 struct dirent *dent;
1558 DIR *dir = opendir(dir_name);
1562 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1566 while ((dent = readdir(dir)) != NULL) {
1567 char path[PATH_MAX];
1570 /*sshfs might return bad dent->d_type, so we have to stat*/
1571 path__join(path, sizeof(path), dir_name, dent->d_name);
1572 if (stat(path, &st))
1575 if (S_ISDIR(st.st_mode)) {
1576 if (!strcmp(dent->d_name, ".") ||
1577 !strcmp(dent->d_name, ".."))
1580 /* Do not follow top-level source and build symlinks */
1582 if (!strcmp(dent->d_name, "source") ||
1583 !strcmp(dent->d_name, "build"))
1587 ret = maps__set_modules_path_dir(maps, path, depth + 1);
1593 ret = kmod_path__parse_name(&m, dent->d_name);
1598 ret = maps__set_module_path(maps, path, &m);
1612 static int machine__set_modules_path(struct machine *machine)
1615 char modules_path[PATH_MAX];
1617 version = get_kernel_version(machine->root_dir);
1621 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1622 machine->root_dir, version);
1625 return maps__set_modules_path_dir(machine__kernel_maps(machine), modules_path, 0);
1627 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1628 u64 *size __maybe_unused,
1629 const char *name __maybe_unused)
1634 static int machine__create_module(void *arg, const char *name, u64 start,
1637 struct machine *machine = arg;
1640 if (arch__fix_module_text_start(&start, &size, name) < 0)
1643 map = machine__addnew_module_map(machine, start, name);
1646 map__set_end(map, start + size);
1648 dso__kernel_module_get_build_id(map__dso(map), machine->root_dir);
1653 static int machine__create_modules(struct machine *machine)
1655 const char *modules;
1656 char path[PATH_MAX];
1658 if (machine__is_default_guest(machine)) {
1659 modules = symbol_conf.default_guest_modules;
1661 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1665 if (symbol__restricted_filename(modules, "/proc/modules"))
1668 if (modules__parse(modules, machine, machine__create_module))
1671 if (!machine__set_modules_path(machine))
1674 pr_debug("Problems setting modules path maps, continuing anyway...\n");
1679 static void machine__set_kernel_mmap(struct machine *machine,
1682 map__set_start(machine->vmlinux_map, start);
1683 map__set_end(machine->vmlinux_map, end);
1685 * Be a bit paranoid here, some perf.data file came with
1686 * a zero sized synthesized MMAP event for the kernel.
1688 if (start == 0 && end == 0)
1689 map__set_end(machine->vmlinux_map, ~0ULL);
1692 static int machine__update_kernel_mmap(struct machine *machine,
1695 struct map *orig, *updated;
1698 orig = machine->vmlinux_map;
1699 updated = map__get(orig);
1701 machine->vmlinux_map = updated;
1702 machine__set_kernel_mmap(machine, start, end);
1703 maps__remove(machine__kernel_maps(machine), orig);
1704 err = maps__insert(machine__kernel_maps(machine), updated);
1710 int machine__create_kernel_maps(struct machine *machine)
1712 struct dso *kernel = machine__get_kernel(machine);
1713 const char *name = NULL;
1714 u64 start = 0, end = ~0ULL;
1720 ret = __machine__create_kernel_maps(machine, kernel);
1724 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1725 if (machine__is_host(machine))
1726 pr_debug("Problems creating module maps, "
1727 "continuing anyway...\n");
1729 pr_debug("Problems creating module maps for guest %d, "
1730 "continuing anyway...\n", machine->pid);
1733 if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1735 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1736 machine__destroy_kernel_maps(machine);
1742 * we have a real start address now, so re-order the kmaps
1743 * assume it's the last in the kmaps
1745 ret = machine__update_kernel_mmap(machine, start, end);
1750 if (machine__create_extra_kernel_maps(machine, kernel))
1751 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1754 /* update end address of the kernel map using adjacent module address */
1755 struct map_rb_node *rb_node = maps__find_node(machine__kernel_maps(machine),
1756 machine__kernel_map(machine));
1757 struct map_rb_node *next = map_rb_node__next(rb_node);
1760 machine__set_kernel_mmap(machine, start, map__start(next->map));
1768 static bool machine__uses_kcore(struct machine *machine)
1772 list_for_each_entry(dso, &machine->dsos.head, node) {
1773 if (dso__is_kcore(dso))
1780 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1781 struct extra_kernel_map *xm)
1783 return machine__is(machine, "x86_64") &&
1784 is_entry_trampoline(xm->name);
1787 static int machine__process_extra_kernel_map(struct machine *machine,
1788 struct extra_kernel_map *xm)
1790 struct dso *kernel = machine__kernel_dso(machine);
1795 return machine__create_extra_kernel_map(machine, kernel, xm);
1798 static int machine__process_kernel_mmap_event(struct machine *machine,
1799 struct extra_kernel_map *xm,
1800 struct build_id *bid)
1802 enum dso_space_type dso_space;
1803 bool is_kernel_mmap;
1804 const char *mmap_name = machine->mmap_name;
1806 /* If we have maps from kcore then we do not need or want any others */
1807 if (machine__uses_kcore(machine))
1810 if (machine__is_host(machine))
1811 dso_space = DSO_SPACE__KERNEL;
1813 dso_space = DSO_SPACE__KERNEL_GUEST;
1815 is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1816 if (!is_kernel_mmap && !machine__is_host(machine)) {
1818 * If the event was recorded inside the guest and injected into
1819 * the host perf.data file, then it will match a host mmap_name,
1820 * so try that - see machine__set_mmap_name().
1822 mmap_name = "[kernel.kallsyms]";
1823 is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1825 if (xm->name[0] == '/' ||
1826 (!is_kernel_mmap && xm->name[0] == '[')) {
1827 struct map *map = machine__addnew_module_map(machine, xm->start, xm->name);
1832 map__set_end(map, map__start(map) + xm->end - xm->start);
1834 if (build_id__is_defined(bid))
1835 dso__set_build_id(map__dso(map), bid);
1838 } else if (is_kernel_mmap) {
1839 const char *symbol_name = xm->name + strlen(mmap_name);
1841 * Should be there already, from the build-id table in
1844 struct dso *kernel = NULL;
1847 down_read(&machine->dsos.lock);
1849 list_for_each_entry(dso, &machine->dsos.head, node) {
1852 * The cpumode passed to is_kernel_module is not the
1853 * cpumode of *this* event. If we insist on passing
1854 * correct cpumode to is_kernel_module, we should
1855 * record the cpumode when we adding this dso to the
1858 * However we don't really need passing correct
1859 * cpumode. We know the correct cpumode must be kernel
1860 * mode (if not, we should not link it onto kernel_dsos
1863 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1864 * is_kernel_module() treats it as a kernel cpumode.
1868 is_kernel_module(dso->long_name,
1869 PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1873 kernel = dso__get(dso);
1877 up_read(&machine->dsos.lock);
1880 kernel = machine__findnew_dso(machine, machine->mmap_name);
1884 kernel->kernel = dso_space;
1885 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1890 if (strstr(kernel->long_name, "vmlinux"))
1891 dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1893 if (machine__update_kernel_mmap(machine, xm->start, xm->end) < 0) {
1898 if (build_id__is_defined(bid))
1899 dso__set_build_id(kernel, bid);
1902 * Avoid using a zero address (kptr_restrict) for the ref reloc
1903 * symbol. Effectively having zero here means that at record
1904 * time /proc/sys/kernel/kptr_restrict was non zero.
1906 if (xm->pgoff != 0) {
1907 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1912 if (machine__is_default_guest(machine)) {
1914 * preload dso of guest kernel and modules
1916 dso__load(kernel, machine__kernel_map(machine));
1919 } else if (perf_event__is_extra_kernel_mmap(machine, xm)) {
1920 return machine__process_extra_kernel_map(machine, xm);
1927 int machine__process_mmap2_event(struct machine *machine,
1928 union perf_event *event,
1929 struct perf_sample *sample)
1931 struct thread *thread;
1933 struct dso_id dso_id = {
1934 .maj = event->mmap2.maj,
1935 .min = event->mmap2.min,
1936 .ino = event->mmap2.ino,
1937 .ino_generation = event->mmap2.ino_generation,
1939 struct build_id __bid, *bid = NULL;
1943 perf_event__fprintf_mmap2(event, stdout);
1945 if (event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID) {
1947 build_id__init(bid, event->mmap2.build_id, event->mmap2.build_id_size);
1950 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1951 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1952 struct extra_kernel_map xm = {
1953 .start = event->mmap2.start,
1954 .end = event->mmap2.start + event->mmap2.len,
1955 .pgoff = event->mmap2.pgoff,
1958 strlcpy(xm.name, event->mmap2.filename, KMAP_NAME_LEN);
1959 ret = machine__process_kernel_mmap_event(machine, &xm, bid);
1965 thread = machine__findnew_thread(machine, event->mmap2.pid,
1970 map = map__new(machine, event->mmap2.start,
1971 event->mmap2.len, event->mmap2.pgoff,
1972 &dso_id, event->mmap2.prot,
1973 event->mmap2.flags, bid,
1974 event->mmap2.filename, thread);
1977 goto out_problem_map;
1979 ret = thread__insert_map(thread, map);
1981 goto out_problem_insert;
1983 thread__put(thread);
1990 thread__put(thread);
1992 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1996 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1997 struct perf_sample *sample)
1999 struct thread *thread;
2005 perf_event__fprintf_mmap(event, stdout);
2007 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
2008 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
2009 struct extra_kernel_map xm = {
2010 .start = event->mmap.start,
2011 .end = event->mmap.start + event->mmap.len,
2012 .pgoff = event->mmap.pgoff,
2015 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
2016 ret = machine__process_kernel_mmap_event(machine, &xm, NULL);
2022 thread = machine__findnew_thread(machine, event->mmap.pid,
2027 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
2030 map = map__new(machine, event->mmap.start,
2031 event->mmap.len, event->mmap.pgoff,
2032 NULL, prot, 0, NULL, event->mmap.filename, thread);
2035 goto out_problem_map;
2037 ret = thread__insert_map(thread, map);
2039 goto out_problem_insert;
2041 thread__put(thread);
2048 thread__put(thread);
2050 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
2054 static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
2055 struct thread *th, bool lock)
2057 struct threads *threads = machine__threads(machine, thread__tid(th));
2060 nd = thread_rb_node__find(th, &threads->entries.rb_root);
2062 if (threads->last_match && RC_CHK_ACCESS(threads->last_match) == RC_CHK_ACCESS(th))
2063 threads__set_last_match(threads, NULL);
2066 down_write(&threads->lock);
2068 BUG_ON(refcount_read(thread__refcnt(th)) == 0);
2070 thread__put(nd->thread);
2071 rb_erase_cached(&nd->rb_node, &threads->entries);
2072 RB_CLEAR_NODE(&nd->rb_node);
2078 up_write(&threads->lock);
2081 void machine__remove_thread(struct machine *machine, struct thread *th)
2083 return __machine__remove_thread(machine, NULL, th, true);
2086 int machine__process_fork_event(struct machine *machine, union perf_event *event,
2087 struct perf_sample *sample)
2089 struct thread *thread = machine__find_thread(machine,
2092 struct thread *parent = machine__findnew_thread(machine,
2095 bool do_maps_clone = true;
2099 perf_event__fprintf_task(event, stdout);
2102 * There may be an existing thread that is not actually the parent,
2103 * either because we are processing events out of order, or because the
2104 * (fork) event that would have removed the thread was lost. Assume the
2105 * latter case and continue on as best we can.
2107 if (thread__pid(parent) != (pid_t)event->fork.ppid) {
2108 dump_printf("removing erroneous parent thread %d/%d\n",
2109 thread__pid(parent), thread__tid(parent));
2110 machine__remove_thread(machine, parent);
2111 thread__put(parent);
2112 parent = machine__findnew_thread(machine, event->fork.ppid,
2116 /* if a thread currently exists for the thread id remove it */
2117 if (thread != NULL) {
2118 machine__remove_thread(machine, thread);
2119 thread__put(thread);
2122 thread = machine__findnew_thread(machine, event->fork.pid,
2125 * When synthesizing FORK events, we are trying to create thread
2126 * objects for the already running tasks on the machine.
2128 * Normally, for a kernel FORK event, we want to clone the parent's
2129 * maps because that is what the kernel just did.
2131 * But when synthesizing, this should not be done. If we do, we end up
2132 * with overlapping maps as we process the synthesized MMAP2 events that
2133 * get delivered shortly thereafter.
2135 * Use the FORK event misc flags in an internal way to signal this
2136 * situation, so we can elide the map clone when appropriate.
2138 if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
2139 do_maps_clone = false;
2141 if (thread == NULL || parent == NULL ||
2142 thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
2143 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
2146 thread__put(thread);
2147 thread__put(parent);
2152 int machine__process_exit_event(struct machine *machine, union perf_event *event,
2153 struct perf_sample *sample __maybe_unused)
2155 struct thread *thread = machine__find_thread(machine,
2160 perf_event__fprintf_task(event, stdout);
2163 thread__put(thread);
2168 int machine__process_event(struct machine *machine, union perf_event *event,
2169 struct perf_sample *sample)
2173 switch (event->header.type) {
2174 case PERF_RECORD_COMM:
2175 ret = machine__process_comm_event(machine, event, sample); break;
2176 case PERF_RECORD_MMAP:
2177 ret = machine__process_mmap_event(machine, event, sample); break;
2178 case PERF_RECORD_NAMESPACES:
2179 ret = machine__process_namespaces_event(machine, event, sample); break;
2180 case PERF_RECORD_CGROUP:
2181 ret = machine__process_cgroup_event(machine, event, sample); break;
2182 case PERF_RECORD_MMAP2:
2183 ret = machine__process_mmap2_event(machine, event, sample); break;
2184 case PERF_RECORD_FORK:
2185 ret = machine__process_fork_event(machine, event, sample); break;
2186 case PERF_RECORD_EXIT:
2187 ret = machine__process_exit_event(machine, event, sample); break;
2188 case PERF_RECORD_LOST:
2189 ret = machine__process_lost_event(machine, event, sample); break;
2190 case PERF_RECORD_AUX:
2191 ret = machine__process_aux_event(machine, event); break;
2192 case PERF_RECORD_ITRACE_START:
2193 ret = machine__process_itrace_start_event(machine, event); break;
2194 case PERF_RECORD_LOST_SAMPLES:
2195 ret = machine__process_lost_samples_event(machine, event, sample); break;
2196 case PERF_RECORD_SWITCH:
2197 case PERF_RECORD_SWITCH_CPU_WIDE:
2198 ret = machine__process_switch_event(machine, event); break;
2199 case PERF_RECORD_KSYMBOL:
2200 ret = machine__process_ksymbol(machine, event, sample); break;
2201 case PERF_RECORD_BPF_EVENT:
2202 ret = machine__process_bpf(machine, event, sample); break;
2203 case PERF_RECORD_TEXT_POKE:
2204 ret = machine__process_text_poke(machine, event, sample); break;
2205 case PERF_RECORD_AUX_OUTPUT_HW_ID:
2206 ret = machine__process_aux_output_hw_id_event(machine, event); break;
2215 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
2217 if (!regexec(regex, sym->name, 0, NULL, 0))
2222 static void ip__resolve_ams(struct thread *thread,
2223 struct addr_map_symbol *ams,
2226 struct addr_location al;
2228 addr_location__init(&al);
2230 * We cannot use the header.misc hint to determine whether a
2231 * branch stack address is user, kernel, guest, hypervisor.
2232 * Branches may straddle the kernel/user/hypervisor boundaries.
2233 * Thus, we have to try consecutively until we find a match
2234 * or else, the symbol is unknown
2236 thread__find_cpumode_addr_location(thread, ip, &al);
2239 ams->al_addr = al.addr;
2240 ams->al_level = al.level;
2241 ams->ms.maps = maps__get(al.maps);
2242 ams->ms.sym = al.sym;
2243 ams->ms.map = map__get(al.map);
2245 ams->data_page_size = 0;
2246 addr_location__exit(&al);
2249 static void ip__resolve_data(struct thread *thread,
2250 u8 m, struct addr_map_symbol *ams,
2251 u64 addr, u64 phys_addr, u64 daddr_page_size)
2253 struct addr_location al;
2255 addr_location__init(&al);
2257 thread__find_symbol(thread, m, addr, &al);
2260 ams->al_addr = al.addr;
2261 ams->al_level = al.level;
2262 ams->ms.maps = maps__get(al.maps);
2263 ams->ms.sym = al.sym;
2264 ams->ms.map = map__get(al.map);
2265 ams->phys_addr = phys_addr;
2266 ams->data_page_size = daddr_page_size;
2267 addr_location__exit(&al);
2270 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
2271 struct addr_location *al)
2273 struct mem_info *mi = mem_info__new();
2278 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
2279 ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
2280 sample->addr, sample->phys_addr,
2281 sample->data_page_size);
2282 mi->data_src.val = sample->data_src;
2287 static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2289 struct map *map = ms->map;
2290 char *srcline = NULL;
2293 if (!map || callchain_param.key == CCKEY_FUNCTION)
2296 dso = map__dso(map);
2297 srcline = srcline__tree_find(&dso->srclines, ip);
2299 bool show_sym = false;
2300 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2302 srcline = get_srcline(dso, map__rip_2objdump(map, ip),
2303 ms->sym, show_sym, show_addr, ip);
2304 srcline__tree_insert(&dso->srclines, ip, srcline);
2315 static int add_callchain_ip(struct thread *thread,
2316 struct callchain_cursor *cursor,
2317 struct symbol **parent,
2318 struct addr_location *root_al,
2322 struct branch_flags *flags,
2323 struct iterations *iter,
2326 struct map_symbol ms = {};
2327 struct addr_location al;
2328 int nr_loop_iter = 0, err = 0;
2329 u64 iter_cycles = 0;
2330 const char *srcline = NULL;
2332 addr_location__init(&al);
2337 thread__find_cpumode_addr_location(thread, ip, &al);
2339 if (ip >= PERF_CONTEXT_MAX) {
2341 case PERF_CONTEXT_HV:
2342 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2344 case PERF_CONTEXT_KERNEL:
2345 *cpumode = PERF_RECORD_MISC_KERNEL;
2347 case PERF_CONTEXT_USER:
2348 *cpumode = PERF_RECORD_MISC_USER;
2351 pr_debug("invalid callchain context: "
2352 "%"PRId64"\n", (s64) ip);
2354 * It seems the callchain is corrupted.
2357 callchain_cursor_reset(cursor);
2363 thread__find_symbol(thread, *cpumode, ip, &al);
2366 if (al.sym != NULL) {
2367 if (perf_hpp_list.parent && !*parent &&
2368 symbol__match_regex(al.sym, &parent_regex))
2370 else if (have_ignore_callees && root_al &&
2371 symbol__match_regex(al.sym, &ignore_callees_regex)) {
2372 /* Treat this symbol as the root,
2373 forgetting its callees. */
2374 addr_location__copy(root_al, &al);
2375 callchain_cursor_reset(cursor);
2379 if (symbol_conf.hide_unresolved && al.sym == NULL)
2383 nr_loop_iter = iter->nr_loop_iter;
2384 iter_cycles = iter->cycles;
2387 ms.maps = maps__get(al.maps);
2388 ms.map = map__get(al.map);
2391 if (!branch && append_inlines(cursor, &ms, ip) == 0)
2394 srcline = callchain_srcline(&ms, al.addr);
2395 err = callchain_cursor_append(cursor, ip, &ms,
2396 branch, flags, nr_loop_iter,
2397 iter_cycles, branch_from, srcline);
2399 addr_location__exit(&al);
2405 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2406 struct addr_location *al)
2409 const struct branch_stack *bs = sample->branch_stack;
2410 struct branch_entry *entries = perf_sample__branch_entries(sample);
2411 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2416 for (i = 0; i < bs->nr; i++) {
2417 ip__resolve_ams(al->thread, &bi[i].to, entries[i].to);
2418 ip__resolve_ams(al->thread, &bi[i].from, entries[i].from);
2419 bi[i].flags = entries[i].flags;
2424 static void save_iterations(struct iterations *iter,
2425 struct branch_entry *be, int nr)
2429 iter->nr_loop_iter++;
2432 for (i = 0; i < nr; i++)
2433 iter->cycles += be[i].flags.cycles;
2438 #define NO_ENTRY 0xff
2440 #define PERF_MAX_BRANCH_DEPTH 127
2443 static int remove_loops(struct branch_entry *l, int nr,
2444 struct iterations *iter)
2447 unsigned char chash[CHASHSZ];
2449 memset(chash, NO_ENTRY, sizeof(chash));
2451 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2453 for (i = 0; i < nr; i++) {
2454 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2456 /* no collision handling for now */
2457 if (chash[h] == NO_ENTRY) {
2459 } else if (l[chash[h]].from == l[i].from) {
2460 bool is_loop = true;
2461 /* check if it is a real loop */
2463 for (j = chash[h]; j < i && i + off < nr; j++, off++)
2464 if (l[j].from != l[i + off].from) {
2471 save_iterations(iter + i + off,
2474 memmove(iter + i, iter + i + off,
2477 memmove(l + i, l + i + off,
2488 static int lbr_callchain_add_kernel_ip(struct thread *thread,
2489 struct callchain_cursor *cursor,
2490 struct perf_sample *sample,
2491 struct symbol **parent,
2492 struct addr_location *root_al,
2494 bool callee, int end)
2496 struct ip_callchain *chain = sample->callchain;
2497 u8 cpumode = PERF_RECORD_MISC_USER;
2501 for (i = 0; i < end + 1; i++) {
2502 err = add_callchain_ip(thread, cursor, parent,
2503 root_al, &cpumode, chain->ips[i],
2504 false, NULL, NULL, branch_from);
2511 for (i = end; i >= 0; i--) {
2512 err = add_callchain_ip(thread, cursor, parent,
2513 root_al, &cpumode, chain->ips[i],
2514 false, NULL, NULL, branch_from);
2522 static void save_lbr_cursor_node(struct thread *thread,
2523 struct callchain_cursor *cursor,
2526 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2531 if (cursor->pos == cursor->nr) {
2532 lbr_stitch->prev_lbr_cursor[idx].valid = false;
2537 cursor->curr = cursor->first;
2539 cursor->curr = cursor->curr->next;
2540 memcpy(&lbr_stitch->prev_lbr_cursor[idx], cursor->curr,
2541 sizeof(struct callchain_cursor_node));
2543 lbr_stitch->prev_lbr_cursor[idx].valid = true;
2547 static int lbr_callchain_add_lbr_ip(struct thread *thread,
2548 struct callchain_cursor *cursor,
2549 struct perf_sample *sample,
2550 struct symbol **parent,
2551 struct addr_location *root_al,
2555 struct branch_stack *lbr_stack = sample->branch_stack;
2556 struct branch_entry *entries = perf_sample__branch_entries(sample);
2557 u8 cpumode = PERF_RECORD_MISC_USER;
2558 int lbr_nr = lbr_stack->nr;
2559 struct branch_flags *flags;
2564 * The curr and pos are not used in writing session. They are cleared
2565 * in callchain_cursor_commit() when the writing session is closed.
2566 * Using curr and pos to track the current cursor node.
2568 if (thread__lbr_stitch(thread)) {
2569 cursor->curr = NULL;
2570 cursor->pos = cursor->nr;
2572 cursor->curr = cursor->first;
2573 for (i = 0; i < (int)(cursor->nr - 1); i++)
2574 cursor->curr = cursor->curr->next;
2579 /* Add LBR ip from first entries.to */
2581 flags = &entries[0].flags;
2582 *branch_from = entries[0].from;
2583 err = add_callchain_ip(thread, cursor, parent,
2584 root_al, &cpumode, ip,
2591 * The number of cursor node increases.
2592 * Move the current cursor node.
2593 * But does not need to save current cursor node for entry 0.
2594 * It's impossible to stitch the whole LBRs of previous sample.
2596 if (thread__lbr_stitch(thread) && (cursor->pos != cursor->nr)) {
2598 cursor->curr = cursor->first;
2600 cursor->curr = cursor->curr->next;
2604 /* Add LBR ip from entries.from one by one. */
2605 for (i = 0; i < lbr_nr; i++) {
2606 ip = entries[i].from;
2607 flags = &entries[i].flags;
2608 err = add_callchain_ip(thread, cursor, parent,
2609 root_al, &cpumode, ip,
2614 save_lbr_cursor_node(thread, cursor, i);
2619 /* Add LBR ip from entries.from one by one. */
2620 for (i = lbr_nr - 1; i >= 0; i--) {
2621 ip = entries[i].from;
2622 flags = &entries[i].flags;
2623 err = add_callchain_ip(thread, cursor, parent,
2624 root_al, &cpumode, ip,
2629 save_lbr_cursor_node(thread, cursor, i);
2632 /* Add LBR ip from first entries.to */
2634 flags = &entries[0].flags;
2635 *branch_from = entries[0].from;
2636 err = add_callchain_ip(thread, cursor, parent,
2637 root_al, &cpumode, ip,
2646 static int lbr_callchain_add_stitched_lbr_ip(struct thread *thread,
2647 struct callchain_cursor *cursor)
2649 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2650 struct callchain_cursor_node *cnode;
2651 struct stitch_list *stitch_node;
2654 list_for_each_entry(stitch_node, &lbr_stitch->lists, node) {
2655 cnode = &stitch_node->cursor;
2657 err = callchain_cursor_append(cursor, cnode->ip,
2660 &cnode->branch_flags,
2661 cnode->nr_loop_iter,
2671 static struct stitch_list *get_stitch_node(struct thread *thread)
2673 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2674 struct stitch_list *stitch_node;
2676 if (!list_empty(&lbr_stitch->free_lists)) {
2677 stitch_node = list_first_entry(&lbr_stitch->free_lists,
2678 struct stitch_list, node);
2679 list_del(&stitch_node->node);
2684 return malloc(sizeof(struct stitch_list));
2687 static bool has_stitched_lbr(struct thread *thread,
2688 struct perf_sample *cur,
2689 struct perf_sample *prev,
2690 unsigned int max_lbr,
2693 struct branch_stack *cur_stack = cur->branch_stack;
2694 struct branch_entry *cur_entries = perf_sample__branch_entries(cur);
2695 struct branch_stack *prev_stack = prev->branch_stack;
2696 struct branch_entry *prev_entries = perf_sample__branch_entries(prev);
2697 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2698 int i, j, nr_identical_branches = 0;
2699 struct stitch_list *stitch_node;
2700 u64 cur_base, distance;
2702 if (!cur_stack || !prev_stack)
2705 /* Find the physical index of the base-of-stack for current sample. */
2706 cur_base = max_lbr - cur_stack->nr + cur_stack->hw_idx + 1;
2708 distance = (prev_stack->hw_idx > cur_base) ? (prev_stack->hw_idx - cur_base) :
2709 (max_lbr + prev_stack->hw_idx - cur_base);
2710 /* Previous sample has shorter stack. Nothing can be stitched. */
2711 if (distance + 1 > prev_stack->nr)
2715 * Check if there are identical LBRs between two samples.
2716 * Identical LBRs must have same from, to and flags values. Also,
2717 * they have to be saved in the same LBR registers (same physical
2720 * Starts from the base-of-stack of current sample.
2722 for (i = distance, j = cur_stack->nr - 1; (i >= 0) && (j >= 0); i--, j--) {
2723 if ((prev_entries[i].from != cur_entries[j].from) ||
2724 (prev_entries[i].to != cur_entries[j].to) ||
2725 (prev_entries[i].flags.value != cur_entries[j].flags.value))
2727 nr_identical_branches++;
2730 if (!nr_identical_branches)
2734 * Save the LBRs between the base-of-stack of previous sample
2735 * and the base-of-stack of current sample into lbr_stitch->lists.
2736 * These LBRs will be stitched later.
2738 for (i = prev_stack->nr - 1; i > (int)distance; i--) {
2740 if (!lbr_stitch->prev_lbr_cursor[i].valid)
2743 stitch_node = get_stitch_node(thread);
2747 memcpy(&stitch_node->cursor, &lbr_stitch->prev_lbr_cursor[i],
2748 sizeof(struct callchain_cursor_node));
2751 list_add(&stitch_node->node, &lbr_stitch->lists);
2753 list_add_tail(&stitch_node->node, &lbr_stitch->lists);
2759 static bool alloc_lbr_stitch(struct thread *thread, unsigned int max_lbr)
2761 if (thread__lbr_stitch(thread))
2764 thread__set_lbr_stitch(thread, zalloc(sizeof(struct lbr_stitch)));
2765 if (!thread__lbr_stitch(thread))
2768 thread__lbr_stitch(thread)->prev_lbr_cursor =
2769 calloc(max_lbr + 1, sizeof(struct callchain_cursor_node));
2770 if (!thread__lbr_stitch(thread)->prev_lbr_cursor)
2771 goto free_lbr_stitch;
2773 INIT_LIST_HEAD(&thread__lbr_stitch(thread)->lists);
2774 INIT_LIST_HEAD(&thread__lbr_stitch(thread)->free_lists);
2779 free(thread__lbr_stitch(thread));
2780 thread__set_lbr_stitch(thread, NULL);
2782 pr_warning("Failed to allocate space for stitched LBRs. Disable LBR stitch\n");
2783 thread__set_lbr_stitch_enable(thread, false);
2788 * Resolve LBR callstack chain sample
2790 * 1 on success get LBR callchain information
2791 * 0 no available LBR callchain information, should try fp
2792 * negative error code on other errors.
2794 static int resolve_lbr_callchain_sample(struct thread *thread,
2795 struct callchain_cursor *cursor,
2796 struct perf_sample *sample,
2797 struct symbol **parent,
2798 struct addr_location *root_al,
2800 unsigned int max_lbr)
2802 bool callee = (callchain_param.order == ORDER_CALLEE);
2803 struct ip_callchain *chain = sample->callchain;
2804 int chain_nr = min(max_stack, (int)chain->nr), i;
2805 struct lbr_stitch *lbr_stitch;
2806 bool stitched_lbr = false;
2807 u64 branch_from = 0;
2810 for (i = 0; i < chain_nr; i++) {
2811 if (chain->ips[i] == PERF_CONTEXT_USER)
2815 /* LBR only affects the user callchain */
2819 if (thread__lbr_stitch_enable(thread) && !sample->no_hw_idx &&
2820 (max_lbr > 0) && alloc_lbr_stitch(thread, max_lbr)) {
2821 lbr_stitch = thread__lbr_stitch(thread);
2823 stitched_lbr = has_stitched_lbr(thread, sample,
2824 &lbr_stitch->prev_sample,
2827 if (!stitched_lbr && !list_empty(&lbr_stitch->lists)) {
2828 list_replace_init(&lbr_stitch->lists,
2829 &lbr_stitch->free_lists);
2831 memcpy(&lbr_stitch->prev_sample, sample, sizeof(*sample));
2836 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2837 parent, root_al, branch_from,
2842 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2843 root_al, &branch_from, true);
2848 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2855 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2859 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2860 root_al, &branch_from, false);
2865 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2866 parent, root_al, branch_from,
2874 return (err < 0) ? err : 0;
2877 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2878 struct callchain_cursor *cursor,
2879 struct symbol **parent,
2880 struct addr_location *root_al,
2881 u8 *cpumode, int ent)
2885 while (--ent >= 0) {
2886 u64 ip = chain->ips[ent];
2888 if (ip >= PERF_CONTEXT_MAX) {
2889 err = add_callchain_ip(thread, cursor, parent,
2890 root_al, cpumode, ip,
2891 false, NULL, NULL, 0);
2898 static u64 get_leaf_frame_caller(struct perf_sample *sample,
2899 struct thread *thread, int usr_idx)
2901 if (machine__normalized_is(maps__machine(thread__maps(thread)), "arm64"))
2902 return get_leaf_frame_caller_aarch64(sample, thread, usr_idx);
2907 static int thread__resolve_callchain_sample(struct thread *thread,
2908 struct callchain_cursor *cursor,
2909 struct evsel *evsel,
2910 struct perf_sample *sample,
2911 struct symbol **parent,
2912 struct addr_location *root_al,
2915 struct branch_stack *branch = sample->branch_stack;
2916 struct branch_entry *entries = perf_sample__branch_entries(sample);
2917 struct ip_callchain *chain = sample->callchain;
2919 u8 cpumode = PERF_RECORD_MISC_USER;
2920 int i, j, err, nr_entries, usr_idx;
2923 u64 leaf_frame_caller;
2926 chain_nr = chain->nr;
2928 if (evsel__has_branch_callstack(evsel)) {
2929 struct perf_env *env = evsel__env(evsel);
2931 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2933 !env ? 0 : env->max_branches);
2935 return (err < 0) ? err : 0;
2939 * Based on DWARF debug information, some architectures skip
2940 * a callchain entry saved by the kernel.
2942 skip_idx = arch_skip_callchain_idx(thread, chain);
2945 * Add branches to call stack for easier browsing. This gives
2946 * more context for a sample than just the callers.
2948 * This uses individual histograms of paths compared to the
2949 * aggregated histograms the normal LBR mode uses.
2951 * Limitations for now:
2952 * - No extra filters
2953 * - No annotations (should annotate somehow)
2956 if (branch && callchain_param.branch_callstack) {
2957 int nr = min(max_stack, (int)branch->nr);
2958 struct branch_entry be[nr];
2959 struct iterations iter[nr];
2961 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2962 pr_warning("corrupted branch chain. skipping...\n");
2966 for (i = 0; i < nr; i++) {
2967 if (callchain_param.order == ORDER_CALLEE) {
2974 * Check for overlap into the callchain.
2975 * The return address is one off compared to
2976 * the branch entry. To adjust for this
2977 * assume the calling instruction is not longer
2980 if (i == skip_idx ||
2981 chain->ips[first_call] >= PERF_CONTEXT_MAX)
2983 else if (be[i].from < chain->ips[first_call] &&
2984 be[i].from >= chain->ips[first_call] - 8)
2987 be[i] = entries[branch->nr - i - 1];
2990 memset(iter, 0, sizeof(struct iterations) * nr);
2991 nr = remove_loops(be, nr, iter);
2993 for (i = 0; i < nr; i++) {
2994 err = add_callchain_ip(thread, cursor, parent,
3001 err = add_callchain_ip(thread, cursor, parent, root_al,
3018 if (chain && callchain_param.order != ORDER_CALLEE) {
3019 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
3020 &cpumode, chain->nr - first_call);
3022 return (err < 0) ? err : 0;
3024 for (i = first_call, nr_entries = 0;
3025 i < chain_nr && nr_entries < max_stack; i++) {
3028 if (callchain_param.order == ORDER_CALLEE)
3031 j = chain->nr - i - 1;
3033 #ifdef HAVE_SKIP_CALLCHAIN_IDX
3038 if (ip < PERF_CONTEXT_MAX)
3040 else if (callchain_param.order != ORDER_CALLEE) {
3041 err = find_prev_cpumode(chain, thread, cursor, parent,
3042 root_al, &cpumode, j);
3044 return (err < 0) ? err : 0;
3049 * PERF_CONTEXT_USER allows us to locate where the user stack ends.
3050 * Depending on callchain_param.order and the position of PERF_CONTEXT_USER,
3051 * the index will be different in order to add the missing frame
3052 * at the right place.
3055 usr_idx = callchain_param.order == ORDER_CALLEE ? j-2 : j-1;
3057 if (usr_idx >= 0 && chain->ips[usr_idx] == PERF_CONTEXT_USER) {
3059 leaf_frame_caller = get_leaf_frame_caller(sample, thread, usr_idx);
3062 * check if leaf_frame_Caller != ip to not add the same
3066 if (leaf_frame_caller && leaf_frame_caller != ip) {
3068 err = add_callchain_ip(thread, cursor, parent,
3069 root_al, &cpumode, leaf_frame_caller,
3070 false, NULL, NULL, 0);
3072 return (err < 0) ? err : 0;
3076 err = add_callchain_ip(thread, cursor, parent,
3077 root_al, &cpumode, ip,
3078 false, NULL, NULL, 0);
3081 return (err < 0) ? err : 0;
3087 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
3089 struct symbol *sym = ms->sym;
3090 struct map *map = ms->map;
3091 struct inline_node *inline_node;
3092 struct inline_list *ilist;
3096 struct map_symbol ilist_ms;
3098 if (!symbol_conf.inline_name || !map || !sym)
3101 addr = map__dso_map_ip(map, ip);
3102 addr = map__rip_2objdump(map, addr);
3103 dso = map__dso(map);
3105 inline_node = inlines__tree_find(&dso->inlined_nodes, addr);
3107 inline_node = dso__parse_addr_inlines(dso, addr, sym);
3110 inlines__tree_insert(&dso->inlined_nodes, inline_node);
3113 ilist_ms = (struct map_symbol) {
3114 .maps = maps__get(ms->maps),
3115 .map = map__get(map),
3117 list_for_each_entry(ilist, &inline_node->val, list) {
3118 ilist_ms.sym = ilist->symbol;
3119 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
3120 NULL, 0, 0, 0, ilist->srcline);
3125 map__put(ilist_ms.map);
3126 maps__put(ilist_ms.maps);
3131 static int unwind_entry(struct unwind_entry *entry, void *arg)
3133 struct callchain_cursor *cursor = arg;
3134 const char *srcline = NULL;
3135 u64 addr = entry->ip;
3137 if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
3140 if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
3144 * Convert entry->ip from a virtual address to an offset in
3145 * its corresponding binary.
3148 addr = map__dso_map_ip(entry->ms.map, entry->ip);
3150 srcline = callchain_srcline(&entry->ms, addr);
3151 return callchain_cursor_append(cursor, entry->ip, &entry->ms,
3152 false, NULL, 0, 0, 0, srcline);
3155 static int thread__resolve_callchain_unwind(struct thread *thread,
3156 struct callchain_cursor *cursor,
3157 struct evsel *evsel,
3158 struct perf_sample *sample,
3161 /* Can we do dwarf post unwind? */
3162 if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
3163 (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
3166 /* Bail out if nothing was captured. */
3167 if ((!sample->user_regs.regs) ||
3168 (!sample->user_stack.size))
3171 return unwind__get_entries(unwind_entry, cursor,
3172 thread, sample, max_stack, false);
3175 int thread__resolve_callchain(struct thread *thread,
3176 struct callchain_cursor *cursor,
3177 struct evsel *evsel,
3178 struct perf_sample *sample,
3179 struct symbol **parent,
3180 struct addr_location *root_al,
3188 callchain_cursor_reset(cursor);
3190 if (callchain_param.order == ORDER_CALLEE) {
3191 ret = thread__resolve_callchain_sample(thread, cursor,
3197 ret = thread__resolve_callchain_unwind(thread, cursor,
3201 ret = thread__resolve_callchain_unwind(thread, cursor,
3206 ret = thread__resolve_callchain_sample(thread, cursor,
3215 int machine__for_each_thread(struct machine *machine,
3216 int (*fn)(struct thread *thread, void *p),
3219 struct threads *threads;
3224 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3225 threads = &machine->threads[i];
3226 for (nd = rb_first_cached(&threads->entries); nd;
3228 struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
3230 rc = fn(trb->thread, priv);
3238 int machines__for_each_thread(struct machines *machines,
3239 int (*fn)(struct thread *thread, void *p),
3245 rc = machine__for_each_thread(&machines->host, fn, priv);
3249 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
3250 struct machine *machine = rb_entry(nd, struct machine, rb_node);
3252 rc = machine__for_each_thread(machine, fn, priv);
3259 pid_t machine__get_current_tid(struct machine *machine, int cpu)
3261 if (cpu < 0 || (size_t)cpu >= machine->current_tid_sz)
3264 return machine->current_tid[cpu];
3267 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
3270 struct thread *thread;
3271 const pid_t init_val = -1;
3276 if (realloc_array_as_needed(machine->current_tid,
3277 machine->current_tid_sz,
3282 machine->current_tid[cpu] = tid;
3284 thread = machine__findnew_thread(machine, pid, tid);
3288 thread__set_cpu(thread, cpu);
3289 thread__put(thread);
3295 * Compares the raw arch string. N.B. see instead perf_env__arch() or
3296 * machine__normalized_is() if a normalized arch is needed.
3298 bool machine__is(struct machine *machine, const char *arch)
3300 return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
3303 bool machine__normalized_is(struct machine *machine, const char *arch)
3305 return machine && !strcmp(perf_env__arch(machine->env), arch);
3308 int machine__nr_cpus_avail(struct machine *machine)
3310 return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
3313 int machine__get_kernel_start(struct machine *machine)
3315 struct map *map = machine__kernel_map(machine);
3319 * The only addresses above 2^63 are kernel addresses of a 64-bit
3320 * kernel. Note that addresses are unsigned so that on a 32-bit system
3321 * all addresses including kernel addresses are less than 2^32. In
3322 * that case (32-bit system), if the kernel mapping is unknown, all
3323 * addresses will be assumed to be in user space - see
3324 * machine__kernel_ip().
3326 machine->kernel_start = 1ULL << 63;
3328 err = map__load(map);
3330 * On x86_64, PTI entry trampolines are less than the
3331 * start of kernel text, but still above 2^63. So leave
3332 * kernel_start = 1ULL << 63 for x86_64.
3334 if (!err && !machine__is(machine, "x86_64"))
3335 machine->kernel_start = map__start(map);
3340 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
3342 u8 addr_cpumode = cpumode;
3345 if (!machine->single_address_space)
3348 kernel_ip = machine__kernel_ip(machine, addr);
3350 case PERF_RECORD_MISC_KERNEL:
3351 case PERF_RECORD_MISC_USER:
3352 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
3353 PERF_RECORD_MISC_USER;
3355 case PERF_RECORD_MISC_GUEST_KERNEL:
3356 case PERF_RECORD_MISC_GUEST_USER:
3357 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
3358 PERF_RECORD_MISC_GUEST_USER;
3364 return addr_cpumode;
3367 struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id)
3369 return dsos__findnew_id(&machine->dsos, filename, id);
3372 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
3374 return machine__findnew_dso_id(machine, filename, NULL);
3377 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
3379 struct machine *machine = vmachine;
3381 struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
3386 *modp = __map__is_kmodule(map) ? (char *)map__dso(map)->short_name : NULL;
3387 *addrp = map__unmap_ip(map, sym->start);
3391 int machine__for_each_dso(struct machine *machine, machine__dso_t fn, void *priv)
3396 list_for_each_entry(pos, &machine->dsos.head, node) {
3397 if (fn(pos, machine, priv))
3403 int machine__for_each_kernel_map(struct machine *machine, machine__map_t fn, void *priv)
3405 struct maps *maps = machine__kernel_maps(machine);
3406 struct map_rb_node *pos;
3409 maps__for_each_entry(maps, pos) {
3410 err = fn(pos->map, priv);
3418 bool machine__is_lock_function(struct machine *machine, u64 addr)
3420 if (!machine->sched.text_start) {
3422 struct symbol *sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_start", &kmap);
3425 /* to avoid retry */
3426 machine->sched.text_start = 1;
3430 machine->sched.text_start = map__unmap_ip(kmap, sym->start);
3432 /* should not fail from here */
3433 sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_end", &kmap);
3434 machine->sched.text_end = map__unmap_ip(kmap, sym->start);
3436 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_start", &kmap);
3437 machine->lock.text_start = map__unmap_ip(kmap, sym->start);
3439 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_end", &kmap);
3440 machine->lock.text_end = map__unmap_ip(kmap, sym->start);
3443 /* failed to get kernel symbols */
3444 if (machine->sched.text_start == 1)
3447 /* mutex and rwsem functions are in sched text section */
3448 if (machine->sched.text_start <= addr && addr < machine->sched.text_end)
3451 /* spinlock functions are in lock text section */
3452 if (machine->lock.text_start <= addr && addr < machine->lock.text_end)