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);
49 static struct dso *machine__kernel_dso(struct machine *machine)
51 return map__dso(machine->vmlinux_map);
54 static void dsos__init(struct dsos *dsos)
56 INIT_LIST_HEAD(&dsos->head);
58 init_rwsem(&dsos->lock);
61 static void machine__threads_init(struct machine *machine)
65 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
66 struct threads *threads = &machine->threads[i];
67 threads->entries = RB_ROOT_CACHED;
68 init_rwsem(&threads->lock);
70 INIT_LIST_HEAD(&threads->dead);
71 threads->last_match = NULL;
75 static int thread_rb_node__cmp_tid(const void *key, const struct rb_node *nd)
77 int to_find = (int) *((pid_t *)key);
79 return to_find - (int)thread__tid(rb_entry(nd, struct thread_rb_node, rb_node)->thread);
82 static struct thread_rb_node *thread_rb_node__find(const struct thread *th,
85 pid_t to_find = thread__tid(th);
86 struct rb_node *nd = rb_find(&to_find, tree, thread_rb_node__cmp_tid);
88 return rb_entry(nd, struct thread_rb_node, rb_node);
91 static int machine__set_mmap_name(struct machine *machine)
93 if (machine__is_host(machine))
94 machine->mmap_name = strdup("[kernel.kallsyms]");
95 else if (machine__is_default_guest(machine))
96 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
97 else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
99 machine->mmap_name = NULL;
101 return machine->mmap_name ? 0 : -ENOMEM;
104 static void thread__set_guest_comm(struct thread *thread, pid_t pid)
108 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
109 thread__set_comm(thread, comm, 0);
112 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
116 memset(machine, 0, sizeof(*machine));
117 machine->kmaps = maps__new(machine);
118 if (machine->kmaps == NULL)
121 RB_CLEAR_NODE(&machine->rb_node);
122 dsos__init(&machine->dsos);
124 machine__threads_init(machine);
126 machine->vdso_info = NULL;
131 machine->id_hdr_size = 0;
132 machine->kptr_restrict_warned = false;
133 machine->comm_exec = false;
134 machine->kernel_start = 0;
135 machine->vmlinux_map = NULL;
137 machine->root_dir = strdup(root_dir);
138 if (machine->root_dir == NULL)
141 if (machine__set_mmap_name(machine))
144 if (pid != HOST_KERNEL_ID) {
145 struct thread *thread = machine__findnew_thread(machine, -1,
151 thread__set_guest_comm(thread, pid);
155 machine->current_tid = NULL;
160 zfree(&machine->kmaps);
161 zfree(&machine->root_dir);
162 zfree(&machine->mmap_name);
167 struct machine *machine__new_host(void)
169 struct machine *machine = malloc(sizeof(*machine));
171 if (machine != NULL) {
172 machine__init(machine, "", HOST_KERNEL_ID);
174 if (machine__create_kernel_maps(machine) < 0)
184 struct machine *machine__new_kallsyms(void)
186 struct machine *machine = machine__new_host();
189 * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
190 * ask for not using the kcore parsing code, once this one is fixed
191 * to create a map per module.
193 if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
194 machine__delete(machine);
201 static void dsos__purge(struct dsos *dsos)
205 down_write(&dsos->lock);
207 list_for_each_entry_safe(pos, n, &dsos->head, node) {
208 RB_CLEAR_NODE(&pos->rb_node);
210 list_del_init(&pos->node);
214 up_write(&dsos->lock);
217 static void dsos__exit(struct dsos *dsos)
220 exit_rwsem(&dsos->lock);
223 void machine__delete_threads(struct machine *machine)
228 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
229 struct threads *threads = &machine->threads[i];
230 down_write(&threads->lock);
231 nd = rb_first_cached(&threads->entries);
233 struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
236 __machine__remove_thread(machine, trb, trb->thread, false);
238 up_write(&threads->lock);
242 void machine__exit(struct machine *machine)
249 machine__destroy_kernel_maps(machine);
250 maps__zput(machine->kmaps);
251 dsos__exit(&machine->dsos);
252 machine__exit_vdso(machine);
253 zfree(&machine->root_dir);
254 zfree(&machine->mmap_name);
255 zfree(&machine->current_tid);
256 zfree(&machine->kallsyms_filename);
258 machine__delete_threads(machine);
259 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
260 struct threads *threads = &machine->threads[i];
262 exit_rwsem(&threads->lock);
266 void machine__delete(struct machine *machine)
269 machine__exit(machine);
274 void machines__init(struct machines *machines)
276 machine__init(&machines->host, "", HOST_KERNEL_ID);
277 machines->guests = RB_ROOT_CACHED;
280 void machines__exit(struct machines *machines)
282 machine__exit(&machines->host);
286 struct machine *machines__add(struct machines *machines, pid_t pid,
287 const char *root_dir)
289 struct rb_node **p = &machines->guests.rb_root.rb_node;
290 struct rb_node *parent = NULL;
291 struct machine *pos, *machine = malloc(sizeof(*machine));
292 bool leftmost = true;
297 if (machine__init(machine, root_dir, pid) != 0) {
304 pos = rb_entry(parent, struct machine, rb_node);
313 rb_link_node(&machine->rb_node, parent, p);
314 rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
316 machine->machines = machines;
321 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
325 machines->host.comm_exec = comm_exec;
327 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
328 struct machine *machine = rb_entry(nd, struct machine, rb_node);
330 machine->comm_exec = comm_exec;
334 struct machine *machines__find(struct machines *machines, pid_t pid)
336 struct rb_node **p = &machines->guests.rb_root.rb_node;
337 struct rb_node *parent = NULL;
338 struct machine *machine;
339 struct machine *default_machine = NULL;
341 if (pid == HOST_KERNEL_ID)
342 return &machines->host;
346 machine = rb_entry(parent, struct machine, rb_node);
347 if (pid < machine->pid)
349 else if (pid > machine->pid)
354 default_machine = machine;
357 return default_machine;
360 struct machine *machines__findnew(struct machines *machines, pid_t pid)
363 const char *root_dir = "";
364 struct machine *machine = machines__find(machines, pid);
366 if (machine && (machine->pid == pid))
369 if ((pid != HOST_KERNEL_ID) &&
370 (pid != DEFAULT_GUEST_KERNEL_ID) &&
371 (symbol_conf.guestmount)) {
372 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
373 if (access(path, R_OK)) {
374 static struct strlist *seen;
377 seen = strlist__new(NULL, NULL);
379 if (!strlist__has_entry(seen, path)) {
380 pr_err("Can't access file %s\n", path);
381 strlist__add(seen, path);
389 machine = machines__add(machines, pid, root_dir);
394 struct machine *machines__find_guest(struct machines *machines, pid_t pid)
396 struct machine *machine = machines__find(machines, pid);
399 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID);
404 * A common case for KVM test programs is that the test program acts as the
405 * hypervisor, creating, running and destroying the virtual machine, and
406 * providing the guest object code from its own object code. In this case,
407 * the VM is not running an OS, but only the functions loaded into it by the
408 * hypervisor test program, and conveniently, loaded at the same virtual
411 * Normally to resolve addresses, MMAP events are needed to map addresses
412 * back to the object code and debug symbols for that object code.
414 * Currently, there is no way to get such mapping information from guests
415 * but, in the scenario described above, the guest has the same mappings
416 * as the hypervisor, so support for that scenario can be achieved.
418 * To support that, copy the host thread's maps to the guest thread's maps.
419 * Note, we do not discover the guest until we encounter a guest event,
420 * which works well because it is not until then that we know that the host
421 * thread's maps have been set up.
423 * This function returns the guest thread. Apart from keeping the data
424 * structures sane, using a thread belonging to the guest machine, instead
425 * of the host thread, allows it to have its own comm (refer
426 * thread__set_guest_comm()).
428 static struct thread *findnew_guest_code(struct machine *machine,
429 struct machine *host_machine,
432 struct thread *host_thread;
433 struct thread *thread;
439 thread = machine__findnew_thread(machine, -1, pid);
443 /* Assume maps are set up if there are any */
444 if (maps__nr_maps(thread__maps(thread)))
447 host_thread = machine__find_thread(host_machine, -1, pid);
451 thread__set_guest_comm(thread, pid);
454 * Guest code can be found in hypervisor process at the same address
457 err = maps__clone(thread, thread__maps(host_thread));
458 thread__put(host_thread);
465 thread__zput(thread);
469 struct thread *machines__findnew_guest_code(struct machines *machines, pid_t pid)
471 struct machine *host_machine = machines__find(machines, HOST_KERNEL_ID);
472 struct machine *machine = machines__findnew(machines, pid);
474 return findnew_guest_code(machine, host_machine, pid);
477 struct thread *machine__findnew_guest_code(struct machine *machine, pid_t pid)
479 struct machines *machines = machine->machines;
480 struct machine *host_machine;
485 host_machine = machines__find(machines, HOST_KERNEL_ID);
487 return findnew_guest_code(machine, host_machine, pid);
490 void machines__process_guests(struct machines *machines,
491 machine__process_t process, void *data)
495 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
496 struct machine *pos = rb_entry(nd, struct machine, rb_node);
501 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
503 struct rb_node *node;
504 struct machine *machine;
506 machines->host.id_hdr_size = id_hdr_size;
508 for (node = rb_first_cached(&machines->guests); node;
509 node = rb_next(node)) {
510 machine = rb_entry(node, struct machine, rb_node);
511 machine->id_hdr_size = id_hdr_size;
517 static void machine__update_thread_pid(struct machine *machine,
518 struct thread *th, pid_t pid)
520 struct thread *leader;
522 if (pid == thread__pid(th) || pid == -1 || thread__pid(th) != -1)
525 thread__set_pid(th, pid);
527 if (thread__pid(th) == thread__tid(th))
530 leader = __machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
534 if (!thread__maps(leader))
535 thread__set_maps(leader, maps__new(machine));
537 if (!thread__maps(leader))
540 if (thread__maps(th) == thread__maps(leader))
543 if (thread__maps(th)) {
545 * Maps are created from MMAP events which provide the pid and
546 * tid. Consequently there never should be any maps on a thread
547 * with an unknown pid. Just print an error if there are.
549 if (!maps__empty(thread__maps(th)))
550 pr_err("Discarding thread maps for %d:%d\n",
551 thread__pid(th), thread__tid(th));
552 maps__put(thread__maps(th));
555 thread__set_maps(th, maps__get(thread__maps(leader)));
560 pr_err("Failed to join map groups for %d:%d\n", thread__pid(th), thread__tid(th));
565 * Front-end cache - TID lookups come in blocks,
566 * so most of the time we dont have to look up
569 static struct thread*
570 __threads__get_last_match(struct threads *threads, struct machine *machine,
575 th = threads->last_match;
577 if (thread__tid(th) == tid) {
578 machine__update_thread_pid(machine, th, pid);
579 return thread__get(th);
581 thread__put(threads->last_match);
582 threads->last_match = NULL;
588 static struct thread*
589 threads__get_last_match(struct threads *threads, struct machine *machine,
592 struct thread *th = NULL;
594 if (perf_singlethreaded)
595 th = __threads__get_last_match(threads, machine, pid, tid);
601 __threads__set_last_match(struct threads *threads, struct thread *th)
603 thread__put(threads->last_match);
604 threads->last_match = thread__get(th);
608 threads__set_last_match(struct threads *threads, struct thread *th)
610 if (perf_singlethreaded)
611 __threads__set_last_match(threads, th);
615 * Caller must eventually drop thread->refcnt returned with a successful
616 * lookup/new thread inserted.
618 static struct thread *____machine__findnew_thread(struct machine *machine,
619 struct threads *threads,
620 pid_t pid, pid_t tid,
623 struct rb_node **p = &threads->entries.rb_root.rb_node;
624 struct rb_node *parent = NULL;
626 struct thread_rb_node *nd;
627 bool leftmost = true;
629 th = threads__get_last_match(threads, machine, pid, tid);
635 th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
637 if (thread__tid(th) == tid) {
638 threads__set_last_match(threads, th);
639 machine__update_thread_pid(machine, th, pid);
640 return thread__get(th);
643 if (tid < thread__tid(th))
654 th = thread__new(pid, tid);
658 nd = malloc(sizeof(*nd));
665 rb_link_node(&nd->rb_node, parent, p);
666 rb_insert_color_cached(&nd->rb_node, &threads->entries, leftmost);
668 * We have to initialize maps separately after rb tree is updated.
670 * The reason is that we call machine__findnew_thread within
671 * thread__init_maps to find the thread leader and that would screwed
674 if (thread__init_maps(th, machine)) {
675 pr_err("Thread init failed thread %d\n", pid);
676 rb_erase_cached(&nd->rb_node, &threads->entries);
677 RB_CLEAR_NODE(&nd->rb_node);
683 * It is now in the rbtree, get a ref
685 threads__set_last_match(threads, th);
688 return thread__get(th);
691 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
693 return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
696 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
699 struct threads *threads = machine__threads(machine, tid);
702 down_write(&threads->lock);
703 th = __machine__findnew_thread(machine, pid, tid);
704 up_write(&threads->lock);
708 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
711 struct threads *threads = machine__threads(machine, tid);
714 down_read(&threads->lock);
715 th = ____machine__findnew_thread(machine, threads, pid, tid, false);
716 up_read(&threads->lock);
721 * Threads are identified by pid and tid, and the idle task has pid == tid == 0.
722 * So here a single thread is created for that, but actually there is a separate
723 * idle task per cpu, so there should be one 'struct thread' per cpu, but there
724 * is only 1. That causes problems for some tools, requiring workarounds. For
725 * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu().
727 struct thread *machine__idle_thread(struct machine *machine)
729 struct thread *thread = machine__findnew_thread(machine, 0, 0);
731 if (!thread || thread__set_comm(thread, "swapper", 0) ||
732 thread__set_namespaces(thread, 0, NULL))
733 pr_err("problem inserting idle task for machine pid %d\n", machine->pid);
738 struct comm *machine__thread_exec_comm(struct machine *machine,
739 struct thread *thread)
741 if (machine->comm_exec)
742 return thread__exec_comm(thread);
744 return thread__comm(thread);
747 int machine__process_comm_event(struct machine *machine, union perf_event *event,
748 struct perf_sample *sample)
750 struct thread *thread = machine__findnew_thread(machine,
753 bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
757 machine->comm_exec = true;
760 perf_event__fprintf_comm(event, stdout);
762 if (thread == NULL ||
763 __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
764 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
773 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
774 union perf_event *event,
775 struct perf_sample *sample __maybe_unused)
777 struct thread *thread = machine__findnew_thread(machine,
778 event->namespaces.pid,
779 event->namespaces.tid);
782 WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
783 "\nWARNING: kernel seems to support more namespaces than perf"
784 " tool.\nTry updating the perf tool..\n\n");
786 WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
787 "\nWARNING: perf tool seems to support more namespaces than"
788 " the kernel.\nTry updating the kernel..\n\n");
791 perf_event__fprintf_namespaces(event, stdout);
793 if (thread == NULL ||
794 thread__set_namespaces(thread, sample->time, &event->namespaces)) {
795 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
804 int machine__process_cgroup_event(struct machine *machine,
805 union perf_event *event,
806 struct perf_sample *sample __maybe_unused)
811 perf_event__fprintf_cgroup(event, stdout);
813 cgrp = cgroup__findnew(machine->env, event->cgroup.id, event->cgroup.path);
820 int machine__process_lost_event(struct machine *machine __maybe_unused,
821 union perf_event *event, struct perf_sample *sample __maybe_unused)
823 dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
824 event->lost.id, event->lost.lost);
828 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
829 union perf_event *event, struct perf_sample *sample)
831 dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
832 sample->id, event->lost_samples.lost);
836 static struct dso *machine__findnew_module_dso(struct machine *machine,
838 const char *filename)
842 down_write(&machine->dsos.lock);
844 dso = __dsos__find(&machine->dsos, m->name, true);
846 dso = __dsos__addnew(&machine->dsos, m->name);
850 dso__set_module_info(dso, m, machine);
851 dso__set_long_name(dso, strdup(filename), true);
852 dso->kernel = DSO_SPACE__KERNEL;
857 up_write(&machine->dsos.lock);
861 int machine__process_aux_event(struct machine *machine __maybe_unused,
862 union perf_event *event)
865 perf_event__fprintf_aux(event, stdout);
869 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
870 union perf_event *event)
873 perf_event__fprintf_itrace_start(event, stdout);
877 int machine__process_aux_output_hw_id_event(struct machine *machine __maybe_unused,
878 union perf_event *event)
881 perf_event__fprintf_aux_output_hw_id(event, stdout);
885 int machine__process_switch_event(struct machine *machine __maybe_unused,
886 union perf_event *event)
889 perf_event__fprintf_switch(event, stdout);
893 static int machine__process_ksymbol_register(struct machine *machine,
894 union perf_event *event,
895 struct perf_sample *sample __maybe_unused)
899 struct map *map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
900 bool put_map = false;
904 dso = dso__new(event->ksymbol.name);
910 dso->kernel = DSO_SPACE__KERNEL;
911 map = map__new2(0, dso);
918 * The inserted map has a get on it, we need to put to release
919 * the reference count here, but do it after all accesses are
923 if (event->ksymbol.ksym_type == PERF_RECORD_KSYMBOL_TYPE_OOL) {
924 dso->binary_type = DSO_BINARY_TYPE__OOL;
925 dso->data.file_size = event->ksymbol.len;
926 dso__set_loaded(dso);
929 map__set_start(map, event->ksymbol.addr);
930 map__set_end(map, map__start(map) + event->ksymbol.len);
931 err = maps__insert(machine__kernel_maps(machine), map);
937 dso__set_loaded(dso);
939 if (is_bpf_image(event->ksymbol.name)) {
940 dso->binary_type = DSO_BINARY_TYPE__BPF_IMAGE;
941 dso__set_long_name(dso, "", false);
947 sym = symbol__new(map__map_ip(map, map__start(map)),
949 0, 0, event->ksymbol.name);
954 dso__insert_symbol(dso, sym);
961 static int machine__process_ksymbol_unregister(struct machine *machine,
962 union perf_event *event,
963 struct perf_sample *sample __maybe_unused)
968 map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
972 if (RC_CHK_ACCESS(map) != RC_CHK_ACCESS(machine->vmlinux_map))
973 maps__remove(machine__kernel_maps(machine), map);
975 struct dso *dso = map__dso(map);
977 sym = dso__find_symbol(dso, map__map_ip(map, map__start(map)));
979 dso__delete_symbol(dso, sym);
985 int machine__process_ksymbol(struct machine *machine __maybe_unused,
986 union perf_event *event,
987 struct perf_sample *sample)
990 perf_event__fprintf_ksymbol(event, stdout);
992 if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
993 return machine__process_ksymbol_unregister(machine, event,
995 return machine__process_ksymbol_register(machine, event, sample);
998 int machine__process_text_poke(struct machine *machine, union perf_event *event,
999 struct perf_sample *sample __maybe_unused)
1001 struct map *map = maps__find(machine__kernel_maps(machine), event->text_poke.addr);
1002 u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1003 struct dso *dso = map ? map__dso(map) : NULL;
1006 perf_event__fprintf_text_poke(event, machine, stdout);
1008 if (!event->text_poke.new_len)
1011 if (cpumode != PERF_RECORD_MISC_KERNEL) {
1012 pr_debug("%s: unsupported cpumode - ignoring\n", __func__);
1017 u8 *new_bytes = event->text_poke.bytes + event->text_poke.old_len;
1021 * Kernel maps might be changed when loading symbols so loading
1022 * must be done prior to using kernel maps.
1025 ret = dso__data_write_cache_addr(dso, map, machine,
1026 event->text_poke.addr,
1028 event->text_poke.new_len);
1029 if (ret != event->text_poke.new_len)
1030 pr_debug("Failed to write kernel text poke at %#" PRI_lx64 "\n",
1031 event->text_poke.addr);
1033 pr_debug("Failed to find kernel text poke address map for %#" PRI_lx64 "\n",
1034 event->text_poke.addr);
1040 static struct map *machine__addnew_module_map(struct machine *machine, u64 start,
1041 const char *filename)
1043 struct map *map = NULL;
1048 if (kmod_path__parse_name(&m, filename))
1051 dso = machine__findnew_module_dso(machine, &m, filename);
1055 map = map__new2(start, dso);
1059 err = maps__insert(machine__kernel_maps(machine), map);
1060 /* If maps__insert failed, return NULL. */
1066 /* put the dso here, corresponding to machine__findnew_module_dso */
1072 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
1075 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
1077 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
1078 struct machine *pos = rb_entry(nd, struct machine, rb_node);
1079 ret += __dsos__fprintf(&pos->dsos.head, fp);
1085 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
1086 bool (skip)(struct dso *dso, int parm), int parm)
1088 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
1091 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
1092 bool (skip)(struct dso *dso, int parm), int parm)
1095 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
1097 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
1098 struct machine *pos = rb_entry(nd, struct machine, rb_node);
1099 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
1104 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
1108 struct dso *kdso = machine__kernel_dso(machine);
1110 if (kdso->has_build_id) {
1111 char filename[PATH_MAX];
1112 if (dso__build_id_filename(kdso, filename, sizeof(filename),
1114 printed += fprintf(fp, "[0] %s\n", filename);
1117 for (i = 0; i < vmlinux_path__nr_entries; ++i)
1118 printed += fprintf(fp, "[%d] %s\n",
1119 i + kdso->has_build_id, vmlinux_path[i]);
1124 size_t machine__fprintf(struct machine *machine, FILE *fp)
1130 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
1131 struct threads *threads = &machine->threads[i];
1133 down_read(&threads->lock);
1135 ret = fprintf(fp, "Threads: %u\n", threads->nr);
1137 for (nd = rb_first_cached(&threads->entries); nd;
1139 struct thread *pos = rb_entry(nd, struct thread_rb_node, rb_node)->thread;
1141 ret += thread__fprintf(pos, fp);
1144 up_read(&threads->lock);
1149 static struct dso *machine__get_kernel(struct machine *machine)
1151 const char *vmlinux_name = machine->mmap_name;
1154 if (machine__is_host(machine)) {
1155 if (symbol_conf.vmlinux_name)
1156 vmlinux_name = symbol_conf.vmlinux_name;
1158 kernel = machine__findnew_kernel(machine, vmlinux_name,
1159 "[kernel]", DSO_SPACE__KERNEL);
1161 if (symbol_conf.default_guest_vmlinux_name)
1162 vmlinux_name = symbol_conf.default_guest_vmlinux_name;
1164 kernel = machine__findnew_kernel(machine, vmlinux_name,
1166 DSO_SPACE__KERNEL_GUEST);
1169 if (kernel != NULL && (!kernel->has_build_id))
1170 dso__read_running_kernel_build_id(kernel, machine);
1175 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
1178 if (machine__is_default_guest(machine))
1179 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
1181 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
1184 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
1186 /* Figure out the start address of kernel map from /proc/kallsyms.
1187 * Returns the name of the start symbol in *symbol_name. Pass in NULL as
1188 * symbol_name if it's not that important.
1190 static int machine__get_running_kernel_start(struct machine *machine,
1191 const char **symbol_name,
1192 u64 *start, u64 *end)
1194 char filename[PATH_MAX];
1199 machine__get_kallsyms_filename(machine, filename, PATH_MAX);
1201 if (symbol__restricted_filename(filename, "/proc/kallsyms"))
1204 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
1205 err = kallsyms__get_function_start(filename, name, &addr);
1214 *symbol_name = name;
1218 err = kallsyms__get_function_start(filename, "_etext", &addr);
1225 int machine__create_extra_kernel_map(struct machine *machine,
1227 struct extra_kernel_map *xm)
1233 map = map__new2(xm->start, kernel);
1237 map__set_end(map, xm->end);
1238 map__set_pgoff(map, xm->pgoff);
1240 kmap = map__kmap(map);
1242 strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1244 err = maps__insert(machine__kernel_maps(machine), map);
1247 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1248 kmap->name, map__start(map), map__end(map));
1256 static u64 find_entry_trampoline(struct dso *dso)
1258 /* Duplicates are removed so lookup all aliases */
1259 const char *syms[] = {
1260 "_entry_trampoline",
1261 "__entry_trampoline_start",
1262 "entry_SYSCALL_64_trampoline",
1264 struct symbol *sym = dso__first_symbol(dso);
1267 for (; sym; sym = dso__next_symbol(sym)) {
1268 if (sym->binding != STB_GLOBAL)
1270 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1271 if (!strcmp(sym->name, syms[i]))
1280 * These values can be used for kernels that do not have symbols for the entry
1281 * trampolines in kallsyms.
1283 #define X86_64_CPU_ENTRY_AREA_PER_CPU 0xfffffe0000000000ULL
1284 #define X86_64_CPU_ENTRY_AREA_SIZE 0x2c000
1285 #define X86_64_ENTRY_TRAMPOLINE 0x6000
1287 /* Map x86_64 PTI entry trampolines */
1288 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1291 struct maps *kmaps = machine__kernel_maps(machine);
1292 int nr_cpus_avail, cpu;
1294 struct map_rb_node *rb_node;
1298 * In the vmlinux case, pgoff is a virtual address which must now be
1299 * mapped to a vmlinux offset.
1301 maps__for_each_entry(kmaps, rb_node) {
1302 struct map *dest_map, *map = rb_node->map;
1303 struct kmap *kmap = __map__kmap(map);
1305 if (!kmap || !is_entry_trampoline(kmap->name))
1308 dest_map = maps__find(kmaps, map__pgoff(map));
1309 if (dest_map != map)
1310 map__set_pgoff(map, map__map_ip(dest_map, map__pgoff(map)));
1313 if (found || machine->trampolines_mapped)
1316 pgoff = find_entry_trampoline(kernel);
1320 nr_cpus_avail = machine__nr_cpus_avail(machine);
1322 /* Add a 1 page map for each CPU's entry trampoline */
1323 for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1324 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1325 cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1326 X86_64_ENTRY_TRAMPOLINE;
1327 struct extra_kernel_map xm = {
1329 .end = va + page_size,
1333 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1335 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1339 machine->trampolines_mapped = nr_cpus_avail;
1344 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1345 struct dso *kernel __maybe_unused)
1351 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1353 /* In case of renewal the kernel map, destroy previous one */
1354 machine__destroy_kernel_maps(machine);
1356 map__put(machine->vmlinux_map);
1357 machine->vmlinux_map = map__new2(0, kernel);
1358 if (machine->vmlinux_map == NULL)
1361 map__set_map_ip(machine->vmlinux_map, identity__map_ip);
1362 map__set_unmap_ip(machine->vmlinux_map, identity__map_ip);
1363 return maps__insert(machine__kernel_maps(machine), machine->vmlinux_map);
1366 void machine__destroy_kernel_maps(struct machine *machine)
1369 struct map *map = machine__kernel_map(machine);
1374 kmap = map__kmap(map);
1375 maps__remove(machine__kernel_maps(machine), map);
1376 if (kmap && kmap->ref_reloc_sym) {
1377 zfree((char **)&kmap->ref_reloc_sym->name);
1378 zfree(&kmap->ref_reloc_sym);
1381 map__zput(machine->vmlinux_map);
1384 int machines__create_guest_kernel_maps(struct machines *machines)
1387 struct dirent **namelist = NULL;
1389 char path[PATH_MAX];
1393 if (symbol_conf.default_guest_vmlinux_name ||
1394 symbol_conf.default_guest_modules ||
1395 symbol_conf.default_guest_kallsyms) {
1396 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1399 if (symbol_conf.guestmount) {
1400 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1403 for (i = 0; i < items; i++) {
1404 if (!isdigit(namelist[i]->d_name[0])) {
1405 /* Filter out . and .. */
1408 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1409 if ((*endp != '\0') ||
1410 (endp == namelist[i]->d_name) ||
1411 (errno == ERANGE)) {
1412 pr_debug("invalid directory (%s). Skipping.\n",
1413 namelist[i]->d_name);
1416 sprintf(path, "%s/%s/proc/kallsyms",
1417 symbol_conf.guestmount,
1418 namelist[i]->d_name);
1419 ret = access(path, R_OK);
1421 pr_debug("Can't access file %s\n", path);
1424 machines__create_kernel_maps(machines, pid);
1433 void machines__destroy_kernel_maps(struct machines *machines)
1435 struct rb_node *next = rb_first_cached(&machines->guests);
1437 machine__destroy_kernel_maps(&machines->host);
1440 struct machine *pos = rb_entry(next, struct machine, rb_node);
1442 next = rb_next(&pos->rb_node);
1443 rb_erase_cached(&pos->rb_node, &machines->guests);
1444 machine__delete(pos);
1448 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1450 struct machine *machine = machines__findnew(machines, pid);
1452 if (machine == NULL)
1455 return machine__create_kernel_maps(machine);
1458 int machine__load_kallsyms(struct machine *machine, const char *filename)
1460 struct map *map = machine__kernel_map(machine);
1461 struct dso *dso = map__dso(map);
1462 int ret = __dso__load_kallsyms(dso, filename, map, true);
1465 dso__set_loaded(dso);
1467 * Since /proc/kallsyms will have multiple sessions for the
1468 * kernel, with modules between them, fixup the end of all
1471 maps__fixup_end(machine__kernel_maps(machine));
1477 int machine__load_vmlinux_path(struct machine *machine)
1479 struct map *map = machine__kernel_map(machine);
1480 struct dso *dso = map__dso(map);
1481 int ret = dso__load_vmlinux_path(dso, map);
1484 dso__set_loaded(dso);
1489 static char *get_kernel_version(const char *root_dir)
1491 char version[PATH_MAX];
1494 const char *prefix = "Linux version ";
1496 sprintf(version, "%s/proc/version", root_dir);
1497 file = fopen(version, "r");
1501 tmp = fgets(version, sizeof(version), file);
1506 name = strstr(version, prefix);
1509 name += strlen(prefix);
1510 tmp = strchr(name, ' ');
1514 return strdup(name);
1517 static bool is_kmod_dso(struct dso *dso)
1519 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1520 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1523 static int maps__set_module_path(struct maps *maps, const char *path, struct kmod_path *m)
1527 struct map *map = maps__find_by_name(maps, m->name);
1532 long_name = strdup(path);
1533 if (long_name == NULL)
1536 dso = map__dso(map);
1537 dso__set_long_name(dso, long_name, true);
1538 dso__kernel_module_get_build_id(dso, "");
1541 * Full name could reveal us kmod compression, so
1542 * we need to update the symtab_type if needed.
1544 if (m->comp && is_kmod_dso(dso)) {
1546 dso->comp = m->comp;
1552 static int maps__set_modules_path_dir(struct maps *maps, const char *dir_name, int depth)
1554 struct dirent *dent;
1555 DIR *dir = opendir(dir_name);
1559 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1563 while ((dent = readdir(dir)) != NULL) {
1564 char path[PATH_MAX];
1567 /*sshfs might return bad dent->d_type, so we have to stat*/
1568 path__join(path, sizeof(path), dir_name, dent->d_name);
1569 if (stat(path, &st))
1572 if (S_ISDIR(st.st_mode)) {
1573 if (!strcmp(dent->d_name, ".") ||
1574 !strcmp(dent->d_name, ".."))
1577 /* Do not follow top-level source and build symlinks */
1579 if (!strcmp(dent->d_name, "source") ||
1580 !strcmp(dent->d_name, "build"))
1584 ret = maps__set_modules_path_dir(maps, path, depth + 1);
1590 ret = kmod_path__parse_name(&m, dent->d_name);
1595 ret = maps__set_module_path(maps, path, &m);
1609 static int machine__set_modules_path(struct machine *machine)
1612 char modules_path[PATH_MAX];
1614 version = get_kernel_version(machine->root_dir);
1618 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1619 machine->root_dir, version);
1622 return maps__set_modules_path_dir(machine__kernel_maps(machine), modules_path, 0);
1624 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1625 u64 *size __maybe_unused,
1626 const char *name __maybe_unused)
1631 static int machine__create_module(void *arg, const char *name, u64 start,
1634 struct machine *machine = arg;
1637 if (arch__fix_module_text_start(&start, &size, name) < 0)
1640 map = machine__addnew_module_map(machine, start, name);
1643 map__set_end(map, start + size);
1645 dso__kernel_module_get_build_id(map__dso(map), machine->root_dir);
1650 static int machine__create_modules(struct machine *machine)
1652 const char *modules;
1653 char path[PATH_MAX];
1655 if (machine__is_default_guest(machine)) {
1656 modules = symbol_conf.default_guest_modules;
1658 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1662 if (symbol__restricted_filename(modules, "/proc/modules"))
1665 if (modules__parse(modules, machine, machine__create_module))
1668 if (!machine__set_modules_path(machine))
1671 pr_debug("Problems setting modules path maps, continuing anyway...\n");
1676 static void machine__set_kernel_mmap(struct machine *machine,
1679 map__set_start(machine->vmlinux_map, start);
1680 map__set_end(machine->vmlinux_map, end);
1682 * Be a bit paranoid here, some perf.data file came with
1683 * a zero sized synthesized MMAP event for the kernel.
1685 if (start == 0 && end == 0)
1686 map__set_end(machine->vmlinux_map, ~0ULL);
1689 static int machine__update_kernel_mmap(struct machine *machine,
1692 struct map *orig, *updated;
1695 orig = machine->vmlinux_map;
1696 updated = map__get(orig);
1698 machine->vmlinux_map = updated;
1699 machine__set_kernel_mmap(machine, start, end);
1700 maps__remove(machine__kernel_maps(machine), orig);
1701 err = maps__insert(machine__kernel_maps(machine), updated);
1707 int machine__create_kernel_maps(struct machine *machine)
1709 struct dso *kernel = machine__get_kernel(machine);
1710 const char *name = NULL;
1711 u64 start = 0, end = ~0ULL;
1717 ret = __machine__create_kernel_maps(machine, kernel);
1721 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1722 if (machine__is_host(machine))
1723 pr_debug("Problems creating module maps, "
1724 "continuing anyway...\n");
1726 pr_debug("Problems creating module maps for guest %d, "
1727 "continuing anyway...\n", machine->pid);
1730 if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1732 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1733 machine__destroy_kernel_maps(machine);
1739 * we have a real start address now, so re-order the kmaps
1740 * assume it's the last in the kmaps
1742 ret = machine__update_kernel_mmap(machine, start, end);
1747 if (machine__create_extra_kernel_maps(machine, kernel))
1748 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1751 /* update end address of the kernel map using adjacent module address */
1752 struct map_rb_node *rb_node = maps__find_node(machine__kernel_maps(machine),
1753 machine__kernel_map(machine));
1754 struct map_rb_node *next = map_rb_node__next(rb_node);
1757 machine__set_kernel_mmap(machine, start, map__start(next->map));
1765 static bool machine__uses_kcore(struct machine *machine)
1769 list_for_each_entry(dso, &machine->dsos.head, node) {
1770 if (dso__is_kcore(dso))
1777 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1778 struct extra_kernel_map *xm)
1780 return machine__is(machine, "x86_64") &&
1781 is_entry_trampoline(xm->name);
1784 static int machine__process_extra_kernel_map(struct machine *machine,
1785 struct extra_kernel_map *xm)
1787 struct dso *kernel = machine__kernel_dso(machine);
1792 return machine__create_extra_kernel_map(machine, kernel, xm);
1795 static int machine__process_kernel_mmap_event(struct machine *machine,
1796 struct extra_kernel_map *xm,
1797 struct build_id *bid)
1799 enum dso_space_type dso_space;
1800 bool is_kernel_mmap;
1801 const char *mmap_name = machine->mmap_name;
1803 /* If we have maps from kcore then we do not need or want any others */
1804 if (machine__uses_kcore(machine))
1807 if (machine__is_host(machine))
1808 dso_space = DSO_SPACE__KERNEL;
1810 dso_space = DSO_SPACE__KERNEL_GUEST;
1812 is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1813 if (!is_kernel_mmap && !machine__is_host(machine)) {
1815 * If the event was recorded inside the guest and injected into
1816 * the host perf.data file, then it will match a host mmap_name,
1817 * so try that - see machine__set_mmap_name().
1819 mmap_name = "[kernel.kallsyms]";
1820 is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1822 if (xm->name[0] == '/' ||
1823 (!is_kernel_mmap && xm->name[0] == '[')) {
1824 struct map *map = machine__addnew_module_map(machine, xm->start, xm->name);
1829 map__set_end(map, map__start(map) + xm->end - xm->start);
1831 if (build_id__is_defined(bid))
1832 dso__set_build_id(map__dso(map), bid);
1835 } else if (is_kernel_mmap) {
1836 const char *symbol_name = xm->name + strlen(mmap_name);
1838 * Should be there already, from the build-id table in
1841 struct dso *kernel = NULL;
1844 down_read(&machine->dsos.lock);
1846 list_for_each_entry(dso, &machine->dsos.head, node) {
1849 * The cpumode passed to is_kernel_module is not the
1850 * cpumode of *this* event. If we insist on passing
1851 * correct cpumode to is_kernel_module, we should
1852 * record the cpumode when we adding this dso to the
1855 * However we don't really need passing correct
1856 * cpumode. We know the correct cpumode must be kernel
1857 * mode (if not, we should not link it onto kernel_dsos
1860 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1861 * is_kernel_module() treats it as a kernel cpumode.
1865 is_kernel_module(dso->long_name,
1866 PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1870 kernel = dso__get(dso);
1874 up_read(&machine->dsos.lock);
1877 kernel = machine__findnew_dso(machine, machine->mmap_name);
1881 kernel->kernel = dso_space;
1882 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1887 if (strstr(kernel->long_name, "vmlinux"))
1888 dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1890 if (machine__update_kernel_mmap(machine, xm->start, xm->end) < 0) {
1895 if (build_id__is_defined(bid))
1896 dso__set_build_id(kernel, bid);
1899 * Avoid using a zero address (kptr_restrict) for the ref reloc
1900 * symbol. Effectively having zero here means that at record
1901 * time /proc/sys/kernel/kptr_restrict was non zero.
1903 if (xm->pgoff != 0) {
1904 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1909 if (machine__is_default_guest(machine)) {
1911 * preload dso of guest kernel and modules
1913 dso__load(kernel, machine__kernel_map(machine));
1916 } else if (perf_event__is_extra_kernel_mmap(machine, xm)) {
1917 return machine__process_extra_kernel_map(machine, xm);
1924 int machine__process_mmap2_event(struct machine *machine,
1925 union perf_event *event,
1926 struct perf_sample *sample)
1928 struct thread *thread;
1930 struct dso_id dso_id = {
1931 .maj = event->mmap2.maj,
1932 .min = event->mmap2.min,
1933 .ino = event->mmap2.ino,
1934 .ino_generation = event->mmap2.ino_generation,
1936 struct build_id __bid, *bid = NULL;
1940 perf_event__fprintf_mmap2(event, stdout);
1942 if (event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID) {
1944 build_id__init(bid, event->mmap2.build_id, event->mmap2.build_id_size);
1947 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1948 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1949 struct extra_kernel_map xm = {
1950 .start = event->mmap2.start,
1951 .end = event->mmap2.start + event->mmap2.len,
1952 .pgoff = event->mmap2.pgoff,
1955 strlcpy(xm.name, event->mmap2.filename, KMAP_NAME_LEN);
1956 ret = machine__process_kernel_mmap_event(machine, &xm, bid);
1962 thread = machine__findnew_thread(machine, event->mmap2.pid,
1967 map = map__new(machine, event->mmap2.start,
1968 event->mmap2.len, event->mmap2.pgoff,
1969 &dso_id, event->mmap2.prot,
1970 event->mmap2.flags, bid,
1971 event->mmap2.filename, thread);
1974 goto out_problem_map;
1976 ret = thread__insert_map(thread, map);
1978 goto out_problem_insert;
1980 thread__put(thread);
1987 thread__put(thread);
1989 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1993 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1994 struct perf_sample *sample)
1996 struct thread *thread;
2002 perf_event__fprintf_mmap(event, stdout);
2004 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
2005 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
2006 struct extra_kernel_map xm = {
2007 .start = event->mmap.start,
2008 .end = event->mmap.start + event->mmap.len,
2009 .pgoff = event->mmap.pgoff,
2012 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
2013 ret = machine__process_kernel_mmap_event(machine, &xm, NULL);
2019 thread = machine__findnew_thread(machine, event->mmap.pid,
2024 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
2027 map = map__new(machine, event->mmap.start,
2028 event->mmap.len, event->mmap.pgoff,
2029 NULL, prot, 0, NULL, event->mmap.filename, thread);
2032 goto out_problem_map;
2034 ret = thread__insert_map(thread, map);
2036 goto out_problem_insert;
2038 thread__put(thread);
2045 thread__put(thread);
2047 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
2051 static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
2052 struct thread *th, bool lock)
2054 struct threads *threads = machine__threads(machine, thread__tid(th));
2057 nd = thread_rb_node__find(th, &threads->entries.rb_root);
2059 if (threads->last_match && RC_CHK_ACCESS(threads->last_match) == RC_CHK_ACCESS(th))
2060 threads__set_last_match(threads, NULL);
2063 down_write(&threads->lock);
2065 BUG_ON(refcount_read(thread__refcnt(th)) == 0);
2067 thread__put(nd->thread);
2068 rb_erase_cached(&nd->rb_node, &threads->entries);
2069 RB_CLEAR_NODE(&nd->rb_node);
2075 up_write(&threads->lock);
2078 void machine__remove_thread(struct machine *machine, struct thread *th)
2080 return __machine__remove_thread(machine, NULL, th, true);
2083 int machine__process_fork_event(struct machine *machine, union perf_event *event,
2084 struct perf_sample *sample)
2086 struct thread *thread = machine__find_thread(machine,
2089 struct thread *parent = machine__findnew_thread(machine,
2092 bool do_maps_clone = true;
2096 perf_event__fprintf_task(event, stdout);
2099 * There may be an existing thread that is not actually the parent,
2100 * either because we are processing events out of order, or because the
2101 * (fork) event that would have removed the thread was lost. Assume the
2102 * latter case and continue on as best we can.
2104 if (thread__pid(parent) != (pid_t)event->fork.ppid) {
2105 dump_printf("removing erroneous parent thread %d/%d\n",
2106 thread__pid(parent), thread__tid(parent));
2107 machine__remove_thread(machine, parent);
2108 thread__put(parent);
2109 parent = machine__findnew_thread(machine, event->fork.ppid,
2113 /* if a thread currently exists for the thread id remove it */
2114 if (thread != NULL) {
2115 machine__remove_thread(machine, thread);
2116 thread__put(thread);
2119 thread = machine__findnew_thread(machine, event->fork.pid,
2122 * When synthesizing FORK events, we are trying to create thread
2123 * objects for the already running tasks on the machine.
2125 * Normally, for a kernel FORK event, we want to clone the parent's
2126 * maps because that is what the kernel just did.
2128 * But when synthesizing, this should not be done. If we do, we end up
2129 * with overlapping maps as we process the synthesized MMAP2 events that
2130 * get delivered shortly thereafter.
2132 * Use the FORK event misc flags in an internal way to signal this
2133 * situation, so we can elide the map clone when appropriate.
2135 if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
2136 do_maps_clone = false;
2138 if (thread == NULL || parent == NULL ||
2139 thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
2140 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
2143 thread__put(thread);
2144 thread__put(parent);
2149 int machine__process_exit_event(struct machine *machine, union perf_event *event,
2150 struct perf_sample *sample __maybe_unused)
2152 struct thread *thread = machine__find_thread(machine,
2157 perf_event__fprintf_task(event, stdout);
2160 thread__put(thread);
2165 int machine__process_event(struct machine *machine, union perf_event *event,
2166 struct perf_sample *sample)
2170 switch (event->header.type) {
2171 case PERF_RECORD_COMM:
2172 ret = machine__process_comm_event(machine, event, sample); break;
2173 case PERF_RECORD_MMAP:
2174 ret = machine__process_mmap_event(machine, event, sample); break;
2175 case PERF_RECORD_NAMESPACES:
2176 ret = machine__process_namespaces_event(machine, event, sample); break;
2177 case PERF_RECORD_CGROUP:
2178 ret = machine__process_cgroup_event(machine, event, sample); break;
2179 case PERF_RECORD_MMAP2:
2180 ret = machine__process_mmap2_event(machine, event, sample); break;
2181 case PERF_RECORD_FORK:
2182 ret = machine__process_fork_event(machine, event, sample); break;
2183 case PERF_RECORD_EXIT:
2184 ret = machine__process_exit_event(machine, event, sample); break;
2185 case PERF_RECORD_LOST:
2186 ret = machine__process_lost_event(machine, event, sample); break;
2187 case PERF_RECORD_AUX:
2188 ret = machine__process_aux_event(machine, event); break;
2189 case PERF_RECORD_ITRACE_START:
2190 ret = machine__process_itrace_start_event(machine, event); break;
2191 case PERF_RECORD_LOST_SAMPLES:
2192 ret = machine__process_lost_samples_event(machine, event, sample); break;
2193 case PERF_RECORD_SWITCH:
2194 case PERF_RECORD_SWITCH_CPU_WIDE:
2195 ret = machine__process_switch_event(machine, event); break;
2196 case PERF_RECORD_KSYMBOL:
2197 ret = machine__process_ksymbol(machine, event, sample); break;
2198 case PERF_RECORD_BPF_EVENT:
2199 ret = machine__process_bpf(machine, event, sample); break;
2200 case PERF_RECORD_TEXT_POKE:
2201 ret = machine__process_text_poke(machine, event, sample); break;
2202 case PERF_RECORD_AUX_OUTPUT_HW_ID:
2203 ret = machine__process_aux_output_hw_id_event(machine, event); break;
2212 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
2214 if (!regexec(regex, sym->name, 0, NULL, 0))
2219 static void ip__resolve_ams(struct thread *thread,
2220 struct addr_map_symbol *ams,
2223 struct addr_location al;
2225 addr_location__init(&al);
2227 * We cannot use the header.misc hint to determine whether a
2228 * branch stack address is user, kernel, guest, hypervisor.
2229 * Branches may straddle the kernel/user/hypervisor boundaries.
2230 * Thus, we have to try consecutively until we find a match
2231 * or else, the symbol is unknown
2233 thread__find_cpumode_addr_location(thread, ip, &al);
2236 ams->al_addr = al.addr;
2237 ams->al_level = al.level;
2238 ams->ms.maps = maps__get(al.maps);
2239 ams->ms.sym = al.sym;
2240 ams->ms.map = map__get(al.map);
2242 ams->data_page_size = 0;
2243 addr_location__exit(&al);
2246 static void ip__resolve_data(struct thread *thread,
2247 u8 m, struct addr_map_symbol *ams,
2248 u64 addr, u64 phys_addr, u64 daddr_page_size)
2250 struct addr_location al;
2252 addr_location__init(&al);
2254 thread__find_symbol(thread, m, addr, &al);
2257 ams->al_addr = al.addr;
2258 ams->al_level = al.level;
2259 ams->ms.maps = maps__get(al.maps);
2260 ams->ms.sym = al.sym;
2261 ams->ms.map = map__get(al.map);
2262 ams->phys_addr = phys_addr;
2263 ams->data_page_size = daddr_page_size;
2264 addr_location__exit(&al);
2267 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
2268 struct addr_location *al)
2270 struct mem_info *mi = mem_info__new();
2275 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
2276 ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
2277 sample->addr, sample->phys_addr,
2278 sample->data_page_size);
2279 mi->data_src.val = sample->data_src;
2284 static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2286 struct map *map = ms->map;
2287 char *srcline = NULL;
2290 if (!map || callchain_param.key == CCKEY_FUNCTION)
2293 dso = map__dso(map);
2294 srcline = srcline__tree_find(&dso->srclines, ip);
2296 bool show_sym = false;
2297 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2299 srcline = get_srcline(dso, map__rip_2objdump(map, ip),
2300 ms->sym, show_sym, show_addr, ip);
2301 srcline__tree_insert(&dso->srclines, ip, srcline);
2312 static int add_callchain_ip(struct thread *thread,
2313 struct callchain_cursor *cursor,
2314 struct symbol **parent,
2315 struct addr_location *root_al,
2319 struct branch_flags *flags,
2320 struct iterations *iter,
2323 struct map_symbol ms = {};
2324 struct addr_location al;
2325 int nr_loop_iter = 0, err = 0;
2326 u64 iter_cycles = 0;
2327 const char *srcline = NULL;
2329 addr_location__init(&al);
2334 thread__find_cpumode_addr_location(thread, ip, &al);
2336 if (ip >= PERF_CONTEXT_MAX) {
2338 case PERF_CONTEXT_HV:
2339 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2341 case PERF_CONTEXT_KERNEL:
2342 *cpumode = PERF_RECORD_MISC_KERNEL;
2344 case PERF_CONTEXT_USER:
2345 *cpumode = PERF_RECORD_MISC_USER;
2348 pr_debug("invalid callchain context: "
2349 "%"PRId64"\n", (s64) ip);
2351 * It seems the callchain is corrupted.
2354 callchain_cursor_reset(cursor);
2360 thread__find_symbol(thread, *cpumode, ip, &al);
2363 if (al.sym != NULL) {
2364 if (perf_hpp_list.parent && !*parent &&
2365 symbol__match_regex(al.sym, &parent_regex))
2367 else if (have_ignore_callees && root_al &&
2368 symbol__match_regex(al.sym, &ignore_callees_regex)) {
2369 /* Treat this symbol as the root,
2370 forgetting its callees. */
2371 addr_location__copy(root_al, &al);
2372 callchain_cursor_reset(cursor);
2376 if (symbol_conf.hide_unresolved && al.sym == NULL)
2380 nr_loop_iter = iter->nr_loop_iter;
2381 iter_cycles = iter->cycles;
2384 ms.maps = maps__get(al.maps);
2385 ms.map = map__get(al.map);
2387 srcline = callchain_srcline(&ms, al.addr);
2388 err = callchain_cursor_append(cursor, ip, &ms,
2389 branch, flags, nr_loop_iter,
2390 iter_cycles, branch_from, srcline);
2392 addr_location__exit(&al);
2398 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2399 struct addr_location *al)
2402 const struct branch_stack *bs = sample->branch_stack;
2403 struct branch_entry *entries = perf_sample__branch_entries(sample);
2404 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2409 for (i = 0; i < bs->nr; i++) {
2410 ip__resolve_ams(al->thread, &bi[i].to, entries[i].to);
2411 ip__resolve_ams(al->thread, &bi[i].from, entries[i].from);
2412 bi[i].flags = entries[i].flags;
2417 static void save_iterations(struct iterations *iter,
2418 struct branch_entry *be, int nr)
2422 iter->nr_loop_iter++;
2425 for (i = 0; i < nr; i++)
2426 iter->cycles += be[i].flags.cycles;
2431 #define NO_ENTRY 0xff
2433 #define PERF_MAX_BRANCH_DEPTH 127
2436 static int remove_loops(struct branch_entry *l, int nr,
2437 struct iterations *iter)
2440 unsigned char chash[CHASHSZ];
2442 memset(chash, NO_ENTRY, sizeof(chash));
2444 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2446 for (i = 0; i < nr; i++) {
2447 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2449 /* no collision handling for now */
2450 if (chash[h] == NO_ENTRY) {
2452 } else if (l[chash[h]].from == l[i].from) {
2453 bool is_loop = true;
2454 /* check if it is a real loop */
2456 for (j = chash[h]; j < i && i + off < nr; j++, off++)
2457 if (l[j].from != l[i + off].from) {
2464 save_iterations(iter + i + off,
2467 memmove(iter + i, iter + i + off,
2470 memmove(l + i, l + i + off,
2481 static int lbr_callchain_add_kernel_ip(struct thread *thread,
2482 struct callchain_cursor *cursor,
2483 struct perf_sample *sample,
2484 struct symbol **parent,
2485 struct addr_location *root_al,
2487 bool callee, int end)
2489 struct ip_callchain *chain = sample->callchain;
2490 u8 cpumode = PERF_RECORD_MISC_USER;
2494 for (i = 0; i < end + 1; i++) {
2495 err = add_callchain_ip(thread, cursor, parent,
2496 root_al, &cpumode, chain->ips[i],
2497 false, NULL, NULL, branch_from);
2504 for (i = end; i >= 0; i--) {
2505 err = add_callchain_ip(thread, cursor, parent,
2506 root_al, &cpumode, chain->ips[i],
2507 false, NULL, NULL, branch_from);
2515 static void save_lbr_cursor_node(struct thread *thread,
2516 struct callchain_cursor *cursor,
2519 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2524 if (cursor->pos == cursor->nr) {
2525 lbr_stitch->prev_lbr_cursor[idx].valid = false;
2530 cursor->curr = cursor->first;
2532 cursor->curr = cursor->curr->next;
2533 memcpy(&lbr_stitch->prev_lbr_cursor[idx], cursor->curr,
2534 sizeof(struct callchain_cursor_node));
2536 lbr_stitch->prev_lbr_cursor[idx].valid = true;
2540 static int lbr_callchain_add_lbr_ip(struct thread *thread,
2541 struct callchain_cursor *cursor,
2542 struct perf_sample *sample,
2543 struct symbol **parent,
2544 struct addr_location *root_al,
2548 struct branch_stack *lbr_stack = sample->branch_stack;
2549 struct branch_entry *entries = perf_sample__branch_entries(sample);
2550 u8 cpumode = PERF_RECORD_MISC_USER;
2551 int lbr_nr = lbr_stack->nr;
2552 struct branch_flags *flags;
2557 * The curr and pos are not used in writing session. They are cleared
2558 * in callchain_cursor_commit() when the writing session is closed.
2559 * Using curr and pos to track the current cursor node.
2561 if (thread__lbr_stitch(thread)) {
2562 cursor->curr = NULL;
2563 cursor->pos = cursor->nr;
2565 cursor->curr = cursor->first;
2566 for (i = 0; i < (int)(cursor->nr - 1); i++)
2567 cursor->curr = cursor->curr->next;
2572 /* Add LBR ip from first entries.to */
2574 flags = &entries[0].flags;
2575 *branch_from = entries[0].from;
2576 err = add_callchain_ip(thread, cursor, parent,
2577 root_al, &cpumode, ip,
2584 * The number of cursor node increases.
2585 * Move the current cursor node.
2586 * But does not need to save current cursor node for entry 0.
2587 * It's impossible to stitch the whole LBRs of previous sample.
2589 if (thread__lbr_stitch(thread) && (cursor->pos != cursor->nr)) {
2591 cursor->curr = cursor->first;
2593 cursor->curr = cursor->curr->next;
2597 /* Add LBR ip from entries.from one by one. */
2598 for (i = 0; i < lbr_nr; i++) {
2599 ip = entries[i].from;
2600 flags = &entries[i].flags;
2601 err = add_callchain_ip(thread, cursor, parent,
2602 root_al, &cpumode, ip,
2607 save_lbr_cursor_node(thread, cursor, i);
2612 /* Add LBR ip from entries.from one by one. */
2613 for (i = lbr_nr - 1; i >= 0; i--) {
2614 ip = entries[i].from;
2615 flags = &entries[i].flags;
2616 err = add_callchain_ip(thread, cursor, parent,
2617 root_al, &cpumode, ip,
2622 save_lbr_cursor_node(thread, cursor, i);
2625 /* Add LBR ip from first entries.to */
2627 flags = &entries[0].flags;
2628 *branch_from = entries[0].from;
2629 err = add_callchain_ip(thread, cursor, parent,
2630 root_al, &cpumode, ip,
2639 static int lbr_callchain_add_stitched_lbr_ip(struct thread *thread,
2640 struct callchain_cursor *cursor)
2642 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2643 struct callchain_cursor_node *cnode;
2644 struct stitch_list *stitch_node;
2647 list_for_each_entry(stitch_node, &lbr_stitch->lists, node) {
2648 cnode = &stitch_node->cursor;
2650 err = callchain_cursor_append(cursor, cnode->ip,
2653 &cnode->branch_flags,
2654 cnode->nr_loop_iter,
2664 static struct stitch_list *get_stitch_node(struct thread *thread)
2666 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2667 struct stitch_list *stitch_node;
2669 if (!list_empty(&lbr_stitch->free_lists)) {
2670 stitch_node = list_first_entry(&lbr_stitch->free_lists,
2671 struct stitch_list, node);
2672 list_del(&stitch_node->node);
2677 return malloc(sizeof(struct stitch_list));
2680 static bool has_stitched_lbr(struct thread *thread,
2681 struct perf_sample *cur,
2682 struct perf_sample *prev,
2683 unsigned int max_lbr,
2686 struct branch_stack *cur_stack = cur->branch_stack;
2687 struct branch_entry *cur_entries = perf_sample__branch_entries(cur);
2688 struct branch_stack *prev_stack = prev->branch_stack;
2689 struct branch_entry *prev_entries = perf_sample__branch_entries(prev);
2690 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2691 int i, j, nr_identical_branches = 0;
2692 struct stitch_list *stitch_node;
2693 u64 cur_base, distance;
2695 if (!cur_stack || !prev_stack)
2698 /* Find the physical index of the base-of-stack for current sample. */
2699 cur_base = max_lbr - cur_stack->nr + cur_stack->hw_idx + 1;
2701 distance = (prev_stack->hw_idx > cur_base) ? (prev_stack->hw_idx - cur_base) :
2702 (max_lbr + prev_stack->hw_idx - cur_base);
2703 /* Previous sample has shorter stack. Nothing can be stitched. */
2704 if (distance + 1 > prev_stack->nr)
2708 * Check if there are identical LBRs between two samples.
2709 * Identical LBRs must have same from, to and flags values. Also,
2710 * they have to be saved in the same LBR registers (same physical
2713 * Starts from the base-of-stack of current sample.
2715 for (i = distance, j = cur_stack->nr - 1; (i >= 0) && (j >= 0); i--, j--) {
2716 if ((prev_entries[i].from != cur_entries[j].from) ||
2717 (prev_entries[i].to != cur_entries[j].to) ||
2718 (prev_entries[i].flags.value != cur_entries[j].flags.value))
2720 nr_identical_branches++;
2723 if (!nr_identical_branches)
2727 * Save the LBRs between the base-of-stack of previous sample
2728 * and the base-of-stack of current sample into lbr_stitch->lists.
2729 * These LBRs will be stitched later.
2731 for (i = prev_stack->nr - 1; i > (int)distance; i--) {
2733 if (!lbr_stitch->prev_lbr_cursor[i].valid)
2736 stitch_node = get_stitch_node(thread);
2740 memcpy(&stitch_node->cursor, &lbr_stitch->prev_lbr_cursor[i],
2741 sizeof(struct callchain_cursor_node));
2744 list_add(&stitch_node->node, &lbr_stitch->lists);
2746 list_add_tail(&stitch_node->node, &lbr_stitch->lists);
2752 static bool alloc_lbr_stitch(struct thread *thread, unsigned int max_lbr)
2754 if (thread__lbr_stitch(thread))
2757 thread__set_lbr_stitch(thread, zalloc(sizeof(struct lbr_stitch)));
2758 if (!thread__lbr_stitch(thread))
2761 thread__lbr_stitch(thread)->prev_lbr_cursor =
2762 calloc(max_lbr + 1, sizeof(struct callchain_cursor_node));
2763 if (!thread__lbr_stitch(thread)->prev_lbr_cursor)
2764 goto free_lbr_stitch;
2766 INIT_LIST_HEAD(&thread__lbr_stitch(thread)->lists);
2767 INIT_LIST_HEAD(&thread__lbr_stitch(thread)->free_lists);
2772 free(thread__lbr_stitch(thread));
2773 thread__set_lbr_stitch(thread, NULL);
2775 pr_warning("Failed to allocate space for stitched LBRs. Disable LBR stitch\n");
2776 thread__set_lbr_stitch_enable(thread, false);
2781 * Resolve LBR callstack chain sample
2783 * 1 on success get LBR callchain information
2784 * 0 no available LBR callchain information, should try fp
2785 * negative error code on other errors.
2787 static int resolve_lbr_callchain_sample(struct thread *thread,
2788 struct callchain_cursor *cursor,
2789 struct perf_sample *sample,
2790 struct symbol **parent,
2791 struct addr_location *root_al,
2793 unsigned int max_lbr)
2795 bool callee = (callchain_param.order == ORDER_CALLEE);
2796 struct ip_callchain *chain = sample->callchain;
2797 int chain_nr = min(max_stack, (int)chain->nr), i;
2798 struct lbr_stitch *lbr_stitch;
2799 bool stitched_lbr = false;
2800 u64 branch_from = 0;
2803 for (i = 0; i < chain_nr; i++) {
2804 if (chain->ips[i] == PERF_CONTEXT_USER)
2808 /* LBR only affects the user callchain */
2812 if (thread__lbr_stitch_enable(thread) && !sample->no_hw_idx &&
2813 (max_lbr > 0) && alloc_lbr_stitch(thread, max_lbr)) {
2814 lbr_stitch = thread__lbr_stitch(thread);
2816 stitched_lbr = has_stitched_lbr(thread, sample,
2817 &lbr_stitch->prev_sample,
2820 if (!stitched_lbr && !list_empty(&lbr_stitch->lists)) {
2821 list_replace_init(&lbr_stitch->lists,
2822 &lbr_stitch->free_lists);
2824 memcpy(&lbr_stitch->prev_sample, sample, sizeof(*sample));
2829 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2830 parent, root_al, branch_from,
2835 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2836 root_al, &branch_from, true);
2841 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2848 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2852 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2853 root_al, &branch_from, false);
2858 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2859 parent, root_al, branch_from,
2867 return (err < 0) ? err : 0;
2870 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2871 struct callchain_cursor *cursor,
2872 struct symbol **parent,
2873 struct addr_location *root_al,
2874 u8 *cpumode, int ent)
2878 while (--ent >= 0) {
2879 u64 ip = chain->ips[ent];
2881 if (ip >= PERF_CONTEXT_MAX) {
2882 err = add_callchain_ip(thread, cursor, parent,
2883 root_al, cpumode, ip,
2884 false, NULL, NULL, 0);
2891 static u64 get_leaf_frame_caller(struct perf_sample *sample,
2892 struct thread *thread, int usr_idx)
2894 if (machine__normalized_is(maps__machine(thread__maps(thread)), "arm64"))
2895 return get_leaf_frame_caller_aarch64(sample, thread, usr_idx);
2900 static int thread__resolve_callchain_sample(struct thread *thread,
2901 struct callchain_cursor *cursor,
2902 struct evsel *evsel,
2903 struct perf_sample *sample,
2904 struct symbol **parent,
2905 struct addr_location *root_al,
2908 struct branch_stack *branch = sample->branch_stack;
2909 struct branch_entry *entries = perf_sample__branch_entries(sample);
2910 struct ip_callchain *chain = sample->callchain;
2912 u8 cpumode = PERF_RECORD_MISC_USER;
2913 int i, j, err, nr_entries, usr_idx;
2916 u64 leaf_frame_caller;
2919 chain_nr = chain->nr;
2921 if (evsel__has_branch_callstack(evsel)) {
2922 struct perf_env *env = evsel__env(evsel);
2924 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2926 !env ? 0 : env->max_branches);
2928 return (err < 0) ? err : 0;
2932 * Based on DWARF debug information, some architectures skip
2933 * a callchain entry saved by the kernel.
2935 skip_idx = arch_skip_callchain_idx(thread, chain);
2938 * Add branches to call stack for easier browsing. This gives
2939 * more context for a sample than just the callers.
2941 * This uses individual histograms of paths compared to the
2942 * aggregated histograms the normal LBR mode uses.
2944 * Limitations for now:
2945 * - No extra filters
2946 * - No annotations (should annotate somehow)
2949 if (branch && callchain_param.branch_callstack) {
2950 int nr = min(max_stack, (int)branch->nr);
2951 struct branch_entry be[nr];
2952 struct iterations iter[nr];
2954 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2955 pr_warning("corrupted branch chain. skipping...\n");
2959 for (i = 0; i < nr; i++) {
2960 if (callchain_param.order == ORDER_CALLEE) {
2967 * Check for overlap into the callchain.
2968 * The return address is one off compared to
2969 * the branch entry. To adjust for this
2970 * assume the calling instruction is not longer
2973 if (i == skip_idx ||
2974 chain->ips[first_call] >= PERF_CONTEXT_MAX)
2976 else if (be[i].from < chain->ips[first_call] &&
2977 be[i].from >= chain->ips[first_call] - 8)
2980 be[i] = entries[branch->nr - i - 1];
2983 memset(iter, 0, sizeof(struct iterations) * nr);
2984 nr = remove_loops(be, nr, iter);
2986 for (i = 0; i < nr; i++) {
2987 err = add_callchain_ip(thread, cursor, parent,
2994 err = add_callchain_ip(thread, cursor, parent, root_al,
3011 if (chain && callchain_param.order != ORDER_CALLEE) {
3012 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
3013 &cpumode, chain->nr - first_call);
3015 return (err < 0) ? err : 0;
3017 for (i = first_call, nr_entries = 0;
3018 i < chain_nr && nr_entries < max_stack; i++) {
3021 if (callchain_param.order == ORDER_CALLEE)
3024 j = chain->nr - i - 1;
3026 #ifdef HAVE_SKIP_CALLCHAIN_IDX
3031 if (ip < PERF_CONTEXT_MAX)
3033 else if (callchain_param.order != ORDER_CALLEE) {
3034 err = find_prev_cpumode(chain, thread, cursor, parent,
3035 root_al, &cpumode, j);
3037 return (err < 0) ? err : 0;
3042 * PERF_CONTEXT_USER allows us to locate where the user stack ends.
3043 * Depending on callchain_param.order and the position of PERF_CONTEXT_USER,
3044 * the index will be different in order to add the missing frame
3045 * at the right place.
3048 usr_idx = callchain_param.order == ORDER_CALLEE ? j-2 : j-1;
3050 if (usr_idx >= 0 && chain->ips[usr_idx] == PERF_CONTEXT_USER) {
3052 leaf_frame_caller = get_leaf_frame_caller(sample, thread, usr_idx);
3055 * check if leaf_frame_Caller != ip to not add the same
3059 if (leaf_frame_caller && leaf_frame_caller != ip) {
3061 err = add_callchain_ip(thread, cursor, parent,
3062 root_al, &cpumode, leaf_frame_caller,
3063 false, NULL, NULL, 0);
3065 return (err < 0) ? err : 0;
3069 err = add_callchain_ip(thread, cursor, parent,
3070 root_al, &cpumode, ip,
3071 false, NULL, NULL, 0);
3074 return (err < 0) ? err : 0;
3080 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
3082 struct symbol *sym = ms->sym;
3083 struct map *map = ms->map;
3084 struct inline_node *inline_node;
3085 struct inline_list *ilist;
3089 struct map_symbol ilist_ms;
3091 if (!symbol_conf.inline_name || !map || !sym)
3094 addr = map__dso_map_ip(map, ip);
3095 addr = map__rip_2objdump(map, addr);
3096 dso = map__dso(map);
3098 inline_node = inlines__tree_find(&dso->inlined_nodes, addr);
3100 inline_node = dso__parse_addr_inlines(dso, addr, sym);
3103 inlines__tree_insert(&dso->inlined_nodes, inline_node);
3106 ilist_ms = (struct map_symbol) {
3107 .maps = maps__get(ms->maps),
3108 .map = map__get(map),
3110 list_for_each_entry(ilist, &inline_node->val, list) {
3111 ilist_ms.sym = ilist->symbol;
3112 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
3113 NULL, 0, 0, 0, ilist->srcline);
3118 map__put(ilist_ms.map);
3119 maps__put(ilist_ms.maps);
3124 static int unwind_entry(struct unwind_entry *entry, void *arg)
3126 struct callchain_cursor *cursor = arg;
3127 const char *srcline = NULL;
3128 u64 addr = entry->ip;
3130 if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
3133 if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
3137 * Convert entry->ip from a virtual address to an offset in
3138 * its corresponding binary.
3141 addr = map__dso_map_ip(entry->ms.map, entry->ip);
3143 srcline = callchain_srcline(&entry->ms, addr);
3144 return callchain_cursor_append(cursor, entry->ip, &entry->ms,
3145 false, NULL, 0, 0, 0, srcline);
3148 static int thread__resolve_callchain_unwind(struct thread *thread,
3149 struct callchain_cursor *cursor,
3150 struct evsel *evsel,
3151 struct perf_sample *sample,
3154 /* Can we do dwarf post unwind? */
3155 if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
3156 (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
3159 /* Bail out if nothing was captured. */
3160 if ((!sample->user_regs.regs) ||
3161 (!sample->user_stack.size))
3164 return unwind__get_entries(unwind_entry, cursor,
3165 thread, sample, max_stack, false);
3168 int thread__resolve_callchain(struct thread *thread,
3169 struct callchain_cursor *cursor,
3170 struct evsel *evsel,
3171 struct perf_sample *sample,
3172 struct symbol **parent,
3173 struct addr_location *root_al,
3181 callchain_cursor_reset(cursor);
3183 if (callchain_param.order == ORDER_CALLEE) {
3184 ret = thread__resolve_callchain_sample(thread, cursor,
3190 ret = thread__resolve_callchain_unwind(thread, cursor,
3194 ret = thread__resolve_callchain_unwind(thread, cursor,
3199 ret = thread__resolve_callchain_sample(thread, cursor,
3208 int machine__for_each_thread(struct machine *machine,
3209 int (*fn)(struct thread *thread, void *p),
3212 struct threads *threads;
3217 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3218 threads = &machine->threads[i];
3219 for (nd = rb_first_cached(&threads->entries); nd;
3221 struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
3223 rc = fn(trb->thread, priv);
3231 int machines__for_each_thread(struct machines *machines,
3232 int (*fn)(struct thread *thread, void *p),
3238 rc = machine__for_each_thread(&machines->host, fn, priv);
3242 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
3243 struct machine *machine = rb_entry(nd, struct machine, rb_node);
3245 rc = machine__for_each_thread(machine, fn, priv);
3252 pid_t machine__get_current_tid(struct machine *machine, int cpu)
3254 if (cpu < 0 || (size_t)cpu >= machine->current_tid_sz)
3257 return machine->current_tid[cpu];
3260 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
3263 struct thread *thread;
3264 const pid_t init_val = -1;
3269 if (realloc_array_as_needed(machine->current_tid,
3270 machine->current_tid_sz,
3275 machine->current_tid[cpu] = tid;
3277 thread = machine__findnew_thread(machine, pid, tid);
3281 thread__set_cpu(thread, cpu);
3282 thread__put(thread);
3288 * Compares the raw arch string. N.B. see instead perf_env__arch() or
3289 * machine__normalized_is() if a normalized arch is needed.
3291 bool machine__is(struct machine *machine, const char *arch)
3293 return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
3296 bool machine__normalized_is(struct machine *machine, const char *arch)
3298 return machine && !strcmp(perf_env__arch(machine->env), arch);
3301 int machine__nr_cpus_avail(struct machine *machine)
3303 return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
3306 int machine__get_kernel_start(struct machine *machine)
3308 struct map *map = machine__kernel_map(machine);
3312 * The only addresses above 2^63 are kernel addresses of a 64-bit
3313 * kernel. Note that addresses are unsigned so that on a 32-bit system
3314 * all addresses including kernel addresses are less than 2^32. In
3315 * that case (32-bit system), if the kernel mapping is unknown, all
3316 * addresses will be assumed to be in user space - see
3317 * machine__kernel_ip().
3319 machine->kernel_start = 1ULL << 63;
3321 err = map__load(map);
3323 * On x86_64, PTI entry trampolines are less than the
3324 * start of kernel text, but still above 2^63. So leave
3325 * kernel_start = 1ULL << 63 for x86_64.
3327 if (!err && !machine__is(machine, "x86_64"))
3328 machine->kernel_start = map__start(map);
3333 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
3335 u8 addr_cpumode = cpumode;
3338 if (!machine->single_address_space)
3341 kernel_ip = machine__kernel_ip(machine, addr);
3343 case PERF_RECORD_MISC_KERNEL:
3344 case PERF_RECORD_MISC_USER:
3345 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
3346 PERF_RECORD_MISC_USER;
3348 case PERF_RECORD_MISC_GUEST_KERNEL:
3349 case PERF_RECORD_MISC_GUEST_USER:
3350 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
3351 PERF_RECORD_MISC_GUEST_USER;
3357 return addr_cpumode;
3360 struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id)
3362 return dsos__findnew_id(&machine->dsos, filename, id);
3365 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
3367 return machine__findnew_dso_id(machine, filename, NULL);
3370 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
3372 struct machine *machine = vmachine;
3374 struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
3379 *modp = __map__is_kmodule(map) ? (char *)map__dso(map)->short_name : NULL;
3380 *addrp = map__unmap_ip(map, sym->start);
3384 int machine__for_each_dso(struct machine *machine, machine__dso_t fn, void *priv)
3389 list_for_each_entry(pos, &machine->dsos.head, node) {
3390 if (fn(pos, machine, priv))
3396 int machine__for_each_kernel_map(struct machine *machine, machine__map_t fn, void *priv)
3398 struct maps *maps = machine__kernel_maps(machine);
3399 struct map_rb_node *pos;
3402 maps__for_each_entry(maps, pos) {
3403 err = fn(pos->map, priv);
3411 bool machine__is_lock_function(struct machine *machine, u64 addr)
3413 if (!machine->sched.text_start) {
3415 struct symbol *sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_start", &kmap);
3418 /* to avoid retry */
3419 machine->sched.text_start = 1;
3423 machine->sched.text_start = map__unmap_ip(kmap, sym->start);
3425 /* should not fail from here */
3426 sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_end", &kmap);
3427 machine->sched.text_end = map__unmap_ip(kmap, sym->start);
3429 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_start", &kmap);
3430 machine->lock.text_start = map__unmap_ip(kmap, sym->start);
3432 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_end", &kmap);
3433 machine->lock.text_end = map__unmap_ip(kmap, sym->start);
3436 /* failed to get kernel symbols */
3437 if (machine->sched.text_start == 1)
3440 /* mutex and rwsem functions are in sched text section */
3441 if (machine->sched.text_start <= addr && addr < machine->sched.text_end)
3444 /* spinlock functions are in lock text section */
3445 if (machine->lock.text_start <= addr && addr < machine->lock.text_end)