1 // SPDX-License-Identifier: GPL-2.0
16 #include "map_symbol.h"
18 #include "mem-events.h"
28 #include <sys/types.h>
32 #include "linux/hash.h"
34 #include "bpf-event.h"
35 #include <internal/lib.h> // page_size
37 #include <linux/ctype.h>
38 #include <symbol/kallsyms.h>
39 #include <linux/mman.h>
40 #include <linux/string.h>
41 #include <linux/zalloc.h>
43 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
45 static struct dso *machine__kernel_dso(struct machine *machine)
47 return machine->vmlinux_map->dso;
50 static void dsos__init(struct dsos *dsos)
52 INIT_LIST_HEAD(&dsos->head);
54 init_rwsem(&dsos->lock);
57 static void machine__threads_init(struct machine *machine)
61 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
62 struct threads *threads = &machine->threads[i];
63 threads->entries = RB_ROOT_CACHED;
64 init_rwsem(&threads->lock);
66 INIT_LIST_HEAD(&threads->dead);
67 threads->last_match = NULL;
71 static int machine__set_mmap_name(struct machine *machine)
73 if (machine__is_host(machine))
74 machine->mmap_name = strdup("[kernel.kallsyms]");
75 else if (machine__is_default_guest(machine))
76 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
77 else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
79 machine->mmap_name = NULL;
81 return machine->mmap_name ? 0 : -ENOMEM;
84 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
88 memset(machine, 0, sizeof(*machine));
89 map_groups__init(&machine->kmaps, machine);
90 RB_CLEAR_NODE(&machine->rb_node);
91 dsos__init(&machine->dsos);
93 machine__threads_init(machine);
95 machine->vdso_info = NULL;
100 machine->id_hdr_size = 0;
101 machine->kptr_restrict_warned = false;
102 machine->comm_exec = false;
103 machine->kernel_start = 0;
104 machine->vmlinux_map = NULL;
106 machine->root_dir = strdup(root_dir);
107 if (machine->root_dir == NULL)
110 if (machine__set_mmap_name(machine))
113 if (pid != HOST_KERNEL_ID) {
114 struct thread *thread = machine__findnew_thread(machine, -1,
121 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
122 thread__set_comm(thread, comm, 0);
126 machine->current_tid = NULL;
131 zfree(&machine->root_dir);
132 zfree(&machine->mmap_name);
137 struct machine *machine__new_host(void)
139 struct machine *machine = malloc(sizeof(*machine));
141 if (machine != NULL) {
142 machine__init(machine, "", HOST_KERNEL_ID);
144 if (machine__create_kernel_maps(machine) < 0)
154 struct machine *machine__new_kallsyms(void)
156 struct machine *machine = machine__new_host();
159 * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
160 * ask for not using the kcore parsing code, once this one is fixed
161 * to create a map per module.
163 if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
164 machine__delete(machine);
171 static void dsos__purge(struct dsos *dsos)
175 down_write(&dsos->lock);
177 list_for_each_entry_safe(pos, n, &dsos->head, node) {
178 RB_CLEAR_NODE(&pos->rb_node);
180 list_del_init(&pos->node);
184 up_write(&dsos->lock);
187 static void dsos__exit(struct dsos *dsos)
190 exit_rwsem(&dsos->lock);
193 void machine__delete_threads(struct machine *machine)
198 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
199 struct threads *threads = &machine->threads[i];
200 down_write(&threads->lock);
201 nd = rb_first_cached(&threads->entries);
203 struct thread *t = rb_entry(nd, struct thread, rb_node);
206 __machine__remove_thread(machine, t, false);
208 up_write(&threads->lock);
212 void machine__exit(struct machine *machine)
219 machine__destroy_kernel_maps(machine);
220 map_groups__exit(&machine->kmaps);
221 dsos__exit(&machine->dsos);
222 machine__exit_vdso(machine);
223 zfree(&machine->root_dir);
224 zfree(&machine->mmap_name);
225 zfree(&machine->current_tid);
227 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
228 struct threads *threads = &machine->threads[i];
229 struct thread *thread, *n;
231 * Forget about the dead, at this point whatever threads were
232 * left in the dead lists better have a reference count taken
233 * by who is using them, and then, when they drop those references
234 * and it finally hits zero, thread__put() will check and see that
235 * its not in the dead threads list and will not try to remove it
236 * from there, just calling thread__delete() straight away.
238 list_for_each_entry_safe(thread, n, &threads->dead, node)
239 list_del_init(&thread->node);
241 exit_rwsem(&threads->lock);
245 void machine__delete(struct machine *machine)
248 machine__exit(machine);
253 void machines__init(struct machines *machines)
255 machine__init(&machines->host, "", HOST_KERNEL_ID);
256 machines->guests = RB_ROOT_CACHED;
259 void machines__exit(struct machines *machines)
261 machine__exit(&machines->host);
265 struct machine *machines__add(struct machines *machines, pid_t pid,
266 const char *root_dir)
268 struct rb_node **p = &machines->guests.rb_root.rb_node;
269 struct rb_node *parent = NULL;
270 struct machine *pos, *machine = malloc(sizeof(*machine));
271 bool leftmost = true;
276 if (machine__init(machine, root_dir, pid) != 0) {
283 pos = rb_entry(parent, struct machine, rb_node);
292 rb_link_node(&machine->rb_node, parent, p);
293 rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
298 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
302 machines->host.comm_exec = comm_exec;
304 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
305 struct machine *machine = rb_entry(nd, struct machine, rb_node);
307 machine->comm_exec = comm_exec;
311 struct machine *machines__find(struct machines *machines, pid_t pid)
313 struct rb_node **p = &machines->guests.rb_root.rb_node;
314 struct rb_node *parent = NULL;
315 struct machine *machine;
316 struct machine *default_machine = NULL;
318 if (pid == HOST_KERNEL_ID)
319 return &machines->host;
323 machine = rb_entry(parent, struct machine, rb_node);
324 if (pid < machine->pid)
326 else if (pid > machine->pid)
331 default_machine = machine;
334 return default_machine;
337 struct machine *machines__findnew(struct machines *machines, pid_t pid)
340 const char *root_dir = "";
341 struct machine *machine = machines__find(machines, pid);
343 if (machine && (machine->pid == pid))
346 if ((pid != HOST_KERNEL_ID) &&
347 (pid != DEFAULT_GUEST_KERNEL_ID) &&
348 (symbol_conf.guestmount)) {
349 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
350 if (access(path, R_OK)) {
351 static struct strlist *seen;
354 seen = strlist__new(NULL, NULL);
356 if (!strlist__has_entry(seen, path)) {
357 pr_err("Can't access file %s\n", path);
358 strlist__add(seen, path);
366 machine = machines__add(machines, pid, root_dir);
371 void machines__process_guests(struct machines *machines,
372 machine__process_t process, void *data)
376 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
377 struct machine *pos = rb_entry(nd, struct machine, rb_node);
382 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
384 struct rb_node *node;
385 struct machine *machine;
387 machines->host.id_hdr_size = id_hdr_size;
389 for (node = rb_first_cached(&machines->guests); node;
390 node = rb_next(node)) {
391 machine = rb_entry(node, struct machine, rb_node);
392 machine->id_hdr_size = id_hdr_size;
398 static void machine__update_thread_pid(struct machine *machine,
399 struct thread *th, pid_t pid)
401 struct thread *leader;
403 if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
408 if (th->pid_ == th->tid)
411 leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
416 leader->mg = map_groups__new(machine);
421 if (th->mg == leader->mg)
426 * Maps are created from MMAP events which provide the pid and
427 * tid. Consequently there never should be any maps on a thread
428 * with an unknown pid. Just print an error if there are.
430 if (!map_groups__empty(th->mg))
431 pr_err("Discarding thread maps for %d:%d\n",
433 map_groups__put(th->mg);
436 th->mg = map_groups__get(leader->mg);
441 pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
446 * Front-end cache - TID lookups come in blocks,
447 * so most of the time we dont have to look up
450 static struct thread*
451 __threads__get_last_match(struct threads *threads, struct machine *machine,
456 th = threads->last_match;
458 if (th->tid == tid) {
459 machine__update_thread_pid(machine, th, pid);
460 return thread__get(th);
463 threads->last_match = NULL;
469 static struct thread*
470 threads__get_last_match(struct threads *threads, struct machine *machine,
473 struct thread *th = NULL;
475 if (perf_singlethreaded)
476 th = __threads__get_last_match(threads, machine, pid, tid);
482 __threads__set_last_match(struct threads *threads, struct thread *th)
484 threads->last_match = th;
488 threads__set_last_match(struct threads *threads, struct thread *th)
490 if (perf_singlethreaded)
491 __threads__set_last_match(threads, th);
495 * Caller must eventually drop thread->refcnt returned with a successful
496 * lookup/new thread inserted.
498 static struct thread *____machine__findnew_thread(struct machine *machine,
499 struct threads *threads,
500 pid_t pid, pid_t tid,
503 struct rb_node **p = &threads->entries.rb_root.rb_node;
504 struct rb_node *parent = NULL;
506 bool leftmost = true;
508 th = threads__get_last_match(threads, machine, pid, tid);
514 th = rb_entry(parent, struct thread, rb_node);
516 if (th->tid == tid) {
517 threads__set_last_match(threads, th);
518 machine__update_thread_pid(machine, th, pid);
519 return thread__get(th);
533 th = thread__new(pid, tid);
535 rb_link_node(&th->rb_node, parent, p);
536 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
539 * We have to initialize map_groups separately
540 * after rb tree is updated.
542 * The reason is that we call machine__findnew_thread
543 * within thread__init_map_groups to find the thread
544 * leader and that would screwed the rb tree.
546 if (thread__init_map_groups(th, machine)) {
547 rb_erase_cached(&th->rb_node, &threads->entries);
548 RB_CLEAR_NODE(&th->rb_node);
553 * It is now in the rbtree, get a ref
556 threads__set_last_match(threads, th);
563 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
565 return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
568 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
571 struct threads *threads = machine__threads(machine, tid);
574 down_write(&threads->lock);
575 th = __machine__findnew_thread(machine, pid, tid);
576 up_write(&threads->lock);
580 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
583 struct threads *threads = machine__threads(machine, tid);
586 down_read(&threads->lock);
587 th = ____machine__findnew_thread(machine, threads, pid, tid, false);
588 up_read(&threads->lock);
592 struct comm *machine__thread_exec_comm(struct machine *machine,
593 struct thread *thread)
595 if (machine->comm_exec)
596 return thread__exec_comm(thread);
598 return thread__comm(thread);
601 int machine__process_comm_event(struct machine *machine, union perf_event *event,
602 struct perf_sample *sample)
604 struct thread *thread = machine__findnew_thread(machine,
607 bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
611 machine->comm_exec = true;
614 perf_event__fprintf_comm(event, stdout);
616 if (thread == NULL ||
617 __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
618 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
627 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
628 union perf_event *event,
629 struct perf_sample *sample __maybe_unused)
631 struct thread *thread = machine__findnew_thread(machine,
632 event->namespaces.pid,
633 event->namespaces.tid);
636 WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
637 "\nWARNING: kernel seems to support more namespaces than perf"
638 " tool.\nTry updating the perf tool..\n\n");
640 WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
641 "\nWARNING: perf tool seems to support more namespaces than"
642 " the kernel.\nTry updating the kernel..\n\n");
645 perf_event__fprintf_namespaces(event, stdout);
647 if (thread == NULL ||
648 thread__set_namespaces(thread, sample->time, &event->namespaces)) {
649 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
658 int machine__process_lost_event(struct machine *machine __maybe_unused,
659 union perf_event *event, struct perf_sample *sample __maybe_unused)
661 dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
662 event->lost.id, event->lost.lost);
666 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
667 union perf_event *event, struct perf_sample *sample)
669 dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
670 sample->id, event->lost_samples.lost);
674 static struct dso *machine__findnew_module_dso(struct machine *machine,
676 const char *filename)
680 down_write(&machine->dsos.lock);
682 dso = __dsos__find(&machine->dsos, m->name, true);
684 dso = __dsos__addnew(&machine->dsos, m->name);
688 dso__set_module_info(dso, m, machine);
689 dso__set_long_name(dso, strdup(filename), true);
694 up_write(&machine->dsos.lock);
698 int machine__process_aux_event(struct machine *machine __maybe_unused,
699 union perf_event *event)
702 perf_event__fprintf_aux(event, stdout);
706 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
707 union perf_event *event)
710 perf_event__fprintf_itrace_start(event, stdout);
714 int machine__process_switch_event(struct machine *machine __maybe_unused,
715 union perf_event *event)
718 perf_event__fprintf_switch(event, stdout);
722 static int machine__process_ksymbol_register(struct machine *machine,
723 union perf_event *event,
724 struct perf_sample *sample __maybe_unused)
729 map = map_groups__find(&machine->kmaps, event->ksymbol.addr);
731 map = dso__new_map(event->ksymbol.name);
735 map->start = event->ksymbol.addr;
736 map->end = map->start + event->ksymbol.len;
737 map_groups__insert(&machine->kmaps, map);
740 sym = symbol__new(map->map_ip(map, map->start),
742 0, 0, event->ksymbol.name);
745 dso__insert_symbol(map->dso, sym);
749 static int machine__process_ksymbol_unregister(struct machine *machine,
750 union perf_event *event,
751 struct perf_sample *sample __maybe_unused)
755 map = map_groups__find(&machine->kmaps, event->ksymbol.addr);
757 map_groups__remove(&machine->kmaps, map);
762 int machine__process_ksymbol(struct machine *machine __maybe_unused,
763 union perf_event *event,
764 struct perf_sample *sample)
767 perf_event__fprintf_ksymbol(event, stdout);
769 if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
770 return machine__process_ksymbol_unregister(machine, event,
772 return machine__process_ksymbol_register(machine, event, sample);
775 static void dso__adjust_kmod_long_name(struct dso *dso, const char *filename)
777 const char *dup_filename;
779 if (!filename || !dso || !dso->long_name)
781 if (dso->long_name[0] != '[')
783 if (!strchr(filename, '/'))
786 dup_filename = strdup(filename);
790 dso__set_long_name(dso, dup_filename, true);
793 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
794 const char *filename)
796 struct map *map = NULL;
797 struct dso *dso = NULL;
800 if (kmod_path__parse_name(&m, filename))
803 map = map_groups__find_by_name(&machine->kmaps, m.name);
806 * If the map's dso is an offline module, give dso__load()
807 * a chance to find the file path of that module by fixing
810 dso__adjust_kmod_long_name(map->dso, filename);
814 dso = machine__findnew_module_dso(machine, &m, filename);
818 map = map__new2(start, dso);
822 map_groups__insert(&machine->kmaps, map);
824 /* Put the map here because map_groups__insert alread got it */
827 /* put the dso here, corresponding to machine__findnew_module_dso */
833 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
836 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
838 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
839 struct machine *pos = rb_entry(nd, struct machine, rb_node);
840 ret += __dsos__fprintf(&pos->dsos.head, fp);
846 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
847 bool (skip)(struct dso *dso, int parm), int parm)
849 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
852 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
853 bool (skip)(struct dso *dso, int parm), int parm)
856 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
858 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
859 struct machine *pos = rb_entry(nd, struct machine, rb_node);
860 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
865 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
869 struct dso *kdso = machine__kernel_dso(machine);
871 if (kdso->has_build_id) {
872 char filename[PATH_MAX];
873 if (dso__build_id_filename(kdso, filename, sizeof(filename),
875 printed += fprintf(fp, "[0] %s\n", filename);
878 for (i = 0; i < vmlinux_path__nr_entries; ++i)
879 printed += fprintf(fp, "[%d] %s\n",
880 i + kdso->has_build_id, vmlinux_path[i]);
885 size_t machine__fprintf(struct machine *machine, FILE *fp)
891 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
892 struct threads *threads = &machine->threads[i];
894 down_read(&threads->lock);
896 ret = fprintf(fp, "Threads: %u\n", threads->nr);
898 for (nd = rb_first_cached(&threads->entries); nd;
900 struct thread *pos = rb_entry(nd, struct thread, rb_node);
902 ret += thread__fprintf(pos, fp);
905 up_read(&threads->lock);
910 static struct dso *machine__get_kernel(struct machine *machine)
912 const char *vmlinux_name = machine->mmap_name;
915 if (machine__is_host(machine)) {
916 if (symbol_conf.vmlinux_name)
917 vmlinux_name = symbol_conf.vmlinux_name;
919 kernel = machine__findnew_kernel(machine, vmlinux_name,
920 "[kernel]", DSO_TYPE_KERNEL);
922 if (symbol_conf.default_guest_vmlinux_name)
923 vmlinux_name = symbol_conf.default_guest_vmlinux_name;
925 kernel = machine__findnew_kernel(machine, vmlinux_name,
927 DSO_TYPE_GUEST_KERNEL);
930 if (kernel != NULL && (!kernel->has_build_id))
931 dso__read_running_kernel_build_id(kernel, machine);
936 struct process_args {
940 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
943 if (machine__is_default_guest(machine))
944 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
946 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
949 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
951 /* Figure out the start address of kernel map from /proc/kallsyms.
952 * Returns the name of the start symbol in *symbol_name. Pass in NULL as
953 * symbol_name if it's not that important.
955 static int machine__get_running_kernel_start(struct machine *machine,
956 const char **symbol_name,
957 u64 *start, u64 *end)
959 char filename[PATH_MAX];
964 machine__get_kallsyms_filename(machine, filename, PATH_MAX);
966 if (symbol__restricted_filename(filename, "/proc/kallsyms"))
969 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
970 err = kallsyms__get_function_start(filename, name, &addr);
983 err = kallsyms__get_function_start(filename, "_etext", &addr);
990 int machine__create_extra_kernel_map(struct machine *machine,
992 struct extra_kernel_map *xm)
997 map = map__new2(xm->start, kernel);
1002 map->pgoff = xm->pgoff;
1004 kmap = map__kmap(map);
1006 kmap->kmaps = &machine->kmaps;
1007 strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1009 map_groups__insert(&machine->kmaps, map);
1011 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1012 kmap->name, map->start, map->end);
1019 static u64 find_entry_trampoline(struct dso *dso)
1021 /* Duplicates are removed so lookup all aliases */
1022 const char *syms[] = {
1023 "_entry_trampoline",
1024 "__entry_trampoline_start",
1025 "entry_SYSCALL_64_trampoline",
1027 struct symbol *sym = dso__first_symbol(dso);
1030 for (; sym; sym = dso__next_symbol(sym)) {
1031 if (sym->binding != STB_GLOBAL)
1033 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1034 if (!strcmp(sym->name, syms[i]))
1043 * These values can be used for kernels that do not have symbols for the entry
1044 * trampolines in kallsyms.
1046 #define X86_64_CPU_ENTRY_AREA_PER_CPU 0xfffffe0000000000ULL
1047 #define X86_64_CPU_ENTRY_AREA_SIZE 0x2c000
1048 #define X86_64_ENTRY_TRAMPOLINE 0x6000
1050 /* Map x86_64 PTI entry trampolines */
1051 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1054 struct map_groups *kmaps = &machine->kmaps;
1055 struct maps *maps = &kmaps->maps;
1056 int nr_cpus_avail, cpu;
1062 * In the vmlinux case, pgoff is a virtual address which must now be
1063 * mapped to a vmlinux offset.
1065 maps__for_each_entry(maps, map) {
1066 struct kmap *kmap = __map__kmap(map);
1067 struct map *dest_map;
1069 if (!kmap || !is_entry_trampoline(kmap->name))
1072 dest_map = map_groups__find(kmaps, map->pgoff);
1073 if (dest_map != map)
1074 map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1077 if (found || machine->trampolines_mapped)
1080 pgoff = find_entry_trampoline(kernel);
1084 nr_cpus_avail = machine__nr_cpus_avail(machine);
1086 /* Add a 1 page map for each CPU's entry trampoline */
1087 for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1088 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1089 cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1090 X86_64_ENTRY_TRAMPOLINE;
1091 struct extra_kernel_map xm = {
1093 .end = va + page_size,
1097 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1099 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1103 machine->trampolines_mapped = nr_cpus_avail;
1108 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1109 struct dso *kernel __maybe_unused)
1115 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1120 /* In case of renewal the kernel map, destroy previous one */
1121 machine__destroy_kernel_maps(machine);
1123 machine->vmlinux_map = map__new2(0, kernel);
1124 if (machine->vmlinux_map == NULL)
1127 machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1128 map = machine__kernel_map(machine);
1129 kmap = map__kmap(map);
1133 kmap->kmaps = &machine->kmaps;
1134 map_groups__insert(&machine->kmaps, map);
1139 void machine__destroy_kernel_maps(struct machine *machine)
1142 struct map *map = machine__kernel_map(machine);
1147 kmap = map__kmap(map);
1148 map_groups__remove(&machine->kmaps, map);
1149 if (kmap && kmap->ref_reloc_sym) {
1150 zfree((char **)&kmap->ref_reloc_sym->name);
1151 zfree(&kmap->ref_reloc_sym);
1154 map__zput(machine->vmlinux_map);
1157 int machines__create_guest_kernel_maps(struct machines *machines)
1160 struct dirent **namelist = NULL;
1162 char path[PATH_MAX];
1166 if (symbol_conf.default_guest_vmlinux_name ||
1167 symbol_conf.default_guest_modules ||
1168 symbol_conf.default_guest_kallsyms) {
1169 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1172 if (symbol_conf.guestmount) {
1173 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1176 for (i = 0; i < items; i++) {
1177 if (!isdigit(namelist[i]->d_name[0])) {
1178 /* Filter out . and .. */
1181 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1182 if ((*endp != '\0') ||
1183 (endp == namelist[i]->d_name) ||
1184 (errno == ERANGE)) {
1185 pr_debug("invalid directory (%s). Skipping.\n",
1186 namelist[i]->d_name);
1189 sprintf(path, "%s/%s/proc/kallsyms",
1190 symbol_conf.guestmount,
1191 namelist[i]->d_name);
1192 ret = access(path, R_OK);
1194 pr_debug("Can't access file %s\n", path);
1197 machines__create_kernel_maps(machines, pid);
1206 void machines__destroy_kernel_maps(struct machines *machines)
1208 struct rb_node *next = rb_first_cached(&machines->guests);
1210 machine__destroy_kernel_maps(&machines->host);
1213 struct machine *pos = rb_entry(next, struct machine, rb_node);
1215 next = rb_next(&pos->rb_node);
1216 rb_erase_cached(&pos->rb_node, &machines->guests);
1217 machine__delete(pos);
1221 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1223 struct machine *machine = machines__findnew(machines, pid);
1225 if (machine == NULL)
1228 return machine__create_kernel_maps(machine);
1231 int machine__load_kallsyms(struct machine *machine, const char *filename)
1233 struct map *map = machine__kernel_map(machine);
1234 int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1237 dso__set_loaded(map->dso);
1239 * Since /proc/kallsyms will have multiple sessions for the
1240 * kernel, with modules between them, fixup the end of all
1243 map_groups__fixup_end(&machine->kmaps);
1249 int machine__load_vmlinux_path(struct machine *machine)
1251 struct map *map = machine__kernel_map(machine);
1252 int ret = dso__load_vmlinux_path(map->dso, map);
1255 dso__set_loaded(map->dso);
1260 static char *get_kernel_version(const char *root_dir)
1262 char version[PATH_MAX];
1265 const char *prefix = "Linux version ";
1267 sprintf(version, "%s/proc/version", root_dir);
1268 file = fopen(version, "r");
1272 tmp = fgets(version, sizeof(version), file);
1277 name = strstr(version, prefix);
1280 name += strlen(prefix);
1281 tmp = strchr(name, ' ');
1285 return strdup(name);
1288 static bool is_kmod_dso(struct dso *dso)
1290 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1291 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1294 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
1295 struct kmod_path *m)
1298 struct map *map = map_groups__find_by_name(mg, m->name);
1303 long_name = strdup(path);
1304 if (long_name == NULL)
1307 dso__set_long_name(map->dso, long_name, true);
1308 dso__kernel_module_get_build_id(map->dso, "");
1311 * Full name could reveal us kmod compression, so
1312 * we need to update the symtab_type if needed.
1314 if (m->comp && is_kmod_dso(map->dso)) {
1315 map->dso->symtab_type++;
1316 map->dso->comp = m->comp;
1322 static int map_groups__set_modules_path_dir(struct map_groups *mg,
1323 const char *dir_name, int depth)
1325 struct dirent *dent;
1326 DIR *dir = opendir(dir_name);
1330 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1334 while ((dent = readdir(dir)) != NULL) {
1335 char path[PATH_MAX];
1338 /*sshfs might return bad dent->d_type, so we have to stat*/
1339 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1340 if (stat(path, &st))
1343 if (S_ISDIR(st.st_mode)) {
1344 if (!strcmp(dent->d_name, ".") ||
1345 !strcmp(dent->d_name, ".."))
1348 /* Do not follow top-level source and build symlinks */
1350 if (!strcmp(dent->d_name, "source") ||
1351 !strcmp(dent->d_name, "build"))
1355 ret = map_groups__set_modules_path_dir(mg, path,
1362 ret = kmod_path__parse_name(&m, dent->d_name);
1367 ret = map_groups__set_module_path(mg, path, &m);
1381 static int machine__set_modules_path(struct machine *machine)
1384 char modules_path[PATH_MAX];
1386 version = get_kernel_version(machine->root_dir);
1390 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1391 machine->root_dir, version);
1394 return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1396 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1397 u64 *size __maybe_unused,
1398 const char *name __maybe_unused)
1403 static int machine__create_module(void *arg, const char *name, u64 start,
1406 struct machine *machine = arg;
1409 if (arch__fix_module_text_start(&start, &size, name) < 0)
1412 map = machine__findnew_module_map(machine, start, name);
1415 map->end = start + size;
1417 dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1422 static int machine__create_modules(struct machine *machine)
1424 const char *modules;
1425 char path[PATH_MAX];
1427 if (machine__is_default_guest(machine)) {
1428 modules = symbol_conf.default_guest_modules;
1430 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1434 if (symbol__restricted_filename(modules, "/proc/modules"))
1437 if (modules__parse(modules, machine, machine__create_module))
1440 if (!machine__set_modules_path(machine))
1443 pr_debug("Problems setting modules path maps, continuing anyway...\n");
1448 static void machine__set_kernel_mmap(struct machine *machine,
1451 machine->vmlinux_map->start = start;
1452 machine->vmlinux_map->end = end;
1454 * Be a bit paranoid here, some perf.data file came with
1455 * a zero sized synthesized MMAP event for the kernel.
1457 if (start == 0 && end == 0)
1458 machine->vmlinux_map->end = ~0ULL;
1461 static void machine__update_kernel_mmap(struct machine *machine,
1464 struct map *map = machine__kernel_map(machine);
1467 map_groups__remove(&machine->kmaps, map);
1469 machine__set_kernel_mmap(machine, start, end);
1471 map_groups__insert(&machine->kmaps, map);
1475 int machine__create_kernel_maps(struct machine *machine)
1477 struct dso *kernel = machine__get_kernel(machine);
1478 const char *name = NULL;
1480 u64 start = 0, end = ~0ULL;
1486 ret = __machine__create_kernel_maps(machine, kernel);
1490 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1491 if (machine__is_host(machine))
1492 pr_debug("Problems creating module maps, "
1493 "continuing anyway...\n");
1495 pr_debug("Problems creating module maps for guest %d, "
1496 "continuing anyway...\n", machine->pid);
1499 if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1501 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1502 machine__destroy_kernel_maps(machine);
1508 * we have a real start address now, so re-order the kmaps
1509 * assume it's the last in the kmaps
1511 machine__update_kernel_mmap(machine, start, end);
1514 if (machine__create_extra_kernel_maps(machine, kernel))
1515 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1518 /* update end address of the kernel map using adjacent module address */
1519 map = map__next(machine__kernel_map(machine));
1521 machine__set_kernel_mmap(machine, start, map->start);
1529 static bool machine__uses_kcore(struct machine *machine)
1533 list_for_each_entry(dso, &machine->dsos.head, node) {
1534 if (dso__is_kcore(dso))
1541 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1542 union perf_event *event)
1544 return machine__is(machine, "x86_64") &&
1545 is_entry_trampoline(event->mmap.filename);
1548 static int machine__process_extra_kernel_map(struct machine *machine,
1549 union perf_event *event)
1551 struct dso *kernel = machine__kernel_dso(machine);
1552 struct extra_kernel_map xm = {
1553 .start = event->mmap.start,
1554 .end = event->mmap.start + event->mmap.len,
1555 .pgoff = event->mmap.pgoff,
1561 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1563 return machine__create_extra_kernel_map(machine, kernel, &xm);
1566 static int machine__process_kernel_mmap_event(struct machine *machine,
1567 union perf_event *event)
1570 enum dso_kernel_type kernel_type;
1571 bool is_kernel_mmap;
1573 /* If we have maps from kcore then we do not need or want any others */
1574 if (machine__uses_kcore(machine))
1577 if (machine__is_host(machine))
1578 kernel_type = DSO_TYPE_KERNEL;
1580 kernel_type = DSO_TYPE_GUEST_KERNEL;
1582 is_kernel_mmap = memcmp(event->mmap.filename,
1584 strlen(machine->mmap_name) - 1) == 0;
1585 if (event->mmap.filename[0] == '/' ||
1586 (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1587 map = machine__findnew_module_map(machine, event->mmap.start,
1588 event->mmap.filename);
1592 map->end = map->start + event->mmap.len;
1593 } else if (is_kernel_mmap) {
1594 const char *symbol_name = (event->mmap.filename +
1595 strlen(machine->mmap_name));
1597 * Should be there already, from the build-id table in
1600 struct dso *kernel = NULL;
1603 down_read(&machine->dsos.lock);
1605 list_for_each_entry(dso, &machine->dsos.head, node) {
1608 * The cpumode passed to is_kernel_module is not the
1609 * cpumode of *this* event. If we insist on passing
1610 * correct cpumode to is_kernel_module, we should
1611 * record the cpumode when we adding this dso to the
1614 * However we don't really need passing correct
1615 * cpumode. We know the correct cpumode must be kernel
1616 * mode (if not, we should not link it onto kernel_dsos
1619 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1620 * is_kernel_module() treats it as a kernel cpumode.
1624 is_kernel_module(dso->long_name,
1625 PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1633 up_read(&machine->dsos.lock);
1636 kernel = machine__findnew_dso(machine, machine->mmap_name);
1640 kernel->kernel = kernel_type;
1641 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1646 if (strstr(kernel->long_name, "vmlinux"))
1647 dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1649 machine__update_kernel_mmap(machine, event->mmap.start,
1650 event->mmap.start + event->mmap.len);
1653 * Avoid using a zero address (kptr_restrict) for the ref reloc
1654 * symbol. Effectively having zero here means that at record
1655 * time /proc/sys/kernel/kptr_restrict was non zero.
1657 if (event->mmap.pgoff != 0) {
1658 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1663 if (machine__is_default_guest(machine)) {
1665 * preload dso of guest kernel and modules
1667 dso__load(kernel, machine__kernel_map(machine));
1669 } else if (perf_event__is_extra_kernel_mmap(machine, event)) {
1670 return machine__process_extra_kernel_map(machine, event);
1677 int machine__process_mmap2_event(struct machine *machine,
1678 union perf_event *event,
1679 struct perf_sample *sample)
1681 struct thread *thread;
1686 perf_event__fprintf_mmap2(event, stdout);
1688 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1689 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1690 ret = machine__process_kernel_mmap_event(machine, event);
1696 thread = machine__findnew_thread(machine, event->mmap2.pid,
1701 map = map__new(machine, event->mmap2.start,
1702 event->mmap2.len, event->mmap2.pgoff,
1704 event->mmap2.min, event->mmap2.ino,
1705 event->mmap2.ino_generation,
1708 event->mmap2.filename, thread);
1711 goto out_problem_map;
1713 ret = thread__insert_map(thread, map);
1715 goto out_problem_insert;
1717 thread__put(thread);
1724 thread__put(thread);
1726 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1730 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1731 struct perf_sample *sample)
1733 struct thread *thread;
1739 perf_event__fprintf_mmap(event, stdout);
1741 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1742 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1743 ret = machine__process_kernel_mmap_event(machine, event);
1749 thread = machine__findnew_thread(machine, event->mmap.pid,
1754 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1757 map = map__new(machine, event->mmap.start,
1758 event->mmap.len, event->mmap.pgoff,
1759 0, 0, 0, 0, prot, 0,
1760 event->mmap.filename,
1764 goto out_problem_map;
1766 ret = thread__insert_map(thread, map);
1768 goto out_problem_insert;
1770 thread__put(thread);
1777 thread__put(thread);
1779 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1783 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1785 struct threads *threads = machine__threads(machine, th->tid);
1787 if (threads->last_match == th)
1788 threads__set_last_match(threads, NULL);
1791 down_write(&threads->lock);
1793 BUG_ON(refcount_read(&th->refcnt) == 0);
1795 rb_erase_cached(&th->rb_node, &threads->entries);
1796 RB_CLEAR_NODE(&th->rb_node);
1799 * Move it first to the dead_threads list, then drop the reference,
1800 * if this is the last reference, then the thread__delete destructor
1801 * will be called and we will remove it from the dead_threads list.
1803 list_add_tail(&th->node, &threads->dead);
1806 * We need to do the put here because if this is the last refcount,
1807 * then we will be touching the threads->dead head when removing the
1813 up_write(&threads->lock);
1816 void machine__remove_thread(struct machine *machine, struct thread *th)
1818 return __machine__remove_thread(machine, th, true);
1821 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1822 struct perf_sample *sample)
1824 struct thread *thread = machine__find_thread(machine,
1827 struct thread *parent = machine__findnew_thread(machine,
1830 bool do_maps_clone = true;
1834 perf_event__fprintf_task(event, stdout);
1837 * There may be an existing thread that is not actually the parent,
1838 * either because we are processing events out of order, or because the
1839 * (fork) event that would have removed the thread was lost. Assume the
1840 * latter case and continue on as best we can.
1842 if (parent->pid_ != (pid_t)event->fork.ppid) {
1843 dump_printf("removing erroneous parent thread %d/%d\n",
1844 parent->pid_, parent->tid);
1845 machine__remove_thread(machine, parent);
1846 thread__put(parent);
1847 parent = machine__findnew_thread(machine, event->fork.ppid,
1851 /* if a thread currently exists for the thread id remove it */
1852 if (thread != NULL) {
1853 machine__remove_thread(machine, thread);
1854 thread__put(thread);
1857 thread = machine__findnew_thread(machine, event->fork.pid,
1860 * When synthesizing FORK events, we are trying to create thread
1861 * objects for the already running tasks on the machine.
1863 * Normally, for a kernel FORK event, we want to clone the parent's
1864 * maps because that is what the kernel just did.
1866 * But when synthesizing, this should not be done. If we do, we end up
1867 * with overlapping maps as we process the sythesized MMAP2 events that
1868 * get delivered shortly thereafter.
1870 * Use the FORK event misc flags in an internal way to signal this
1871 * situation, so we can elide the map clone when appropriate.
1873 if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1874 do_maps_clone = false;
1876 if (thread == NULL || parent == NULL ||
1877 thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1878 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1881 thread__put(thread);
1882 thread__put(parent);
1887 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1888 struct perf_sample *sample __maybe_unused)
1890 struct thread *thread = machine__find_thread(machine,
1895 perf_event__fprintf_task(event, stdout);
1897 if (thread != NULL) {
1898 thread__exited(thread);
1899 thread__put(thread);
1905 int machine__process_event(struct machine *machine, union perf_event *event,
1906 struct perf_sample *sample)
1910 switch (event->header.type) {
1911 case PERF_RECORD_COMM:
1912 ret = machine__process_comm_event(machine, event, sample); break;
1913 case PERF_RECORD_MMAP:
1914 ret = machine__process_mmap_event(machine, event, sample); break;
1915 case PERF_RECORD_NAMESPACES:
1916 ret = machine__process_namespaces_event(machine, event, sample); break;
1917 case PERF_RECORD_MMAP2:
1918 ret = machine__process_mmap2_event(machine, event, sample); break;
1919 case PERF_RECORD_FORK:
1920 ret = machine__process_fork_event(machine, event, sample); break;
1921 case PERF_RECORD_EXIT:
1922 ret = machine__process_exit_event(machine, event, sample); break;
1923 case PERF_RECORD_LOST:
1924 ret = machine__process_lost_event(machine, event, sample); break;
1925 case PERF_RECORD_AUX:
1926 ret = machine__process_aux_event(machine, event); break;
1927 case PERF_RECORD_ITRACE_START:
1928 ret = machine__process_itrace_start_event(machine, event); break;
1929 case PERF_RECORD_LOST_SAMPLES:
1930 ret = machine__process_lost_samples_event(machine, event, sample); break;
1931 case PERF_RECORD_SWITCH:
1932 case PERF_RECORD_SWITCH_CPU_WIDE:
1933 ret = machine__process_switch_event(machine, event); break;
1934 case PERF_RECORD_KSYMBOL:
1935 ret = machine__process_ksymbol(machine, event, sample); break;
1936 case PERF_RECORD_BPF_EVENT:
1937 ret = machine__process_bpf(machine, event, sample); break;
1946 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1948 if (!regexec(regex, sym->name, 0, NULL, 0))
1953 static void ip__resolve_ams(struct thread *thread,
1954 struct addr_map_symbol *ams,
1957 struct addr_location al;
1959 memset(&al, 0, sizeof(al));
1961 * We cannot use the header.misc hint to determine whether a
1962 * branch stack address is user, kernel, guest, hypervisor.
1963 * Branches may straddle the kernel/user/hypervisor boundaries.
1964 * Thus, we have to try consecutively until we find a match
1965 * or else, the symbol is unknown
1967 thread__find_cpumode_addr_location(thread, ip, &al);
1970 ams->al_addr = al.addr;
1972 ams->ms.sym = al.sym;
1973 ams->ms.map = al.map;
1977 static void ip__resolve_data(struct thread *thread,
1978 u8 m, struct addr_map_symbol *ams,
1979 u64 addr, u64 phys_addr)
1981 struct addr_location al;
1983 memset(&al, 0, sizeof(al));
1985 thread__find_symbol(thread, m, addr, &al);
1988 ams->al_addr = al.addr;
1990 ams->ms.sym = al.sym;
1991 ams->ms.map = al.map;
1992 ams->phys_addr = phys_addr;
1995 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1996 struct addr_location *al)
1998 struct mem_info *mi = mem_info__new();
2003 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
2004 ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
2005 sample->addr, sample->phys_addr);
2006 mi->data_src.val = sample->data_src;
2011 static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2013 struct map *map = ms->map;
2014 char *srcline = NULL;
2016 if (!map || callchain_param.key == CCKEY_FUNCTION)
2019 srcline = srcline__tree_find(&map->dso->srclines, ip);
2021 bool show_sym = false;
2022 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2024 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
2025 ms->sym, show_sym, show_addr, ip);
2026 srcline__tree_insert(&map->dso->srclines, ip, srcline);
2037 static int add_callchain_ip(struct thread *thread,
2038 struct callchain_cursor *cursor,
2039 struct symbol **parent,
2040 struct addr_location *root_al,
2044 struct branch_flags *flags,
2045 struct iterations *iter,
2048 struct map_symbol ms;
2049 struct addr_location al;
2050 int nr_loop_iter = 0;
2051 u64 iter_cycles = 0;
2052 const char *srcline = NULL;
2057 thread__find_cpumode_addr_location(thread, ip, &al);
2059 if (ip >= PERF_CONTEXT_MAX) {
2061 case PERF_CONTEXT_HV:
2062 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2064 case PERF_CONTEXT_KERNEL:
2065 *cpumode = PERF_RECORD_MISC_KERNEL;
2067 case PERF_CONTEXT_USER:
2068 *cpumode = PERF_RECORD_MISC_USER;
2071 pr_debug("invalid callchain context: "
2072 "%"PRId64"\n", (s64) ip);
2074 * It seems the callchain is corrupted.
2077 callchain_cursor_reset(cursor);
2082 thread__find_symbol(thread, *cpumode, ip, &al);
2085 if (al.sym != NULL) {
2086 if (perf_hpp_list.parent && !*parent &&
2087 symbol__match_regex(al.sym, &parent_regex))
2089 else if (have_ignore_callees && root_al &&
2090 symbol__match_regex(al.sym, &ignore_callees_regex)) {
2091 /* Treat this symbol as the root,
2092 forgetting its callees. */
2094 callchain_cursor_reset(cursor);
2098 if (symbol_conf.hide_unresolved && al.sym == NULL)
2102 nr_loop_iter = iter->nr_loop_iter;
2103 iter_cycles = iter->cycles;
2109 srcline = callchain_srcline(&ms, al.addr);
2110 return callchain_cursor_append(cursor, ip, &ms,
2111 branch, flags, nr_loop_iter,
2112 iter_cycles, branch_from, srcline);
2115 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2116 struct addr_location *al)
2119 const struct branch_stack *bs = sample->branch_stack;
2120 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2125 for (i = 0; i < bs->nr; i++) {
2126 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
2127 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
2128 bi[i].flags = bs->entries[i].flags;
2133 static void save_iterations(struct iterations *iter,
2134 struct branch_entry *be, int nr)
2138 iter->nr_loop_iter++;
2141 for (i = 0; i < nr; i++)
2142 iter->cycles += be[i].flags.cycles;
2147 #define NO_ENTRY 0xff
2149 #define PERF_MAX_BRANCH_DEPTH 127
2152 static int remove_loops(struct branch_entry *l, int nr,
2153 struct iterations *iter)
2156 unsigned char chash[CHASHSZ];
2158 memset(chash, NO_ENTRY, sizeof(chash));
2160 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2162 for (i = 0; i < nr; i++) {
2163 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2165 /* no collision handling for now */
2166 if (chash[h] == NO_ENTRY) {
2168 } else if (l[chash[h]].from == l[i].from) {
2169 bool is_loop = true;
2170 /* check if it is a real loop */
2172 for (j = chash[h]; j < i && i + off < nr; j++, off++)
2173 if (l[j].from != l[i + off].from) {
2180 save_iterations(iter + i + off,
2183 memmove(iter + i, iter + i + off,
2186 memmove(l + i, l + i + off,
2198 * Recolve LBR callstack chain sample
2200 * 1 on success get LBR callchain information
2201 * 0 no available LBR callchain information, should try fp
2202 * negative error code on other errors.
2204 static int resolve_lbr_callchain_sample(struct thread *thread,
2205 struct callchain_cursor *cursor,
2206 struct perf_sample *sample,
2207 struct symbol **parent,
2208 struct addr_location *root_al,
2211 struct ip_callchain *chain = sample->callchain;
2212 int chain_nr = min(max_stack, (int)chain->nr), i;
2213 u8 cpumode = PERF_RECORD_MISC_USER;
2214 u64 ip, branch_from = 0;
2216 for (i = 0; i < chain_nr; i++) {
2217 if (chain->ips[i] == PERF_CONTEXT_USER)
2221 /* LBR only affects the user callchain */
2222 if (i != chain_nr) {
2223 struct branch_stack *lbr_stack = sample->branch_stack;
2224 int lbr_nr = lbr_stack->nr, j, k;
2226 struct branch_flags *flags;
2228 * LBR callstack can only get user call chain.
2229 * The mix_chain_nr is kernel call chain
2230 * number plus LBR user call chain number.
2231 * i is kernel call chain number,
2232 * 1 is PERF_CONTEXT_USER,
2233 * lbr_nr + 1 is the user call chain number.
2234 * For details, please refer to the comments
2235 * in callchain__printf
2237 int mix_chain_nr = i + 1 + lbr_nr + 1;
2239 for (j = 0; j < mix_chain_nr; j++) {
2244 if (callchain_param.order == ORDER_CALLEE) {
2247 else if (j > i + 1) {
2249 ip = lbr_stack->entries[k].from;
2251 flags = &lbr_stack->entries[k].flags;
2253 ip = lbr_stack->entries[0].to;
2255 flags = &lbr_stack->entries[0].flags;
2257 lbr_stack->entries[0].from;
2262 ip = lbr_stack->entries[k].from;
2264 flags = &lbr_stack->entries[k].flags;
2266 else if (j > lbr_nr)
2267 ip = chain->ips[i + 1 - (j - lbr_nr)];
2269 ip = lbr_stack->entries[0].to;
2271 flags = &lbr_stack->entries[0].flags;
2273 lbr_stack->entries[0].from;
2277 err = add_callchain_ip(thread, cursor, parent,
2278 root_al, &cpumode, ip,
2279 branch, flags, NULL,
2282 return (err < 0) ? err : 0;
2290 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2291 struct callchain_cursor *cursor,
2292 struct symbol **parent,
2293 struct addr_location *root_al,
2294 u8 *cpumode, int ent)
2298 while (--ent >= 0) {
2299 u64 ip = chain->ips[ent];
2301 if (ip >= PERF_CONTEXT_MAX) {
2302 err = add_callchain_ip(thread, cursor, parent,
2303 root_al, cpumode, ip,
2304 false, NULL, NULL, 0);
2311 static int thread__resolve_callchain_sample(struct thread *thread,
2312 struct callchain_cursor *cursor,
2313 struct evsel *evsel,
2314 struct perf_sample *sample,
2315 struct symbol **parent,
2316 struct addr_location *root_al,
2319 struct branch_stack *branch = sample->branch_stack;
2320 struct ip_callchain *chain = sample->callchain;
2322 u8 cpumode = PERF_RECORD_MISC_USER;
2323 int i, j, err, nr_entries;
2328 chain_nr = chain->nr;
2330 if (perf_evsel__has_branch_callstack(evsel)) {
2331 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2332 root_al, max_stack);
2334 return (err < 0) ? err : 0;
2338 * Based on DWARF debug information, some architectures skip
2339 * a callchain entry saved by the kernel.
2341 skip_idx = arch_skip_callchain_idx(thread, chain);
2344 * Add branches to call stack for easier browsing. This gives
2345 * more context for a sample than just the callers.
2347 * This uses individual histograms of paths compared to the
2348 * aggregated histograms the normal LBR mode uses.
2350 * Limitations for now:
2351 * - No extra filters
2352 * - No annotations (should annotate somehow)
2355 if (branch && callchain_param.branch_callstack) {
2356 int nr = min(max_stack, (int)branch->nr);
2357 struct branch_entry be[nr];
2358 struct iterations iter[nr];
2360 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2361 pr_warning("corrupted branch chain. skipping...\n");
2365 for (i = 0; i < nr; i++) {
2366 if (callchain_param.order == ORDER_CALLEE) {
2367 be[i] = branch->entries[i];
2373 * Check for overlap into the callchain.
2374 * The return address is one off compared to
2375 * the branch entry. To adjust for this
2376 * assume the calling instruction is not longer
2379 if (i == skip_idx ||
2380 chain->ips[first_call] >= PERF_CONTEXT_MAX)
2382 else if (be[i].from < chain->ips[first_call] &&
2383 be[i].from >= chain->ips[first_call] - 8)
2386 be[i] = branch->entries[branch->nr - i - 1];
2389 memset(iter, 0, sizeof(struct iterations) * nr);
2390 nr = remove_loops(be, nr, iter);
2392 for (i = 0; i < nr; i++) {
2393 err = add_callchain_ip(thread, cursor, parent,
2400 err = add_callchain_ip(thread, cursor, parent, root_al,
2417 if (callchain_param.order != ORDER_CALLEE) {
2418 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2419 &cpumode, chain->nr - first_call);
2421 return (err < 0) ? err : 0;
2423 for (i = first_call, nr_entries = 0;
2424 i < chain_nr && nr_entries < max_stack; i++) {
2427 if (callchain_param.order == ORDER_CALLEE)
2430 j = chain->nr - i - 1;
2432 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2437 if (ip < PERF_CONTEXT_MAX)
2439 else if (callchain_param.order != ORDER_CALLEE) {
2440 err = find_prev_cpumode(chain, thread, cursor, parent,
2441 root_al, &cpumode, j);
2443 return (err < 0) ? err : 0;
2447 err = add_callchain_ip(thread, cursor, parent,
2448 root_al, &cpumode, ip,
2449 false, NULL, NULL, 0);
2452 return (err < 0) ? err : 0;
2458 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
2460 struct symbol *sym = ms->sym;
2461 struct map *map = ms->map;
2462 struct inline_node *inline_node;
2463 struct inline_list *ilist;
2467 if (!symbol_conf.inline_name || !map || !sym)
2470 addr = map__map_ip(map, ip);
2471 addr = map__rip_2objdump(map, addr);
2473 inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2475 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2478 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2481 list_for_each_entry(ilist, &inline_node->val, list) {
2482 struct map_symbol ilist_ms = {
2484 .sym = ilist->symbol,
2486 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
2487 NULL, 0, 0, 0, ilist->srcline);
2496 static int unwind_entry(struct unwind_entry *entry, void *arg)
2498 struct callchain_cursor *cursor = arg;
2499 const char *srcline = NULL;
2500 u64 addr = entry->ip;
2502 if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
2505 if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
2509 * Convert entry->ip from a virtual address to an offset in
2510 * its corresponding binary.
2513 addr = map__map_ip(entry->ms.map, entry->ip);
2515 srcline = callchain_srcline(&entry->ms, addr);
2516 return callchain_cursor_append(cursor, entry->ip, &entry->ms,
2517 false, NULL, 0, 0, 0, srcline);
2520 static int thread__resolve_callchain_unwind(struct thread *thread,
2521 struct callchain_cursor *cursor,
2522 struct evsel *evsel,
2523 struct perf_sample *sample,
2526 /* Can we do dwarf post unwind? */
2527 if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2528 (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
2531 /* Bail out if nothing was captured. */
2532 if ((!sample->user_regs.regs) ||
2533 (!sample->user_stack.size))
2536 return unwind__get_entries(unwind_entry, cursor,
2537 thread, sample, max_stack);
2540 int thread__resolve_callchain(struct thread *thread,
2541 struct callchain_cursor *cursor,
2542 struct evsel *evsel,
2543 struct perf_sample *sample,
2544 struct symbol **parent,
2545 struct addr_location *root_al,
2550 callchain_cursor_reset(cursor);
2552 if (callchain_param.order == ORDER_CALLEE) {
2553 ret = thread__resolve_callchain_sample(thread, cursor,
2559 ret = thread__resolve_callchain_unwind(thread, cursor,
2563 ret = thread__resolve_callchain_unwind(thread, cursor,
2568 ret = thread__resolve_callchain_sample(thread, cursor,
2577 int machine__for_each_thread(struct machine *machine,
2578 int (*fn)(struct thread *thread, void *p),
2581 struct threads *threads;
2583 struct thread *thread;
2587 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2588 threads = &machine->threads[i];
2589 for (nd = rb_first_cached(&threads->entries); nd;
2591 thread = rb_entry(nd, struct thread, rb_node);
2592 rc = fn(thread, priv);
2597 list_for_each_entry(thread, &threads->dead, node) {
2598 rc = fn(thread, priv);
2606 int machines__for_each_thread(struct machines *machines,
2607 int (*fn)(struct thread *thread, void *p),
2613 rc = machine__for_each_thread(&machines->host, fn, priv);
2617 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2618 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2620 rc = machine__for_each_thread(machine, fn, priv);
2627 pid_t machine__get_current_tid(struct machine *machine, int cpu)
2629 int nr_cpus = min(machine->env->nr_cpus_online, MAX_NR_CPUS);
2631 if (cpu < 0 || cpu >= nr_cpus || !machine->current_tid)
2634 return machine->current_tid[cpu];
2637 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
2640 struct thread *thread;
2641 int nr_cpus = min(machine->env->nr_cpus_online, MAX_NR_CPUS);
2646 if (!machine->current_tid) {
2649 machine->current_tid = calloc(nr_cpus, sizeof(pid_t));
2650 if (!machine->current_tid)
2652 for (i = 0; i < nr_cpus; i++)
2653 machine->current_tid[i] = -1;
2656 if (cpu >= nr_cpus) {
2657 pr_err("Requested CPU %d too large. ", cpu);
2658 pr_err("Consider raising MAX_NR_CPUS\n");
2662 machine->current_tid[cpu] = tid;
2664 thread = machine__findnew_thread(machine, pid, tid);
2669 thread__put(thread);
2675 * Compares the raw arch string. N.B. see instead perf_env__arch() if a
2676 * normalized arch is needed.
2678 bool machine__is(struct machine *machine, const char *arch)
2680 return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
2683 int machine__nr_cpus_avail(struct machine *machine)
2685 return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
2688 int machine__get_kernel_start(struct machine *machine)
2690 struct map *map = machine__kernel_map(machine);
2694 * The only addresses above 2^63 are kernel addresses of a 64-bit
2695 * kernel. Note that addresses are unsigned so that on a 32-bit system
2696 * all addresses including kernel addresses are less than 2^32. In
2697 * that case (32-bit system), if the kernel mapping is unknown, all
2698 * addresses will be assumed to be in user space - see
2699 * machine__kernel_ip().
2701 machine->kernel_start = 1ULL << 63;
2703 err = map__load(map);
2705 * On x86_64, PTI entry trampolines are less than the
2706 * start of kernel text, but still above 2^63. So leave
2707 * kernel_start = 1ULL << 63 for x86_64.
2709 if (!err && !machine__is(machine, "x86_64"))
2710 machine->kernel_start = map->start;
2715 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
2717 u8 addr_cpumode = cpumode;
2720 if (!machine->single_address_space)
2723 kernel_ip = machine__kernel_ip(machine, addr);
2725 case PERF_RECORD_MISC_KERNEL:
2726 case PERF_RECORD_MISC_USER:
2727 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
2728 PERF_RECORD_MISC_USER;
2730 case PERF_RECORD_MISC_GUEST_KERNEL:
2731 case PERF_RECORD_MISC_GUEST_USER:
2732 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
2733 PERF_RECORD_MISC_GUEST_USER;
2739 return addr_cpumode;
2742 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2744 return dsos__findnew(&machine->dsos, filename);
2747 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2749 struct machine *machine = vmachine;
2751 struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
2756 *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2757 *addrp = map->unmap_ip(map, sym->start);