Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include <dirent.h>
3#include <errno.h>
4#include <inttypes.h>
5#include <regex.h>
6#include <stdlib.h>
7#include "callchain.h"
8#include "debug.h"
9#include "dso.h"
10#include "env.h"
11#include "event.h"
12#include "evsel.h"
13#include "hist.h"
14#include "machine.h"
15#include "map.h"
16#include "map_symbol.h"
17#include "branch.h"
18#include "mem-events.h"
19#include "path.h"
20#include "srcline.h"
21#include "symbol.h"
22#include "sort.h"
23#include "strlist.h"
24#include "target.h"
25#include "thread.h"
26#include "util.h"
27#include "vdso.h"
28#include <stdbool.h>
29#include <sys/types.h>
30#include <sys/stat.h>
31#include <unistd.h>
32#include "unwind.h"
33#include "linux/hash.h"
34#include "asm/bug.h"
35#include "bpf-event.h"
36#include <internal/lib.h> // page_size
37#include "cgroup.h"
38#include "arm64-frame-pointer-unwind-support.h"
39
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>
45
46static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
47 struct thread *th, bool lock);
48
49static struct dso *machine__kernel_dso(struct machine *machine)
50{
51 return map__dso(machine->vmlinux_map);
52}
53
54static void dsos__init(struct dsos *dsos)
55{
56 INIT_LIST_HEAD(&dsos->head);
57 dsos->root = RB_ROOT;
58 init_rwsem(&dsos->lock);
59}
60
61static void machine__threads_init(struct machine *machine)
62{
63 int i;
64
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);
69 threads->nr = 0;
70 threads->last_match = NULL;
71 }
72}
73
74static int thread_rb_node__cmp_tid(const void *key, const struct rb_node *nd)
75{
76 int to_find = (int) *((pid_t *)key);
77
78 return to_find - (int)thread__tid(rb_entry(nd, struct thread_rb_node, rb_node)->thread);
79}
80
81static struct thread_rb_node *thread_rb_node__find(const struct thread *th,
82 struct rb_root *tree)
83{
84 pid_t to_find = thread__tid(th);
85 struct rb_node *nd = rb_find(&to_find, tree, thread_rb_node__cmp_tid);
86
87 return rb_entry(nd, struct thread_rb_node, rb_node);
88}
89
90static int machine__set_mmap_name(struct machine *machine)
91{
92 if (machine__is_host(machine))
93 machine->mmap_name = strdup("[kernel.kallsyms]");
94 else if (machine__is_default_guest(machine))
95 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
96 else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
97 machine->pid) < 0)
98 machine->mmap_name = NULL;
99
100 return machine->mmap_name ? 0 : -ENOMEM;
101}
102
103static void thread__set_guest_comm(struct thread *thread, pid_t pid)
104{
105 char comm[64];
106
107 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
108 thread__set_comm(thread, comm, 0);
109}
110
111int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
112{
113 int err = -ENOMEM;
114
115 memset(machine, 0, sizeof(*machine));
116 machine->kmaps = maps__new(machine);
117 if (machine->kmaps == NULL)
118 return -ENOMEM;
119
120 RB_CLEAR_NODE(&machine->rb_node);
121 dsos__init(&machine->dsos);
122
123 machine__threads_init(machine);
124
125 machine->vdso_info = NULL;
126 machine->env = NULL;
127
128 machine->pid = pid;
129
130 machine->id_hdr_size = 0;
131 machine->kptr_restrict_warned = false;
132 machine->comm_exec = false;
133 machine->kernel_start = 0;
134 machine->vmlinux_map = NULL;
135
136 machine->root_dir = strdup(root_dir);
137 if (machine->root_dir == NULL)
138 goto out;
139
140 if (machine__set_mmap_name(machine))
141 goto out;
142
143 if (pid != HOST_KERNEL_ID) {
144 struct thread *thread = machine__findnew_thread(machine, -1,
145 pid);
146
147 if (thread == NULL)
148 goto out;
149
150 thread__set_guest_comm(thread, pid);
151 thread__put(thread);
152 }
153
154 machine->current_tid = NULL;
155 err = 0;
156
157out:
158 if (err) {
159 zfree(&machine->kmaps);
160 zfree(&machine->root_dir);
161 zfree(&machine->mmap_name);
162 }
163 return 0;
164}
165
166struct machine *machine__new_host(void)
167{
168 struct machine *machine = malloc(sizeof(*machine));
169
170 if (machine != NULL) {
171 machine__init(machine, "", HOST_KERNEL_ID);
172
173 if (machine__create_kernel_maps(machine) < 0)
174 goto out_delete;
175 }
176
177 return machine;
178out_delete:
179 free(machine);
180 return NULL;
181}
182
183struct machine *machine__new_kallsyms(void)
184{
185 struct machine *machine = machine__new_host();
186 /*
187 * FIXME:
188 * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
189 * ask for not using the kcore parsing code, once this one is fixed
190 * to create a map per module.
191 */
192 if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
193 machine__delete(machine);
194 machine = NULL;
195 }
196
197 return machine;
198}
199
200static void dsos__purge(struct dsos *dsos)
201{
202 struct dso *pos, *n;
203
204 down_write(&dsos->lock);
205
206 list_for_each_entry_safe(pos, n, &dsos->head, node) {
207 RB_CLEAR_NODE(&pos->rb_node);
208 pos->root = NULL;
209 list_del_init(&pos->node);
210 dso__put(pos);
211 }
212
213 up_write(&dsos->lock);
214}
215
216static void dsos__exit(struct dsos *dsos)
217{
218 dsos__purge(dsos);
219 exit_rwsem(&dsos->lock);
220}
221
222void machine__delete_threads(struct machine *machine)
223{
224 struct rb_node *nd;
225 int i;
226
227 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
228 struct threads *threads = &machine->threads[i];
229 down_write(&threads->lock);
230 nd = rb_first_cached(&threads->entries);
231 while (nd) {
232 struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
233
234 nd = rb_next(nd);
235 __machine__remove_thread(machine, trb, trb->thread, false);
236 }
237 up_write(&threads->lock);
238 }
239}
240
241void machine__exit(struct machine *machine)
242{
243 int i;
244
245 if (machine == NULL)
246 return;
247
248 machine__destroy_kernel_maps(machine);
249 maps__zput(machine->kmaps);
250 dsos__exit(&machine->dsos);
251 machine__exit_vdso(machine);
252 zfree(&machine->root_dir);
253 zfree(&machine->mmap_name);
254 zfree(&machine->current_tid);
255 zfree(&machine->kallsyms_filename);
256
257 machine__delete_threads(machine);
258 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
259 struct threads *threads = &machine->threads[i];
260
261 exit_rwsem(&threads->lock);
262 }
263}
264
265void machine__delete(struct machine *machine)
266{
267 if (machine) {
268 machine__exit(machine);
269 free(machine);
270 }
271}
272
273void machines__init(struct machines *machines)
274{
275 machine__init(&machines->host, "", HOST_KERNEL_ID);
276 machines->guests = RB_ROOT_CACHED;
277}
278
279void machines__exit(struct machines *machines)
280{
281 machine__exit(&machines->host);
282 /* XXX exit guest */
283}
284
285struct machine *machines__add(struct machines *machines, pid_t pid,
286 const char *root_dir)
287{
288 struct rb_node **p = &machines->guests.rb_root.rb_node;
289 struct rb_node *parent = NULL;
290 struct machine *pos, *machine = malloc(sizeof(*machine));
291 bool leftmost = true;
292
293 if (machine == NULL)
294 return NULL;
295
296 if (machine__init(machine, root_dir, pid) != 0) {
297 free(machine);
298 return NULL;
299 }
300
301 while (*p != NULL) {
302 parent = *p;
303 pos = rb_entry(parent, struct machine, rb_node);
304 if (pid < pos->pid)
305 p = &(*p)->rb_left;
306 else {
307 p = &(*p)->rb_right;
308 leftmost = false;
309 }
310 }
311
312 rb_link_node(&machine->rb_node, parent, p);
313 rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
314
315 machine->machines = machines;
316
317 return machine;
318}
319
320void machines__set_comm_exec(struct machines *machines, bool comm_exec)
321{
322 struct rb_node *nd;
323
324 machines->host.comm_exec = comm_exec;
325
326 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
327 struct machine *machine = rb_entry(nd, struct machine, rb_node);
328
329 machine->comm_exec = comm_exec;
330 }
331}
332
333struct machine *machines__find(struct machines *machines, pid_t pid)
334{
335 struct rb_node **p = &machines->guests.rb_root.rb_node;
336 struct rb_node *parent = NULL;
337 struct machine *machine;
338 struct machine *default_machine = NULL;
339
340 if (pid == HOST_KERNEL_ID)
341 return &machines->host;
342
343 while (*p != NULL) {
344 parent = *p;
345 machine = rb_entry(parent, struct machine, rb_node);
346 if (pid < machine->pid)
347 p = &(*p)->rb_left;
348 else if (pid > machine->pid)
349 p = &(*p)->rb_right;
350 else
351 return machine;
352 if (!machine->pid)
353 default_machine = machine;
354 }
355
356 return default_machine;
357}
358
359struct machine *machines__findnew(struct machines *machines, pid_t pid)
360{
361 char path[PATH_MAX];
362 const char *root_dir = "";
363 struct machine *machine = machines__find(machines, pid);
364
365 if (machine && (machine->pid == pid))
366 goto out;
367
368 if ((pid != HOST_KERNEL_ID) &&
369 (pid != DEFAULT_GUEST_KERNEL_ID) &&
370 (symbol_conf.guestmount)) {
371 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
372 if (access(path, R_OK)) {
373 static struct strlist *seen;
374
375 if (!seen)
376 seen = strlist__new(NULL, NULL);
377
378 if (!strlist__has_entry(seen, path)) {
379 pr_err("Can't access file %s\n", path);
380 strlist__add(seen, path);
381 }
382 machine = NULL;
383 goto out;
384 }
385 root_dir = path;
386 }
387
388 machine = machines__add(machines, pid, root_dir);
389out:
390 return machine;
391}
392
393struct machine *machines__find_guest(struct machines *machines, pid_t pid)
394{
395 struct machine *machine = machines__find(machines, pid);
396
397 if (!machine)
398 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID);
399 return machine;
400}
401
402/*
403 * A common case for KVM test programs is that the test program acts as the
404 * hypervisor, creating, running and destroying the virtual machine, and
405 * providing the guest object code from its own object code. In this case,
406 * the VM is not running an OS, but only the functions loaded into it by the
407 * hypervisor test program, and conveniently, loaded at the same virtual
408 * addresses.
409 *
410 * Normally to resolve addresses, MMAP events are needed to map addresses
411 * back to the object code and debug symbols for that object code.
412 *
413 * Currently, there is no way to get such mapping information from guests
414 * but, in the scenario described above, the guest has the same mappings
415 * as the hypervisor, so support for that scenario can be achieved.
416 *
417 * To support that, copy the host thread's maps to the guest thread's maps.
418 * Note, we do not discover the guest until we encounter a guest event,
419 * which works well because it is not until then that we know that the host
420 * thread's maps have been set up.
421 *
422 * This function returns the guest thread. Apart from keeping the data
423 * structures sane, using a thread belonging to the guest machine, instead
424 * of the host thread, allows it to have its own comm (refer
425 * thread__set_guest_comm()).
426 */
427static struct thread *findnew_guest_code(struct machine *machine,
428 struct machine *host_machine,
429 pid_t pid)
430{
431 struct thread *host_thread;
432 struct thread *thread;
433 int err;
434
435 if (!machine)
436 return NULL;
437
438 thread = machine__findnew_thread(machine, -1, pid);
439 if (!thread)
440 return NULL;
441
442 /* Assume maps are set up if there are any */
443 if (maps__nr_maps(thread__maps(thread)))
444 return thread;
445
446 host_thread = machine__find_thread(host_machine, -1, pid);
447 if (!host_thread)
448 goto out_err;
449
450 thread__set_guest_comm(thread, pid);
451
452 /*
453 * Guest code can be found in hypervisor process at the same address
454 * so copy host maps.
455 */
456 err = maps__copy_from(thread__maps(thread), thread__maps(host_thread));
457 thread__put(host_thread);
458 if (err)
459 goto out_err;
460
461 return thread;
462
463out_err:
464 thread__zput(thread);
465 return NULL;
466}
467
468struct thread *machines__findnew_guest_code(struct machines *machines, pid_t pid)
469{
470 struct machine *host_machine = machines__find(machines, HOST_KERNEL_ID);
471 struct machine *machine = machines__findnew(machines, pid);
472
473 return findnew_guest_code(machine, host_machine, pid);
474}
475
476struct thread *machine__findnew_guest_code(struct machine *machine, pid_t pid)
477{
478 struct machines *machines = machine->machines;
479 struct machine *host_machine;
480
481 if (!machines)
482 return NULL;
483
484 host_machine = machines__find(machines, HOST_KERNEL_ID);
485
486 return findnew_guest_code(machine, host_machine, pid);
487}
488
489void machines__process_guests(struct machines *machines,
490 machine__process_t process, void *data)
491{
492 struct rb_node *nd;
493
494 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
495 struct machine *pos = rb_entry(nd, struct machine, rb_node);
496 process(pos, data);
497 }
498}
499
500void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
501{
502 struct rb_node *node;
503 struct machine *machine;
504
505 machines->host.id_hdr_size = id_hdr_size;
506
507 for (node = rb_first_cached(&machines->guests); node;
508 node = rb_next(node)) {
509 machine = rb_entry(node, struct machine, rb_node);
510 machine->id_hdr_size = id_hdr_size;
511 }
512
513 return;
514}
515
516static void machine__update_thread_pid(struct machine *machine,
517 struct thread *th, pid_t pid)
518{
519 struct thread *leader;
520
521 if (pid == thread__pid(th) || pid == -1 || thread__pid(th) != -1)
522 return;
523
524 thread__set_pid(th, pid);
525
526 if (thread__pid(th) == thread__tid(th))
527 return;
528
529 leader = __machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
530 if (!leader)
531 goto out_err;
532
533 if (!thread__maps(leader))
534 thread__set_maps(leader, maps__new(machine));
535
536 if (!thread__maps(leader))
537 goto out_err;
538
539 if (thread__maps(th) == thread__maps(leader))
540 goto out_put;
541
542 if (thread__maps(th)) {
543 /*
544 * Maps are created from MMAP events which provide the pid and
545 * tid. Consequently there never should be any maps on a thread
546 * with an unknown pid. Just print an error if there are.
547 */
548 if (!maps__empty(thread__maps(th)))
549 pr_err("Discarding thread maps for %d:%d\n",
550 thread__pid(th), thread__tid(th));
551 maps__put(thread__maps(th));
552 }
553
554 thread__set_maps(th, maps__get(thread__maps(leader)));
555out_put:
556 thread__put(leader);
557 return;
558out_err:
559 pr_err("Failed to join map groups for %d:%d\n", thread__pid(th), thread__tid(th));
560 goto out_put;
561}
562
563/*
564 * Front-end cache - TID lookups come in blocks,
565 * so most of the time we dont have to look up
566 * the full rbtree:
567 */
568static struct thread*
569__threads__get_last_match(struct threads *threads, struct machine *machine,
570 int pid, int tid)
571{
572 struct thread *th;
573
574 th = threads->last_match;
575 if (th != NULL) {
576 if (thread__tid(th) == tid) {
577 machine__update_thread_pid(machine, th, pid);
578 return thread__get(th);
579 }
580 thread__put(threads->last_match);
581 threads->last_match = NULL;
582 }
583
584 return NULL;
585}
586
587static struct thread*
588threads__get_last_match(struct threads *threads, struct machine *machine,
589 int pid, int tid)
590{
591 struct thread *th = NULL;
592
593 if (perf_singlethreaded)
594 th = __threads__get_last_match(threads, machine, pid, tid);
595
596 return th;
597}
598
599static void
600__threads__set_last_match(struct threads *threads, struct thread *th)
601{
602 thread__put(threads->last_match);
603 threads->last_match = thread__get(th);
604}
605
606static void
607threads__set_last_match(struct threads *threads, struct thread *th)
608{
609 if (perf_singlethreaded)
610 __threads__set_last_match(threads, th);
611}
612
613/*
614 * Caller must eventually drop thread->refcnt returned with a successful
615 * lookup/new thread inserted.
616 */
617static struct thread *____machine__findnew_thread(struct machine *machine,
618 struct threads *threads,
619 pid_t pid, pid_t tid,
620 bool create)
621{
622 struct rb_node **p = &threads->entries.rb_root.rb_node;
623 struct rb_node *parent = NULL;
624 struct thread *th;
625 struct thread_rb_node *nd;
626 bool leftmost = true;
627
628 th = threads__get_last_match(threads, machine, pid, tid);
629 if (th)
630 return th;
631
632 while (*p != NULL) {
633 parent = *p;
634 th = rb_entry(parent, struct thread_rb_node, rb_node)->thread;
635
636 if (thread__tid(th) == tid) {
637 threads__set_last_match(threads, th);
638 machine__update_thread_pid(machine, th, pid);
639 return thread__get(th);
640 }
641
642 if (tid < thread__tid(th))
643 p = &(*p)->rb_left;
644 else {
645 p = &(*p)->rb_right;
646 leftmost = false;
647 }
648 }
649
650 if (!create)
651 return NULL;
652
653 th = thread__new(pid, tid);
654 if (th == NULL)
655 return NULL;
656
657 nd = malloc(sizeof(*nd));
658 if (nd == NULL) {
659 thread__put(th);
660 return NULL;
661 }
662 nd->thread = th;
663
664 rb_link_node(&nd->rb_node, parent, p);
665 rb_insert_color_cached(&nd->rb_node, &threads->entries, leftmost);
666 /*
667 * We have to initialize maps separately after rb tree is updated.
668 *
669 * The reason is that we call machine__findnew_thread within
670 * thread__init_maps to find the thread leader and that would screwed
671 * the rb tree.
672 */
673 if (thread__init_maps(th, machine)) {
674 pr_err("Thread init failed thread %d\n", pid);
675 rb_erase_cached(&nd->rb_node, &threads->entries);
676 RB_CLEAR_NODE(&nd->rb_node);
677 free(nd);
678 thread__put(th);
679 return NULL;
680 }
681 /*
682 * It is now in the rbtree, get a ref
683 */
684 threads__set_last_match(threads, th);
685 ++threads->nr;
686
687 return thread__get(th);
688}
689
690struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
691{
692 return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
693}
694
695struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
696 pid_t tid)
697{
698 struct threads *threads = machine__threads(machine, tid);
699 struct thread *th;
700
701 down_write(&threads->lock);
702 th = __machine__findnew_thread(machine, pid, tid);
703 up_write(&threads->lock);
704 return th;
705}
706
707struct thread *machine__find_thread(struct machine *machine, pid_t pid,
708 pid_t tid)
709{
710 struct threads *threads = machine__threads(machine, tid);
711 struct thread *th;
712
713 down_read(&threads->lock);
714 th = ____machine__findnew_thread(machine, threads, pid, tid, false);
715 up_read(&threads->lock);
716 return th;
717}
718
719/*
720 * Threads are identified by pid and tid, and the idle task has pid == tid == 0.
721 * So here a single thread is created for that, but actually there is a separate
722 * idle task per cpu, so there should be one 'struct thread' per cpu, but there
723 * is only 1. That causes problems for some tools, requiring workarounds. For
724 * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu().
725 */
726struct thread *machine__idle_thread(struct machine *machine)
727{
728 struct thread *thread = machine__findnew_thread(machine, 0, 0);
729
730 if (!thread || thread__set_comm(thread, "swapper", 0) ||
731 thread__set_namespaces(thread, 0, NULL))
732 pr_err("problem inserting idle task for machine pid %d\n", machine->pid);
733
734 return thread;
735}
736
737struct comm *machine__thread_exec_comm(struct machine *machine,
738 struct thread *thread)
739{
740 if (machine->comm_exec)
741 return thread__exec_comm(thread);
742 else
743 return thread__comm(thread);
744}
745
746int machine__process_comm_event(struct machine *machine, union perf_event *event,
747 struct perf_sample *sample)
748{
749 struct thread *thread = machine__findnew_thread(machine,
750 event->comm.pid,
751 event->comm.tid);
752 bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
753 int err = 0;
754
755 if (exec)
756 machine->comm_exec = true;
757
758 if (dump_trace)
759 perf_event__fprintf_comm(event, stdout);
760
761 if (thread == NULL ||
762 __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
763 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
764 err = -1;
765 }
766
767 thread__put(thread);
768
769 return err;
770}
771
772int machine__process_namespaces_event(struct machine *machine __maybe_unused,
773 union perf_event *event,
774 struct perf_sample *sample __maybe_unused)
775{
776 struct thread *thread = machine__findnew_thread(machine,
777 event->namespaces.pid,
778 event->namespaces.tid);
779 int err = 0;
780
781 WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
782 "\nWARNING: kernel seems to support more namespaces than perf"
783 " tool.\nTry updating the perf tool..\n\n");
784
785 WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
786 "\nWARNING: perf tool seems to support more namespaces than"
787 " the kernel.\nTry updating the kernel..\n\n");
788
789 if (dump_trace)
790 perf_event__fprintf_namespaces(event, stdout);
791
792 if (thread == NULL ||
793 thread__set_namespaces(thread, sample->time, &event->namespaces)) {
794 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
795 err = -1;
796 }
797
798 thread__put(thread);
799
800 return err;
801}
802
803int machine__process_cgroup_event(struct machine *machine,
804 union perf_event *event,
805 struct perf_sample *sample __maybe_unused)
806{
807 struct cgroup *cgrp;
808
809 if (dump_trace)
810 perf_event__fprintf_cgroup(event, stdout);
811
812 cgrp = cgroup__findnew(machine->env, event->cgroup.id, event->cgroup.path);
813 if (cgrp == NULL)
814 return -ENOMEM;
815
816 return 0;
817}
818
819int machine__process_lost_event(struct machine *machine __maybe_unused,
820 union perf_event *event, struct perf_sample *sample __maybe_unused)
821{
822 dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
823 event->lost.id, event->lost.lost);
824 return 0;
825}
826
827int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
828 union perf_event *event, struct perf_sample *sample)
829{
830 dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
831 sample->id, event->lost_samples.lost);
832 return 0;
833}
834
835static struct dso *machine__findnew_module_dso(struct machine *machine,
836 struct kmod_path *m,
837 const char *filename)
838{
839 struct dso *dso;
840
841 down_write(&machine->dsos.lock);
842
843 dso = __dsos__find(&machine->dsos, m->name, true);
844 if (!dso) {
845 dso = __dsos__addnew(&machine->dsos, m->name);
846 if (dso == NULL)
847 goto out_unlock;
848
849 dso__set_module_info(dso, m, machine);
850 dso__set_long_name(dso, strdup(filename), true);
851 dso->kernel = DSO_SPACE__KERNEL;
852 }
853
854 dso__get(dso);
855out_unlock:
856 up_write(&machine->dsos.lock);
857 return dso;
858}
859
860int machine__process_aux_event(struct machine *machine __maybe_unused,
861 union perf_event *event)
862{
863 if (dump_trace)
864 perf_event__fprintf_aux(event, stdout);
865 return 0;
866}
867
868int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
869 union perf_event *event)
870{
871 if (dump_trace)
872 perf_event__fprintf_itrace_start(event, stdout);
873 return 0;
874}
875
876int machine__process_aux_output_hw_id_event(struct machine *machine __maybe_unused,
877 union perf_event *event)
878{
879 if (dump_trace)
880 perf_event__fprintf_aux_output_hw_id(event, stdout);
881 return 0;
882}
883
884int machine__process_switch_event(struct machine *machine __maybe_unused,
885 union perf_event *event)
886{
887 if (dump_trace)
888 perf_event__fprintf_switch(event, stdout);
889 return 0;
890}
891
892static int machine__process_ksymbol_register(struct machine *machine,
893 union perf_event *event,
894 struct perf_sample *sample __maybe_unused)
895{
896 struct symbol *sym;
897 struct dso *dso;
898 struct map *map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
899 bool put_map = false;
900 int err = 0;
901
902 if (!map) {
903 dso = dso__new(event->ksymbol.name);
904
905 if (!dso) {
906 err = -ENOMEM;
907 goto out;
908 }
909 dso->kernel = DSO_SPACE__KERNEL;
910 map = map__new2(0, dso);
911 dso__put(dso);
912 if (!map) {
913 err = -ENOMEM;
914 goto out;
915 }
916 /*
917 * The inserted map has a get on it, we need to put to release
918 * the reference count here, but do it after all accesses are
919 * done.
920 */
921 put_map = true;
922 if (event->ksymbol.ksym_type == PERF_RECORD_KSYMBOL_TYPE_OOL) {
923 dso->binary_type = DSO_BINARY_TYPE__OOL;
924 dso->data.file_size = event->ksymbol.len;
925 dso__set_loaded(dso);
926 }
927
928 map__set_start(map, event->ksymbol.addr);
929 map__set_end(map, map__start(map) + event->ksymbol.len);
930 err = maps__insert(machine__kernel_maps(machine), map);
931 if (err) {
932 err = -ENOMEM;
933 goto out;
934 }
935
936 dso__set_loaded(dso);
937
938 if (is_bpf_image(event->ksymbol.name)) {
939 dso->binary_type = DSO_BINARY_TYPE__BPF_IMAGE;
940 dso__set_long_name(dso, "", false);
941 }
942 } else {
943 dso = map__dso(map);
944 }
945
946 sym = symbol__new(map__map_ip(map, map__start(map)),
947 event->ksymbol.len,
948 0, 0, event->ksymbol.name);
949 if (!sym) {
950 err = -ENOMEM;
951 goto out;
952 }
953 dso__insert_symbol(dso, sym);
954out:
955 if (put_map)
956 map__put(map);
957 return err;
958}
959
960static int machine__process_ksymbol_unregister(struct machine *machine,
961 union perf_event *event,
962 struct perf_sample *sample __maybe_unused)
963{
964 struct symbol *sym;
965 struct map *map;
966
967 map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
968 if (!map)
969 return 0;
970
971 if (!RC_CHK_EQUAL(map, machine->vmlinux_map))
972 maps__remove(machine__kernel_maps(machine), map);
973 else {
974 struct dso *dso = map__dso(map);
975
976 sym = dso__find_symbol(dso, map__map_ip(map, map__start(map)));
977 if (sym)
978 dso__delete_symbol(dso, sym);
979 }
980
981 return 0;
982}
983
984int machine__process_ksymbol(struct machine *machine __maybe_unused,
985 union perf_event *event,
986 struct perf_sample *sample)
987{
988 if (dump_trace)
989 perf_event__fprintf_ksymbol(event, stdout);
990
991 if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
992 return machine__process_ksymbol_unregister(machine, event,
993 sample);
994 return machine__process_ksymbol_register(machine, event, sample);
995}
996
997int machine__process_text_poke(struct machine *machine, union perf_event *event,
998 struct perf_sample *sample __maybe_unused)
999{
1000 struct map *map = maps__find(machine__kernel_maps(machine), event->text_poke.addr);
1001 u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1002 struct dso *dso = map ? map__dso(map) : NULL;
1003
1004 if (dump_trace)
1005 perf_event__fprintf_text_poke(event, machine, stdout);
1006
1007 if (!event->text_poke.new_len)
1008 return 0;
1009
1010 if (cpumode != PERF_RECORD_MISC_KERNEL) {
1011 pr_debug("%s: unsupported cpumode - ignoring\n", __func__);
1012 return 0;
1013 }
1014
1015 if (dso) {
1016 u8 *new_bytes = event->text_poke.bytes + event->text_poke.old_len;
1017 int ret;
1018
1019 /*
1020 * Kernel maps might be changed when loading symbols so loading
1021 * must be done prior to using kernel maps.
1022 */
1023 map__load(map);
1024 ret = dso__data_write_cache_addr(dso, map, machine,
1025 event->text_poke.addr,
1026 new_bytes,
1027 event->text_poke.new_len);
1028 if (ret != event->text_poke.new_len)
1029 pr_debug("Failed to write kernel text poke at %#" PRI_lx64 "\n",
1030 event->text_poke.addr);
1031 } else {
1032 pr_debug("Failed to find kernel text poke address map for %#" PRI_lx64 "\n",
1033 event->text_poke.addr);
1034 }
1035
1036 return 0;
1037}
1038
1039static struct map *machine__addnew_module_map(struct machine *machine, u64 start,
1040 const char *filename)
1041{
1042 struct map *map = NULL;
1043 struct kmod_path m;
1044 struct dso *dso;
1045 int err;
1046
1047 if (kmod_path__parse_name(&m, filename))
1048 return NULL;
1049
1050 dso = machine__findnew_module_dso(machine, &m, filename);
1051 if (dso == NULL)
1052 goto out;
1053
1054 map = map__new2(start, dso);
1055 if (map == NULL)
1056 goto out;
1057
1058 err = maps__insert(machine__kernel_maps(machine), map);
1059 /* If maps__insert failed, return NULL. */
1060 if (err) {
1061 map__put(map);
1062 map = NULL;
1063 }
1064out:
1065 /* put the dso here, corresponding to machine__findnew_module_dso */
1066 dso__put(dso);
1067 zfree(&m.name);
1068 return map;
1069}
1070
1071size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
1072{
1073 struct rb_node *nd;
1074 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
1075
1076 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
1077 struct machine *pos = rb_entry(nd, struct machine, rb_node);
1078 ret += __dsos__fprintf(&pos->dsos.head, fp);
1079 }
1080
1081 return ret;
1082}
1083
1084size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
1085 bool (skip)(struct dso *dso, int parm), int parm)
1086{
1087 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
1088}
1089
1090size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
1091 bool (skip)(struct dso *dso, int parm), int parm)
1092{
1093 struct rb_node *nd;
1094 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
1095
1096 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
1097 struct machine *pos = rb_entry(nd, struct machine, rb_node);
1098 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
1099 }
1100 return ret;
1101}
1102
1103size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
1104{
1105 int i;
1106 size_t printed = 0;
1107 struct dso *kdso = machine__kernel_dso(machine);
1108
1109 if (kdso->has_build_id) {
1110 char filename[PATH_MAX];
1111 if (dso__build_id_filename(kdso, filename, sizeof(filename),
1112 false))
1113 printed += fprintf(fp, "[0] %s\n", filename);
1114 }
1115
1116 for (i = 0; i < vmlinux_path__nr_entries; ++i)
1117 printed += fprintf(fp, "[%d] %s\n",
1118 i + kdso->has_build_id, vmlinux_path[i]);
1119
1120 return printed;
1121}
1122
1123size_t machine__fprintf(struct machine *machine, FILE *fp)
1124{
1125 struct rb_node *nd;
1126 size_t ret;
1127 int i;
1128
1129 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
1130 struct threads *threads = &machine->threads[i];
1131
1132 down_read(&threads->lock);
1133
1134 ret = fprintf(fp, "Threads: %u\n", threads->nr);
1135
1136 for (nd = rb_first_cached(&threads->entries); nd;
1137 nd = rb_next(nd)) {
1138 struct thread *pos = rb_entry(nd, struct thread_rb_node, rb_node)->thread;
1139
1140 ret += thread__fprintf(pos, fp);
1141 }
1142
1143 up_read(&threads->lock);
1144 }
1145 return ret;
1146}
1147
1148static struct dso *machine__get_kernel(struct machine *machine)
1149{
1150 const char *vmlinux_name = machine->mmap_name;
1151 struct dso *kernel;
1152
1153 if (machine__is_host(machine)) {
1154 if (symbol_conf.vmlinux_name)
1155 vmlinux_name = symbol_conf.vmlinux_name;
1156
1157 kernel = machine__findnew_kernel(machine, vmlinux_name,
1158 "[kernel]", DSO_SPACE__KERNEL);
1159 } else {
1160 if (symbol_conf.default_guest_vmlinux_name)
1161 vmlinux_name = symbol_conf.default_guest_vmlinux_name;
1162
1163 kernel = machine__findnew_kernel(machine, vmlinux_name,
1164 "[guest.kernel]",
1165 DSO_SPACE__KERNEL_GUEST);
1166 }
1167
1168 if (kernel != NULL && (!kernel->has_build_id))
1169 dso__read_running_kernel_build_id(kernel, machine);
1170
1171 return kernel;
1172}
1173
1174void machine__get_kallsyms_filename(struct machine *machine, char *buf,
1175 size_t bufsz)
1176{
1177 if (machine__is_default_guest(machine))
1178 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
1179 else
1180 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
1181}
1182
1183const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
1184
1185/* Figure out the start address of kernel map from /proc/kallsyms.
1186 * Returns the name of the start symbol in *symbol_name. Pass in NULL as
1187 * symbol_name if it's not that important.
1188 */
1189static int machine__get_running_kernel_start(struct machine *machine,
1190 const char **symbol_name,
1191 u64 *start, u64 *end)
1192{
1193 char filename[PATH_MAX];
1194 int i, err = -1;
1195 const char *name;
1196 u64 addr = 0;
1197
1198 machine__get_kallsyms_filename(machine, filename, PATH_MAX);
1199
1200 if (symbol__restricted_filename(filename, "/proc/kallsyms"))
1201 return 0;
1202
1203 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
1204 err = kallsyms__get_function_start(filename, name, &addr);
1205 if (!err)
1206 break;
1207 }
1208
1209 if (err)
1210 return -1;
1211
1212 if (symbol_name)
1213 *symbol_name = name;
1214
1215 *start = addr;
1216
1217 err = kallsyms__get_symbol_start(filename, "_edata", &addr);
1218 if (err)
1219 err = kallsyms__get_function_start(filename, "_etext", &addr);
1220 if (!err)
1221 *end = addr;
1222
1223 return 0;
1224}
1225
1226int machine__create_extra_kernel_map(struct machine *machine,
1227 struct dso *kernel,
1228 struct extra_kernel_map *xm)
1229{
1230 struct kmap *kmap;
1231 struct map *map;
1232 int err;
1233
1234 map = map__new2(xm->start, kernel);
1235 if (!map)
1236 return -ENOMEM;
1237
1238 map__set_end(map, xm->end);
1239 map__set_pgoff(map, xm->pgoff);
1240
1241 kmap = map__kmap(map);
1242
1243 strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1244
1245 err = maps__insert(machine__kernel_maps(machine), map);
1246
1247 if (!err) {
1248 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1249 kmap->name, map__start(map), map__end(map));
1250 }
1251
1252 map__put(map);
1253
1254 return err;
1255}
1256
1257static u64 find_entry_trampoline(struct dso *dso)
1258{
1259 /* Duplicates are removed so lookup all aliases */
1260 const char *syms[] = {
1261 "_entry_trampoline",
1262 "__entry_trampoline_start",
1263 "entry_SYSCALL_64_trampoline",
1264 };
1265 struct symbol *sym = dso__first_symbol(dso);
1266 unsigned int i;
1267
1268 for (; sym; sym = dso__next_symbol(sym)) {
1269 if (sym->binding != STB_GLOBAL)
1270 continue;
1271 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1272 if (!strcmp(sym->name, syms[i]))
1273 return sym->start;
1274 }
1275 }
1276
1277 return 0;
1278}
1279
1280/*
1281 * These values can be used for kernels that do not have symbols for the entry
1282 * trampolines in kallsyms.
1283 */
1284#define X86_64_CPU_ENTRY_AREA_PER_CPU 0xfffffe0000000000ULL
1285#define X86_64_CPU_ENTRY_AREA_SIZE 0x2c000
1286#define X86_64_ENTRY_TRAMPOLINE 0x6000
1287
1288struct machine__map_x86_64_entry_trampolines_args {
1289 struct maps *kmaps;
1290 bool found;
1291};
1292
1293static int machine__map_x86_64_entry_trampolines_cb(struct map *map, void *data)
1294{
1295 struct machine__map_x86_64_entry_trampolines_args *args = data;
1296 struct map *dest_map;
1297 struct kmap *kmap = __map__kmap(map);
1298
1299 if (!kmap || !is_entry_trampoline(kmap->name))
1300 return 0;
1301
1302 dest_map = maps__find(args->kmaps, map__pgoff(map));
1303 if (dest_map != map)
1304 map__set_pgoff(map, map__map_ip(dest_map, map__pgoff(map)));
1305
1306 args->found = true;
1307 return 0;
1308}
1309
1310/* Map x86_64 PTI entry trampolines */
1311int machine__map_x86_64_entry_trampolines(struct machine *machine,
1312 struct dso *kernel)
1313{
1314 struct machine__map_x86_64_entry_trampolines_args args = {
1315 .kmaps = machine__kernel_maps(machine),
1316 .found = false,
1317 };
1318 int nr_cpus_avail, cpu;
1319 u64 pgoff;
1320
1321 /*
1322 * In the vmlinux case, pgoff is a virtual address which must now be
1323 * mapped to a vmlinux offset.
1324 */
1325 maps__for_each_map(args.kmaps, machine__map_x86_64_entry_trampolines_cb, &args);
1326
1327 if (args.found || machine->trampolines_mapped)
1328 return 0;
1329
1330 pgoff = find_entry_trampoline(kernel);
1331 if (!pgoff)
1332 return 0;
1333
1334 nr_cpus_avail = machine__nr_cpus_avail(machine);
1335
1336 /* Add a 1 page map for each CPU's entry trampoline */
1337 for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1338 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1339 cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1340 X86_64_ENTRY_TRAMPOLINE;
1341 struct extra_kernel_map xm = {
1342 .start = va,
1343 .end = va + page_size,
1344 .pgoff = pgoff,
1345 };
1346
1347 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1348
1349 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1350 return -1;
1351 }
1352
1353 machine->trampolines_mapped = nr_cpus_avail;
1354
1355 return 0;
1356}
1357
1358int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1359 struct dso *kernel __maybe_unused)
1360{
1361 return 0;
1362}
1363
1364static int
1365__machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1366{
1367 /* In case of renewal the kernel map, destroy previous one */
1368 machine__destroy_kernel_maps(machine);
1369
1370 map__put(machine->vmlinux_map);
1371 machine->vmlinux_map = map__new2(0, kernel);
1372 if (machine->vmlinux_map == NULL)
1373 return -ENOMEM;
1374
1375 map__set_mapping_type(machine->vmlinux_map, MAPPING_TYPE__IDENTITY);
1376 return maps__insert(machine__kernel_maps(machine), machine->vmlinux_map);
1377}
1378
1379void machine__destroy_kernel_maps(struct machine *machine)
1380{
1381 struct kmap *kmap;
1382 struct map *map = machine__kernel_map(machine);
1383
1384 if (map == NULL)
1385 return;
1386
1387 kmap = map__kmap(map);
1388 maps__remove(machine__kernel_maps(machine), map);
1389 if (kmap && kmap->ref_reloc_sym) {
1390 zfree((char **)&kmap->ref_reloc_sym->name);
1391 zfree(&kmap->ref_reloc_sym);
1392 }
1393
1394 map__zput(machine->vmlinux_map);
1395}
1396
1397int machines__create_guest_kernel_maps(struct machines *machines)
1398{
1399 int ret = 0;
1400 struct dirent **namelist = NULL;
1401 int i, items = 0;
1402 char path[PATH_MAX];
1403 pid_t pid;
1404 char *endp;
1405
1406 if (symbol_conf.default_guest_vmlinux_name ||
1407 symbol_conf.default_guest_modules ||
1408 symbol_conf.default_guest_kallsyms) {
1409 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1410 }
1411
1412 if (symbol_conf.guestmount) {
1413 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1414 if (items <= 0)
1415 return -ENOENT;
1416 for (i = 0; i < items; i++) {
1417 if (!isdigit(namelist[i]->d_name[0])) {
1418 /* Filter out . and .. */
1419 continue;
1420 }
1421 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1422 if ((*endp != '\0') ||
1423 (endp == namelist[i]->d_name) ||
1424 (errno == ERANGE)) {
1425 pr_debug("invalid directory (%s). Skipping.\n",
1426 namelist[i]->d_name);
1427 continue;
1428 }
1429 sprintf(path, "%s/%s/proc/kallsyms",
1430 symbol_conf.guestmount,
1431 namelist[i]->d_name);
1432 ret = access(path, R_OK);
1433 if (ret) {
1434 pr_debug("Can't access file %s\n", path);
1435 goto failure;
1436 }
1437 machines__create_kernel_maps(machines, pid);
1438 }
1439failure:
1440 free(namelist);
1441 }
1442
1443 return ret;
1444}
1445
1446void machines__destroy_kernel_maps(struct machines *machines)
1447{
1448 struct rb_node *next = rb_first_cached(&machines->guests);
1449
1450 machine__destroy_kernel_maps(&machines->host);
1451
1452 while (next) {
1453 struct machine *pos = rb_entry(next, struct machine, rb_node);
1454
1455 next = rb_next(&pos->rb_node);
1456 rb_erase_cached(&pos->rb_node, &machines->guests);
1457 machine__delete(pos);
1458 }
1459}
1460
1461int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1462{
1463 struct machine *machine = machines__findnew(machines, pid);
1464
1465 if (machine == NULL)
1466 return -1;
1467
1468 return machine__create_kernel_maps(machine);
1469}
1470
1471int machine__load_kallsyms(struct machine *machine, const char *filename)
1472{
1473 struct map *map = machine__kernel_map(machine);
1474 struct dso *dso = map__dso(map);
1475 int ret = __dso__load_kallsyms(dso, filename, map, true);
1476
1477 if (ret > 0) {
1478 dso__set_loaded(dso);
1479 /*
1480 * Since /proc/kallsyms will have multiple sessions for the
1481 * kernel, with modules between them, fixup the end of all
1482 * sections.
1483 */
1484 maps__fixup_end(machine__kernel_maps(machine));
1485 }
1486
1487 return ret;
1488}
1489
1490int machine__load_vmlinux_path(struct machine *machine)
1491{
1492 struct map *map = machine__kernel_map(machine);
1493 struct dso *dso = map__dso(map);
1494 int ret = dso__load_vmlinux_path(dso, map);
1495
1496 if (ret > 0)
1497 dso__set_loaded(dso);
1498
1499 return ret;
1500}
1501
1502static char *get_kernel_version(const char *root_dir)
1503{
1504 char version[PATH_MAX];
1505 FILE *file;
1506 char *name, *tmp;
1507 const char *prefix = "Linux version ";
1508
1509 sprintf(version, "%s/proc/version", root_dir);
1510 file = fopen(version, "r");
1511 if (!file)
1512 return NULL;
1513
1514 tmp = fgets(version, sizeof(version), file);
1515 fclose(file);
1516 if (!tmp)
1517 return NULL;
1518
1519 name = strstr(version, prefix);
1520 if (!name)
1521 return NULL;
1522 name += strlen(prefix);
1523 tmp = strchr(name, ' ');
1524 if (tmp)
1525 *tmp = '\0';
1526
1527 return strdup(name);
1528}
1529
1530static bool is_kmod_dso(struct dso *dso)
1531{
1532 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1533 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1534}
1535
1536static int maps__set_module_path(struct maps *maps, const char *path, struct kmod_path *m)
1537{
1538 char *long_name;
1539 struct dso *dso;
1540 struct map *map = maps__find_by_name(maps, m->name);
1541
1542 if (map == NULL)
1543 return 0;
1544
1545 long_name = strdup(path);
1546 if (long_name == NULL)
1547 return -ENOMEM;
1548
1549 dso = map__dso(map);
1550 dso__set_long_name(dso, long_name, true);
1551 dso__kernel_module_get_build_id(dso, "");
1552
1553 /*
1554 * Full name could reveal us kmod compression, so
1555 * we need to update the symtab_type if needed.
1556 */
1557 if (m->comp && is_kmod_dso(dso)) {
1558 dso->symtab_type++;
1559 dso->comp = m->comp;
1560 }
1561
1562 return 0;
1563}
1564
1565static int maps__set_modules_path_dir(struct maps *maps, const char *dir_name, int depth)
1566{
1567 struct dirent *dent;
1568 DIR *dir = opendir(dir_name);
1569 int ret = 0;
1570
1571 if (!dir) {
1572 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1573 return -1;
1574 }
1575
1576 while ((dent = readdir(dir)) != NULL) {
1577 char path[PATH_MAX];
1578 struct stat st;
1579
1580 /*sshfs might return bad dent->d_type, so we have to stat*/
1581 path__join(path, sizeof(path), dir_name, dent->d_name);
1582 if (stat(path, &st))
1583 continue;
1584
1585 if (S_ISDIR(st.st_mode)) {
1586 if (!strcmp(dent->d_name, ".") ||
1587 !strcmp(dent->d_name, ".."))
1588 continue;
1589
1590 /* Do not follow top-level source and build symlinks */
1591 if (depth == 0) {
1592 if (!strcmp(dent->d_name, "source") ||
1593 !strcmp(dent->d_name, "build"))
1594 continue;
1595 }
1596
1597 ret = maps__set_modules_path_dir(maps, path, depth + 1);
1598 if (ret < 0)
1599 goto out;
1600 } else {
1601 struct kmod_path m;
1602
1603 ret = kmod_path__parse_name(&m, dent->d_name);
1604 if (ret)
1605 goto out;
1606
1607 if (m.kmod)
1608 ret = maps__set_module_path(maps, path, &m);
1609
1610 zfree(&m.name);
1611
1612 if (ret)
1613 goto out;
1614 }
1615 }
1616
1617out:
1618 closedir(dir);
1619 return ret;
1620}
1621
1622static int machine__set_modules_path(struct machine *machine)
1623{
1624 char *version;
1625 char modules_path[PATH_MAX];
1626
1627 version = get_kernel_version(machine->root_dir);
1628 if (!version)
1629 return -1;
1630
1631 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1632 machine->root_dir, version);
1633 free(version);
1634
1635 return maps__set_modules_path_dir(machine__kernel_maps(machine), modules_path, 0);
1636}
1637int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1638 u64 *size __maybe_unused,
1639 const char *name __maybe_unused)
1640{
1641 return 0;
1642}
1643
1644static int machine__create_module(void *arg, const char *name, u64 start,
1645 u64 size)
1646{
1647 struct machine *machine = arg;
1648 struct map *map;
1649
1650 if (arch__fix_module_text_start(&start, &size, name) < 0)
1651 return -1;
1652
1653 map = machine__addnew_module_map(machine, start, name);
1654 if (map == NULL)
1655 return -1;
1656 map__set_end(map, start + size);
1657
1658 dso__kernel_module_get_build_id(map__dso(map), machine->root_dir);
1659 map__put(map);
1660 return 0;
1661}
1662
1663static int machine__create_modules(struct machine *machine)
1664{
1665 const char *modules;
1666 char path[PATH_MAX];
1667
1668 if (machine__is_default_guest(machine)) {
1669 modules = symbol_conf.default_guest_modules;
1670 } else {
1671 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1672 modules = path;
1673 }
1674
1675 if (symbol__restricted_filename(modules, "/proc/modules"))
1676 return -1;
1677
1678 if (modules__parse(modules, machine, machine__create_module))
1679 return -1;
1680
1681 if (!machine__set_modules_path(machine))
1682 return 0;
1683
1684 pr_debug("Problems setting modules path maps, continuing anyway...\n");
1685
1686 return 0;
1687}
1688
1689static void machine__set_kernel_mmap(struct machine *machine,
1690 u64 start, u64 end)
1691{
1692 map__set_start(machine->vmlinux_map, start);
1693 map__set_end(machine->vmlinux_map, end);
1694 /*
1695 * Be a bit paranoid here, some perf.data file came with
1696 * a zero sized synthesized MMAP event for the kernel.
1697 */
1698 if (start == 0 && end == 0)
1699 map__set_end(machine->vmlinux_map, ~0ULL);
1700}
1701
1702static int machine__update_kernel_mmap(struct machine *machine,
1703 u64 start, u64 end)
1704{
1705 struct map *orig, *updated;
1706 int err;
1707
1708 orig = machine->vmlinux_map;
1709 updated = map__get(orig);
1710
1711 machine->vmlinux_map = updated;
1712 machine__set_kernel_mmap(machine, start, end);
1713 maps__remove(machine__kernel_maps(machine), orig);
1714 err = maps__insert(machine__kernel_maps(machine), updated);
1715 map__put(orig);
1716
1717 return err;
1718}
1719
1720int machine__create_kernel_maps(struct machine *machine)
1721{
1722 struct dso *kernel = machine__get_kernel(machine);
1723 const char *name = NULL;
1724 u64 start = 0, end = ~0ULL;
1725 int ret;
1726
1727 if (kernel == NULL)
1728 return -1;
1729
1730 ret = __machine__create_kernel_maps(machine, kernel);
1731 if (ret < 0)
1732 goto out_put;
1733
1734 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1735 if (machine__is_host(machine))
1736 pr_debug("Problems creating module maps, "
1737 "continuing anyway...\n");
1738 else
1739 pr_debug("Problems creating module maps for guest %d, "
1740 "continuing anyway...\n", machine->pid);
1741 }
1742
1743 if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1744 if (name &&
1745 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1746 machine__destroy_kernel_maps(machine);
1747 ret = -1;
1748 goto out_put;
1749 }
1750
1751 /*
1752 * we have a real start address now, so re-order the kmaps
1753 * assume it's the last in the kmaps
1754 */
1755 ret = machine__update_kernel_mmap(machine, start, end);
1756 if (ret < 0)
1757 goto out_put;
1758 }
1759
1760 if (machine__create_extra_kernel_maps(machine, kernel))
1761 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1762
1763 if (end == ~0ULL) {
1764 /* update end address of the kernel map using adjacent module address */
1765 struct map *next = maps__find_next_entry(machine__kernel_maps(machine),
1766 machine__kernel_map(machine));
1767
1768 if (next)
1769 machine__set_kernel_mmap(machine, start, map__start(next));
1770 }
1771
1772out_put:
1773 dso__put(kernel);
1774 return ret;
1775}
1776
1777static bool machine__uses_kcore(struct machine *machine)
1778{
1779 struct dso *dso;
1780
1781 list_for_each_entry(dso, &machine->dsos.head, node) {
1782 if (dso__is_kcore(dso))
1783 return true;
1784 }
1785
1786 return false;
1787}
1788
1789static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1790 struct extra_kernel_map *xm)
1791{
1792 return machine__is(machine, "x86_64") &&
1793 is_entry_trampoline(xm->name);
1794}
1795
1796static int machine__process_extra_kernel_map(struct machine *machine,
1797 struct extra_kernel_map *xm)
1798{
1799 struct dso *kernel = machine__kernel_dso(machine);
1800
1801 if (kernel == NULL)
1802 return -1;
1803
1804 return machine__create_extra_kernel_map(machine, kernel, xm);
1805}
1806
1807static int machine__process_kernel_mmap_event(struct machine *machine,
1808 struct extra_kernel_map *xm,
1809 struct build_id *bid)
1810{
1811 enum dso_space_type dso_space;
1812 bool is_kernel_mmap;
1813 const char *mmap_name = machine->mmap_name;
1814
1815 /* If we have maps from kcore then we do not need or want any others */
1816 if (machine__uses_kcore(machine))
1817 return 0;
1818
1819 if (machine__is_host(machine))
1820 dso_space = DSO_SPACE__KERNEL;
1821 else
1822 dso_space = DSO_SPACE__KERNEL_GUEST;
1823
1824 is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1825 if (!is_kernel_mmap && !machine__is_host(machine)) {
1826 /*
1827 * If the event was recorded inside the guest and injected into
1828 * the host perf.data file, then it will match a host mmap_name,
1829 * so try that - see machine__set_mmap_name().
1830 */
1831 mmap_name = "[kernel.kallsyms]";
1832 is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1833 }
1834 if (xm->name[0] == '/' ||
1835 (!is_kernel_mmap && xm->name[0] == '[')) {
1836 struct map *map = machine__addnew_module_map(machine, xm->start, xm->name);
1837
1838 if (map == NULL)
1839 goto out_problem;
1840
1841 map__set_end(map, map__start(map) + xm->end - xm->start);
1842
1843 if (build_id__is_defined(bid))
1844 dso__set_build_id(map__dso(map), bid);
1845
1846 map__put(map);
1847 } else if (is_kernel_mmap) {
1848 const char *symbol_name = xm->name + strlen(mmap_name);
1849 /*
1850 * Should be there already, from the build-id table in
1851 * the header.
1852 */
1853 struct dso *kernel = NULL;
1854 struct dso *dso;
1855
1856 down_read(&machine->dsos.lock);
1857
1858 list_for_each_entry(dso, &machine->dsos.head, node) {
1859
1860 /*
1861 * The cpumode passed to is_kernel_module is not the
1862 * cpumode of *this* event. If we insist on passing
1863 * correct cpumode to is_kernel_module, we should
1864 * record the cpumode when we adding this dso to the
1865 * linked list.
1866 *
1867 * However we don't really need passing correct
1868 * cpumode. We know the correct cpumode must be kernel
1869 * mode (if not, we should not link it onto kernel_dsos
1870 * list).
1871 *
1872 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1873 * is_kernel_module() treats it as a kernel cpumode.
1874 */
1875
1876 if (!dso->kernel ||
1877 is_kernel_module(dso->long_name,
1878 PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1879 continue;
1880
1881
1882 kernel = dso__get(dso);
1883 break;
1884 }
1885
1886 up_read(&machine->dsos.lock);
1887
1888 if (kernel == NULL)
1889 kernel = machine__findnew_dso(machine, machine->mmap_name);
1890 if (kernel == NULL)
1891 goto out_problem;
1892
1893 kernel->kernel = dso_space;
1894 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1895 dso__put(kernel);
1896 goto out_problem;
1897 }
1898
1899 if (strstr(kernel->long_name, "vmlinux"))
1900 dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1901
1902 if (machine__update_kernel_mmap(machine, xm->start, xm->end) < 0) {
1903 dso__put(kernel);
1904 goto out_problem;
1905 }
1906
1907 if (build_id__is_defined(bid))
1908 dso__set_build_id(kernel, bid);
1909
1910 /*
1911 * Avoid using a zero address (kptr_restrict) for the ref reloc
1912 * symbol. Effectively having zero here means that at record
1913 * time /proc/sys/kernel/kptr_restrict was non zero.
1914 */
1915 if (xm->pgoff != 0) {
1916 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1917 symbol_name,
1918 xm->pgoff);
1919 }
1920
1921 if (machine__is_default_guest(machine)) {
1922 /*
1923 * preload dso of guest kernel and modules
1924 */
1925 dso__load(kernel, machine__kernel_map(machine));
1926 }
1927 dso__put(kernel);
1928 } else if (perf_event__is_extra_kernel_mmap(machine, xm)) {
1929 return machine__process_extra_kernel_map(machine, xm);
1930 }
1931 return 0;
1932out_problem:
1933 return -1;
1934}
1935
1936int machine__process_mmap2_event(struct machine *machine,
1937 union perf_event *event,
1938 struct perf_sample *sample)
1939{
1940 struct thread *thread;
1941 struct map *map;
1942 struct dso_id dso_id = {
1943 .maj = event->mmap2.maj,
1944 .min = event->mmap2.min,
1945 .ino = event->mmap2.ino,
1946 .ino_generation = event->mmap2.ino_generation,
1947 };
1948 struct build_id __bid, *bid = NULL;
1949 int ret = 0;
1950
1951 if (dump_trace)
1952 perf_event__fprintf_mmap2(event, stdout);
1953
1954 if (event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID) {
1955 bid = &__bid;
1956 build_id__init(bid, event->mmap2.build_id, event->mmap2.build_id_size);
1957 }
1958
1959 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1960 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1961 struct extra_kernel_map xm = {
1962 .start = event->mmap2.start,
1963 .end = event->mmap2.start + event->mmap2.len,
1964 .pgoff = event->mmap2.pgoff,
1965 };
1966
1967 strlcpy(xm.name, event->mmap2.filename, KMAP_NAME_LEN);
1968 ret = machine__process_kernel_mmap_event(machine, &xm, bid);
1969 if (ret < 0)
1970 goto out_problem;
1971 return 0;
1972 }
1973
1974 thread = machine__findnew_thread(machine, event->mmap2.pid,
1975 event->mmap2.tid);
1976 if (thread == NULL)
1977 goto out_problem;
1978
1979 map = map__new(machine, event->mmap2.start,
1980 event->mmap2.len, event->mmap2.pgoff,
1981 &dso_id, event->mmap2.prot,
1982 event->mmap2.flags, bid,
1983 event->mmap2.filename, thread);
1984
1985 if (map == NULL)
1986 goto out_problem_map;
1987
1988 ret = thread__insert_map(thread, map);
1989 if (ret)
1990 goto out_problem_insert;
1991
1992 thread__put(thread);
1993 map__put(map);
1994 return 0;
1995
1996out_problem_insert:
1997 map__put(map);
1998out_problem_map:
1999 thread__put(thread);
2000out_problem:
2001 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
2002 return 0;
2003}
2004
2005int machine__process_mmap_event(struct machine *machine, union perf_event *event,
2006 struct perf_sample *sample)
2007{
2008 struct thread *thread;
2009 struct map *map;
2010 u32 prot = 0;
2011 int ret = 0;
2012
2013 if (dump_trace)
2014 perf_event__fprintf_mmap(event, stdout);
2015
2016 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
2017 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
2018 struct extra_kernel_map xm = {
2019 .start = event->mmap.start,
2020 .end = event->mmap.start + event->mmap.len,
2021 .pgoff = event->mmap.pgoff,
2022 };
2023
2024 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
2025 ret = machine__process_kernel_mmap_event(machine, &xm, NULL);
2026 if (ret < 0)
2027 goto out_problem;
2028 return 0;
2029 }
2030
2031 thread = machine__findnew_thread(machine, event->mmap.pid,
2032 event->mmap.tid);
2033 if (thread == NULL)
2034 goto out_problem;
2035
2036 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
2037 prot = PROT_EXEC;
2038
2039 map = map__new(machine, event->mmap.start,
2040 event->mmap.len, event->mmap.pgoff,
2041 NULL, prot, 0, NULL, event->mmap.filename, thread);
2042
2043 if (map == NULL)
2044 goto out_problem_map;
2045
2046 ret = thread__insert_map(thread, map);
2047 if (ret)
2048 goto out_problem_insert;
2049
2050 thread__put(thread);
2051 map__put(map);
2052 return 0;
2053
2054out_problem_insert:
2055 map__put(map);
2056out_problem_map:
2057 thread__put(thread);
2058out_problem:
2059 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
2060 return 0;
2061}
2062
2063static void __machine__remove_thread(struct machine *machine, struct thread_rb_node *nd,
2064 struct thread *th, bool lock)
2065{
2066 struct threads *threads = machine__threads(machine, thread__tid(th));
2067
2068 if (!nd)
2069 nd = thread_rb_node__find(th, &threads->entries.rb_root);
2070
2071 if (threads->last_match && RC_CHK_EQUAL(threads->last_match, th))
2072 threads__set_last_match(threads, NULL);
2073
2074 if (lock)
2075 down_write(&threads->lock);
2076
2077 BUG_ON(refcount_read(thread__refcnt(th)) == 0);
2078
2079 thread__put(nd->thread);
2080 rb_erase_cached(&nd->rb_node, &threads->entries);
2081 RB_CLEAR_NODE(&nd->rb_node);
2082 --threads->nr;
2083
2084 free(nd);
2085
2086 if (lock)
2087 up_write(&threads->lock);
2088}
2089
2090void machine__remove_thread(struct machine *machine, struct thread *th)
2091{
2092 return __machine__remove_thread(machine, NULL, th, true);
2093}
2094
2095int machine__process_fork_event(struct machine *machine, union perf_event *event,
2096 struct perf_sample *sample)
2097{
2098 struct thread *thread = machine__find_thread(machine,
2099 event->fork.pid,
2100 event->fork.tid);
2101 struct thread *parent = machine__findnew_thread(machine,
2102 event->fork.ppid,
2103 event->fork.ptid);
2104 bool do_maps_clone = true;
2105 int err = 0;
2106
2107 if (dump_trace)
2108 perf_event__fprintf_task(event, stdout);
2109
2110 /*
2111 * There may be an existing thread that is not actually the parent,
2112 * either because we are processing events out of order, or because the
2113 * (fork) event that would have removed the thread was lost. Assume the
2114 * latter case and continue on as best we can.
2115 */
2116 if (thread__pid(parent) != (pid_t)event->fork.ppid) {
2117 dump_printf("removing erroneous parent thread %d/%d\n",
2118 thread__pid(parent), thread__tid(parent));
2119 machine__remove_thread(machine, parent);
2120 thread__put(parent);
2121 parent = machine__findnew_thread(machine, event->fork.ppid,
2122 event->fork.ptid);
2123 }
2124
2125 /* if a thread currently exists for the thread id remove it */
2126 if (thread != NULL) {
2127 machine__remove_thread(machine, thread);
2128 thread__put(thread);
2129 }
2130
2131 thread = machine__findnew_thread(machine, event->fork.pid,
2132 event->fork.tid);
2133 /*
2134 * When synthesizing FORK events, we are trying to create thread
2135 * objects for the already running tasks on the machine.
2136 *
2137 * Normally, for a kernel FORK event, we want to clone the parent's
2138 * maps because that is what the kernel just did.
2139 *
2140 * But when synthesizing, this should not be done. If we do, we end up
2141 * with overlapping maps as we process the synthesized MMAP2 events that
2142 * get delivered shortly thereafter.
2143 *
2144 * Use the FORK event misc flags in an internal way to signal this
2145 * situation, so we can elide the map clone when appropriate.
2146 */
2147 if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
2148 do_maps_clone = false;
2149
2150 if (thread == NULL || parent == NULL ||
2151 thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
2152 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
2153 err = -1;
2154 }
2155 thread__put(thread);
2156 thread__put(parent);
2157
2158 return err;
2159}
2160
2161int machine__process_exit_event(struct machine *machine, union perf_event *event,
2162 struct perf_sample *sample __maybe_unused)
2163{
2164 struct thread *thread = machine__find_thread(machine,
2165 event->fork.pid,
2166 event->fork.tid);
2167
2168 if (dump_trace)
2169 perf_event__fprintf_task(event, stdout);
2170
2171 if (thread != NULL) {
2172 if (symbol_conf.keep_exited_threads)
2173 thread__set_exited(thread, /*exited=*/true);
2174 else
2175 machine__remove_thread(machine, thread);
2176 }
2177 thread__put(thread);
2178 return 0;
2179}
2180
2181int machine__process_event(struct machine *machine, union perf_event *event,
2182 struct perf_sample *sample)
2183{
2184 int ret;
2185
2186 switch (event->header.type) {
2187 case PERF_RECORD_COMM:
2188 ret = machine__process_comm_event(machine, event, sample); break;
2189 case PERF_RECORD_MMAP:
2190 ret = machine__process_mmap_event(machine, event, sample); break;
2191 case PERF_RECORD_NAMESPACES:
2192 ret = machine__process_namespaces_event(machine, event, sample); break;
2193 case PERF_RECORD_CGROUP:
2194 ret = machine__process_cgroup_event(machine, event, sample); break;
2195 case PERF_RECORD_MMAP2:
2196 ret = machine__process_mmap2_event(machine, event, sample); break;
2197 case PERF_RECORD_FORK:
2198 ret = machine__process_fork_event(machine, event, sample); break;
2199 case PERF_RECORD_EXIT:
2200 ret = machine__process_exit_event(machine, event, sample); break;
2201 case PERF_RECORD_LOST:
2202 ret = machine__process_lost_event(machine, event, sample); break;
2203 case PERF_RECORD_AUX:
2204 ret = machine__process_aux_event(machine, event); break;
2205 case PERF_RECORD_ITRACE_START:
2206 ret = machine__process_itrace_start_event(machine, event); break;
2207 case PERF_RECORD_LOST_SAMPLES:
2208 ret = machine__process_lost_samples_event(machine, event, sample); break;
2209 case PERF_RECORD_SWITCH:
2210 case PERF_RECORD_SWITCH_CPU_WIDE:
2211 ret = machine__process_switch_event(machine, event); break;
2212 case PERF_RECORD_KSYMBOL:
2213 ret = machine__process_ksymbol(machine, event, sample); break;
2214 case PERF_RECORD_BPF_EVENT:
2215 ret = machine__process_bpf(machine, event, sample); break;
2216 case PERF_RECORD_TEXT_POKE:
2217 ret = machine__process_text_poke(machine, event, sample); break;
2218 case PERF_RECORD_AUX_OUTPUT_HW_ID:
2219 ret = machine__process_aux_output_hw_id_event(machine, event); break;
2220 default:
2221 ret = -1;
2222 break;
2223 }
2224
2225 return ret;
2226}
2227
2228static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
2229{
2230 return regexec(regex, sym->name, 0, NULL, 0) == 0;
2231}
2232
2233static void ip__resolve_ams(struct thread *thread,
2234 struct addr_map_symbol *ams,
2235 u64 ip)
2236{
2237 struct addr_location al;
2238
2239 addr_location__init(&al);
2240 /*
2241 * We cannot use the header.misc hint to determine whether a
2242 * branch stack address is user, kernel, guest, hypervisor.
2243 * Branches may straddle the kernel/user/hypervisor boundaries.
2244 * Thus, we have to try consecutively until we find a match
2245 * or else, the symbol is unknown
2246 */
2247 thread__find_cpumode_addr_location(thread, ip, &al);
2248
2249 ams->addr = ip;
2250 ams->al_addr = al.addr;
2251 ams->al_level = al.level;
2252 ams->ms.maps = maps__get(al.maps);
2253 ams->ms.sym = al.sym;
2254 ams->ms.map = map__get(al.map);
2255 ams->phys_addr = 0;
2256 ams->data_page_size = 0;
2257 addr_location__exit(&al);
2258}
2259
2260static void ip__resolve_data(struct thread *thread,
2261 u8 m, struct addr_map_symbol *ams,
2262 u64 addr, u64 phys_addr, u64 daddr_page_size)
2263{
2264 struct addr_location al;
2265
2266 addr_location__init(&al);
2267
2268 thread__find_symbol(thread, m, addr, &al);
2269
2270 ams->addr = addr;
2271 ams->al_addr = al.addr;
2272 ams->al_level = al.level;
2273 ams->ms.maps = maps__get(al.maps);
2274 ams->ms.sym = al.sym;
2275 ams->ms.map = map__get(al.map);
2276 ams->phys_addr = phys_addr;
2277 ams->data_page_size = daddr_page_size;
2278 addr_location__exit(&al);
2279}
2280
2281struct mem_info *sample__resolve_mem(struct perf_sample *sample,
2282 struct addr_location *al)
2283{
2284 struct mem_info *mi = mem_info__new();
2285
2286 if (!mi)
2287 return NULL;
2288
2289 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
2290 ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
2291 sample->addr, sample->phys_addr,
2292 sample->data_page_size);
2293 mi->data_src.val = sample->data_src;
2294
2295 return mi;
2296}
2297
2298static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2299{
2300 struct map *map = ms->map;
2301 char *srcline = NULL;
2302 struct dso *dso;
2303
2304 if (!map || callchain_param.key == CCKEY_FUNCTION)
2305 return srcline;
2306
2307 dso = map__dso(map);
2308 srcline = srcline__tree_find(&dso->srclines, ip);
2309 if (!srcline) {
2310 bool show_sym = false;
2311 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2312
2313 srcline = get_srcline(dso, map__rip_2objdump(map, ip),
2314 ms->sym, show_sym, show_addr, ip);
2315 srcline__tree_insert(&dso->srclines, ip, srcline);
2316 }
2317
2318 return srcline;
2319}
2320
2321struct iterations {
2322 int nr_loop_iter;
2323 u64 cycles;
2324};
2325
2326static int add_callchain_ip(struct thread *thread,
2327 struct callchain_cursor *cursor,
2328 struct symbol **parent,
2329 struct addr_location *root_al,
2330 u8 *cpumode,
2331 u64 ip,
2332 bool branch,
2333 struct branch_flags *flags,
2334 struct iterations *iter,
2335 u64 branch_from)
2336{
2337 struct map_symbol ms = {};
2338 struct addr_location al;
2339 int nr_loop_iter = 0, err = 0;
2340 u64 iter_cycles = 0;
2341 const char *srcline = NULL;
2342
2343 addr_location__init(&al);
2344 al.filtered = 0;
2345 al.sym = NULL;
2346 al.srcline = NULL;
2347 if (!cpumode) {
2348 thread__find_cpumode_addr_location(thread, ip, &al);
2349 } else {
2350 if (ip >= PERF_CONTEXT_MAX) {
2351 switch (ip) {
2352 case PERF_CONTEXT_HV:
2353 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2354 break;
2355 case PERF_CONTEXT_KERNEL:
2356 *cpumode = PERF_RECORD_MISC_KERNEL;
2357 break;
2358 case PERF_CONTEXT_USER:
2359 *cpumode = PERF_RECORD_MISC_USER;
2360 break;
2361 default:
2362 pr_debug("invalid callchain context: "
2363 "%"PRId64"\n", (s64) ip);
2364 /*
2365 * It seems the callchain is corrupted.
2366 * Discard all.
2367 */
2368 callchain_cursor_reset(cursor);
2369 err = 1;
2370 goto out;
2371 }
2372 goto out;
2373 }
2374 thread__find_symbol(thread, *cpumode, ip, &al);
2375 }
2376
2377 if (al.sym != NULL) {
2378 if (perf_hpp_list.parent && !*parent &&
2379 symbol__match_regex(al.sym, &parent_regex))
2380 *parent = al.sym;
2381 else if (have_ignore_callees && root_al &&
2382 symbol__match_regex(al.sym, &ignore_callees_regex)) {
2383 /* Treat this symbol as the root,
2384 forgetting its callees. */
2385 addr_location__copy(root_al, &al);
2386 callchain_cursor_reset(cursor);
2387 }
2388 }
2389
2390 if (symbol_conf.hide_unresolved && al.sym == NULL)
2391 goto out;
2392
2393 if (iter) {
2394 nr_loop_iter = iter->nr_loop_iter;
2395 iter_cycles = iter->cycles;
2396 }
2397
2398 ms.maps = maps__get(al.maps);
2399 ms.map = map__get(al.map);
2400 ms.sym = al.sym;
2401 srcline = callchain_srcline(&ms, al.addr);
2402 err = callchain_cursor_append(cursor, ip, &ms,
2403 branch, flags, nr_loop_iter,
2404 iter_cycles, branch_from, srcline);
2405out:
2406 addr_location__exit(&al);
2407 map_symbol__exit(&ms);
2408 return err;
2409}
2410
2411struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2412 struct addr_location *al)
2413{
2414 unsigned int i;
2415 const struct branch_stack *bs = sample->branch_stack;
2416 struct branch_entry *entries = perf_sample__branch_entries(sample);
2417 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2418
2419 if (!bi)
2420 return NULL;
2421
2422 for (i = 0; i < bs->nr; i++) {
2423 ip__resolve_ams(al->thread, &bi[i].to, entries[i].to);
2424 ip__resolve_ams(al->thread, &bi[i].from, entries[i].from);
2425 bi[i].flags = entries[i].flags;
2426 }
2427 return bi;
2428}
2429
2430static void save_iterations(struct iterations *iter,
2431 struct branch_entry *be, int nr)
2432{
2433 int i;
2434
2435 iter->nr_loop_iter++;
2436 iter->cycles = 0;
2437
2438 for (i = 0; i < nr; i++)
2439 iter->cycles += be[i].flags.cycles;
2440}
2441
2442#define CHASHSZ 127
2443#define CHASHBITS 7
2444#define NO_ENTRY 0xff
2445
2446#define PERF_MAX_BRANCH_DEPTH 127
2447
2448/* Remove loops. */
2449static int remove_loops(struct branch_entry *l, int nr,
2450 struct iterations *iter)
2451{
2452 int i, j, off;
2453 unsigned char chash[CHASHSZ];
2454
2455 memset(chash, NO_ENTRY, sizeof(chash));
2456
2457 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2458
2459 for (i = 0; i < nr; i++) {
2460 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2461
2462 /* no collision handling for now */
2463 if (chash[h] == NO_ENTRY) {
2464 chash[h] = i;
2465 } else if (l[chash[h]].from == l[i].from) {
2466 bool is_loop = true;
2467 /* check if it is a real loop */
2468 off = 0;
2469 for (j = chash[h]; j < i && i + off < nr; j++, off++)
2470 if (l[j].from != l[i + off].from) {
2471 is_loop = false;
2472 break;
2473 }
2474 if (is_loop) {
2475 j = nr - (i + off);
2476 if (j > 0) {
2477 save_iterations(iter + i + off,
2478 l + i, off);
2479
2480 memmove(iter + i, iter + i + off,
2481 j * sizeof(*iter));
2482
2483 memmove(l + i, l + i + off,
2484 j * sizeof(*l));
2485 }
2486
2487 nr -= off;
2488 }
2489 }
2490 }
2491 return nr;
2492}
2493
2494static int lbr_callchain_add_kernel_ip(struct thread *thread,
2495 struct callchain_cursor *cursor,
2496 struct perf_sample *sample,
2497 struct symbol **parent,
2498 struct addr_location *root_al,
2499 u64 branch_from,
2500 bool callee, int end)
2501{
2502 struct ip_callchain *chain = sample->callchain;
2503 u8 cpumode = PERF_RECORD_MISC_USER;
2504 int err, i;
2505
2506 if (callee) {
2507 for (i = 0; i < end + 1; i++) {
2508 err = add_callchain_ip(thread, cursor, parent,
2509 root_al, &cpumode, chain->ips[i],
2510 false, NULL, NULL, branch_from);
2511 if (err)
2512 return err;
2513 }
2514 return 0;
2515 }
2516
2517 for (i = end; i >= 0; i--) {
2518 err = add_callchain_ip(thread, cursor, parent,
2519 root_al, &cpumode, chain->ips[i],
2520 false, NULL, NULL, branch_from);
2521 if (err)
2522 return err;
2523 }
2524
2525 return 0;
2526}
2527
2528static void save_lbr_cursor_node(struct thread *thread,
2529 struct callchain_cursor *cursor,
2530 int idx)
2531{
2532 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2533
2534 if (!lbr_stitch)
2535 return;
2536
2537 if (cursor->pos == cursor->nr) {
2538 lbr_stitch->prev_lbr_cursor[idx].valid = false;
2539 return;
2540 }
2541
2542 if (!cursor->curr)
2543 cursor->curr = cursor->first;
2544 else
2545 cursor->curr = cursor->curr->next;
2546 memcpy(&lbr_stitch->prev_lbr_cursor[idx], cursor->curr,
2547 sizeof(struct callchain_cursor_node));
2548
2549 lbr_stitch->prev_lbr_cursor[idx].valid = true;
2550 cursor->pos++;
2551}
2552
2553static int lbr_callchain_add_lbr_ip(struct thread *thread,
2554 struct callchain_cursor *cursor,
2555 struct perf_sample *sample,
2556 struct symbol **parent,
2557 struct addr_location *root_al,
2558 u64 *branch_from,
2559 bool callee)
2560{
2561 struct branch_stack *lbr_stack = sample->branch_stack;
2562 struct branch_entry *entries = perf_sample__branch_entries(sample);
2563 u8 cpumode = PERF_RECORD_MISC_USER;
2564 int lbr_nr = lbr_stack->nr;
2565 struct branch_flags *flags;
2566 int err, i;
2567 u64 ip;
2568
2569 /*
2570 * The curr and pos are not used in writing session. They are cleared
2571 * in callchain_cursor_commit() when the writing session is closed.
2572 * Using curr and pos to track the current cursor node.
2573 */
2574 if (thread__lbr_stitch(thread)) {
2575 cursor->curr = NULL;
2576 cursor->pos = cursor->nr;
2577 if (cursor->nr) {
2578 cursor->curr = cursor->first;
2579 for (i = 0; i < (int)(cursor->nr - 1); i++)
2580 cursor->curr = cursor->curr->next;
2581 }
2582 }
2583
2584 if (callee) {
2585 /* Add LBR ip from first entries.to */
2586 ip = entries[0].to;
2587 flags = &entries[0].flags;
2588 *branch_from = entries[0].from;
2589 err = add_callchain_ip(thread, cursor, parent,
2590 root_al, &cpumode, ip,
2591 true, flags, NULL,
2592 *branch_from);
2593 if (err)
2594 return err;
2595
2596 /*
2597 * The number of cursor node increases.
2598 * Move the current cursor node.
2599 * But does not need to save current cursor node for entry 0.
2600 * It's impossible to stitch the whole LBRs of previous sample.
2601 */
2602 if (thread__lbr_stitch(thread) && (cursor->pos != cursor->nr)) {
2603 if (!cursor->curr)
2604 cursor->curr = cursor->first;
2605 else
2606 cursor->curr = cursor->curr->next;
2607 cursor->pos++;
2608 }
2609
2610 /* Add LBR ip from entries.from one by one. */
2611 for (i = 0; i < lbr_nr; i++) {
2612 ip = entries[i].from;
2613 flags = &entries[i].flags;
2614 err = add_callchain_ip(thread, cursor, parent,
2615 root_al, &cpumode, ip,
2616 true, flags, NULL,
2617 *branch_from);
2618 if (err)
2619 return err;
2620 save_lbr_cursor_node(thread, cursor, i);
2621 }
2622 return 0;
2623 }
2624
2625 /* Add LBR ip from entries.from one by one. */
2626 for (i = lbr_nr - 1; i >= 0; i--) {
2627 ip = entries[i].from;
2628 flags = &entries[i].flags;
2629 err = add_callchain_ip(thread, cursor, parent,
2630 root_al, &cpumode, ip,
2631 true, flags, NULL,
2632 *branch_from);
2633 if (err)
2634 return err;
2635 save_lbr_cursor_node(thread, cursor, i);
2636 }
2637
2638 if (lbr_nr > 0) {
2639 /* Add LBR ip from first entries.to */
2640 ip = entries[0].to;
2641 flags = &entries[0].flags;
2642 *branch_from = entries[0].from;
2643 err = add_callchain_ip(thread, cursor, parent,
2644 root_al, &cpumode, ip,
2645 true, flags, NULL,
2646 *branch_from);
2647 if (err)
2648 return err;
2649 }
2650
2651 return 0;
2652}
2653
2654static int lbr_callchain_add_stitched_lbr_ip(struct thread *thread,
2655 struct callchain_cursor *cursor)
2656{
2657 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2658 struct callchain_cursor_node *cnode;
2659 struct stitch_list *stitch_node;
2660 int err;
2661
2662 list_for_each_entry(stitch_node, &lbr_stitch->lists, node) {
2663 cnode = &stitch_node->cursor;
2664
2665 err = callchain_cursor_append(cursor, cnode->ip,
2666 &cnode->ms,
2667 cnode->branch,
2668 &cnode->branch_flags,
2669 cnode->nr_loop_iter,
2670 cnode->iter_cycles,
2671 cnode->branch_from,
2672 cnode->srcline);
2673 if (err)
2674 return err;
2675 }
2676 return 0;
2677}
2678
2679static struct stitch_list *get_stitch_node(struct thread *thread)
2680{
2681 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2682 struct stitch_list *stitch_node;
2683
2684 if (!list_empty(&lbr_stitch->free_lists)) {
2685 stitch_node = list_first_entry(&lbr_stitch->free_lists,
2686 struct stitch_list, node);
2687 list_del(&stitch_node->node);
2688
2689 return stitch_node;
2690 }
2691
2692 return malloc(sizeof(struct stitch_list));
2693}
2694
2695static bool has_stitched_lbr(struct thread *thread,
2696 struct perf_sample *cur,
2697 struct perf_sample *prev,
2698 unsigned int max_lbr,
2699 bool callee)
2700{
2701 struct branch_stack *cur_stack = cur->branch_stack;
2702 struct branch_entry *cur_entries = perf_sample__branch_entries(cur);
2703 struct branch_stack *prev_stack = prev->branch_stack;
2704 struct branch_entry *prev_entries = perf_sample__branch_entries(prev);
2705 struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2706 int i, j, nr_identical_branches = 0;
2707 struct stitch_list *stitch_node;
2708 u64 cur_base, distance;
2709
2710 if (!cur_stack || !prev_stack)
2711 return false;
2712
2713 /* Find the physical index of the base-of-stack for current sample. */
2714 cur_base = max_lbr - cur_stack->nr + cur_stack->hw_idx + 1;
2715
2716 distance = (prev_stack->hw_idx > cur_base) ? (prev_stack->hw_idx - cur_base) :
2717 (max_lbr + prev_stack->hw_idx - cur_base);
2718 /* Previous sample has shorter stack. Nothing can be stitched. */
2719 if (distance + 1 > prev_stack->nr)
2720 return false;
2721
2722 /*
2723 * Check if there are identical LBRs between two samples.
2724 * Identical LBRs must have same from, to and flags values. Also,
2725 * they have to be saved in the same LBR registers (same physical
2726 * index).
2727 *
2728 * Starts from the base-of-stack of current sample.
2729 */
2730 for (i = distance, j = cur_stack->nr - 1; (i >= 0) && (j >= 0); i--, j--) {
2731 if ((prev_entries[i].from != cur_entries[j].from) ||
2732 (prev_entries[i].to != cur_entries[j].to) ||
2733 (prev_entries[i].flags.value != cur_entries[j].flags.value))
2734 break;
2735 nr_identical_branches++;
2736 }
2737
2738 if (!nr_identical_branches)
2739 return false;
2740
2741 /*
2742 * Save the LBRs between the base-of-stack of previous sample
2743 * and the base-of-stack of current sample into lbr_stitch->lists.
2744 * These LBRs will be stitched later.
2745 */
2746 for (i = prev_stack->nr - 1; i > (int)distance; i--) {
2747
2748 if (!lbr_stitch->prev_lbr_cursor[i].valid)
2749 continue;
2750
2751 stitch_node = get_stitch_node(thread);
2752 if (!stitch_node)
2753 return false;
2754
2755 memcpy(&stitch_node->cursor, &lbr_stitch->prev_lbr_cursor[i],
2756 sizeof(struct callchain_cursor_node));
2757
2758 if (callee)
2759 list_add(&stitch_node->node, &lbr_stitch->lists);
2760 else
2761 list_add_tail(&stitch_node->node, &lbr_stitch->lists);
2762 }
2763
2764 return true;
2765}
2766
2767static bool alloc_lbr_stitch(struct thread *thread, unsigned int max_lbr)
2768{
2769 if (thread__lbr_stitch(thread))
2770 return true;
2771
2772 thread__set_lbr_stitch(thread, zalloc(sizeof(struct lbr_stitch)));
2773 if (!thread__lbr_stitch(thread))
2774 goto err;
2775
2776 thread__lbr_stitch(thread)->prev_lbr_cursor =
2777 calloc(max_lbr + 1, sizeof(struct callchain_cursor_node));
2778 if (!thread__lbr_stitch(thread)->prev_lbr_cursor)
2779 goto free_lbr_stitch;
2780
2781 INIT_LIST_HEAD(&thread__lbr_stitch(thread)->lists);
2782 INIT_LIST_HEAD(&thread__lbr_stitch(thread)->free_lists);
2783
2784 return true;
2785
2786free_lbr_stitch:
2787 free(thread__lbr_stitch(thread));
2788 thread__set_lbr_stitch(thread, NULL);
2789err:
2790 pr_warning("Failed to allocate space for stitched LBRs. Disable LBR stitch\n");
2791 thread__set_lbr_stitch_enable(thread, false);
2792 return false;
2793}
2794
2795/*
2796 * Resolve LBR callstack chain sample
2797 * Return:
2798 * 1 on success get LBR callchain information
2799 * 0 no available LBR callchain information, should try fp
2800 * negative error code on other errors.
2801 */
2802static int resolve_lbr_callchain_sample(struct thread *thread,
2803 struct callchain_cursor *cursor,
2804 struct perf_sample *sample,
2805 struct symbol **parent,
2806 struct addr_location *root_al,
2807 int max_stack,
2808 unsigned int max_lbr)
2809{
2810 bool callee = (callchain_param.order == ORDER_CALLEE);
2811 struct ip_callchain *chain = sample->callchain;
2812 int chain_nr = min(max_stack, (int)chain->nr), i;
2813 struct lbr_stitch *lbr_stitch;
2814 bool stitched_lbr = false;
2815 u64 branch_from = 0;
2816 int err;
2817
2818 for (i = 0; i < chain_nr; i++) {
2819 if (chain->ips[i] == PERF_CONTEXT_USER)
2820 break;
2821 }
2822
2823 /* LBR only affects the user callchain */
2824 if (i == chain_nr)
2825 return 0;
2826
2827 if (thread__lbr_stitch_enable(thread) && !sample->no_hw_idx &&
2828 (max_lbr > 0) && alloc_lbr_stitch(thread, max_lbr)) {
2829 lbr_stitch = thread__lbr_stitch(thread);
2830
2831 stitched_lbr = has_stitched_lbr(thread, sample,
2832 &lbr_stitch->prev_sample,
2833 max_lbr, callee);
2834
2835 if (!stitched_lbr && !list_empty(&lbr_stitch->lists)) {
2836 list_replace_init(&lbr_stitch->lists,
2837 &lbr_stitch->free_lists);
2838 }
2839 memcpy(&lbr_stitch->prev_sample, sample, sizeof(*sample));
2840 }
2841
2842 if (callee) {
2843 /* Add kernel ip */
2844 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2845 parent, root_al, branch_from,
2846 true, i);
2847 if (err)
2848 goto error;
2849
2850 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2851 root_al, &branch_from, true);
2852 if (err)
2853 goto error;
2854
2855 if (stitched_lbr) {
2856 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2857 if (err)
2858 goto error;
2859 }
2860
2861 } else {
2862 if (stitched_lbr) {
2863 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2864 if (err)
2865 goto error;
2866 }
2867 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2868 root_al, &branch_from, false);
2869 if (err)
2870 goto error;
2871
2872 /* Add kernel ip */
2873 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2874 parent, root_al, branch_from,
2875 false, i);
2876 if (err)
2877 goto error;
2878 }
2879 return 1;
2880
2881error:
2882 return (err < 0) ? err : 0;
2883}
2884
2885static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2886 struct callchain_cursor *cursor,
2887 struct symbol **parent,
2888 struct addr_location *root_al,
2889 u8 *cpumode, int ent)
2890{
2891 int err = 0;
2892
2893 while (--ent >= 0) {
2894 u64 ip = chain->ips[ent];
2895
2896 if (ip >= PERF_CONTEXT_MAX) {
2897 err = add_callchain_ip(thread, cursor, parent,
2898 root_al, cpumode, ip,
2899 false, NULL, NULL, 0);
2900 break;
2901 }
2902 }
2903 return err;
2904}
2905
2906static u64 get_leaf_frame_caller(struct perf_sample *sample,
2907 struct thread *thread, int usr_idx)
2908{
2909 if (machine__normalized_is(maps__machine(thread__maps(thread)), "arm64"))
2910 return get_leaf_frame_caller_aarch64(sample, thread, usr_idx);
2911 else
2912 return 0;
2913}
2914
2915static int thread__resolve_callchain_sample(struct thread *thread,
2916 struct callchain_cursor *cursor,
2917 struct evsel *evsel,
2918 struct perf_sample *sample,
2919 struct symbol **parent,
2920 struct addr_location *root_al,
2921 int max_stack)
2922{
2923 struct branch_stack *branch = sample->branch_stack;
2924 struct branch_entry *entries = perf_sample__branch_entries(sample);
2925 struct ip_callchain *chain = sample->callchain;
2926 int chain_nr = 0;
2927 u8 cpumode = PERF_RECORD_MISC_USER;
2928 int i, j, err, nr_entries, usr_idx;
2929 int skip_idx = -1;
2930 int first_call = 0;
2931 u64 leaf_frame_caller;
2932
2933 if (chain)
2934 chain_nr = chain->nr;
2935
2936 if (evsel__has_branch_callstack(evsel)) {
2937 struct perf_env *env = evsel__env(evsel);
2938
2939 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2940 root_al, max_stack,
2941 !env ? 0 : env->max_branches);
2942 if (err)
2943 return (err < 0) ? err : 0;
2944 }
2945
2946 /*
2947 * Based on DWARF debug information, some architectures skip
2948 * a callchain entry saved by the kernel.
2949 */
2950 skip_idx = arch_skip_callchain_idx(thread, chain);
2951
2952 /*
2953 * Add branches to call stack for easier browsing. This gives
2954 * more context for a sample than just the callers.
2955 *
2956 * This uses individual histograms of paths compared to the
2957 * aggregated histograms the normal LBR mode uses.
2958 *
2959 * Limitations for now:
2960 * - No extra filters
2961 * - No annotations (should annotate somehow)
2962 */
2963
2964 if (branch && callchain_param.branch_callstack) {
2965 int nr = min(max_stack, (int)branch->nr);
2966 struct branch_entry be[nr];
2967 struct iterations iter[nr];
2968
2969 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2970 pr_warning("corrupted branch chain. skipping...\n");
2971 goto check_calls;
2972 }
2973
2974 for (i = 0; i < nr; i++) {
2975 if (callchain_param.order == ORDER_CALLEE) {
2976 be[i] = entries[i];
2977
2978 if (chain == NULL)
2979 continue;
2980
2981 /*
2982 * Check for overlap into the callchain.
2983 * The return address is one off compared to
2984 * the branch entry. To adjust for this
2985 * assume the calling instruction is not longer
2986 * than 8 bytes.
2987 */
2988 if (i == skip_idx ||
2989 chain->ips[first_call] >= PERF_CONTEXT_MAX)
2990 first_call++;
2991 else if (be[i].from < chain->ips[first_call] &&
2992 be[i].from >= chain->ips[first_call] - 8)
2993 first_call++;
2994 } else
2995 be[i] = entries[branch->nr - i - 1];
2996 }
2997
2998 memset(iter, 0, sizeof(struct iterations) * nr);
2999 nr = remove_loops(be, nr, iter);
3000
3001 for (i = 0; i < nr; i++) {
3002 err = add_callchain_ip(thread, cursor, parent,
3003 root_al,
3004 NULL, be[i].to,
3005 true, &be[i].flags,
3006 NULL, be[i].from);
3007
3008 if (!err)
3009 err = add_callchain_ip(thread, cursor, parent, root_al,
3010 NULL, be[i].from,
3011 true, &be[i].flags,
3012 &iter[i], 0);
3013 if (err == -EINVAL)
3014 break;
3015 if (err)
3016 return err;
3017 }
3018
3019 if (chain_nr == 0)
3020 return 0;
3021
3022 chain_nr -= nr;
3023 }
3024
3025check_calls:
3026 if (chain && callchain_param.order != ORDER_CALLEE) {
3027 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
3028 &cpumode, chain->nr - first_call);
3029 if (err)
3030 return (err < 0) ? err : 0;
3031 }
3032 for (i = first_call, nr_entries = 0;
3033 i < chain_nr && nr_entries < max_stack; i++) {
3034 u64 ip;
3035
3036 if (callchain_param.order == ORDER_CALLEE)
3037 j = i;
3038 else
3039 j = chain->nr - i - 1;
3040
3041#ifdef HAVE_SKIP_CALLCHAIN_IDX
3042 if (j == skip_idx)
3043 continue;
3044#endif
3045 ip = chain->ips[j];
3046 if (ip < PERF_CONTEXT_MAX)
3047 ++nr_entries;
3048 else if (callchain_param.order != ORDER_CALLEE) {
3049 err = find_prev_cpumode(chain, thread, cursor, parent,
3050 root_al, &cpumode, j);
3051 if (err)
3052 return (err < 0) ? err : 0;
3053 continue;
3054 }
3055
3056 /*
3057 * PERF_CONTEXT_USER allows us to locate where the user stack ends.
3058 * Depending on callchain_param.order and the position of PERF_CONTEXT_USER,
3059 * the index will be different in order to add the missing frame
3060 * at the right place.
3061 */
3062
3063 usr_idx = callchain_param.order == ORDER_CALLEE ? j-2 : j-1;
3064
3065 if (usr_idx >= 0 && chain->ips[usr_idx] == PERF_CONTEXT_USER) {
3066
3067 leaf_frame_caller = get_leaf_frame_caller(sample, thread, usr_idx);
3068
3069 /*
3070 * check if leaf_frame_Caller != ip to not add the same
3071 * value twice.
3072 */
3073
3074 if (leaf_frame_caller && leaf_frame_caller != ip) {
3075
3076 err = add_callchain_ip(thread, cursor, parent,
3077 root_al, &cpumode, leaf_frame_caller,
3078 false, NULL, NULL, 0);
3079 if (err)
3080 return (err < 0) ? err : 0;
3081 }
3082 }
3083
3084 err = add_callchain_ip(thread, cursor, parent,
3085 root_al, &cpumode, ip,
3086 false, NULL, NULL, 0);
3087
3088 if (err)
3089 return (err < 0) ? err : 0;
3090 }
3091
3092 return 0;
3093}
3094
3095static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
3096{
3097 struct symbol *sym = ms->sym;
3098 struct map *map = ms->map;
3099 struct inline_node *inline_node;
3100 struct inline_list *ilist;
3101 struct dso *dso;
3102 u64 addr;
3103 int ret = 1;
3104 struct map_symbol ilist_ms;
3105
3106 if (!symbol_conf.inline_name || !map || !sym)
3107 return ret;
3108
3109 addr = map__dso_map_ip(map, ip);
3110 addr = map__rip_2objdump(map, addr);
3111 dso = map__dso(map);
3112
3113 inline_node = inlines__tree_find(&dso->inlined_nodes, addr);
3114 if (!inline_node) {
3115 inline_node = dso__parse_addr_inlines(dso, addr, sym);
3116 if (!inline_node)
3117 return ret;
3118 inlines__tree_insert(&dso->inlined_nodes, inline_node);
3119 }
3120
3121 ilist_ms = (struct map_symbol) {
3122 .maps = maps__get(ms->maps),
3123 .map = map__get(map),
3124 };
3125 list_for_each_entry(ilist, &inline_node->val, list) {
3126 ilist_ms.sym = ilist->symbol;
3127 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
3128 NULL, 0, 0, 0, ilist->srcline);
3129
3130 if (ret != 0)
3131 return ret;
3132 }
3133 map_symbol__exit(&ilist_ms);
3134
3135 return ret;
3136}
3137
3138static int unwind_entry(struct unwind_entry *entry, void *arg)
3139{
3140 struct callchain_cursor *cursor = arg;
3141 const char *srcline = NULL;
3142 u64 addr = entry->ip;
3143
3144 if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
3145 return 0;
3146
3147 if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
3148 return 0;
3149
3150 /*
3151 * Convert entry->ip from a virtual address to an offset in
3152 * its corresponding binary.
3153 */
3154 if (entry->ms.map)
3155 addr = map__dso_map_ip(entry->ms.map, entry->ip);
3156
3157 srcline = callchain_srcline(&entry->ms, addr);
3158 return callchain_cursor_append(cursor, entry->ip, &entry->ms,
3159 false, NULL, 0, 0, 0, srcline);
3160}
3161
3162static int thread__resolve_callchain_unwind(struct thread *thread,
3163 struct callchain_cursor *cursor,
3164 struct evsel *evsel,
3165 struct perf_sample *sample,
3166 int max_stack)
3167{
3168 /* Can we do dwarf post unwind? */
3169 if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
3170 (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
3171 return 0;
3172
3173 /* Bail out if nothing was captured. */
3174 if ((!sample->user_regs.regs) ||
3175 (!sample->user_stack.size))
3176 return 0;
3177
3178 return unwind__get_entries(unwind_entry, cursor,
3179 thread, sample, max_stack, false);
3180}
3181
3182int thread__resolve_callchain(struct thread *thread,
3183 struct callchain_cursor *cursor,
3184 struct evsel *evsel,
3185 struct perf_sample *sample,
3186 struct symbol **parent,
3187 struct addr_location *root_al,
3188 int max_stack)
3189{
3190 int ret = 0;
3191
3192 if (cursor == NULL)
3193 return -ENOMEM;
3194
3195 callchain_cursor_reset(cursor);
3196
3197 if (callchain_param.order == ORDER_CALLEE) {
3198 ret = thread__resolve_callchain_sample(thread, cursor,
3199 evsel, sample,
3200 parent, root_al,
3201 max_stack);
3202 if (ret)
3203 return ret;
3204 ret = thread__resolve_callchain_unwind(thread, cursor,
3205 evsel, sample,
3206 max_stack);
3207 } else {
3208 ret = thread__resolve_callchain_unwind(thread, cursor,
3209 evsel, sample,
3210 max_stack);
3211 if (ret)
3212 return ret;
3213 ret = thread__resolve_callchain_sample(thread, cursor,
3214 evsel, sample,
3215 parent, root_al,
3216 max_stack);
3217 }
3218
3219 return ret;
3220}
3221
3222int machine__for_each_thread(struct machine *machine,
3223 int (*fn)(struct thread *thread, void *p),
3224 void *priv)
3225{
3226 struct threads *threads;
3227 struct rb_node *nd;
3228 int rc = 0;
3229 int i;
3230
3231 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3232 threads = &machine->threads[i];
3233 for (nd = rb_first_cached(&threads->entries); nd;
3234 nd = rb_next(nd)) {
3235 struct thread_rb_node *trb = rb_entry(nd, struct thread_rb_node, rb_node);
3236
3237 rc = fn(trb->thread, priv);
3238 if (rc != 0)
3239 return rc;
3240 }
3241 }
3242 return rc;
3243}
3244
3245int machines__for_each_thread(struct machines *machines,
3246 int (*fn)(struct thread *thread, void *p),
3247 void *priv)
3248{
3249 struct rb_node *nd;
3250 int rc = 0;
3251
3252 rc = machine__for_each_thread(&machines->host, fn, priv);
3253 if (rc != 0)
3254 return rc;
3255
3256 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
3257 struct machine *machine = rb_entry(nd, struct machine, rb_node);
3258
3259 rc = machine__for_each_thread(machine, fn, priv);
3260 if (rc != 0)
3261 return rc;
3262 }
3263 return rc;
3264}
3265
3266pid_t machine__get_current_tid(struct machine *machine, int cpu)
3267{
3268 if (cpu < 0 || (size_t)cpu >= machine->current_tid_sz)
3269 return -1;
3270
3271 return machine->current_tid[cpu];
3272}
3273
3274int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
3275 pid_t tid)
3276{
3277 struct thread *thread;
3278 const pid_t init_val = -1;
3279
3280 if (cpu < 0)
3281 return -EINVAL;
3282
3283 if (realloc_array_as_needed(machine->current_tid,
3284 machine->current_tid_sz,
3285 (unsigned int)cpu,
3286 &init_val))
3287 return -ENOMEM;
3288
3289 machine->current_tid[cpu] = tid;
3290
3291 thread = machine__findnew_thread(machine, pid, tid);
3292 if (!thread)
3293 return -ENOMEM;
3294
3295 thread__set_cpu(thread, cpu);
3296 thread__put(thread);
3297
3298 return 0;
3299}
3300
3301/*
3302 * Compares the raw arch string. N.B. see instead perf_env__arch() or
3303 * machine__normalized_is() if a normalized arch is needed.
3304 */
3305bool machine__is(struct machine *machine, const char *arch)
3306{
3307 return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
3308}
3309
3310bool machine__normalized_is(struct machine *machine, const char *arch)
3311{
3312 return machine && !strcmp(perf_env__arch(machine->env), arch);
3313}
3314
3315int machine__nr_cpus_avail(struct machine *machine)
3316{
3317 return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
3318}
3319
3320int machine__get_kernel_start(struct machine *machine)
3321{
3322 struct map *map = machine__kernel_map(machine);
3323 int err = 0;
3324
3325 /*
3326 * The only addresses above 2^63 are kernel addresses of a 64-bit
3327 * kernel. Note that addresses are unsigned so that on a 32-bit system
3328 * all addresses including kernel addresses are less than 2^32. In
3329 * that case (32-bit system), if the kernel mapping is unknown, all
3330 * addresses will be assumed to be in user space - see
3331 * machine__kernel_ip().
3332 */
3333 machine->kernel_start = 1ULL << 63;
3334 if (map) {
3335 err = map__load(map);
3336 /*
3337 * On x86_64, PTI entry trampolines are less than the
3338 * start of kernel text, but still above 2^63. So leave
3339 * kernel_start = 1ULL << 63 for x86_64.
3340 */
3341 if (!err && !machine__is(machine, "x86_64"))
3342 machine->kernel_start = map__start(map);
3343 }
3344 return err;
3345}
3346
3347u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
3348{
3349 u8 addr_cpumode = cpumode;
3350 bool kernel_ip;
3351
3352 if (!machine->single_address_space)
3353 goto out;
3354
3355 kernel_ip = machine__kernel_ip(machine, addr);
3356 switch (cpumode) {
3357 case PERF_RECORD_MISC_KERNEL:
3358 case PERF_RECORD_MISC_USER:
3359 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
3360 PERF_RECORD_MISC_USER;
3361 break;
3362 case PERF_RECORD_MISC_GUEST_KERNEL:
3363 case PERF_RECORD_MISC_GUEST_USER:
3364 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
3365 PERF_RECORD_MISC_GUEST_USER;
3366 break;
3367 default:
3368 break;
3369 }
3370out:
3371 return addr_cpumode;
3372}
3373
3374struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id)
3375{
3376 return dsos__findnew_id(&machine->dsos, filename, id);
3377}
3378
3379struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
3380{
3381 return machine__findnew_dso_id(machine, filename, NULL);
3382}
3383
3384char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
3385{
3386 struct machine *machine = vmachine;
3387 struct map *map;
3388 struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
3389
3390 if (sym == NULL)
3391 return NULL;
3392
3393 *modp = __map__is_kmodule(map) ? (char *)map__dso(map)->short_name : NULL;
3394 *addrp = map__unmap_ip(map, sym->start);
3395 return sym->name;
3396}
3397
3398int machine__for_each_dso(struct machine *machine, machine__dso_t fn, void *priv)
3399{
3400 struct dso *pos;
3401 int err = 0;
3402
3403 list_for_each_entry(pos, &machine->dsos.head, node) {
3404 if (fn(pos, machine, priv))
3405 err = -1;
3406 }
3407 return err;
3408}
3409
3410int machine__for_each_kernel_map(struct machine *machine, machine__map_t fn, void *priv)
3411{
3412 struct maps *maps = machine__kernel_maps(machine);
3413
3414 return maps__for_each_map(maps, fn, priv);
3415}
3416
3417bool machine__is_lock_function(struct machine *machine, u64 addr)
3418{
3419 if (!machine->sched.text_start) {
3420 struct map *kmap;
3421 struct symbol *sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_start", &kmap);
3422
3423 if (!sym) {
3424 /* to avoid retry */
3425 machine->sched.text_start = 1;
3426 return false;
3427 }
3428
3429 machine->sched.text_start = map__unmap_ip(kmap, sym->start);
3430
3431 /* should not fail from here */
3432 sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_end", &kmap);
3433 machine->sched.text_end = map__unmap_ip(kmap, sym->start);
3434
3435 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_start", &kmap);
3436 machine->lock.text_start = map__unmap_ip(kmap, sym->start);
3437
3438 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_end", &kmap);
3439 machine->lock.text_end = map__unmap_ip(kmap, sym->start);
3440 }
3441
3442 /* failed to get kernel symbols */
3443 if (machine->sched.text_start == 1)
3444 return false;
3445
3446 /* mutex and rwsem functions are in sched text section */
3447 if (machine->sched.text_start <= addr && addr < machine->sched.text_end)
3448 return true;
3449
3450 /* spinlock functions are in lock text section */
3451 if (machine->lock.text_start <= addr && addr < machine->lock.text_end)
3452 return true;
3453
3454 return false;
3455}
1// SPDX-License-Identifier: GPL-2.0
2#include <dirent.h>
3#include <errno.h>
4#include <inttypes.h>
5#include <regex.h>
6#include <stdlib.h>
7#include "callchain.h"
8#include "debug.h"
9#include "dso.h"
10#include "env.h"
11#include "event.h"
12#include "evsel.h"
13#include "hist.h"
14#include "machine.h"
15#include "map.h"
16#include "map_symbol.h"
17#include "branch.h"
18#include "mem-events.h"
19#include "srcline.h"
20#include "symbol.h"
21#include "sort.h"
22#include "strlist.h"
23#include "target.h"
24#include "thread.h"
25#include "util.h"
26#include "vdso.h"
27#include <stdbool.h>
28#include <sys/types.h>
29#include <sys/stat.h>
30#include <unistd.h>
31#include "unwind.h"
32#include "linux/hash.h"
33#include "asm/bug.h"
34#include "bpf-event.h"
35#include <internal/lib.h> // page_size
36#include "cgroup.h"
37
38#include <linux/ctype.h>
39#include <symbol/kallsyms.h>
40#include <linux/mman.h>
41#include <linux/string.h>
42#include <linux/zalloc.h>
43
44static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
45
46static struct dso *machine__kernel_dso(struct machine *machine)
47{
48 return machine->vmlinux_map->dso;
49}
50
51static void dsos__init(struct dsos *dsos)
52{
53 INIT_LIST_HEAD(&dsos->head);
54 dsos->root = RB_ROOT;
55 init_rwsem(&dsos->lock);
56}
57
58static void machine__threads_init(struct machine *machine)
59{
60 int i;
61
62 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
63 struct threads *threads = &machine->threads[i];
64 threads->entries = RB_ROOT_CACHED;
65 init_rwsem(&threads->lock);
66 threads->nr = 0;
67 INIT_LIST_HEAD(&threads->dead);
68 threads->last_match = NULL;
69 }
70}
71
72static int machine__set_mmap_name(struct machine *machine)
73{
74 if (machine__is_host(machine))
75 machine->mmap_name = strdup("[kernel.kallsyms]");
76 else if (machine__is_default_guest(machine))
77 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
78 else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
79 machine->pid) < 0)
80 machine->mmap_name = NULL;
81
82 return machine->mmap_name ? 0 : -ENOMEM;
83}
84
85int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
86{
87 int err = -ENOMEM;
88
89 memset(machine, 0, sizeof(*machine));
90 maps__init(&machine->kmaps, machine);
91 RB_CLEAR_NODE(&machine->rb_node);
92 dsos__init(&machine->dsos);
93
94 machine__threads_init(machine);
95
96 machine->vdso_info = NULL;
97 machine->env = NULL;
98
99 machine->pid = pid;
100
101 machine->id_hdr_size = 0;
102 machine->kptr_restrict_warned = false;
103 machine->comm_exec = false;
104 machine->kernel_start = 0;
105 machine->vmlinux_map = NULL;
106
107 machine->root_dir = strdup(root_dir);
108 if (machine->root_dir == NULL)
109 return -ENOMEM;
110
111 if (machine__set_mmap_name(machine))
112 goto out;
113
114 if (pid != HOST_KERNEL_ID) {
115 struct thread *thread = machine__findnew_thread(machine, -1,
116 pid);
117 char comm[64];
118
119 if (thread == NULL)
120 goto out;
121
122 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
123 thread__set_comm(thread, comm, 0);
124 thread__put(thread);
125 }
126
127 machine->current_tid = NULL;
128 err = 0;
129
130out:
131 if (err) {
132 zfree(&machine->root_dir);
133 zfree(&machine->mmap_name);
134 }
135 return 0;
136}
137
138struct machine *machine__new_host(void)
139{
140 struct machine *machine = malloc(sizeof(*machine));
141
142 if (machine != NULL) {
143 machine__init(machine, "", HOST_KERNEL_ID);
144
145 if (machine__create_kernel_maps(machine) < 0)
146 goto out_delete;
147 }
148
149 return machine;
150out_delete:
151 free(machine);
152 return NULL;
153}
154
155struct machine *machine__new_kallsyms(void)
156{
157 struct machine *machine = machine__new_host();
158 /*
159 * FIXME:
160 * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
161 * ask for not using the kcore parsing code, once this one is fixed
162 * to create a map per module.
163 */
164 if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
165 machine__delete(machine);
166 machine = NULL;
167 }
168
169 return machine;
170}
171
172static void dsos__purge(struct dsos *dsos)
173{
174 struct dso *pos, *n;
175
176 down_write(&dsos->lock);
177
178 list_for_each_entry_safe(pos, n, &dsos->head, node) {
179 RB_CLEAR_NODE(&pos->rb_node);
180 pos->root = NULL;
181 list_del_init(&pos->node);
182 dso__put(pos);
183 }
184
185 up_write(&dsos->lock);
186}
187
188static void dsos__exit(struct dsos *dsos)
189{
190 dsos__purge(dsos);
191 exit_rwsem(&dsos->lock);
192}
193
194void machine__delete_threads(struct machine *machine)
195{
196 struct rb_node *nd;
197 int i;
198
199 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
200 struct threads *threads = &machine->threads[i];
201 down_write(&threads->lock);
202 nd = rb_first_cached(&threads->entries);
203 while (nd) {
204 struct thread *t = rb_entry(nd, struct thread, rb_node);
205
206 nd = rb_next(nd);
207 __machine__remove_thread(machine, t, false);
208 }
209 up_write(&threads->lock);
210 }
211}
212
213void machine__exit(struct machine *machine)
214{
215 int i;
216
217 if (machine == NULL)
218 return;
219
220 machine__destroy_kernel_maps(machine);
221 maps__exit(&machine->kmaps);
222 dsos__exit(&machine->dsos);
223 machine__exit_vdso(machine);
224 zfree(&machine->root_dir);
225 zfree(&machine->mmap_name);
226 zfree(&machine->current_tid);
227
228 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
229 struct threads *threads = &machine->threads[i];
230 struct thread *thread, *n;
231 /*
232 * Forget about the dead, at this point whatever threads were
233 * left in the dead lists better have a reference count taken
234 * by who is using them, and then, when they drop those references
235 * and it finally hits zero, thread__put() will check and see that
236 * its not in the dead threads list and will not try to remove it
237 * from there, just calling thread__delete() straight away.
238 */
239 list_for_each_entry_safe(thread, n, &threads->dead, node)
240 list_del_init(&thread->node);
241
242 exit_rwsem(&threads->lock);
243 }
244}
245
246void machine__delete(struct machine *machine)
247{
248 if (machine) {
249 machine__exit(machine);
250 free(machine);
251 }
252}
253
254void machines__init(struct machines *machines)
255{
256 machine__init(&machines->host, "", HOST_KERNEL_ID);
257 machines->guests = RB_ROOT_CACHED;
258}
259
260void machines__exit(struct machines *machines)
261{
262 machine__exit(&machines->host);
263 /* XXX exit guest */
264}
265
266struct machine *machines__add(struct machines *machines, pid_t pid,
267 const char *root_dir)
268{
269 struct rb_node **p = &machines->guests.rb_root.rb_node;
270 struct rb_node *parent = NULL;
271 struct machine *pos, *machine = malloc(sizeof(*machine));
272 bool leftmost = true;
273
274 if (machine == NULL)
275 return NULL;
276
277 if (machine__init(machine, root_dir, pid) != 0) {
278 free(machine);
279 return NULL;
280 }
281
282 while (*p != NULL) {
283 parent = *p;
284 pos = rb_entry(parent, struct machine, rb_node);
285 if (pid < pos->pid)
286 p = &(*p)->rb_left;
287 else {
288 p = &(*p)->rb_right;
289 leftmost = false;
290 }
291 }
292
293 rb_link_node(&machine->rb_node, parent, p);
294 rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
295
296 return machine;
297}
298
299void machines__set_comm_exec(struct machines *machines, bool comm_exec)
300{
301 struct rb_node *nd;
302
303 machines->host.comm_exec = comm_exec;
304
305 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
306 struct machine *machine = rb_entry(nd, struct machine, rb_node);
307
308 machine->comm_exec = comm_exec;
309 }
310}
311
312struct machine *machines__find(struct machines *machines, pid_t pid)
313{
314 struct rb_node **p = &machines->guests.rb_root.rb_node;
315 struct rb_node *parent = NULL;
316 struct machine *machine;
317 struct machine *default_machine = NULL;
318
319 if (pid == HOST_KERNEL_ID)
320 return &machines->host;
321
322 while (*p != NULL) {
323 parent = *p;
324 machine = rb_entry(parent, struct machine, rb_node);
325 if (pid < machine->pid)
326 p = &(*p)->rb_left;
327 else if (pid > machine->pid)
328 p = &(*p)->rb_right;
329 else
330 return machine;
331 if (!machine->pid)
332 default_machine = machine;
333 }
334
335 return default_machine;
336}
337
338struct machine *machines__findnew(struct machines *machines, pid_t pid)
339{
340 char path[PATH_MAX];
341 const char *root_dir = "";
342 struct machine *machine = machines__find(machines, pid);
343
344 if (machine && (machine->pid == pid))
345 goto out;
346
347 if ((pid != HOST_KERNEL_ID) &&
348 (pid != DEFAULT_GUEST_KERNEL_ID) &&
349 (symbol_conf.guestmount)) {
350 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
351 if (access(path, R_OK)) {
352 static struct strlist *seen;
353
354 if (!seen)
355 seen = strlist__new(NULL, NULL);
356
357 if (!strlist__has_entry(seen, path)) {
358 pr_err("Can't access file %s\n", path);
359 strlist__add(seen, path);
360 }
361 machine = NULL;
362 goto out;
363 }
364 root_dir = path;
365 }
366
367 machine = machines__add(machines, pid, root_dir);
368out:
369 return machine;
370}
371
372struct machine *machines__find_guest(struct machines *machines, pid_t pid)
373{
374 struct machine *machine = machines__find(machines, pid);
375
376 if (!machine)
377 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID);
378 return machine;
379}
380
381void machines__process_guests(struct machines *machines,
382 machine__process_t process, void *data)
383{
384 struct rb_node *nd;
385
386 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
387 struct machine *pos = rb_entry(nd, struct machine, rb_node);
388 process(pos, data);
389 }
390}
391
392void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
393{
394 struct rb_node *node;
395 struct machine *machine;
396
397 machines->host.id_hdr_size = id_hdr_size;
398
399 for (node = rb_first_cached(&machines->guests); node;
400 node = rb_next(node)) {
401 machine = rb_entry(node, struct machine, rb_node);
402 machine->id_hdr_size = id_hdr_size;
403 }
404
405 return;
406}
407
408static void machine__update_thread_pid(struct machine *machine,
409 struct thread *th, pid_t pid)
410{
411 struct thread *leader;
412
413 if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
414 return;
415
416 th->pid_ = pid;
417
418 if (th->pid_ == th->tid)
419 return;
420
421 leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
422 if (!leader)
423 goto out_err;
424
425 if (!leader->maps)
426 leader->maps = maps__new(machine);
427
428 if (!leader->maps)
429 goto out_err;
430
431 if (th->maps == leader->maps)
432 return;
433
434 if (th->maps) {
435 /*
436 * Maps are created from MMAP events which provide the pid and
437 * tid. Consequently there never should be any maps on a thread
438 * with an unknown pid. Just print an error if there are.
439 */
440 if (!maps__empty(th->maps))
441 pr_err("Discarding thread maps for %d:%d\n",
442 th->pid_, th->tid);
443 maps__put(th->maps);
444 }
445
446 th->maps = maps__get(leader->maps);
447out_put:
448 thread__put(leader);
449 return;
450out_err:
451 pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
452 goto out_put;
453}
454
455/*
456 * Front-end cache - TID lookups come in blocks,
457 * so most of the time we dont have to look up
458 * the full rbtree:
459 */
460static struct thread*
461__threads__get_last_match(struct threads *threads, struct machine *machine,
462 int pid, int tid)
463{
464 struct thread *th;
465
466 th = threads->last_match;
467 if (th != NULL) {
468 if (th->tid == tid) {
469 machine__update_thread_pid(machine, th, pid);
470 return thread__get(th);
471 }
472
473 threads->last_match = NULL;
474 }
475
476 return NULL;
477}
478
479static struct thread*
480threads__get_last_match(struct threads *threads, struct machine *machine,
481 int pid, int tid)
482{
483 struct thread *th = NULL;
484
485 if (perf_singlethreaded)
486 th = __threads__get_last_match(threads, machine, pid, tid);
487
488 return th;
489}
490
491static void
492__threads__set_last_match(struct threads *threads, struct thread *th)
493{
494 threads->last_match = th;
495}
496
497static void
498threads__set_last_match(struct threads *threads, struct thread *th)
499{
500 if (perf_singlethreaded)
501 __threads__set_last_match(threads, th);
502}
503
504/*
505 * Caller must eventually drop thread->refcnt returned with a successful
506 * lookup/new thread inserted.
507 */
508static struct thread *____machine__findnew_thread(struct machine *machine,
509 struct threads *threads,
510 pid_t pid, pid_t tid,
511 bool create)
512{
513 struct rb_node **p = &threads->entries.rb_root.rb_node;
514 struct rb_node *parent = NULL;
515 struct thread *th;
516 bool leftmost = true;
517
518 th = threads__get_last_match(threads, machine, pid, tid);
519 if (th)
520 return th;
521
522 while (*p != NULL) {
523 parent = *p;
524 th = rb_entry(parent, struct thread, rb_node);
525
526 if (th->tid == tid) {
527 threads__set_last_match(threads, th);
528 machine__update_thread_pid(machine, th, pid);
529 return thread__get(th);
530 }
531
532 if (tid < th->tid)
533 p = &(*p)->rb_left;
534 else {
535 p = &(*p)->rb_right;
536 leftmost = false;
537 }
538 }
539
540 if (!create)
541 return NULL;
542
543 th = thread__new(pid, tid);
544 if (th != NULL) {
545 rb_link_node(&th->rb_node, parent, p);
546 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
547
548 /*
549 * We have to initialize maps separately after rb tree is updated.
550 *
551 * The reason is that we call machine__findnew_thread
552 * within thread__init_maps to find the thread
553 * leader and that would screwed the rb tree.
554 */
555 if (thread__init_maps(th, machine)) {
556 rb_erase_cached(&th->rb_node, &threads->entries);
557 RB_CLEAR_NODE(&th->rb_node);
558 thread__put(th);
559 return NULL;
560 }
561 /*
562 * It is now in the rbtree, get a ref
563 */
564 thread__get(th);
565 threads__set_last_match(threads, th);
566 ++threads->nr;
567 }
568
569 return th;
570}
571
572struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
573{
574 return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
575}
576
577struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
578 pid_t tid)
579{
580 struct threads *threads = machine__threads(machine, tid);
581 struct thread *th;
582
583 down_write(&threads->lock);
584 th = __machine__findnew_thread(machine, pid, tid);
585 up_write(&threads->lock);
586 return th;
587}
588
589struct thread *machine__find_thread(struct machine *machine, pid_t pid,
590 pid_t tid)
591{
592 struct threads *threads = machine__threads(machine, tid);
593 struct thread *th;
594
595 down_read(&threads->lock);
596 th = ____machine__findnew_thread(machine, threads, pid, tid, false);
597 up_read(&threads->lock);
598 return th;
599}
600
601/*
602 * Threads are identified by pid and tid, and the idle task has pid == tid == 0.
603 * So here a single thread is created for that, but actually there is a separate
604 * idle task per cpu, so there should be one 'struct thread' per cpu, but there
605 * is only 1. That causes problems for some tools, requiring workarounds. For
606 * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu().
607 */
608struct thread *machine__idle_thread(struct machine *machine)
609{
610 struct thread *thread = machine__findnew_thread(machine, 0, 0);
611
612 if (!thread || thread__set_comm(thread, "swapper", 0) ||
613 thread__set_namespaces(thread, 0, NULL))
614 pr_err("problem inserting idle task for machine pid %d\n", machine->pid);
615
616 return thread;
617}
618
619struct comm *machine__thread_exec_comm(struct machine *machine,
620 struct thread *thread)
621{
622 if (machine->comm_exec)
623 return thread__exec_comm(thread);
624 else
625 return thread__comm(thread);
626}
627
628int machine__process_comm_event(struct machine *machine, union perf_event *event,
629 struct perf_sample *sample)
630{
631 struct thread *thread = machine__findnew_thread(machine,
632 event->comm.pid,
633 event->comm.tid);
634 bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
635 int err = 0;
636
637 if (exec)
638 machine->comm_exec = true;
639
640 if (dump_trace)
641 perf_event__fprintf_comm(event, stdout);
642
643 if (thread == NULL ||
644 __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
645 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
646 err = -1;
647 }
648
649 thread__put(thread);
650
651 return err;
652}
653
654int machine__process_namespaces_event(struct machine *machine __maybe_unused,
655 union perf_event *event,
656 struct perf_sample *sample __maybe_unused)
657{
658 struct thread *thread = machine__findnew_thread(machine,
659 event->namespaces.pid,
660 event->namespaces.tid);
661 int err = 0;
662
663 WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
664 "\nWARNING: kernel seems to support more namespaces than perf"
665 " tool.\nTry updating the perf tool..\n\n");
666
667 WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
668 "\nWARNING: perf tool seems to support more namespaces than"
669 " the kernel.\nTry updating the kernel..\n\n");
670
671 if (dump_trace)
672 perf_event__fprintf_namespaces(event, stdout);
673
674 if (thread == NULL ||
675 thread__set_namespaces(thread, sample->time, &event->namespaces)) {
676 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
677 err = -1;
678 }
679
680 thread__put(thread);
681
682 return err;
683}
684
685int machine__process_cgroup_event(struct machine *machine,
686 union perf_event *event,
687 struct perf_sample *sample __maybe_unused)
688{
689 struct cgroup *cgrp;
690
691 if (dump_trace)
692 perf_event__fprintf_cgroup(event, stdout);
693
694 cgrp = cgroup__findnew(machine->env, event->cgroup.id, event->cgroup.path);
695 if (cgrp == NULL)
696 return -ENOMEM;
697
698 return 0;
699}
700
701int machine__process_lost_event(struct machine *machine __maybe_unused,
702 union perf_event *event, struct perf_sample *sample __maybe_unused)
703{
704 dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
705 event->lost.id, event->lost.lost);
706 return 0;
707}
708
709int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
710 union perf_event *event, struct perf_sample *sample)
711{
712 dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
713 sample->id, event->lost_samples.lost);
714 return 0;
715}
716
717static struct dso *machine__findnew_module_dso(struct machine *machine,
718 struct kmod_path *m,
719 const char *filename)
720{
721 struct dso *dso;
722
723 down_write(&machine->dsos.lock);
724
725 dso = __dsos__find(&machine->dsos, m->name, true);
726 if (!dso) {
727 dso = __dsos__addnew(&machine->dsos, m->name);
728 if (dso == NULL)
729 goto out_unlock;
730
731 dso__set_module_info(dso, m, machine);
732 dso__set_long_name(dso, strdup(filename), true);
733 dso->kernel = DSO_SPACE__KERNEL;
734 }
735
736 dso__get(dso);
737out_unlock:
738 up_write(&machine->dsos.lock);
739 return dso;
740}
741
742int machine__process_aux_event(struct machine *machine __maybe_unused,
743 union perf_event *event)
744{
745 if (dump_trace)
746 perf_event__fprintf_aux(event, stdout);
747 return 0;
748}
749
750int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
751 union perf_event *event)
752{
753 if (dump_trace)
754 perf_event__fprintf_itrace_start(event, stdout);
755 return 0;
756}
757
758int machine__process_switch_event(struct machine *machine __maybe_unused,
759 union perf_event *event)
760{
761 if (dump_trace)
762 perf_event__fprintf_switch(event, stdout);
763 return 0;
764}
765
766static int machine__process_ksymbol_register(struct machine *machine,
767 union perf_event *event,
768 struct perf_sample *sample __maybe_unused)
769{
770 struct symbol *sym;
771 struct map *map = maps__find(&machine->kmaps, event->ksymbol.addr);
772
773 if (!map) {
774 struct dso *dso = dso__new(event->ksymbol.name);
775
776 if (dso) {
777 dso->kernel = DSO_SPACE__KERNEL;
778 map = map__new2(0, dso);
779 dso__put(dso);
780 }
781
782 if (!dso || !map) {
783 return -ENOMEM;
784 }
785
786 if (event->ksymbol.ksym_type == PERF_RECORD_KSYMBOL_TYPE_OOL) {
787 map->dso->binary_type = DSO_BINARY_TYPE__OOL;
788 map->dso->data.file_size = event->ksymbol.len;
789 dso__set_loaded(map->dso);
790 }
791
792 map->start = event->ksymbol.addr;
793 map->end = map->start + event->ksymbol.len;
794 maps__insert(&machine->kmaps, map);
795 map__put(map);
796 dso__set_loaded(dso);
797
798 if (is_bpf_image(event->ksymbol.name)) {
799 dso->binary_type = DSO_BINARY_TYPE__BPF_IMAGE;
800 dso__set_long_name(dso, "", false);
801 }
802 }
803
804 sym = symbol__new(map->map_ip(map, map->start),
805 event->ksymbol.len,
806 0, 0, event->ksymbol.name);
807 if (!sym)
808 return -ENOMEM;
809 dso__insert_symbol(map->dso, sym);
810 return 0;
811}
812
813static int machine__process_ksymbol_unregister(struct machine *machine,
814 union perf_event *event,
815 struct perf_sample *sample __maybe_unused)
816{
817 struct symbol *sym;
818 struct map *map;
819
820 map = maps__find(&machine->kmaps, event->ksymbol.addr);
821 if (!map)
822 return 0;
823
824 if (map != machine->vmlinux_map)
825 maps__remove(&machine->kmaps, map);
826 else {
827 sym = dso__find_symbol(map->dso, map->map_ip(map, map->start));
828 if (sym)
829 dso__delete_symbol(map->dso, sym);
830 }
831
832 return 0;
833}
834
835int machine__process_ksymbol(struct machine *machine __maybe_unused,
836 union perf_event *event,
837 struct perf_sample *sample)
838{
839 if (dump_trace)
840 perf_event__fprintf_ksymbol(event, stdout);
841
842 if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
843 return machine__process_ksymbol_unregister(machine, event,
844 sample);
845 return machine__process_ksymbol_register(machine, event, sample);
846}
847
848int machine__process_text_poke(struct machine *machine, union perf_event *event,
849 struct perf_sample *sample __maybe_unused)
850{
851 struct map *map = maps__find(&machine->kmaps, event->text_poke.addr);
852 u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
853
854 if (dump_trace)
855 perf_event__fprintf_text_poke(event, machine, stdout);
856
857 if (!event->text_poke.new_len)
858 return 0;
859
860 if (cpumode != PERF_RECORD_MISC_KERNEL) {
861 pr_debug("%s: unsupported cpumode - ignoring\n", __func__);
862 return 0;
863 }
864
865 if (map && map->dso) {
866 u8 *new_bytes = event->text_poke.bytes + event->text_poke.old_len;
867 int ret;
868
869 /*
870 * Kernel maps might be changed when loading symbols so loading
871 * must be done prior to using kernel maps.
872 */
873 map__load(map);
874 ret = dso__data_write_cache_addr(map->dso, map, machine,
875 event->text_poke.addr,
876 new_bytes,
877 event->text_poke.new_len);
878 if (ret != event->text_poke.new_len)
879 pr_debug("Failed to write kernel text poke at %#" PRI_lx64 "\n",
880 event->text_poke.addr);
881 } else {
882 pr_debug("Failed to find kernel text poke address map for %#" PRI_lx64 "\n",
883 event->text_poke.addr);
884 }
885
886 return 0;
887}
888
889static struct map *machine__addnew_module_map(struct machine *machine, u64 start,
890 const char *filename)
891{
892 struct map *map = NULL;
893 struct kmod_path m;
894 struct dso *dso;
895
896 if (kmod_path__parse_name(&m, filename))
897 return NULL;
898
899 dso = machine__findnew_module_dso(machine, &m, filename);
900 if (dso == NULL)
901 goto out;
902
903 map = map__new2(start, dso);
904 if (map == NULL)
905 goto out;
906
907 maps__insert(&machine->kmaps, map);
908
909 /* Put the map here because maps__insert already got it */
910 map__put(map);
911out:
912 /* put the dso here, corresponding to machine__findnew_module_dso */
913 dso__put(dso);
914 zfree(&m.name);
915 return map;
916}
917
918size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
919{
920 struct rb_node *nd;
921 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
922
923 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
924 struct machine *pos = rb_entry(nd, struct machine, rb_node);
925 ret += __dsos__fprintf(&pos->dsos.head, fp);
926 }
927
928 return ret;
929}
930
931size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
932 bool (skip)(struct dso *dso, int parm), int parm)
933{
934 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
935}
936
937size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
938 bool (skip)(struct dso *dso, int parm), int parm)
939{
940 struct rb_node *nd;
941 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
942
943 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
944 struct machine *pos = rb_entry(nd, struct machine, rb_node);
945 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
946 }
947 return ret;
948}
949
950size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
951{
952 int i;
953 size_t printed = 0;
954 struct dso *kdso = machine__kernel_dso(machine);
955
956 if (kdso->has_build_id) {
957 char filename[PATH_MAX];
958 if (dso__build_id_filename(kdso, filename, sizeof(filename),
959 false))
960 printed += fprintf(fp, "[0] %s\n", filename);
961 }
962
963 for (i = 0; i < vmlinux_path__nr_entries; ++i)
964 printed += fprintf(fp, "[%d] %s\n",
965 i + kdso->has_build_id, vmlinux_path[i]);
966
967 return printed;
968}
969
970size_t machine__fprintf(struct machine *machine, FILE *fp)
971{
972 struct rb_node *nd;
973 size_t ret;
974 int i;
975
976 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
977 struct threads *threads = &machine->threads[i];
978
979 down_read(&threads->lock);
980
981 ret = fprintf(fp, "Threads: %u\n", threads->nr);
982
983 for (nd = rb_first_cached(&threads->entries); nd;
984 nd = rb_next(nd)) {
985 struct thread *pos = rb_entry(nd, struct thread, rb_node);
986
987 ret += thread__fprintf(pos, fp);
988 }
989
990 up_read(&threads->lock);
991 }
992 return ret;
993}
994
995static struct dso *machine__get_kernel(struct machine *machine)
996{
997 const char *vmlinux_name = machine->mmap_name;
998 struct dso *kernel;
999
1000 if (machine__is_host(machine)) {
1001 if (symbol_conf.vmlinux_name)
1002 vmlinux_name = symbol_conf.vmlinux_name;
1003
1004 kernel = machine__findnew_kernel(machine, vmlinux_name,
1005 "[kernel]", DSO_SPACE__KERNEL);
1006 } else {
1007 if (symbol_conf.default_guest_vmlinux_name)
1008 vmlinux_name = symbol_conf.default_guest_vmlinux_name;
1009
1010 kernel = machine__findnew_kernel(machine, vmlinux_name,
1011 "[guest.kernel]",
1012 DSO_SPACE__KERNEL_GUEST);
1013 }
1014
1015 if (kernel != NULL && (!kernel->has_build_id))
1016 dso__read_running_kernel_build_id(kernel, machine);
1017
1018 return kernel;
1019}
1020
1021struct process_args {
1022 u64 start;
1023};
1024
1025void machine__get_kallsyms_filename(struct machine *machine, char *buf,
1026 size_t bufsz)
1027{
1028 if (machine__is_default_guest(machine))
1029 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
1030 else
1031 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
1032}
1033
1034const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
1035
1036/* Figure out the start address of kernel map from /proc/kallsyms.
1037 * Returns the name of the start symbol in *symbol_name. Pass in NULL as
1038 * symbol_name if it's not that important.
1039 */
1040static int machine__get_running_kernel_start(struct machine *machine,
1041 const char **symbol_name,
1042 u64 *start, u64 *end)
1043{
1044 char filename[PATH_MAX];
1045 int i, err = -1;
1046 const char *name;
1047 u64 addr = 0;
1048
1049 machine__get_kallsyms_filename(machine, filename, PATH_MAX);
1050
1051 if (symbol__restricted_filename(filename, "/proc/kallsyms"))
1052 return 0;
1053
1054 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
1055 err = kallsyms__get_function_start(filename, name, &addr);
1056 if (!err)
1057 break;
1058 }
1059
1060 if (err)
1061 return -1;
1062
1063 if (symbol_name)
1064 *symbol_name = name;
1065
1066 *start = addr;
1067
1068 err = kallsyms__get_function_start(filename, "_etext", &addr);
1069 if (!err)
1070 *end = addr;
1071
1072 return 0;
1073}
1074
1075int machine__create_extra_kernel_map(struct machine *machine,
1076 struct dso *kernel,
1077 struct extra_kernel_map *xm)
1078{
1079 struct kmap *kmap;
1080 struct map *map;
1081
1082 map = map__new2(xm->start, kernel);
1083 if (!map)
1084 return -1;
1085
1086 map->end = xm->end;
1087 map->pgoff = xm->pgoff;
1088
1089 kmap = map__kmap(map);
1090
1091 strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1092
1093 maps__insert(&machine->kmaps, map);
1094
1095 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1096 kmap->name, map->start, map->end);
1097
1098 map__put(map);
1099
1100 return 0;
1101}
1102
1103static u64 find_entry_trampoline(struct dso *dso)
1104{
1105 /* Duplicates are removed so lookup all aliases */
1106 const char *syms[] = {
1107 "_entry_trampoline",
1108 "__entry_trampoline_start",
1109 "entry_SYSCALL_64_trampoline",
1110 };
1111 struct symbol *sym = dso__first_symbol(dso);
1112 unsigned int i;
1113
1114 for (; sym; sym = dso__next_symbol(sym)) {
1115 if (sym->binding != STB_GLOBAL)
1116 continue;
1117 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1118 if (!strcmp(sym->name, syms[i]))
1119 return sym->start;
1120 }
1121 }
1122
1123 return 0;
1124}
1125
1126/*
1127 * These values can be used for kernels that do not have symbols for the entry
1128 * trampolines in kallsyms.
1129 */
1130#define X86_64_CPU_ENTRY_AREA_PER_CPU 0xfffffe0000000000ULL
1131#define X86_64_CPU_ENTRY_AREA_SIZE 0x2c000
1132#define X86_64_ENTRY_TRAMPOLINE 0x6000
1133
1134/* Map x86_64 PTI entry trampolines */
1135int machine__map_x86_64_entry_trampolines(struct machine *machine,
1136 struct dso *kernel)
1137{
1138 struct maps *kmaps = &machine->kmaps;
1139 int nr_cpus_avail, cpu;
1140 bool found = false;
1141 struct map *map;
1142 u64 pgoff;
1143
1144 /*
1145 * In the vmlinux case, pgoff is a virtual address which must now be
1146 * mapped to a vmlinux offset.
1147 */
1148 maps__for_each_entry(kmaps, map) {
1149 struct kmap *kmap = __map__kmap(map);
1150 struct map *dest_map;
1151
1152 if (!kmap || !is_entry_trampoline(kmap->name))
1153 continue;
1154
1155 dest_map = maps__find(kmaps, map->pgoff);
1156 if (dest_map != map)
1157 map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1158 found = true;
1159 }
1160 if (found || machine->trampolines_mapped)
1161 return 0;
1162
1163 pgoff = find_entry_trampoline(kernel);
1164 if (!pgoff)
1165 return 0;
1166
1167 nr_cpus_avail = machine__nr_cpus_avail(machine);
1168
1169 /* Add a 1 page map for each CPU's entry trampoline */
1170 for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1171 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1172 cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1173 X86_64_ENTRY_TRAMPOLINE;
1174 struct extra_kernel_map xm = {
1175 .start = va,
1176 .end = va + page_size,
1177 .pgoff = pgoff,
1178 };
1179
1180 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1181
1182 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1183 return -1;
1184 }
1185
1186 machine->trampolines_mapped = nr_cpus_avail;
1187
1188 return 0;
1189}
1190
1191int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1192 struct dso *kernel __maybe_unused)
1193{
1194 return 0;
1195}
1196
1197static int
1198__machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1199{
1200 /* In case of renewal the kernel map, destroy previous one */
1201 machine__destroy_kernel_maps(machine);
1202
1203 machine->vmlinux_map = map__new2(0, kernel);
1204 if (machine->vmlinux_map == NULL)
1205 return -1;
1206
1207 machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1208 maps__insert(&machine->kmaps, machine->vmlinux_map);
1209 return 0;
1210}
1211
1212void machine__destroy_kernel_maps(struct machine *machine)
1213{
1214 struct kmap *kmap;
1215 struct map *map = machine__kernel_map(machine);
1216
1217 if (map == NULL)
1218 return;
1219
1220 kmap = map__kmap(map);
1221 maps__remove(&machine->kmaps, map);
1222 if (kmap && kmap->ref_reloc_sym) {
1223 zfree((char **)&kmap->ref_reloc_sym->name);
1224 zfree(&kmap->ref_reloc_sym);
1225 }
1226
1227 map__zput(machine->vmlinux_map);
1228}
1229
1230int machines__create_guest_kernel_maps(struct machines *machines)
1231{
1232 int ret = 0;
1233 struct dirent **namelist = NULL;
1234 int i, items = 0;
1235 char path[PATH_MAX];
1236 pid_t pid;
1237 char *endp;
1238
1239 if (symbol_conf.default_guest_vmlinux_name ||
1240 symbol_conf.default_guest_modules ||
1241 symbol_conf.default_guest_kallsyms) {
1242 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1243 }
1244
1245 if (symbol_conf.guestmount) {
1246 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1247 if (items <= 0)
1248 return -ENOENT;
1249 for (i = 0; i < items; i++) {
1250 if (!isdigit(namelist[i]->d_name[0])) {
1251 /* Filter out . and .. */
1252 continue;
1253 }
1254 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1255 if ((*endp != '\0') ||
1256 (endp == namelist[i]->d_name) ||
1257 (errno == ERANGE)) {
1258 pr_debug("invalid directory (%s). Skipping.\n",
1259 namelist[i]->d_name);
1260 continue;
1261 }
1262 sprintf(path, "%s/%s/proc/kallsyms",
1263 symbol_conf.guestmount,
1264 namelist[i]->d_name);
1265 ret = access(path, R_OK);
1266 if (ret) {
1267 pr_debug("Can't access file %s\n", path);
1268 goto failure;
1269 }
1270 machines__create_kernel_maps(machines, pid);
1271 }
1272failure:
1273 free(namelist);
1274 }
1275
1276 return ret;
1277}
1278
1279void machines__destroy_kernel_maps(struct machines *machines)
1280{
1281 struct rb_node *next = rb_first_cached(&machines->guests);
1282
1283 machine__destroy_kernel_maps(&machines->host);
1284
1285 while (next) {
1286 struct machine *pos = rb_entry(next, struct machine, rb_node);
1287
1288 next = rb_next(&pos->rb_node);
1289 rb_erase_cached(&pos->rb_node, &machines->guests);
1290 machine__delete(pos);
1291 }
1292}
1293
1294int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1295{
1296 struct machine *machine = machines__findnew(machines, pid);
1297
1298 if (machine == NULL)
1299 return -1;
1300
1301 return machine__create_kernel_maps(machine);
1302}
1303
1304int machine__load_kallsyms(struct machine *machine, const char *filename)
1305{
1306 struct map *map = machine__kernel_map(machine);
1307 int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1308
1309 if (ret > 0) {
1310 dso__set_loaded(map->dso);
1311 /*
1312 * Since /proc/kallsyms will have multiple sessions for the
1313 * kernel, with modules between them, fixup the end of all
1314 * sections.
1315 */
1316 maps__fixup_end(&machine->kmaps);
1317 }
1318
1319 return ret;
1320}
1321
1322int machine__load_vmlinux_path(struct machine *machine)
1323{
1324 struct map *map = machine__kernel_map(machine);
1325 int ret = dso__load_vmlinux_path(map->dso, map);
1326
1327 if (ret > 0)
1328 dso__set_loaded(map->dso);
1329
1330 return ret;
1331}
1332
1333static char *get_kernel_version(const char *root_dir)
1334{
1335 char version[PATH_MAX];
1336 FILE *file;
1337 char *name, *tmp;
1338 const char *prefix = "Linux version ";
1339
1340 sprintf(version, "%s/proc/version", root_dir);
1341 file = fopen(version, "r");
1342 if (!file)
1343 return NULL;
1344
1345 tmp = fgets(version, sizeof(version), file);
1346 fclose(file);
1347 if (!tmp)
1348 return NULL;
1349
1350 name = strstr(version, prefix);
1351 if (!name)
1352 return NULL;
1353 name += strlen(prefix);
1354 tmp = strchr(name, ' ');
1355 if (tmp)
1356 *tmp = '\0';
1357
1358 return strdup(name);
1359}
1360
1361static bool is_kmod_dso(struct dso *dso)
1362{
1363 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1364 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1365}
1366
1367static int maps__set_module_path(struct maps *maps, const char *path, struct kmod_path *m)
1368{
1369 char *long_name;
1370 struct map *map = maps__find_by_name(maps, m->name);
1371
1372 if (map == NULL)
1373 return 0;
1374
1375 long_name = strdup(path);
1376 if (long_name == NULL)
1377 return -ENOMEM;
1378
1379 dso__set_long_name(map->dso, long_name, true);
1380 dso__kernel_module_get_build_id(map->dso, "");
1381
1382 /*
1383 * Full name could reveal us kmod compression, so
1384 * we need to update the symtab_type if needed.
1385 */
1386 if (m->comp && is_kmod_dso(map->dso)) {
1387 map->dso->symtab_type++;
1388 map->dso->comp = m->comp;
1389 }
1390
1391 return 0;
1392}
1393
1394static int maps__set_modules_path_dir(struct maps *maps, const char *dir_name, int depth)
1395{
1396 struct dirent *dent;
1397 DIR *dir = opendir(dir_name);
1398 int ret = 0;
1399
1400 if (!dir) {
1401 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1402 return -1;
1403 }
1404
1405 while ((dent = readdir(dir)) != NULL) {
1406 char path[PATH_MAX];
1407 struct stat st;
1408
1409 /*sshfs might return bad dent->d_type, so we have to stat*/
1410 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1411 if (stat(path, &st))
1412 continue;
1413
1414 if (S_ISDIR(st.st_mode)) {
1415 if (!strcmp(dent->d_name, ".") ||
1416 !strcmp(dent->d_name, ".."))
1417 continue;
1418
1419 /* Do not follow top-level source and build symlinks */
1420 if (depth == 0) {
1421 if (!strcmp(dent->d_name, "source") ||
1422 !strcmp(dent->d_name, "build"))
1423 continue;
1424 }
1425
1426 ret = maps__set_modules_path_dir(maps, path, depth + 1);
1427 if (ret < 0)
1428 goto out;
1429 } else {
1430 struct kmod_path m;
1431
1432 ret = kmod_path__parse_name(&m, dent->d_name);
1433 if (ret)
1434 goto out;
1435
1436 if (m.kmod)
1437 ret = maps__set_module_path(maps, path, &m);
1438
1439 zfree(&m.name);
1440
1441 if (ret)
1442 goto out;
1443 }
1444 }
1445
1446out:
1447 closedir(dir);
1448 return ret;
1449}
1450
1451static int machine__set_modules_path(struct machine *machine)
1452{
1453 char *version;
1454 char modules_path[PATH_MAX];
1455
1456 version = get_kernel_version(machine->root_dir);
1457 if (!version)
1458 return -1;
1459
1460 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1461 machine->root_dir, version);
1462 free(version);
1463
1464 return maps__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1465}
1466int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1467 u64 *size __maybe_unused,
1468 const char *name __maybe_unused)
1469{
1470 return 0;
1471}
1472
1473static int machine__create_module(void *arg, const char *name, u64 start,
1474 u64 size)
1475{
1476 struct machine *machine = arg;
1477 struct map *map;
1478
1479 if (arch__fix_module_text_start(&start, &size, name) < 0)
1480 return -1;
1481
1482 map = machine__addnew_module_map(machine, start, name);
1483 if (map == NULL)
1484 return -1;
1485 map->end = start + size;
1486
1487 dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1488
1489 return 0;
1490}
1491
1492static int machine__create_modules(struct machine *machine)
1493{
1494 const char *modules;
1495 char path[PATH_MAX];
1496
1497 if (machine__is_default_guest(machine)) {
1498 modules = symbol_conf.default_guest_modules;
1499 } else {
1500 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1501 modules = path;
1502 }
1503
1504 if (symbol__restricted_filename(modules, "/proc/modules"))
1505 return -1;
1506
1507 if (modules__parse(modules, machine, machine__create_module))
1508 return -1;
1509
1510 if (!machine__set_modules_path(machine))
1511 return 0;
1512
1513 pr_debug("Problems setting modules path maps, continuing anyway...\n");
1514
1515 return 0;
1516}
1517
1518static void machine__set_kernel_mmap(struct machine *machine,
1519 u64 start, u64 end)
1520{
1521 machine->vmlinux_map->start = start;
1522 machine->vmlinux_map->end = end;
1523 /*
1524 * Be a bit paranoid here, some perf.data file came with
1525 * a zero sized synthesized MMAP event for the kernel.
1526 */
1527 if (start == 0 && end == 0)
1528 machine->vmlinux_map->end = ~0ULL;
1529}
1530
1531static void machine__update_kernel_mmap(struct machine *machine,
1532 u64 start, u64 end)
1533{
1534 struct map *map = machine__kernel_map(machine);
1535
1536 map__get(map);
1537 maps__remove(&machine->kmaps, map);
1538
1539 machine__set_kernel_mmap(machine, start, end);
1540
1541 maps__insert(&machine->kmaps, map);
1542 map__put(map);
1543}
1544
1545int machine__create_kernel_maps(struct machine *machine)
1546{
1547 struct dso *kernel = machine__get_kernel(machine);
1548 const char *name = NULL;
1549 struct map *map;
1550 u64 start = 0, end = ~0ULL;
1551 int ret;
1552
1553 if (kernel == NULL)
1554 return -1;
1555
1556 ret = __machine__create_kernel_maps(machine, kernel);
1557 if (ret < 0)
1558 goto out_put;
1559
1560 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1561 if (machine__is_host(machine))
1562 pr_debug("Problems creating module maps, "
1563 "continuing anyway...\n");
1564 else
1565 pr_debug("Problems creating module maps for guest %d, "
1566 "continuing anyway...\n", machine->pid);
1567 }
1568
1569 if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1570 if (name &&
1571 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1572 machine__destroy_kernel_maps(machine);
1573 ret = -1;
1574 goto out_put;
1575 }
1576
1577 /*
1578 * we have a real start address now, so re-order the kmaps
1579 * assume it's the last in the kmaps
1580 */
1581 machine__update_kernel_mmap(machine, start, end);
1582 }
1583
1584 if (machine__create_extra_kernel_maps(machine, kernel))
1585 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1586
1587 if (end == ~0ULL) {
1588 /* update end address of the kernel map using adjacent module address */
1589 map = map__next(machine__kernel_map(machine));
1590 if (map)
1591 machine__set_kernel_mmap(machine, start, map->start);
1592 }
1593
1594out_put:
1595 dso__put(kernel);
1596 return ret;
1597}
1598
1599static bool machine__uses_kcore(struct machine *machine)
1600{
1601 struct dso *dso;
1602
1603 list_for_each_entry(dso, &machine->dsos.head, node) {
1604 if (dso__is_kcore(dso))
1605 return true;
1606 }
1607
1608 return false;
1609}
1610
1611static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1612 struct extra_kernel_map *xm)
1613{
1614 return machine__is(machine, "x86_64") &&
1615 is_entry_trampoline(xm->name);
1616}
1617
1618static int machine__process_extra_kernel_map(struct machine *machine,
1619 struct extra_kernel_map *xm)
1620{
1621 struct dso *kernel = machine__kernel_dso(machine);
1622
1623 if (kernel == NULL)
1624 return -1;
1625
1626 return machine__create_extra_kernel_map(machine, kernel, xm);
1627}
1628
1629static int machine__process_kernel_mmap_event(struct machine *machine,
1630 struct extra_kernel_map *xm,
1631 struct build_id *bid)
1632{
1633 struct map *map;
1634 enum dso_space_type dso_space;
1635 bool is_kernel_mmap;
1636
1637 /* If we have maps from kcore then we do not need or want any others */
1638 if (machine__uses_kcore(machine))
1639 return 0;
1640
1641 if (machine__is_host(machine))
1642 dso_space = DSO_SPACE__KERNEL;
1643 else
1644 dso_space = DSO_SPACE__KERNEL_GUEST;
1645
1646 is_kernel_mmap = memcmp(xm->name, machine->mmap_name,
1647 strlen(machine->mmap_name) - 1) == 0;
1648 if (xm->name[0] == '/' ||
1649 (!is_kernel_mmap && xm->name[0] == '[')) {
1650 map = machine__addnew_module_map(machine, xm->start,
1651 xm->name);
1652 if (map == NULL)
1653 goto out_problem;
1654
1655 map->end = map->start + xm->end - xm->start;
1656
1657 if (build_id__is_defined(bid))
1658 dso__set_build_id(map->dso, bid);
1659
1660 } else if (is_kernel_mmap) {
1661 const char *symbol_name = (xm->name + strlen(machine->mmap_name));
1662 /*
1663 * Should be there already, from the build-id table in
1664 * the header.
1665 */
1666 struct dso *kernel = NULL;
1667 struct dso *dso;
1668
1669 down_read(&machine->dsos.lock);
1670
1671 list_for_each_entry(dso, &machine->dsos.head, node) {
1672
1673 /*
1674 * The cpumode passed to is_kernel_module is not the
1675 * cpumode of *this* event. If we insist on passing
1676 * correct cpumode to is_kernel_module, we should
1677 * record the cpumode when we adding this dso to the
1678 * linked list.
1679 *
1680 * However we don't really need passing correct
1681 * cpumode. We know the correct cpumode must be kernel
1682 * mode (if not, we should not link it onto kernel_dsos
1683 * list).
1684 *
1685 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1686 * is_kernel_module() treats it as a kernel cpumode.
1687 */
1688
1689 if (!dso->kernel ||
1690 is_kernel_module(dso->long_name,
1691 PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1692 continue;
1693
1694
1695 kernel = dso;
1696 break;
1697 }
1698
1699 up_read(&machine->dsos.lock);
1700
1701 if (kernel == NULL)
1702 kernel = machine__findnew_dso(machine, machine->mmap_name);
1703 if (kernel == NULL)
1704 goto out_problem;
1705
1706 kernel->kernel = dso_space;
1707 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1708 dso__put(kernel);
1709 goto out_problem;
1710 }
1711
1712 if (strstr(kernel->long_name, "vmlinux"))
1713 dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1714
1715 machine__update_kernel_mmap(machine, xm->start, xm->end);
1716
1717 if (build_id__is_defined(bid))
1718 dso__set_build_id(kernel, bid);
1719
1720 /*
1721 * Avoid using a zero address (kptr_restrict) for the ref reloc
1722 * symbol. Effectively having zero here means that at record
1723 * time /proc/sys/kernel/kptr_restrict was non zero.
1724 */
1725 if (xm->pgoff != 0) {
1726 map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1727 symbol_name,
1728 xm->pgoff);
1729 }
1730
1731 if (machine__is_default_guest(machine)) {
1732 /*
1733 * preload dso of guest kernel and modules
1734 */
1735 dso__load(kernel, machine__kernel_map(machine));
1736 }
1737 } else if (perf_event__is_extra_kernel_mmap(machine, xm)) {
1738 return machine__process_extra_kernel_map(machine, xm);
1739 }
1740 return 0;
1741out_problem:
1742 return -1;
1743}
1744
1745int machine__process_mmap2_event(struct machine *machine,
1746 union perf_event *event,
1747 struct perf_sample *sample)
1748{
1749 struct thread *thread;
1750 struct map *map;
1751 struct dso_id dso_id = {
1752 .maj = event->mmap2.maj,
1753 .min = event->mmap2.min,
1754 .ino = event->mmap2.ino,
1755 .ino_generation = event->mmap2.ino_generation,
1756 };
1757 struct build_id __bid, *bid = NULL;
1758 int ret = 0;
1759
1760 if (dump_trace)
1761 perf_event__fprintf_mmap2(event, stdout);
1762
1763 if (event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID) {
1764 bid = &__bid;
1765 build_id__init(bid, event->mmap2.build_id, event->mmap2.build_id_size);
1766 }
1767
1768 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1769 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1770 struct extra_kernel_map xm = {
1771 .start = event->mmap2.start,
1772 .end = event->mmap2.start + event->mmap2.len,
1773 .pgoff = event->mmap2.pgoff,
1774 };
1775
1776 strlcpy(xm.name, event->mmap2.filename, KMAP_NAME_LEN);
1777 ret = machine__process_kernel_mmap_event(machine, &xm, bid);
1778 if (ret < 0)
1779 goto out_problem;
1780 return 0;
1781 }
1782
1783 thread = machine__findnew_thread(machine, event->mmap2.pid,
1784 event->mmap2.tid);
1785 if (thread == NULL)
1786 goto out_problem;
1787
1788 map = map__new(machine, event->mmap2.start,
1789 event->mmap2.len, event->mmap2.pgoff,
1790 &dso_id, event->mmap2.prot,
1791 event->mmap2.flags, bid,
1792 event->mmap2.filename, thread);
1793
1794 if (map == NULL)
1795 goto out_problem_map;
1796
1797 ret = thread__insert_map(thread, map);
1798 if (ret)
1799 goto out_problem_insert;
1800
1801 thread__put(thread);
1802 map__put(map);
1803 return 0;
1804
1805out_problem_insert:
1806 map__put(map);
1807out_problem_map:
1808 thread__put(thread);
1809out_problem:
1810 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1811 return 0;
1812}
1813
1814int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1815 struct perf_sample *sample)
1816{
1817 struct thread *thread;
1818 struct map *map;
1819 u32 prot = 0;
1820 int ret = 0;
1821
1822 if (dump_trace)
1823 perf_event__fprintf_mmap(event, stdout);
1824
1825 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1826 sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1827 struct extra_kernel_map xm = {
1828 .start = event->mmap.start,
1829 .end = event->mmap.start + event->mmap.len,
1830 .pgoff = event->mmap.pgoff,
1831 };
1832
1833 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1834 ret = machine__process_kernel_mmap_event(machine, &xm, NULL);
1835 if (ret < 0)
1836 goto out_problem;
1837 return 0;
1838 }
1839
1840 thread = machine__findnew_thread(machine, event->mmap.pid,
1841 event->mmap.tid);
1842 if (thread == NULL)
1843 goto out_problem;
1844
1845 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1846 prot = PROT_EXEC;
1847
1848 map = map__new(machine, event->mmap.start,
1849 event->mmap.len, event->mmap.pgoff,
1850 NULL, prot, 0, NULL, event->mmap.filename, thread);
1851
1852 if (map == NULL)
1853 goto out_problem_map;
1854
1855 ret = thread__insert_map(thread, map);
1856 if (ret)
1857 goto out_problem_insert;
1858
1859 thread__put(thread);
1860 map__put(map);
1861 return 0;
1862
1863out_problem_insert:
1864 map__put(map);
1865out_problem_map:
1866 thread__put(thread);
1867out_problem:
1868 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1869 return 0;
1870}
1871
1872static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1873{
1874 struct threads *threads = machine__threads(machine, th->tid);
1875
1876 if (threads->last_match == th)
1877 threads__set_last_match(threads, NULL);
1878
1879 if (lock)
1880 down_write(&threads->lock);
1881
1882 BUG_ON(refcount_read(&th->refcnt) == 0);
1883
1884 rb_erase_cached(&th->rb_node, &threads->entries);
1885 RB_CLEAR_NODE(&th->rb_node);
1886 --threads->nr;
1887 /*
1888 * Move it first to the dead_threads list, then drop the reference,
1889 * if this is the last reference, then the thread__delete destructor
1890 * will be called and we will remove it from the dead_threads list.
1891 */
1892 list_add_tail(&th->node, &threads->dead);
1893
1894 /*
1895 * We need to do the put here because if this is the last refcount,
1896 * then we will be touching the threads->dead head when removing the
1897 * thread.
1898 */
1899 thread__put(th);
1900
1901 if (lock)
1902 up_write(&threads->lock);
1903}
1904
1905void machine__remove_thread(struct machine *machine, struct thread *th)
1906{
1907 return __machine__remove_thread(machine, th, true);
1908}
1909
1910int machine__process_fork_event(struct machine *machine, union perf_event *event,
1911 struct perf_sample *sample)
1912{
1913 struct thread *thread = machine__find_thread(machine,
1914 event->fork.pid,
1915 event->fork.tid);
1916 struct thread *parent = machine__findnew_thread(machine,
1917 event->fork.ppid,
1918 event->fork.ptid);
1919 bool do_maps_clone = true;
1920 int err = 0;
1921
1922 if (dump_trace)
1923 perf_event__fprintf_task(event, stdout);
1924
1925 /*
1926 * There may be an existing thread that is not actually the parent,
1927 * either because we are processing events out of order, or because the
1928 * (fork) event that would have removed the thread was lost. Assume the
1929 * latter case and continue on as best we can.
1930 */
1931 if (parent->pid_ != (pid_t)event->fork.ppid) {
1932 dump_printf("removing erroneous parent thread %d/%d\n",
1933 parent->pid_, parent->tid);
1934 machine__remove_thread(machine, parent);
1935 thread__put(parent);
1936 parent = machine__findnew_thread(machine, event->fork.ppid,
1937 event->fork.ptid);
1938 }
1939
1940 /* if a thread currently exists for the thread id remove it */
1941 if (thread != NULL) {
1942 machine__remove_thread(machine, thread);
1943 thread__put(thread);
1944 }
1945
1946 thread = machine__findnew_thread(machine, event->fork.pid,
1947 event->fork.tid);
1948 /*
1949 * When synthesizing FORK events, we are trying to create thread
1950 * objects for the already running tasks on the machine.
1951 *
1952 * Normally, for a kernel FORK event, we want to clone the parent's
1953 * maps because that is what the kernel just did.
1954 *
1955 * But when synthesizing, this should not be done. If we do, we end up
1956 * with overlapping maps as we process the synthesized MMAP2 events that
1957 * get delivered shortly thereafter.
1958 *
1959 * Use the FORK event misc flags in an internal way to signal this
1960 * situation, so we can elide the map clone when appropriate.
1961 */
1962 if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1963 do_maps_clone = false;
1964
1965 if (thread == NULL || parent == NULL ||
1966 thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1967 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1968 err = -1;
1969 }
1970 thread__put(thread);
1971 thread__put(parent);
1972
1973 return err;
1974}
1975
1976int machine__process_exit_event(struct machine *machine, union perf_event *event,
1977 struct perf_sample *sample __maybe_unused)
1978{
1979 struct thread *thread = machine__find_thread(machine,
1980 event->fork.pid,
1981 event->fork.tid);
1982
1983 if (dump_trace)
1984 perf_event__fprintf_task(event, stdout);
1985
1986 if (thread != NULL) {
1987 thread__exited(thread);
1988 thread__put(thread);
1989 }
1990
1991 return 0;
1992}
1993
1994int machine__process_event(struct machine *machine, union perf_event *event,
1995 struct perf_sample *sample)
1996{
1997 int ret;
1998
1999 switch (event->header.type) {
2000 case PERF_RECORD_COMM:
2001 ret = machine__process_comm_event(machine, event, sample); break;
2002 case PERF_RECORD_MMAP:
2003 ret = machine__process_mmap_event(machine, event, sample); break;
2004 case PERF_RECORD_NAMESPACES:
2005 ret = machine__process_namespaces_event(machine, event, sample); break;
2006 case PERF_RECORD_CGROUP:
2007 ret = machine__process_cgroup_event(machine, event, sample); break;
2008 case PERF_RECORD_MMAP2:
2009 ret = machine__process_mmap2_event(machine, event, sample); break;
2010 case PERF_RECORD_FORK:
2011 ret = machine__process_fork_event(machine, event, sample); break;
2012 case PERF_RECORD_EXIT:
2013 ret = machine__process_exit_event(machine, event, sample); break;
2014 case PERF_RECORD_LOST:
2015 ret = machine__process_lost_event(machine, event, sample); break;
2016 case PERF_RECORD_AUX:
2017 ret = machine__process_aux_event(machine, event); break;
2018 case PERF_RECORD_ITRACE_START:
2019 ret = machine__process_itrace_start_event(machine, event); break;
2020 case PERF_RECORD_LOST_SAMPLES:
2021 ret = machine__process_lost_samples_event(machine, event, sample); break;
2022 case PERF_RECORD_SWITCH:
2023 case PERF_RECORD_SWITCH_CPU_WIDE:
2024 ret = machine__process_switch_event(machine, event); break;
2025 case PERF_RECORD_KSYMBOL:
2026 ret = machine__process_ksymbol(machine, event, sample); break;
2027 case PERF_RECORD_BPF_EVENT:
2028 ret = machine__process_bpf(machine, event, sample); break;
2029 case PERF_RECORD_TEXT_POKE:
2030 ret = machine__process_text_poke(machine, event, sample); break;
2031 default:
2032 ret = -1;
2033 break;
2034 }
2035
2036 return ret;
2037}
2038
2039static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
2040{
2041 if (!regexec(regex, sym->name, 0, NULL, 0))
2042 return true;
2043 return false;
2044}
2045
2046static void ip__resolve_ams(struct thread *thread,
2047 struct addr_map_symbol *ams,
2048 u64 ip)
2049{
2050 struct addr_location al;
2051
2052 memset(&al, 0, sizeof(al));
2053 /*
2054 * We cannot use the header.misc hint to determine whether a
2055 * branch stack address is user, kernel, guest, hypervisor.
2056 * Branches may straddle the kernel/user/hypervisor boundaries.
2057 * Thus, we have to try consecutively until we find a match
2058 * or else, the symbol is unknown
2059 */
2060 thread__find_cpumode_addr_location(thread, ip, &al);
2061
2062 ams->addr = ip;
2063 ams->al_addr = al.addr;
2064 ams->ms.maps = al.maps;
2065 ams->ms.sym = al.sym;
2066 ams->ms.map = al.map;
2067 ams->phys_addr = 0;
2068 ams->data_page_size = 0;
2069}
2070
2071static void ip__resolve_data(struct thread *thread,
2072 u8 m, struct addr_map_symbol *ams,
2073 u64 addr, u64 phys_addr, u64 daddr_page_size)
2074{
2075 struct addr_location al;
2076
2077 memset(&al, 0, sizeof(al));
2078
2079 thread__find_symbol(thread, m, addr, &al);
2080
2081 ams->addr = addr;
2082 ams->al_addr = al.addr;
2083 ams->ms.maps = al.maps;
2084 ams->ms.sym = al.sym;
2085 ams->ms.map = al.map;
2086 ams->phys_addr = phys_addr;
2087 ams->data_page_size = daddr_page_size;
2088}
2089
2090struct mem_info *sample__resolve_mem(struct perf_sample *sample,
2091 struct addr_location *al)
2092{
2093 struct mem_info *mi = mem_info__new();
2094
2095 if (!mi)
2096 return NULL;
2097
2098 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
2099 ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
2100 sample->addr, sample->phys_addr,
2101 sample->data_page_size);
2102 mi->data_src.val = sample->data_src;
2103
2104 return mi;
2105}
2106
2107static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2108{
2109 struct map *map = ms->map;
2110 char *srcline = NULL;
2111
2112 if (!map || callchain_param.key == CCKEY_FUNCTION)
2113 return srcline;
2114
2115 srcline = srcline__tree_find(&map->dso->srclines, ip);
2116 if (!srcline) {
2117 bool show_sym = false;
2118 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2119
2120 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
2121 ms->sym, show_sym, show_addr, ip);
2122 srcline__tree_insert(&map->dso->srclines, ip, srcline);
2123 }
2124
2125 return srcline;
2126}
2127
2128struct iterations {
2129 int nr_loop_iter;
2130 u64 cycles;
2131};
2132
2133static int add_callchain_ip(struct thread *thread,
2134 struct callchain_cursor *cursor,
2135 struct symbol **parent,
2136 struct addr_location *root_al,
2137 u8 *cpumode,
2138 u64 ip,
2139 bool branch,
2140 struct branch_flags *flags,
2141 struct iterations *iter,
2142 u64 branch_from)
2143{
2144 struct map_symbol ms;
2145 struct addr_location al;
2146 int nr_loop_iter = 0;
2147 u64 iter_cycles = 0;
2148 const char *srcline = NULL;
2149
2150 al.filtered = 0;
2151 al.sym = NULL;
2152 al.srcline = NULL;
2153 if (!cpumode) {
2154 thread__find_cpumode_addr_location(thread, ip, &al);
2155 } else {
2156 if (ip >= PERF_CONTEXT_MAX) {
2157 switch (ip) {
2158 case PERF_CONTEXT_HV:
2159 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2160 break;
2161 case PERF_CONTEXT_KERNEL:
2162 *cpumode = PERF_RECORD_MISC_KERNEL;
2163 break;
2164 case PERF_CONTEXT_USER:
2165 *cpumode = PERF_RECORD_MISC_USER;
2166 break;
2167 default:
2168 pr_debug("invalid callchain context: "
2169 "%"PRId64"\n", (s64) ip);
2170 /*
2171 * It seems the callchain is corrupted.
2172 * Discard all.
2173 */
2174 callchain_cursor_reset(cursor);
2175 return 1;
2176 }
2177 return 0;
2178 }
2179 thread__find_symbol(thread, *cpumode, ip, &al);
2180 }
2181
2182 if (al.sym != NULL) {
2183 if (perf_hpp_list.parent && !*parent &&
2184 symbol__match_regex(al.sym, &parent_regex))
2185 *parent = al.sym;
2186 else if (have_ignore_callees && root_al &&
2187 symbol__match_regex(al.sym, &ignore_callees_regex)) {
2188 /* Treat this symbol as the root,
2189 forgetting its callees. */
2190 *root_al = al;
2191 callchain_cursor_reset(cursor);
2192 }
2193 }
2194
2195 if (symbol_conf.hide_unresolved && al.sym == NULL)
2196 return 0;
2197
2198 if (iter) {
2199 nr_loop_iter = iter->nr_loop_iter;
2200 iter_cycles = iter->cycles;
2201 }
2202
2203 ms.maps = al.maps;
2204 ms.map = al.map;
2205 ms.sym = al.sym;
2206 srcline = callchain_srcline(&ms, al.addr);
2207 return callchain_cursor_append(cursor, ip, &ms,
2208 branch, flags, nr_loop_iter,
2209 iter_cycles, branch_from, srcline);
2210}
2211
2212struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2213 struct addr_location *al)
2214{
2215 unsigned int i;
2216 const struct branch_stack *bs = sample->branch_stack;
2217 struct branch_entry *entries = perf_sample__branch_entries(sample);
2218 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2219
2220 if (!bi)
2221 return NULL;
2222
2223 for (i = 0; i < bs->nr; i++) {
2224 ip__resolve_ams(al->thread, &bi[i].to, entries[i].to);
2225 ip__resolve_ams(al->thread, &bi[i].from, entries[i].from);
2226 bi[i].flags = entries[i].flags;
2227 }
2228 return bi;
2229}
2230
2231static void save_iterations(struct iterations *iter,
2232 struct branch_entry *be, int nr)
2233{
2234 int i;
2235
2236 iter->nr_loop_iter++;
2237 iter->cycles = 0;
2238
2239 for (i = 0; i < nr; i++)
2240 iter->cycles += be[i].flags.cycles;
2241}
2242
2243#define CHASHSZ 127
2244#define CHASHBITS 7
2245#define NO_ENTRY 0xff
2246
2247#define PERF_MAX_BRANCH_DEPTH 127
2248
2249/* Remove loops. */
2250static int remove_loops(struct branch_entry *l, int nr,
2251 struct iterations *iter)
2252{
2253 int i, j, off;
2254 unsigned char chash[CHASHSZ];
2255
2256 memset(chash, NO_ENTRY, sizeof(chash));
2257
2258 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2259
2260 for (i = 0; i < nr; i++) {
2261 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2262
2263 /* no collision handling for now */
2264 if (chash[h] == NO_ENTRY) {
2265 chash[h] = i;
2266 } else if (l[chash[h]].from == l[i].from) {
2267 bool is_loop = true;
2268 /* check if it is a real loop */
2269 off = 0;
2270 for (j = chash[h]; j < i && i + off < nr; j++, off++)
2271 if (l[j].from != l[i + off].from) {
2272 is_loop = false;
2273 break;
2274 }
2275 if (is_loop) {
2276 j = nr - (i + off);
2277 if (j > 0) {
2278 save_iterations(iter + i + off,
2279 l + i, off);
2280
2281 memmove(iter + i, iter + i + off,
2282 j * sizeof(*iter));
2283
2284 memmove(l + i, l + i + off,
2285 j * sizeof(*l));
2286 }
2287
2288 nr -= off;
2289 }
2290 }
2291 }
2292 return nr;
2293}
2294
2295static int lbr_callchain_add_kernel_ip(struct thread *thread,
2296 struct callchain_cursor *cursor,
2297 struct perf_sample *sample,
2298 struct symbol **parent,
2299 struct addr_location *root_al,
2300 u64 branch_from,
2301 bool callee, int end)
2302{
2303 struct ip_callchain *chain = sample->callchain;
2304 u8 cpumode = PERF_RECORD_MISC_USER;
2305 int err, i;
2306
2307 if (callee) {
2308 for (i = 0; i < end + 1; i++) {
2309 err = add_callchain_ip(thread, cursor, parent,
2310 root_al, &cpumode, chain->ips[i],
2311 false, NULL, NULL, branch_from);
2312 if (err)
2313 return err;
2314 }
2315 return 0;
2316 }
2317
2318 for (i = end; i >= 0; i--) {
2319 err = add_callchain_ip(thread, cursor, parent,
2320 root_al, &cpumode, chain->ips[i],
2321 false, NULL, NULL, branch_from);
2322 if (err)
2323 return err;
2324 }
2325
2326 return 0;
2327}
2328
2329static void save_lbr_cursor_node(struct thread *thread,
2330 struct callchain_cursor *cursor,
2331 int idx)
2332{
2333 struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2334
2335 if (!lbr_stitch)
2336 return;
2337
2338 if (cursor->pos == cursor->nr) {
2339 lbr_stitch->prev_lbr_cursor[idx].valid = false;
2340 return;
2341 }
2342
2343 if (!cursor->curr)
2344 cursor->curr = cursor->first;
2345 else
2346 cursor->curr = cursor->curr->next;
2347 memcpy(&lbr_stitch->prev_lbr_cursor[idx], cursor->curr,
2348 sizeof(struct callchain_cursor_node));
2349
2350 lbr_stitch->prev_lbr_cursor[idx].valid = true;
2351 cursor->pos++;
2352}
2353
2354static int lbr_callchain_add_lbr_ip(struct thread *thread,
2355 struct callchain_cursor *cursor,
2356 struct perf_sample *sample,
2357 struct symbol **parent,
2358 struct addr_location *root_al,
2359 u64 *branch_from,
2360 bool callee)
2361{
2362 struct branch_stack *lbr_stack = sample->branch_stack;
2363 struct branch_entry *entries = perf_sample__branch_entries(sample);
2364 u8 cpumode = PERF_RECORD_MISC_USER;
2365 int lbr_nr = lbr_stack->nr;
2366 struct branch_flags *flags;
2367 int err, i;
2368 u64 ip;
2369
2370 /*
2371 * The curr and pos are not used in writing session. They are cleared
2372 * in callchain_cursor_commit() when the writing session is closed.
2373 * Using curr and pos to track the current cursor node.
2374 */
2375 if (thread->lbr_stitch) {
2376 cursor->curr = NULL;
2377 cursor->pos = cursor->nr;
2378 if (cursor->nr) {
2379 cursor->curr = cursor->first;
2380 for (i = 0; i < (int)(cursor->nr - 1); i++)
2381 cursor->curr = cursor->curr->next;
2382 }
2383 }
2384
2385 if (callee) {
2386 /* Add LBR ip from first entries.to */
2387 ip = entries[0].to;
2388 flags = &entries[0].flags;
2389 *branch_from = entries[0].from;
2390 err = add_callchain_ip(thread, cursor, parent,
2391 root_al, &cpumode, ip,
2392 true, flags, NULL,
2393 *branch_from);
2394 if (err)
2395 return err;
2396
2397 /*
2398 * The number of cursor node increases.
2399 * Move the current cursor node.
2400 * But does not need to save current cursor node for entry 0.
2401 * It's impossible to stitch the whole LBRs of previous sample.
2402 */
2403 if (thread->lbr_stitch && (cursor->pos != cursor->nr)) {
2404 if (!cursor->curr)
2405 cursor->curr = cursor->first;
2406 else
2407 cursor->curr = cursor->curr->next;
2408 cursor->pos++;
2409 }
2410
2411 /* Add LBR ip from entries.from one by one. */
2412 for (i = 0; i < lbr_nr; i++) {
2413 ip = entries[i].from;
2414 flags = &entries[i].flags;
2415 err = add_callchain_ip(thread, cursor, parent,
2416 root_al, &cpumode, ip,
2417 true, flags, NULL,
2418 *branch_from);
2419 if (err)
2420 return err;
2421 save_lbr_cursor_node(thread, cursor, i);
2422 }
2423 return 0;
2424 }
2425
2426 /* Add LBR ip from entries.from one by one. */
2427 for (i = lbr_nr - 1; i >= 0; i--) {
2428 ip = entries[i].from;
2429 flags = &entries[i].flags;
2430 err = add_callchain_ip(thread, cursor, parent,
2431 root_al, &cpumode, ip,
2432 true, flags, NULL,
2433 *branch_from);
2434 if (err)
2435 return err;
2436 save_lbr_cursor_node(thread, cursor, i);
2437 }
2438
2439 /* Add LBR ip from first entries.to */
2440 ip = entries[0].to;
2441 flags = &entries[0].flags;
2442 *branch_from = entries[0].from;
2443 err = add_callchain_ip(thread, cursor, parent,
2444 root_al, &cpumode, ip,
2445 true, flags, NULL,
2446 *branch_from);
2447 if (err)
2448 return err;
2449
2450 return 0;
2451}
2452
2453static int lbr_callchain_add_stitched_lbr_ip(struct thread *thread,
2454 struct callchain_cursor *cursor)
2455{
2456 struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2457 struct callchain_cursor_node *cnode;
2458 struct stitch_list *stitch_node;
2459 int err;
2460
2461 list_for_each_entry(stitch_node, &lbr_stitch->lists, node) {
2462 cnode = &stitch_node->cursor;
2463
2464 err = callchain_cursor_append(cursor, cnode->ip,
2465 &cnode->ms,
2466 cnode->branch,
2467 &cnode->branch_flags,
2468 cnode->nr_loop_iter,
2469 cnode->iter_cycles,
2470 cnode->branch_from,
2471 cnode->srcline);
2472 if (err)
2473 return err;
2474 }
2475 return 0;
2476}
2477
2478static struct stitch_list *get_stitch_node(struct thread *thread)
2479{
2480 struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2481 struct stitch_list *stitch_node;
2482
2483 if (!list_empty(&lbr_stitch->free_lists)) {
2484 stitch_node = list_first_entry(&lbr_stitch->free_lists,
2485 struct stitch_list, node);
2486 list_del(&stitch_node->node);
2487
2488 return stitch_node;
2489 }
2490
2491 return malloc(sizeof(struct stitch_list));
2492}
2493
2494static bool has_stitched_lbr(struct thread *thread,
2495 struct perf_sample *cur,
2496 struct perf_sample *prev,
2497 unsigned int max_lbr,
2498 bool callee)
2499{
2500 struct branch_stack *cur_stack = cur->branch_stack;
2501 struct branch_entry *cur_entries = perf_sample__branch_entries(cur);
2502 struct branch_stack *prev_stack = prev->branch_stack;
2503 struct branch_entry *prev_entries = perf_sample__branch_entries(prev);
2504 struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2505 int i, j, nr_identical_branches = 0;
2506 struct stitch_list *stitch_node;
2507 u64 cur_base, distance;
2508
2509 if (!cur_stack || !prev_stack)
2510 return false;
2511
2512 /* Find the physical index of the base-of-stack for current sample. */
2513 cur_base = max_lbr - cur_stack->nr + cur_stack->hw_idx + 1;
2514
2515 distance = (prev_stack->hw_idx > cur_base) ? (prev_stack->hw_idx - cur_base) :
2516 (max_lbr + prev_stack->hw_idx - cur_base);
2517 /* Previous sample has shorter stack. Nothing can be stitched. */
2518 if (distance + 1 > prev_stack->nr)
2519 return false;
2520
2521 /*
2522 * Check if there are identical LBRs between two samples.
2523 * Identical LBRs must have same from, to and flags values. Also,
2524 * they have to be saved in the same LBR registers (same physical
2525 * index).
2526 *
2527 * Starts from the base-of-stack of current sample.
2528 */
2529 for (i = distance, j = cur_stack->nr - 1; (i >= 0) && (j >= 0); i--, j--) {
2530 if ((prev_entries[i].from != cur_entries[j].from) ||
2531 (prev_entries[i].to != cur_entries[j].to) ||
2532 (prev_entries[i].flags.value != cur_entries[j].flags.value))
2533 break;
2534 nr_identical_branches++;
2535 }
2536
2537 if (!nr_identical_branches)
2538 return false;
2539
2540 /*
2541 * Save the LBRs between the base-of-stack of previous sample
2542 * and the base-of-stack of current sample into lbr_stitch->lists.
2543 * These LBRs will be stitched later.
2544 */
2545 for (i = prev_stack->nr - 1; i > (int)distance; i--) {
2546
2547 if (!lbr_stitch->prev_lbr_cursor[i].valid)
2548 continue;
2549
2550 stitch_node = get_stitch_node(thread);
2551 if (!stitch_node)
2552 return false;
2553
2554 memcpy(&stitch_node->cursor, &lbr_stitch->prev_lbr_cursor[i],
2555 sizeof(struct callchain_cursor_node));
2556
2557 if (callee)
2558 list_add(&stitch_node->node, &lbr_stitch->lists);
2559 else
2560 list_add_tail(&stitch_node->node, &lbr_stitch->lists);
2561 }
2562
2563 return true;
2564}
2565
2566static bool alloc_lbr_stitch(struct thread *thread, unsigned int max_lbr)
2567{
2568 if (thread->lbr_stitch)
2569 return true;
2570
2571 thread->lbr_stitch = zalloc(sizeof(*thread->lbr_stitch));
2572 if (!thread->lbr_stitch)
2573 goto err;
2574
2575 thread->lbr_stitch->prev_lbr_cursor = calloc(max_lbr + 1, sizeof(struct callchain_cursor_node));
2576 if (!thread->lbr_stitch->prev_lbr_cursor)
2577 goto free_lbr_stitch;
2578
2579 INIT_LIST_HEAD(&thread->lbr_stitch->lists);
2580 INIT_LIST_HEAD(&thread->lbr_stitch->free_lists);
2581
2582 return true;
2583
2584free_lbr_stitch:
2585 zfree(&thread->lbr_stitch);
2586err:
2587 pr_warning("Failed to allocate space for stitched LBRs. Disable LBR stitch\n");
2588 thread->lbr_stitch_enable = false;
2589 return false;
2590}
2591
2592/*
2593 * Resolve LBR callstack chain sample
2594 * Return:
2595 * 1 on success get LBR callchain information
2596 * 0 no available LBR callchain information, should try fp
2597 * negative error code on other errors.
2598 */
2599static int resolve_lbr_callchain_sample(struct thread *thread,
2600 struct callchain_cursor *cursor,
2601 struct perf_sample *sample,
2602 struct symbol **parent,
2603 struct addr_location *root_al,
2604 int max_stack,
2605 unsigned int max_lbr)
2606{
2607 bool callee = (callchain_param.order == ORDER_CALLEE);
2608 struct ip_callchain *chain = sample->callchain;
2609 int chain_nr = min(max_stack, (int)chain->nr), i;
2610 struct lbr_stitch *lbr_stitch;
2611 bool stitched_lbr = false;
2612 u64 branch_from = 0;
2613 int err;
2614
2615 for (i = 0; i < chain_nr; i++) {
2616 if (chain->ips[i] == PERF_CONTEXT_USER)
2617 break;
2618 }
2619
2620 /* LBR only affects the user callchain */
2621 if (i == chain_nr)
2622 return 0;
2623
2624 if (thread->lbr_stitch_enable && !sample->no_hw_idx &&
2625 (max_lbr > 0) && alloc_lbr_stitch(thread, max_lbr)) {
2626 lbr_stitch = thread->lbr_stitch;
2627
2628 stitched_lbr = has_stitched_lbr(thread, sample,
2629 &lbr_stitch->prev_sample,
2630 max_lbr, callee);
2631
2632 if (!stitched_lbr && !list_empty(&lbr_stitch->lists)) {
2633 list_replace_init(&lbr_stitch->lists,
2634 &lbr_stitch->free_lists);
2635 }
2636 memcpy(&lbr_stitch->prev_sample, sample, sizeof(*sample));
2637 }
2638
2639 if (callee) {
2640 /* Add kernel ip */
2641 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2642 parent, root_al, branch_from,
2643 true, i);
2644 if (err)
2645 goto error;
2646
2647 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2648 root_al, &branch_from, true);
2649 if (err)
2650 goto error;
2651
2652 if (stitched_lbr) {
2653 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2654 if (err)
2655 goto error;
2656 }
2657
2658 } else {
2659 if (stitched_lbr) {
2660 err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2661 if (err)
2662 goto error;
2663 }
2664 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2665 root_al, &branch_from, false);
2666 if (err)
2667 goto error;
2668
2669 /* Add kernel ip */
2670 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2671 parent, root_al, branch_from,
2672 false, i);
2673 if (err)
2674 goto error;
2675 }
2676 return 1;
2677
2678error:
2679 return (err < 0) ? err : 0;
2680}
2681
2682static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2683 struct callchain_cursor *cursor,
2684 struct symbol **parent,
2685 struct addr_location *root_al,
2686 u8 *cpumode, int ent)
2687{
2688 int err = 0;
2689
2690 while (--ent >= 0) {
2691 u64 ip = chain->ips[ent];
2692
2693 if (ip >= PERF_CONTEXT_MAX) {
2694 err = add_callchain_ip(thread, cursor, parent,
2695 root_al, cpumode, ip,
2696 false, NULL, NULL, 0);
2697 break;
2698 }
2699 }
2700 return err;
2701}
2702
2703static int thread__resolve_callchain_sample(struct thread *thread,
2704 struct callchain_cursor *cursor,
2705 struct evsel *evsel,
2706 struct perf_sample *sample,
2707 struct symbol **parent,
2708 struct addr_location *root_al,
2709 int max_stack)
2710{
2711 struct branch_stack *branch = sample->branch_stack;
2712 struct branch_entry *entries = perf_sample__branch_entries(sample);
2713 struct ip_callchain *chain = sample->callchain;
2714 int chain_nr = 0;
2715 u8 cpumode = PERF_RECORD_MISC_USER;
2716 int i, j, err, nr_entries;
2717 int skip_idx = -1;
2718 int first_call = 0;
2719
2720 if (chain)
2721 chain_nr = chain->nr;
2722
2723 if (evsel__has_branch_callstack(evsel)) {
2724 struct perf_env *env = evsel__env(evsel);
2725
2726 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2727 root_al, max_stack,
2728 !env ? 0 : env->max_branches);
2729 if (err)
2730 return (err < 0) ? err : 0;
2731 }
2732
2733 /*
2734 * Based on DWARF debug information, some architectures skip
2735 * a callchain entry saved by the kernel.
2736 */
2737 skip_idx = arch_skip_callchain_idx(thread, chain);
2738
2739 /*
2740 * Add branches to call stack for easier browsing. This gives
2741 * more context for a sample than just the callers.
2742 *
2743 * This uses individual histograms of paths compared to the
2744 * aggregated histograms the normal LBR mode uses.
2745 *
2746 * Limitations for now:
2747 * - No extra filters
2748 * - No annotations (should annotate somehow)
2749 */
2750
2751 if (branch && callchain_param.branch_callstack) {
2752 int nr = min(max_stack, (int)branch->nr);
2753 struct branch_entry be[nr];
2754 struct iterations iter[nr];
2755
2756 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2757 pr_warning("corrupted branch chain. skipping...\n");
2758 goto check_calls;
2759 }
2760
2761 for (i = 0; i < nr; i++) {
2762 if (callchain_param.order == ORDER_CALLEE) {
2763 be[i] = entries[i];
2764
2765 if (chain == NULL)
2766 continue;
2767
2768 /*
2769 * Check for overlap into the callchain.
2770 * The return address is one off compared to
2771 * the branch entry. To adjust for this
2772 * assume the calling instruction is not longer
2773 * than 8 bytes.
2774 */
2775 if (i == skip_idx ||
2776 chain->ips[first_call] >= PERF_CONTEXT_MAX)
2777 first_call++;
2778 else if (be[i].from < chain->ips[first_call] &&
2779 be[i].from >= chain->ips[first_call] - 8)
2780 first_call++;
2781 } else
2782 be[i] = entries[branch->nr - i - 1];
2783 }
2784
2785 memset(iter, 0, sizeof(struct iterations) * nr);
2786 nr = remove_loops(be, nr, iter);
2787
2788 for (i = 0; i < nr; i++) {
2789 err = add_callchain_ip(thread, cursor, parent,
2790 root_al,
2791 NULL, be[i].to,
2792 true, &be[i].flags,
2793 NULL, be[i].from);
2794
2795 if (!err)
2796 err = add_callchain_ip(thread, cursor, parent, root_al,
2797 NULL, be[i].from,
2798 true, &be[i].flags,
2799 &iter[i], 0);
2800 if (err == -EINVAL)
2801 break;
2802 if (err)
2803 return err;
2804 }
2805
2806 if (chain_nr == 0)
2807 return 0;
2808
2809 chain_nr -= nr;
2810 }
2811
2812check_calls:
2813 if (chain && callchain_param.order != ORDER_CALLEE) {
2814 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2815 &cpumode, chain->nr - first_call);
2816 if (err)
2817 return (err < 0) ? err : 0;
2818 }
2819 for (i = first_call, nr_entries = 0;
2820 i < chain_nr && nr_entries < max_stack; i++) {
2821 u64 ip;
2822
2823 if (callchain_param.order == ORDER_CALLEE)
2824 j = i;
2825 else
2826 j = chain->nr - i - 1;
2827
2828#ifdef HAVE_SKIP_CALLCHAIN_IDX
2829 if (j == skip_idx)
2830 continue;
2831#endif
2832 ip = chain->ips[j];
2833 if (ip < PERF_CONTEXT_MAX)
2834 ++nr_entries;
2835 else if (callchain_param.order != ORDER_CALLEE) {
2836 err = find_prev_cpumode(chain, thread, cursor, parent,
2837 root_al, &cpumode, j);
2838 if (err)
2839 return (err < 0) ? err : 0;
2840 continue;
2841 }
2842
2843 err = add_callchain_ip(thread, cursor, parent,
2844 root_al, &cpumode, ip,
2845 false, NULL, NULL, 0);
2846
2847 if (err)
2848 return (err < 0) ? err : 0;
2849 }
2850
2851 return 0;
2852}
2853
2854static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
2855{
2856 struct symbol *sym = ms->sym;
2857 struct map *map = ms->map;
2858 struct inline_node *inline_node;
2859 struct inline_list *ilist;
2860 u64 addr;
2861 int ret = 1;
2862
2863 if (!symbol_conf.inline_name || !map || !sym)
2864 return ret;
2865
2866 addr = map__map_ip(map, ip);
2867 addr = map__rip_2objdump(map, addr);
2868
2869 inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2870 if (!inline_node) {
2871 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2872 if (!inline_node)
2873 return ret;
2874 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2875 }
2876
2877 list_for_each_entry(ilist, &inline_node->val, list) {
2878 struct map_symbol ilist_ms = {
2879 .maps = ms->maps,
2880 .map = map,
2881 .sym = ilist->symbol,
2882 };
2883 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
2884 NULL, 0, 0, 0, ilist->srcline);
2885
2886 if (ret != 0)
2887 return ret;
2888 }
2889
2890 return ret;
2891}
2892
2893static int unwind_entry(struct unwind_entry *entry, void *arg)
2894{
2895 struct callchain_cursor *cursor = arg;
2896 const char *srcline = NULL;
2897 u64 addr = entry->ip;
2898
2899 if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
2900 return 0;
2901
2902 if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
2903 return 0;
2904
2905 /*
2906 * Convert entry->ip from a virtual address to an offset in
2907 * its corresponding binary.
2908 */
2909 if (entry->ms.map)
2910 addr = map__map_ip(entry->ms.map, entry->ip);
2911
2912 srcline = callchain_srcline(&entry->ms, addr);
2913 return callchain_cursor_append(cursor, entry->ip, &entry->ms,
2914 false, NULL, 0, 0, 0, srcline);
2915}
2916
2917static int thread__resolve_callchain_unwind(struct thread *thread,
2918 struct callchain_cursor *cursor,
2919 struct evsel *evsel,
2920 struct perf_sample *sample,
2921 int max_stack)
2922{
2923 /* Can we do dwarf post unwind? */
2924 if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2925 (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
2926 return 0;
2927
2928 /* Bail out if nothing was captured. */
2929 if ((!sample->user_regs.regs) ||
2930 (!sample->user_stack.size))
2931 return 0;
2932
2933 return unwind__get_entries(unwind_entry, cursor,
2934 thread, sample, max_stack);
2935}
2936
2937int thread__resolve_callchain(struct thread *thread,
2938 struct callchain_cursor *cursor,
2939 struct evsel *evsel,
2940 struct perf_sample *sample,
2941 struct symbol **parent,
2942 struct addr_location *root_al,
2943 int max_stack)
2944{
2945 int ret = 0;
2946
2947 callchain_cursor_reset(cursor);
2948
2949 if (callchain_param.order == ORDER_CALLEE) {
2950 ret = thread__resolve_callchain_sample(thread, cursor,
2951 evsel, sample,
2952 parent, root_al,
2953 max_stack);
2954 if (ret)
2955 return ret;
2956 ret = thread__resolve_callchain_unwind(thread, cursor,
2957 evsel, sample,
2958 max_stack);
2959 } else {
2960 ret = thread__resolve_callchain_unwind(thread, cursor,
2961 evsel, sample,
2962 max_stack);
2963 if (ret)
2964 return ret;
2965 ret = thread__resolve_callchain_sample(thread, cursor,
2966 evsel, sample,
2967 parent, root_al,
2968 max_stack);
2969 }
2970
2971 return ret;
2972}
2973
2974int machine__for_each_thread(struct machine *machine,
2975 int (*fn)(struct thread *thread, void *p),
2976 void *priv)
2977{
2978 struct threads *threads;
2979 struct rb_node *nd;
2980 struct thread *thread;
2981 int rc = 0;
2982 int i;
2983
2984 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2985 threads = &machine->threads[i];
2986 for (nd = rb_first_cached(&threads->entries); nd;
2987 nd = rb_next(nd)) {
2988 thread = rb_entry(nd, struct thread, rb_node);
2989 rc = fn(thread, priv);
2990 if (rc != 0)
2991 return rc;
2992 }
2993
2994 list_for_each_entry(thread, &threads->dead, node) {
2995 rc = fn(thread, priv);
2996 if (rc != 0)
2997 return rc;
2998 }
2999 }
3000 return rc;
3001}
3002
3003int machines__for_each_thread(struct machines *machines,
3004 int (*fn)(struct thread *thread, void *p),
3005 void *priv)
3006{
3007 struct rb_node *nd;
3008 int rc = 0;
3009
3010 rc = machine__for_each_thread(&machines->host, fn, priv);
3011 if (rc != 0)
3012 return rc;
3013
3014 for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
3015 struct machine *machine = rb_entry(nd, struct machine, rb_node);
3016
3017 rc = machine__for_each_thread(machine, fn, priv);
3018 if (rc != 0)
3019 return rc;
3020 }
3021 return rc;
3022}
3023
3024pid_t machine__get_current_tid(struct machine *machine, int cpu)
3025{
3026 int nr_cpus = min(machine->env->nr_cpus_avail, MAX_NR_CPUS);
3027
3028 if (cpu < 0 || cpu >= nr_cpus || !machine->current_tid)
3029 return -1;
3030
3031 return machine->current_tid[cpu];
3032}
3033
3034int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
3035 pid_t tid)
3036{
3037 struct thread *thread;
3038 int nr_cpus = min(machine->env->nr_cpus_avail, MAX_NR_CPUS);
3039
3040 if (cpu < 0)
3041 return -EINVAL;
3042
3043 if (!machine->current_tid) {
3044 int i;
3045
3046 machine->current_tid = calloc(nr_cpus, sizeof(pid_t));
3047 if (!machine->current_tid)
3048 return -ENOMEM;
3049 for (i = 0; i < nr_cpus; i++)
3050 machine->current_tid[i] = -1;
3051 }
3052
3053 if (cpu >= nr_cpus) {
3054 pr_err("Requested CPU %d too large. ", cpu);
3055 pr_err("Consider raising MAX_NR_CPUS\n");
3056 return -EINVAL;
3057 }
3058
3059 machine->current_tid[cpu] = tid;
3060
3061 thread = machine__findnew_thread(machine, pid, tid);
3062 if (!thread)
3063 return -ENOMEM;
3064
3065 thread->cpu = cpu;
3066 thread__put(thread);
3067
3068 return 0;
3069}
3070
3071/*
3072 * Compares the raw arch string. N.B. see instead perf_env__arch() if a
3073 * normalized arch is needed.
3074 */
3075bool machine__is(struct machine *machine, const char *arch)
3076{
3077 return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
3078}
3079
3080int machine__nr_cpus_avail(struct machine *machine)
3081{
3082 return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
3083}
3084
3085int machine__get_kernel_start(struct machine *machine)
3086{
3087 struct map *map = machine__kernel_map(machine);
3088 int err = 0;
3089
3090 /*
3091 * The only addresses above 2^63 are kernel addresses of a 64-bit
3092 * kernel. Note that addresses are unsigned so that on a 32-bit system
3093 * all addresses including kernel addresses are less than 2^32. In
3094 * that case (32-bit system), if the kernel mapping is unknown, all
3095 * addresses will be assumed to be in user space - see
3096 * machine__kernel_ip().
3097 */
3098 machine->kernel_start = 1ULL << 63;
3099 if (map) {
3100 err = map__load(map);
3101 /*
3102 * On x86_64, PTI entry trampolines are less than the
3103 * start of kernel text, but still above 2^63. So leave
3104 * kernel_start = 1ULL << 63 for x86_64.
3105 */
3106 if (!err && !machine__is(machine, "x86_64"))
3107 machine->kernel_start = map->start;
3108 }
3109 return err;
3110}
3111
3112u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
3113{
3114 u8 addr_cpumode = cpumode;
3115 bool kernel_ip;
3116
3117 if (!machine->single_address_space)
3118 goto out;
3119
3120 kernel_ip = machine__kernel_ip(machine, addr);
3121 switch (cpumode) {
3122 case PERF_RECORD_MISC_KERNEL:
3123 case PERF_RECORD_MISC_USER:
3124 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
3125 PERF_RECORD_MISC_USER;
3126 break;
3127 case PERF_RECORD_MISC_GUEST_KERNEL:
3128 case PERF_RECORD_MISC_GUEST_USER:
3129 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
3130 PERF_RECORD_MISC_GUEST_USER;
3131 break;
3132 default:
3133 break;
3134 }
3135out:
3136 return addr_cpumode;
3137}
3138
3139struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id)
3140{
3141 return dsos__findnew_id(&machine->dsos, filename, id);
3142}
3143
3144struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
3145{
3146 return machine__findnew_dso_id(machine, filename, NULL);
3147}
3148
3149char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
3150{
3151 struct machine *machine = vmachine;
3152 struct map *map;
3153 struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
3154
3155 if (sym == NULL)
3156 return NULL;
3157
3158 *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
3159 *addrp = map->unmap_ip(map, sym->start);
3160 return sym->name;
3161}
3162
3163int machine__for_each_dso(struct machine *machine, machine__dso_t fn, void *priv)
3164{
3165 struct dso *pos;
3166 int err = 0;
3167
3168 list_for_each_entry(pos, &machine->dsos.head, node) {
3169 if (fn(pos, machine, priv))
3170 err = -1;
3171 }
3172 return err;
3173}