]> git.ipfire.org Git - thirdparty/linux.git/blob - tools/perf/util/thread-stack.c
Merge branch 'mlx5-next' into rdma.git for-next
[thirdparty/linux.git] / tools / perf / util / thread-stack.c
1 /*
2 * thread-stack.c: Synthesize a thread's stack using call / return events
3 * Copyright (c) 2014, Intel Corporation.
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 *
14 */
15
16 #include <linux/rbtree.h>
17 #include <linux/list.h>
18 #include <linux/log2.h>
19 #include <errno.h>
20 #include "thread.h"
21 #include "event.h"
22 #include "machine.h"
23 #include "env.h"
24 #include "util.h"
25 #include "debug.h"
26 #include "symbol.h"
27 #include "comm.h"
28 #include "call-path.h"
29 #include "thread-stack.h"
30
31 #define STACK_GROWTH 2048
32
33 /*
34 * State of retpoline detection.
35 *
36 * RETPOLINE_NONE: no retpoline detection
37 * X86_RETPOLINE_POSSIBLE: x86 retpoline possible
38 * X86_RETPOLINE_DETECTED: x86 retpoline detected
39 */
40 enum retpoline_state_t {
41 RETPOLINE_NONE,
42 X86_RETPOLINE_POSSIBLE,
43 X86_RETPOLINE_DETECTED,
44 };
45
46 /**
47 * struct thread_stack_entry - thread stack entry.
48 * @ret_addr: return address
49 * @timestamp: timestamp (if known)
50 * @ref: external reference (e.g. db_id of sample)
51 * @branch_count: the branch count when the entry was created
52 * @db_id: id used for db-export
53 * @cp: call path
54 * @no_call: a 'call' was not seen
55 * @trace_end: a 'call' but trace ended
56 * @non_call: a branch but not a 'call' to the start of a different symbol
57 */
58 struct thread_stack_entry {
59 u64 ret_addr;
60 u64 timestamp;
61 u64 ref;
62 u64 branch_count;
63 u64 db_id;
64 struct call_path *cp;
65 bool no_call;
66 bool trace_end;
67 bool non_call;
68 };
69
70 /**
71 * struct thread_stack - thread stack constructed from 'call' and 'return'
72 * branch samples.
73 * @stack: array that holds the stack
74 * @cnt: number of entries in the stack
75 * @sz: current maximum stack size
76 * @trace_nr: current trace number
77 * @branch_count: running branch count
78 * @kernel_start: kernel start address
79 * @last_time: last timestamp
80 * @crp: call/return processor
81 * @comm: current comm
82 * @arr_sz: size of array if this is the first element of an array
83 * @rstate: used to detect retpolines
84 */
85 struct thread_stack {
86 struct thread_stack_entry *stack;
87 size_t cnt;
88 size_t sz;
89 u64 trace_nr;
90 u64 branch_count;
91 u64 kernel_start;
92 u64 last_time;
93 struct call_return_processor *crp;
94 struct comm *comm;
95 unsigned int arr_sz;
96 enum retpoline_state_t rstate;
97 };
98
99 /*
100 * Assume pid == tid == 0 identifies the idle task as defined by
101 * perf_session__register_idle_thread(). The idle task is really 1 task per cpu,
102 * and therefore requires a stack for each cpu.
103 */
104 static inline bool thread_stack__per_cpu(struct thread *thread)
105 {
106 return !(thread->tid || thread->pid_);
107 }
108
109 static int thread_stack__grow(struct thread_stack *ts)
110 {
111 struct thread_stack_entry *new_stack;
112 size_t sz, new_sz;
113
114 new_sz = ts->sz + STACK_GROWTH;
115 sz = new_sz * sizeof(struct thread_stack_entry);
116
117 new_stack = realloc(ts->stack, sz);
118 if (!new_stack)
119 return -ENOMEM;
120
121 ts->stack = new_stack;
122 ts->sz = new_sz;
123
124 return 0;
125 }
126
127 static int thread_stack__init(struct thread_stack *ts, struct thread *thread,
128 struct call_return_processor *crp)
129 {
130 int err;
131
132 err = thread_stack__grow(ts);
133 if (err)
134 return err;
135
136 if (thread->mg && thread->mg->machine) {
137 struct machine *machine = thread->mg->machine;
138 const char *arch = perf_env__arch(machine->env);
139
140 ts->kernel_start = machine__kernel_start(machine);
141 if (!strcmp(arch, "x86"))
142 ts->rstate = X86_RETPOLINE_POSSIBLE;
143 } else {
144 ts->kernel_start = 1ULL << 63;
145 }
146 ts->crp = crp;
147
148 return 0;
149 }
150
151 static struct thread_stack *thread_stack__new(struct thread *thread, int cpu,
152 struct call_return_processor *crp)
153 {
154 struct thread_stack *ts = thread->ts, *new_ts;
155 unsigned int old_sz = ts ? ts->arr_sz : 0;
156 unsigned int new_sz = 1;
157
158 if (thread_stack__per_cpu(thread) && cpu > 0)
159 new_sz = roundup_pow_of_two(cpu + 1);
160
161 if (!ts || new_sz > old_sz) {
162 new_ts = calloc(new_sz, sizeof(*ts));
163 if (!new_ts)
164 return NULL;
165 if (ts)
166 memcpy(new_ts, ts, old_sz * sizeof(*ts));
167 new_ts->arr_sz = new_sz;
168 zfree(&thread->ts);
169 thread->ts = new_ts;
170 ts = new_ts;
171 }
172
173 if (thread_stack__per_cpu(thread) && cpu > 0 &&
174 (unsigned int)cpu < ts->arr_sz)
175 ts += cpu;
176
177 if (!ts->stack &&
178 thread_stack__init(ts, thread, crp))
179 return NULL;
180
181 return ts;
182 }
183
184 static struct thread_stack *thread__cpu_stack(struct thread *thread, int cpu)
185 {
186 struct thread_stack *ts = thread->ts;
187
188 if (cpu < 0)
189 cpu = 0;
190
191 if (!ts || (unsigned int)cpu >= ts->arr_sz)
192 return NULL;
193
194 ts += cpu;
195
196 if (!ts->stack)
197 return NULL;
198
199 return ts;
200 }
201
202 static inline struct thread_stack *thread__stack(struct thread *thread,
203 int cpu)
204 {
205 if (!thread)
206 return NULL;
207
208 if (thread_stack__per_cpu(thread))
209 return thread__cpu_stack(thread, cpu);
210
211 return thread->ts;
212 }
213
214 static int thread_stack__push(struct thread_stack *ts, u64 ret_addr,
215 bool trace_end)
216 {
217 int err = 0;
218
219 if (ts->cnt == ts->sz) {
220 err = thread_stack__grow(ts);
221 if (err) {
222 pr_warning("Out of memory: discarding thread stack\n");
223 ts->cnt = 0;
224 }
225 }
226
227 ts->stack[ts->cnt].trace_end = trace_end;
228 ts->stack[ts->cnt++].ret_addr = ret_addr;
229
230 return err;
231 }
232
233 static void thread_stack__pop(struct thread_stack *ts, u64 ret_addr)
234 {
235 size_t i;
236
237 /*
238 * In some cases there may be functions which are not seen to return.
239 * For example when setjmp / longjmp has been used. Or the perf context
240 * switch in the kernel which doesn't stop and start tracing in exactly
241 * the same code path. When that happens the return address will be
242 * further down the stack. If the return address is not found at all,
243 * we assume the opposite (i.e. this is a return for a call that wasn't
244 * seen for some reason) and leave the stack alone.
245 */
246 for (i = ts->cnt; i; ) {
247 if (ts->stack[--i].ret_addr == ret_addr) {
248 ts->cnt = i;
249 return;
250 }
251 }
252 }
253
254 static void thread_stack__pop_trace_end(struct thread_stack *ts)
255 {
256 size_t i;
257
258 for (i = ts->cnt; i; ) {
259 if (ts->stack[--i].trace_end)
260 ts->cnt = i;
261 else
262 return;
263 }
264 }
265
266 static bool thread_stack__in_kernel(struct thread_stack *ts)
267 {
268 if (!ts->cnt)
269 return false;
270
271 return ts->stack[ts->cnt - 1].cp->in_kernel;
272 }
273
274 static int thread_stack__call_return(struct thread *thread,
275 struct thread_stack *ts, size_t idx,
276 u64 timestamp, u64 ref, bool no_return)
277 {
278 struct call_return_processor *crp = ts->crp;
279 struct thread_stack_entry *tse;
280 struct call_return cr = {
281 .thread = thread,
282 .comm = ts->comm,
283 .db_id = 0,
284 };
285 u64 *parent_db_id;
286
287 tse = &ts->stack[idx];
288 cr.cp = tse->cp;
289 cr.call_time = tse->timestamp;
290 cr.return_time = timestamp;
291 cr.branch_count = ts->branch_count - tse->branch_count;
292 cr.db_id = tse->db_id;
293 cr.call_ref = tse->ref;
294 cr.return_ref = ref;
295 if (tse->no_call)
296 cr.flags |= CALL_RETURN_NO_CALL;
297 if (no_return)
298 cr.flags |= CALL_RETURN_NO_RETURN;
299 if (tse->non_call)
300 cr.flags |= CALL_RETURN_NON_CALL;
301
302 /*
303 * The parent db_id must be assigned before exporting the child. Note
304 * it is not possible to export the parent first because its information
305 * is not yet complete because its 'return' has not yet been processed.
306 */
307 parent_db_id = idx ? &(tse - 1)->db_id : NULL;
308
309 return crp->process(&cr, parent_db_id, crp->data);
310 }
311
312 static int __thread_stack__flush(struct thread *thread, struct thread_stack *ts)
313 {
314 struct call_return_processor *crp = ts->crp;
315 int err;
316
317 if (!crp) {
318 ts->cnt = 0;
319 return 0;
320 }
321
322 while (ts->cnt) {
323 err = thread_stack__call_return(thread, ts, --ts->cnt,
324 ts->last_time, 0, true);
325 if (err) {
326 pr_err("Error flushing thread stack!\n");
327 ts->cnt = 0;
328 return err;
329 }
330 }
331
332 return 0;
333 }
334
335 int thread_stack__flush(struct thread *thread)
336 {
337 struct thread_stack *ts = thread->ts;
338 unsigned int pos;
339 int err = 0;
340
341 if (ts) {
342 for (pos = 0; pos < ts->arr_sz; pos++) {
343 int ret = __thread_stack__flush(thread, ts + pos);
344
345 if (ret)
346 err = ret;
347 }
348 }
349
350 return err;
351 }
352
353 int thread_stack__event(struct thread *thread, int cpu, u32 flags, u64 from_ip,
354 u64 to_ip, u16 insn_len, u64 trace_nr)
355 {
356 struct thread_stack *ts = thread__stack(thread, cpu);
357
358 if (!thread)
359 return -EINVAL;
360
361 if (!ts) {
362 ts = thread_stack__new(thread, cpu, NULL);
363 if (!ts) {
364 pr_warning("Out of memory: no thread stack\n");
365 return -ENOMEM;
366 }
367 ts->trace_nr = trace_nr;
368 }
369
370 /*
371 * When the trace is discontinuous, the trace_nr changes. In that case
372 * the stack might be completely invalid. Better to report nothing than
373 * to report something misleading, so flush the stack.
374 */
375 if (trace_nr != ts->trace_nr) {
376 if (ts->trace_nr)
377 __thread_stack__flush(thread, ts);
378 ts->trace_nr = trace_nr;
379 }
380
381 /* Stop here if thread_stack__process() is in use */
382 if (ts->crp)
383 return 0;
384
385 if (flags & PERF_IP_FLAG_CALL) {
386 u64 ret_addr;
387
388 if (!to_ip)
389 return 0;
390 ret_addr = from_ip + insn_len;
391 if (ret_addr == to_ip)
392 return 0; /* Zero-length calls are excluded */
393 return thread_stack__push(ts, ret_addr,
394 flags & PERF_IP_FLAG_TRACE_END);
395 } else if (flags & PERF_IP_FLAG_TRACE_BEGIN) {
396 /*
397 * If the caller did not change the trace number (which would
398 * have flushed the stack) then try to make sense of the stack.
399 * Possibly, tracing began after returning to the current
400 * address, so try to pop that. Also, do not expect a call made
401 * when the trace ended, to return, so pop that.
402 */
403 thread_stack__pop(ts, to_ip);
404 thread_stack__pop_trace_end(ts);
405 } else if ((flags & PERF_IP_FLAG_RETURN) && from_ip) {
406 thread_stack__pop(ts, to_ip);
407 }
408
409 return 0;
410 }
411
412 void thread_stack__set_trace_nr(struct thread *thread, int cpu, u64 trace_nr)
413 {
414 struct thread_stack *ts = thread__stack(thread, cpu);
415
416 if (!ts)
417 return;
418
419 if (trace_nr != ts->trace_nr) {
420 if (ts->trace_nr)
421 __thread_stack__flush(thread, ts);
422 ts->trace_nr = trace_nr;
423 }
424 }
425
426 static void __thread_stack__free(struct thread *thread, struct thread_stack *ts)
427 {
428 __thread_stack__flush(thread, ts);
429 zfree(&ts->stack);
430 }
431
432 static void thread_stack__reset(struct thread *thread, struct thread_stack *ts)
433 {
434 unsigned int arr_sz = ts->arr_sz;
435
436 __thread_stack__free(thread, ts);
437 memset(ts, 0, sizeof(*ts));
438 ts->arr_sz = arr_sz;
439 }
440
441 void thread_stack__free(struct thread *thread)
442 {
443 struct thread_stack *ts = thread->ts;
444 unsigned int pos;
445
446 if (ts) {
447 for (pos = 0; pos < ts->arr_sz; pos++)
448 __thread_stack__free(thread, ts + pos);
449 zfree(&thread->ts);
450 }
451 }
452
453 static inline u64 callchain_context(u64 ip, u64 kernel_start)
454 {
455 return ip < kernel_start ? PERF_CONTEXT_USER : PERF_CONTEXT_KERNEL;
456 }
457
458 void thread_stack__sample(struct thread *thread, int cpu,
459 struct ip_callchain *chain,
460 size_t sz, u64 ip, u64 kernel_start)
461 {
462 struct thread_stack *ts = thread__stack(thread, cpu);
463 u64 context = callchain_context(ip, kernel_start);
464 u64 last_context;
465 size_t i, j;
466
467 if (sz < 2) {
468 chain->nr = 0;
469 return;
470 }
471
472 chain->ips[0] = context;
473 chain->ips[1] = ip;
474
475 if (!ts) {
476 chain->nr = 2;
477 return;
478 }
479
480 last_context = context;
481
482 for (i = 2, j = 1; i < sz && j <= ts->cnt; i++, j++) {
483 ip = ts->stack[ts->cnt - j].ret_addr;
484 context = callchain_context(ip, kernel_start);
485 if (context != last_context) {
486 if (i >= sz - 1)
487 break;
488 chain->ips[i++] = context;
489 last_context = context;
490 }
491 chain->ips[i] = ip;
492 }
493
494 chain->nr = i;
495 }
496
497 struct call_return_processor *
498 call_return_processor__new(int (*process)(struct call_return *cr, u64 *parent_db_id, void *data),
499 void *data)
500 {
501 struct call_return_processor *crp;
502
503 crp = zalloc(sizeof(struct call_return_processor));
504 if (!crp)
505 return NULL;
506 crp->cpr = call_path_root__new();
507 if (!crp->cpr)
508 goto out_free;
509 crp->process = process;
510 crp->data = data;
511 return crp;
512
513 out_free:
514 free(crp);
515 return NULL;
516 }
517
518 void call_return_processor__free(struct call_return_processor *crp)
519 {
520 if (crp) {
521 call_path_root__free(crp->cpr);
522 free(crp);
523 }
524 }
525
526 static int thread_stack__push_cp(struct thread_stack *ts, u64 ret_addr,
527 u64 timestamp, u64 ref, struct call_path *cp,
528 bool no_call, bool trace_end)
529 {
530 struct thread_stack_entry *tse;
531 int err;
532
533 if (!cp)
534 return -ENOMEM;
535
536 if (ts->cnt == ts->sz) {
537 err = thread_stack__grow(ts);
538 if (err)
539 return err;
540 }
541
542 tse = &ts->stack[ts->cnt++];
543 tse->ret_addr = ret_addr;
544 tse->timestamp = timestamp;
545 tse->ref = ref;
546 tse->branch_count = ts->branch_count;
547 tse->cp = cp;
548 tse->no_call = no_call;
549 tse->trace_end = trace_end;
550 tse->non_call = false;
551 tse->db_id = 0;
552
553 return 0;
554 }
555
556 static int thread_stack__pop_cp(struct thread *thread, struct thread_stack *ts,
557 u64 ret_addr, u64 timestamp, u64 ref,
558 struct symbol *sym)
559 {
560 int err;
561
562 if (!ts->cnt)
563 return 1;
564
565 if (ts->cnt == 1) {
566 struct thread_stack_entry *tse = &ts->stack[0];
567
568 if (tse->cp->sym == sym)
569 return thread_stack__call_return(thread, ts, --ts->cnt,
570 timestamp, ref, false);
571 }
572
573 if (ts->stack[ts->cnt - 1].ret_addr == ret_addr &&
574 !ts->stack[ts->cnt - 1].non_call) {
575 return thread_stack__call_return(thread, ts, --ts->cnt,
576 timestamp, ref, false);
577 } else {
578 size_t i = ts->cnt - 1;
579
580 while (i--) {
581 if (ts->stack[i].ret_addr != ret_addr ||
582 ts->stack[i].non_call)
583 continue;
584 i += 1;
585 while (ts->cnt > i) {
586 err = thread_stack__call_return(thread, ts,
587 --ts->cnt,
588 timestamp, ref,
589 true);
590 if (err)
591 return err;
592 }
593 return thread_stack__call_return(thread, ts, --ts->cnt,
594 timestamp, ref, false);
595 }
596 }
597
598 return 1;
599 }
600
601 static int thread_stack__bottom(struct thread_stack *ts,
602 struct perf_sample *sample,
603 struct addr_location *from_al,
604 struct addr_location *to_al, u64 ref)
605 {
606 struct call_path_root *cpr = ts->crp->cpr;
607 struct call_path *cp;
608 struct symbol *sym;
609 u64 ip;
610
611 if (sample->ip) {
612 ip = sample->ip;
613 sym = from_al->sym;
614 } else if (sample->addr) {
615 ip = sample->addr;
616 sym = to_al->sym;
617 } else {
618 return 0;
619 }
620
621 cp = call_path__findnew(cpr, &cpr->call_path, sym, ip,
622 ts->kernel_start);
623
624 return thread_stack__push_cp(ts, ip, sample->time, ref, cp,
625 true, false);
626 }
627
628 static int thread_stack__no_call_return(struct thread *thread,
629 struct thread_stack *ts,
630 struct perf_sample *sample,
631 struct addr_location *from_al,
632 struct addr_location *to_al, u64 ref)
633 {
634 struct call_path_root *cpr = ts->crp->cpr;
635 struct call_path *root = &cpr->call_path;
636 struct symbol *fsym = from_al->sym;
637 struct symbol *tsym = to_al->sym;
638 struct call_path *cp, *parent;
639 u64 ks = ts->kernel_start;
640 u64 addr = sample->addr;
641 u64 tm = sample->time;
642 u64 ip = sample->ip;
643 int err;
644
645 if (ip >= ks && addr < ks) {
646 /* Return to userspace, so pop all kernel addresses */
647 while (thread_stack__in_kernel(ts)) {
648 err = thread_stack__call_return(thread, ts, --ts->cnt,
649 tm, ref, true);
650 if (err)
651 return err;
652 }
653
654 /* If the stack is empty, push the userspace address */
655 if (!ts->cnt) {
656 cp = call_path__findnew(cpr, root, tsym, addr, ks);
657 return thread_stack__push_cp(ts, 0, tm, ref, cp, true,
658 false);
659 }
660 } else if (thread_stack__in_kernel(ts) && ip < ks) {
661 /* Return to userspace, so pop all kernel addresses */
662 while (thread_stack__in_kernel(ts)) {
663 err = thread_stack__call_return(thread, ts, --ts->cnt,
664 tm, ref, true);
665 if (err)
666 return err;
667 }
668 }
669
670 if (ts->cnt)
671 parent = ts->stack[ts->cnt - 1].cp;
672 else
673 parent = root;
674
675 if (parent->sym == from_al->sym) {
676 /*
677 * At the bottom of the stack, assume the missing 'call' was
678 * before the trace started. So, pop the current symbol and push
679 * the 'to' symbol.
680 */
681 if (ts->cnt == 1) {
682 err = thread_stack__call_return(thread, ts, --ts->cnt,
683 tm, ref, false);
684 if (err)
685 return err;
686 }
687
688 if (!ts->cnt) {
689 cp = call_path__findnew(cpr, root, tsym, addr, ks);
690
691 return thread_stack__push_cp(ts, addr, tm, ref, cp,
692 true, false);
693 }
694
695 /*
696 * Otherwise assume the 'return' is being used as a jump (e.g.
697 * retpoline) and just push the 'to' symbol.
698 */
699 cp = call_path__findnew(cpr, parent, tsym, addr, ks);
700
701 err = thread_stack__push_cp(ts, 0, tm, ref, cp, true, false);
702 if (!err)
703 ts->stack[ts->cnt - 1].non_call = true;
704
705 return err;
706 }
707
708 /*
709 * Assume 'parent' has not yet returned, so push 'to', and then push and
710 * pop 'from'.
711 */
712
713 cp = call_path__findnew(cpr, parent, tsym, addr, ks);
714
715 err = thread_stack__push_cp(ts, addr, tm, ref, cp, true, false);
716 if (err)
717 return err;
718
719 cp = call_path__findnew(cpr, cp, fsym, ip, ks);
720
721 err = thread_stack__push_cp(ts, ip, tm, ref, cp, true, false);
722 if (err)
723 return err;
724
725 return thread_stack__call_return(thread, ts, --ts->cnt, tm, ref, false);
726 }
727
728 static int thread_stack__trace_begin(struct thread *thread,
729 struct thread_stack *ts, u64 timestamp,
730 u64 ref)
731 {
732 struct thread_stack_entry *tse;
733 int err;
734
735 if (!ts->cnt)
736 return 0;
737
738 /* Pop trace end */
739 tse = &ts->stack[ts->cnt - 1];
740 if (tse->trace_end) {
741 err = thread_stack__call_return(thread, ts, --ts->cnt,
742 timestamp, ref, false);
743 if (err)
744 return err;
745 }
746
747 return 0;
748 }
749
750 static int thread_stack__trace_end(struct thread_stack *ts,
751 struct perf_sample *sample, u64 ref)
752 {
753 struct call_path_root *cpr = ts->crp->cpr;
754 struct call_path *cp;
755 u64 ret_addr;
756
757 /* No point having 'trace end' on the bottom of the stack */
758 if (!ts->cnt || (ts->cnt == 1 && ts->stack[0].ref == ref))
759 return 0;
760
761 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp, NULL, 0,
762 ts->kernel_start);
763
764 ret_addr = sample->ip + sample->insn_len;
765
766 return thread_stack__push_cp(ts, ret_addr, sample->time, ref, cp,
767 false, true);
768 }
769
770 static bool is_x86_retpoline(const char *name)
771 {
772 const char *p = strstr(name, "__x86_indirect_thunk_");
773
774 return p == name || !strcmp(name, "__indirect_thunk_start");
775 }
776
777 /*
778 * x86 retpoline functions pollute the call graph. This function removes them.
779 * This does not handle function return thunks, nor is there any improvement
780 * for the handling of inline thunks or extern thunks.
781 */
782 static int thread_stack__x86_retpoline(struct thread_stack *ts,
783 struct perf_sample *sample,
784 struct addr_location *to_al)
785 {
786 struct thread_stack_entry *tse = &ts->stack[ts->cnt - 1];
787 struct call_path_root *cpr = ts->crp->cpr;
788 struct symbol *sym = tse->cp->sym;
789 struct symbol *tsym = to_al->sym;
790 struct call_path *cp;
791
792 if (sym && is_x86_retpoline(sym->name)) {
793 /*
794 * This is a x86 retpoline fn. It pollutes the call graph by
795 * showing up everywhere there is an indirect branch, but does
796 * not itself mean anything. Here the top-of-stack is removed,
797 * by decrementing the stack count, and then further down, the
798 * resulting top-of-stack is replaced with the actual target.
799 * The result is that the retpoline functions will no longer
800 * appear in the call graph. Note this only affects the call
801 * graph, since all the original branches are left unchanged.
802 */
803 ts->cnt -= 1;
804 sym = ts->stack[ts->cnt - 2].cp->sym;
805 if (sym && sym == tsym && to_al->addr != tsym->start) {
806 /*
807 * Target is back to the middle of the symbol we came
808 * from so assume it is an indirect jmp and forget it
809 * altogether.
810 */
811 ts->cnt -= 1;
812 return 0;
813 }
814 } else if (sym && sym == tsym) {
815 /*
816 * Target is back to the symbol we came from so assume it is an
817 * indirect jmp and forget it altogether.
818 */
819 ts->cnt -= 1;
820 return 0;
821 }
822
823 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 2].cp, tsym,
824 sample->addr, ts->kernel_start);
825 if (!cp)
826 return -ENOMEM;
827
828 /* Replace the top-of-stack with the actual target */
829 ts->stack[ts->cnt - 1].cp = cp;
830
831 return 0;
832 }
833
834 int thread_stack__process(struct thread *thread, struct comm *comm,
835 struct perf_sample *sample,
836 struct addr_location *from_al,
837 struct addr_location *to_al, u64 ref,
838 struct call_return_processor *crp)
839 {
840 struct thread_stack *ts = thread__stack(thread, sample->cpu);
841 enum retpoline_state_t rstate;
842 int err = 0;
843
844 if (ts && !ts->crp) {
845 /* Supersede thread_stack__event() */
846 thread_stack__reset(thread, ts);
847 ts = NULL;
848 }
849
850 if (!ts) {
851 ts = thread_stack__new(thread, sample->cpu, crp);
852 if (!ts)
853 return -ENOMEM;
854 ts->comm = comm;
855 }
856
857 rstate = ts->rstate;
858 if (rstate == X86_RETPOLINE_DETECTED)
859 ts->rstate = X86_RETPOLINE_POSSIBLE;
860
861 /* Flush stack on exec */
862 if (ts->comm != comm && thread->pid_ == thread->tid) {
863 err = __thread_stack__flush(thread, ts);
864 if (err)
865 return err;
866 ts->comm = comm;
867 }
868
869 /* If the stack is empty, put the current symbol on the stack */
870 if (!ts->cnt) {
871 err = thread_stack__bottom(ts, sample, from_al, to_al, ref);
872 if (err)
873 return err;
874 }
875
876 ts->branch_count += 1;
877 ts->last_time = sample->time;
878
879 if (sample->flags & PERF_IP_FLAG_CALL) {
880 bool trace_end = sample->flags & PERF_IP_FLAG_TRACE_END;
881 struct call_path_root *cpr = ts->crp->cpr;
882 struct call_path *cp;
883 u64 ret_addr;
884
885 if (!sample->ip || !sample->addr)
886 return 0;
887
888 ret_addr = sample->ip + sample->insn_len;
889 if (ret_addr == sample->addr)
890 return 0; /* Zero-length calls are excluded */
891
892 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp,
893 to_al->sym, sample->addr,
894 ts->kernel_start);
895 err = thread_stack__push_cp(ts, ret_addr, sample->time, ref,
896 cp, false, trace_end);
897
898 /*
899 * A call to the same symbol but not the start of the symbol,
900 * may be the start of a x86 retpoline.
901 */
902 if (!err && rstate == X86_RETPOLINE_POSSIBLE && to_al->sym &&
903 from_al->sym == to_al->sym &&
904 to_al->addr != to_al->sym->start)
905 ts->rstate = X86_RETPOLINE_DETECTED;
906
907 } else if (sample->flags & PERF_IP_FLAG_RETURN) {
908 if (!sample->ip || !sample->addr)
909 return 0;
910
911 /* x86 retpoline 'return' doesn't match the stack */
912 if (rstate == X86_RETPOLINE_DETECTED && ts->cnt > 2 &&
913 ts->stack[ts->cnt - 1].ret_addr != sample->addr)
914 return thread_stack__x86_retpoline(ts, sample, to_al);
915
916 err = thread_stack__pop_cp(thread, ts, sample->addr,
917 sample->time, ref, from_al->sym);
918 if (err) {
919 if (err < 0)
920 return err;
921 err = thread_stack__no_call_return(thread, ts, sample,
922 from_al, to_al, ref);
923 }
924 } else if (sample->flags & PERF_IP_FLAG_TRACE_BEGIN) {
925 err = thread_stack__trace_begin(thread, ts, sample->time, ref);
926 } else if (sample->flags & PERF_IP_FLAG_TRACE_END) {
927 err = thread_stack__trace_end(ts, sample, ref);
928 } else if (sample->flags & PERF_IP_FLAG_BRANCH &&
929 from_al->sym != to_al->sym && to_al->sym &&
930 to_al->addr == to_al->sym->start) {
931 struct call_path_root *cpr = ts->crp->cpr;
932 struct call_path *cp;
933
934 /*
935 * The compiler might optimize a call/ret combination by making
936 * it a jmp. Make that visible by recording on the stack a
937 * branch to the start of a different symbol. Note, that means
938 * when a ret pops the stack, all jmps must be popped off first.
939 */
940 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp,
941 to_al->sym, sample->addr,
942 ts->kernel_start);
943 err = thread_stack__push_cp(ts, 0, sample->time, ref, cp, false,
944 false);
945 if (!err)
946 ts->stack[ts->cnt - 1].non_call = true;
947 }
948
949 return err;
950 }
951
952 size_t thread_stack__depth(struct thread *thread, int cpu)
953 {
954 struct thread_stack *ts = thread__stack(thread, cpu);
955
956 if (!ts)
957 return 0;
958 return ts->cnt;
959 }