xref: /linux/tools/perf/util/thread-stack.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * thread-stack.c: Synthesize a thread's stack using call / return events
4  * Copyright (c) 2014, Intel Corporation.
5  */
6 
7 #include <linux/rbtree.h>
8 #include <linux/list.h>
9 #include <linux/log2.h>
10 #include <linux/zalloc.h>
11 #include <errno.h>
12 #include <stdlib.h>
13 #include <string.h>
14 #include "thread.h"
15 #include "event.h"
16 #include "machine.h"
17 #include "env.h"
18 #include "debug.h"
19 #include "symbol.h"
20 #include "comm.h"
21 #include "call-path.h"
22 #include "thread-stack.h"
23 
24 #define STACK_GROWTH 2048
25 
26 /*
27  * State of retpoline detection.
28  *
29  * RETPOLINE_NONE: no retpoline detection
30  * X86_RETPOLINE_POSSIBLE: x86 retpoline possible
31  * X86_RETPOLINE_DETECTED: x86 retpoline detected
32  */
33 enum retpoline_state_t {
34 	RETPOLINE_NONE,
35 	X86_RETPOLINE_POSSIBLE,
36 	X86_RETPOLINE_DETECTED,
37 };
38 
39 /**
40  * struct thread_stack_entry - thread stack entry.
41  * @ret_addr: return address
42  * @timestamp: timestamp (if known)
43  * @ref: external reference (e.g. db_id of sample)
44  * @branch_count: the branch count when the entry was created
45  * @insn_count: the instruction count when the entry was created
46  * @cyc_count the cycle count when the entry was created
47  * @db_id: id used for db-export
48  * @cp: call path
49  * @no_call: a 'call' was not seen
50  * @trace_end: a 'call' but trace ended
51  * @non_call: a branch but not a 'call' to the start of a different symbol
52  */
53 struct thread_stack_entry {
54 	u64 ret_addr;
55 	u64 timestamp;
56 	u64 ref;
57 	u64 branch_count;
58 	u64 insn_count;
59 	u64 cyc_count;
60 	u64 db_id;
61 	struct call_path *cp;
62 	bool no_call;
63 	bool trace_end;
64 	bool non_call;
65 };
66 
67 /**
68  * struct thread_stack - thread stack constructed from 'call' and 'return'
69  *                       branch samples.
70  * @stack: array that holds the stack
71  * @cnt: number of entries in the stack
72  * @sz: current maximum stack size
73  * @trace_nr: current trace number
74  * @branch_count: running branch count
75  * @insn_count: running  instruction count
76  * @cyc_count running  cycle count
77  * @kernel_start: kernel start address
78  * @last_time: last timestamp
79  * @crp: call/return processor
80  * @comm: current comm
81  * @arr_sz: size of array if this is the first element of an array
82  * @rstate: used to detect retpolines
83  * @br_stack_rb: branch stack (ring buffer)
84  * @br_stack_sz: maximum branch stack size
85  * @br_stack_pos: current position in @br_stack_rb
86  * @mispred_all: mark all branches as mispredicted
87  */
88 struct thread_stack {
89 	struct thread_stack_entry *stack;
90 	size_t cnt;
91 	size_t sz;
92 	u64 trace_nr;
93 	u64 branch_count;
94 	u64 insn_count;
95 	u64 cyc_count;
96 	u64 kernel_start;
97 	u64 last_time;
98 	struct call_return_processor *crp;
99 	struct comm *comm;
100 	unsigned int arr_sz;
101 	enum retpoline_state_t rstate;
102 	struct branch_stack *br_stack_rb;
103 	unsigned int br_stack_sz;
104 	unsigned int br_stack_pos;
105 	bool mispred_all;
106 };
107 
108 /*
109  * Assume pid == tid == 0 identifies the idle task as defined by
110  * perf_session__register_idle_thread(). The idle task is really 1 task per cpu,
111  * and therefore requires a stack for each cpu.
112  */
113 static inline bool thread_stack__per_cpu(struct thread *thread)
114 {
115 	return !(thread__tid(thread) || thread__pid(thread));
116 }
117 
118 static int thread_stack__grow(struct thread_stack *ts)
119 {
120 	struct thread_stack_entry *new_stack;
121 	size_t sz, new_sz;
122 
123 	new_sz = ts->sz + STACK_GROWTH;
124 	sz = new_sz * sizeof(struct thread_stack_entry);
125 
126 	new_stack = realloc(ts->stack, sz);
127 	if (!new_stack)
128 		return -ENOMEM;
129 
130 	ts->stack = new_stack;
131 	ts->sz = new_sz;
132 
133 	return 0;
134 }
135 
136 static int thread_stack__init(struct thread_stack *ts, struct thread *thread,
137 			      struct call_return_processor *crp,
138 			      bool callstack, unsigned int br_stack_sz)
139 {
140 	int err;
141 
142 	if (callstack) {
143 		err = thread_stack__grow(ts);
144 		if (err)
145 			return err;
146 	}
147 
148 	if (br_stack_sz) {
149 		size_t sz = sizeof(struct branch_stack);
150 
151 		sz += br_stack_sz * sizeof(struct branch_entry);
152 		ts->br_stack_rb = zalloc(sz);
153 		if (!ts->br_stack_rb)
154 			return -ENOMEM;
155 		ts->br_stack_sz = br_stack_sz;
156 	}
157 
158 	if (thread__maps(thread) && maps__machine(thread__maps(thread))) {
159 		struct machine *machine = maps__machine(thread__maps(thread));
160 		const char *arch = perf_env__arch(machine->env);
161 
162 		ts->kernel_start = machine__kernel_start(machine);
163 		if (!strcmp(arch, "x86"))
164 			ts->rstate = X86_RETPOLINE_POSSIBLE;
165 	} else {
166 		ts->kernel_start = 1ULL << 63;
167 	}
168 	ts->crp = crp;
169 
170 	return 0;
171 }
172 
173 static struct thread_stack *thread_stack__new(struct thread *thread, int cpu,
174 					      struct call_return_processor *crp,
175 					      bool callstack,
176 					      unsigned int br_stack_sz)
177 {
178 	struct thread_stack *ts = thread__ts(thread), *new_ts;
179 	unsigned int old_sz = ts ? ts->arr_sz : 0;
180 	unsigned int new_sz = 1;
181 
182 	if (thread_stack__per_cpu(thread) && cpu > 0)
183 		new_sz = roundup_pow_of_two(cpu + 1);
184 
185 	if (!ts || new_sz > old_sz) {
186 		new_ts = calloc(new_sz, sizeof(*ts));
187 		if (!new_ts)
188 			return NULL;
189 		if (ts)
190 			memcpy(new_ts, ts, old_sz * sizeof(*ts));
191 		new_ts->arr_sz = new_sz;
192 		free(thread__ts(thread));
193 		thread__set_ts(thread, new_ts);
194 		ts = new_ts;
195 	}
196 
197 	if (thread_stack__per_cpu(thread) && cpu > 0 &&
198 	    (unsigned int)cpu < ts->arr_sz)
199 		ts += cpu;
200 
201 	if (!ts->stack &&
202 	    thread_stack__init(ts, thread, crp, callstack, br_stack_sz))
203 		return NULL;
204 
205 	return ts;
206 }
207 
208 static struct thread_stack *thread__cpu_stack(struct thread *thread, int cpu)
209 {
210 	struct thread_stack *ts = thread__ts(thread);
211 
212 	if (cpu < 0)
213 		cpu = 0;
214 
215 	if (!ts || (unsigned int)cpu >= ts->arr_sz)
216 		return NULL;
217 
218 	ts += cpu;
219 
220 	if (!ts->stack)
221 		return NULL;
222 
223 	return ts;
224 }
225 
226 static inline struct thread_stack *thread__stack(struct thread *thread,
227 						    int cpu)
228 {
229 	if (!thread)
230 		return NULL;
231 
232 	if (thread_stack__per_cpu(thread))
233 		return thread__cpu_stack(thread, cpu);
234 
235 	return thread__ts(thread);
236 }
237 
238 static int thread_stack__push(struct thread_stack *ts, u64 ret_addr,
239 			      bool trace_end)
240 {
241 	int err = 0;
242 
243 	if (ts->cnt == ts->sz) {
244 		err = thread_stack__grow(ts);
245 		if (err) {
246 			pr_warning("Out of memory: discarding thread stack\n");
247 			ts->cnt = 0;
248 		}
249 	}
250 
251 	ts->stack[ts->cnt].trace_end = trace_end;
252 	ts->stack[ts->cnt++].ret_addr = ret_addr;
253 
254 	return err;
255 }
256 
257 static void thread_stack__pop(struct thread_stack *ts, u64 ret_addr)
258 {
259 	size_t i;
260 
261 	/*
262 	 * In some cases there may be functions which are not seen to return.
263 	 * For example when setjmp / longjmp has been used.  Or the perf context
264 	 * switch in the kernel which doesn't stop and start tracing in exactly
265 	 * the same code path.  When that happens the return address will be
266 	 * further down the stack.  If the return address is not found at all,
267 	 * we assume the opposite (i.e. this is a return for a call that wasn't
268 	 * seen for some reason) and leave the stack alone.
269 	 */
270 	for (i = ts->cnt; i; ) {
271 		if (ts->stack[--i].ret_addr == ret_addr) {
272 			ts->cnt = i;
273 			return;
274 		}
275 	}
276 }
277 
278 static void thread_stack__pop_trace_end(struct thread_stack *ts)
279 {
280 	size_t i;
281 
282 	for (i = ts->cnt; i; ) {
283 		if (ts->stack[--i].trace_end)
284 			ts->cnt = i;
285 		else
286 			return;
287 	}
288 }
289 
290 static bool thread_stack__in_kernel(struct thread_stack *ts)
291 {
292 	if (!ts->cnt)
293 		return false;
294 
295 	return ts->stack[ts->cnt - 1].cp->in_kernel;
296 }
297 
298 static int thread_stack__call_return(struct thread *thread,
299 				     struct thread_stack *ts, size_t idx,
300 				     u64 timestamp, u64 ref, bool no_return)
301 {
302 	struct call_return_processor *crp = ts->crp;
303 	struct thread_stack_entry *tse;
304 	struct call_return cr = {
305 		.thread = thread,
306 		.comm = ts->comm,
307 		.db_id = 0,
308 	};
309 	u64 *parent_db_id;
310 
311 	tse = &ts->stack[idx];
312 	cr.cp = tse->cp;
313 	cr.call_time = tse->timestamp;
314 	cr.return_time = timestamp;
315 	cr.branch_count = ts->branch_count - tse->branch_count;
316 	cr.insn_count = ts->insn_count - tse->insn_count;
317 	cr.cyc_count = ts->cyc_count - tse->cyc_count;
318 	cr.db_id = tse->db_id;
319 	cr.call_ref = tse->ref;
320 	cr.return_ref = ref;
321 	if (tse->no_call)
322 		cr.flags |= CALL_RETURN_NO_CALL;
323 	if (no_return)
324 		cr.flags |= CALL_RETURN_NO_RETURN;
325 	if (tse->non_call)
326 		cr.flags |= CALL_RETURN_NON_CALL;
327 
328 	/*
329 	 * The parent db_id must be assigned before exporting the child. Note
330 	 * it is not possible to export the parent first because its information
331 	 * is not yet complete because its 'return' has not yet been processed.
332 	 */
333 	parent_db_id = idx ? &(tse - 1)->db_id : NULL;
334 
335 	return crp->process(&cr, parent_db_id, crp->data);
336 }
337 
338 static int __thread_stack__flush(struct thread *thread, struct thread_stack *ts)
339 {
340 	struct call_return_processor *crp = ts->crp;
341 	int err;
342 
343 	if (!crp) {
344 		ts->cnt = 0;
345 		ts->br_stack_pos = 0;
346 		if (ts->br_stack_rb)
347 			ts->br_stack_rb->nr = 0;
348 		return 0;
349 	}
350 
351 	while (ts->cnt) {
352 		err = thread_stack__call_return(thread, ts, --ts->cnt,
353 						ts->last_time, 0, true);
354 		if (err) {
355 			pr_err("Error flushing thread stack!\n");
356 			ts->cnt = 0;
357 			return err;
358 		}
359 	}
360 
361 	return 0;
362 }
363 
364 int thread_stack__flush(struct thread *thread)
365 {
366 	struct thread_stack *ts = thread__ts(thread);
367 	unsigned int pos;
368 	int err = 0;
369 
370 	if (ts) {
371 		for (pos = 0; pos < ts->arr_sz; pos++) {
372 			int ret = __thread_stack__flush(thread, ts + pos);
373 
374 			if (ret)
375 				err = ret;
376 		}
377 	}
378 
379 	return err;
380 }
381 
382 static void thread_stack__update_br_stack(struct thread_stack *ts, u32 flags,
383 					  u64 from_ip, u64 to_ip)
384 {
385 	struct branch_stack *bs = ts->br_stack_rb;
386 	struct branch_entry *be;
387 
388 	if (!ts->br_stack_pos)
389 		ts->br_stack_pos = ts->br_stack_sz;
390 
391 	ts->br_stack_pos -= 1;
392 
393 	be              = &bs->entries[ts->br_stack_pos];
394 	be->from        = from_ip;
395 	be->to          = to_ip;
396 	be->flags.value = 0;
397 	be->flags.abort = !!(flags & PERF_IP_FLAG_TX_ABORT);
398 	be->flags.in_tx = !!(flags & PERF_IP_FLAG_IN_TX);
399 	/* No support for mispredict */
400 	be->flags.mispred = ts->mispred_all;
401 
402 	if (bs->nr < ts->br_stack_sz)
403 		bs->nr += 1;
404 }
405 
406 int thread_stack__event(struct thread *thread, int cpu, u32 flags, u64 from_ip,
407 			u64 to_ip, u16 insn_len, u64 trace_nr, bool callstack,
408 			unsigned int br_stack_sz, bool mispred_all)
409 {
410 	struct thread_stack *ts = thread__stack(thread, cpu);
411 
412 	if (!thread)
413 		return -EINVAL;
414 
415 	if (!ts) {
416 		ts = thread_stack__new(thread, cpu, NULL, callstack, br_stack_sz);
417 		if (!ts) {
418 			pr_warning("Out of memory: no thread stack\n");
419 			return -ENOMEM;
420 		}
421 		ts->trace_nr = trace_nr;
422 		ts->mispred_all = mispred_all;
423 	}
424 
425 	/*
426 	 * When the trace is discontinuous, the trace_nr changes.  In that case
427 	 * the stack might be completely invalid.  Better to report nothing than
428 	 * to report something misleading, so flush the stack.
429 	 */
430 	if (trace_nr != ts->trace_nr) {
431 		if (ts->trace_nr)
432 			__thread_stack__flush(thread, ts);
433 		ts->trace_nr = trace_nr;
434 	}
435 
436 	if (br_stack_sz)
437 		thread_stack__update_br_stack(ts, flags, from_ip, to_ip);
438 
439 	/*
440 	 * Stop here if thread_stack__process() is in use, or not recording call
441 	 * stack.
442 	 */
443 	if (ts->crp || !callstack)
444 		return 0;
445 
446 	if (flags & PERF_IP_FLAG_CALL) {
447 		u64 ret_addr;
448 
449 		if (!to_ip)
450 			return 0;
451 		ret_addr = from_ip + insn_len;
452 		if (ret_addr == to_ip)
453 			return 0; /* Zero-length calls are excluded */
454 		return thread_stack__push(ts, ret_addr,
455 					  flags & PERF_IP_FLAG_TRACE_END);
456 	} else if (flags & PERF_IP_FLAG_TRACE_BEGIN) {
457 		/*
458 		 * If the caller did not change the trace number (which would
459 		 * have flushed the stack) then try to make sense of the stack.
460 		 * Possibly, tracing began after returning to the current
461 		 * address, so try to pop that. Also, do not expect a call made
462 		 * when the trace ended, to return, so pop that.
463 		 */
464 		thread_stack__pop(ts, to_ip);
465 		thread_stack__pop_trace_end(ts);
466 	} else if ((flags & PERF_IP_FLAG_RETURN) && from_ip) {
467 		thread_stack__pop(ts, to_ip);
468 	}
469 
470 	return 0;
471 }
472 
473 void thread_stack__set_trace_nr(struct thread *thread, int cpu, u64 trace_nr)
474 {
475 	struct thread_stack *ts = thread__stack(thread, cpu);
476 
477 	if (!ts)
478 		return;
479 
480 	if (trace_nr != ts->trace_nr) {
481 		if (ts->trace_nr)
482 			__thread_stack__flush(thread, ts);
483 		ts->trace_nr = trace_nr;
484 	}
485 }
486 
487 static void __thread_stack__free(struct thread *thread, struct thread_stack *ts)
488 {
489 	__thread_stack__flush(thread, ts);
490 	zfree(&ts->stack);
491 	zfree(&ts->br_stack_rb);
492 }
493 
494 static void thread_stack__reset(struct thread *thread, struct thread_stack *ts)
495 {
496 	unsigned int arr_sz = ts->arr_sz;
497 
498 	__thread_stack__free(thread, ts);
499 	memset(ts, 0, sizeof(*ts));
500 	ts->arr_sz = arr_sz;
501 }
502 
503 void thread_stack__free(struct thread *thread)
504 {
505 	struct thread_stack *ts = thread__ts(thread);
506 	unsigned int pos;
507 
508 	if (ts) {
509 		for (pos = 0; pos < ts->arr_sz; pos++)
510 			__thread_stack__free(thread, ts + pos);
511 		free(thread__ts(thread));
512 		thread__set_ts(thread, NULL);
513 	}
514 }
515 
516 static inline u64 callchain_context(u64 ip, u64 kernel_start)
517 {
518 	return ip < kernel_start ? PERF_CONTEXT_USER : PERF_CONTEXT_KERNEL;
519 }
520 
521 void thread_stack__sample(struct thread *thread, int cpu,
522 			  struct ip_callchain *chain,
523 			  size_t sz, u64 ip, u64 kernel_start)
524 {
525 	struct thread_stack *ts = thread__stack(thread, cpu);
526 	u64 context = callchain_context(ip, kernel_start);
527 	u64 last_context;
528 	size_t i, j;
529 
530 	if (sz < 2) {
531 		chain->nr = 0;
532 		return;
533 	}
534 
535 	chain->ips[0] = context;
536 	chain->ips[1] = ip;
537 
538 	if (!ts) {
539 		chain->nr = 2;
540 		return;
541 	}
542 
543 	last_context = context;
544 
545 	for (i = 2, j = 1; i < sz && j <= ts->cnt; i++, j++) {
546 		ip = ts->stack[ts->cnt - j].ret_addr;
547 		context = callchain_context(ip, kernel_start);
548 		if (context != last_context) {
549 			if (i >= sz - 1)
550 				break;
551 			chain->ips[i++] = context;
552 			last_context = context;
553 		}
554 		chain->ips[i] = ip;
555 	}
556 
557 	chain->nr = i;
558 }
559 
560 /*
561  * Hardware sample records, created some time after the event occurred, need to
562  * have subsequent addresses removed from the call chain.
563  */
564 void thread_stack__sample_late(struct thread *thread, int cpu,
565 			       struct ip_callchain *chain, size_t sz,
566 			       u64 sample_ip, u64 kernel_start)
567 {
568 	struct thread_stack *ts = thread__stack(thread, cpu);
569 	u64 sample_context = callchain_context(sample_ip, kernel_start);
570 	u64 last_context, context, ip;
571 	size_t nr = 0, j;
572 
573 	if (sz < 2) {
574 		chain->nr = 0;
575 		return;
576 	}
577 
578 	if (!ts)
579 		goto out;
580 
581 	/*
582 	 * When tracing kernel space, kernel addresses occur at the top of the
583 	 * call chain after the event occurred but before tracing stopped.
584 	 * Skip them.
585 	 */
586 	for (j = 1; j <= ts->cnt; j++) {
587 		ip = ts->stack[ts->cnt - j].ret_addr;
588 		context = callchain_context(ip, kernel_start);
589 		if (context == PERF_CONTEXT_USER ||
590 		    (context == sample_context && ip == sample_ip))
591 			break;
592 	}
593 
594 	last_context = sample_ip; /* Use sample_ip as an invalid context */
595 
596 	for (; nr < sz && j <= ts->cnt; nr++, j++) {
597 		ip = ts->stack[ts->cnt - j].ret_addr;
598 		context = callchain_context(ip, kernel_start);
599 		if (context != last_context) {
600 			if (nr >= sz - 1)
601 				break;
602 			chain->ips[nr++] = context;
603 			last_context = context;
604 		}
605 		chain->ips[nr] = ip;
606 	}
607 out:
608 	if (nr) {
609 		chain->nr = nr;
610 	} else {
611 		chain->ips[0] = sample_context;
612 		chain->ips[1] = sample_ip;
613 		chain->nr = 2;
614 	}
615 }
616 
617 void thread_stack__br_sample(struct thread *thread, int cpu,
618 			     struct branch_stack *dst, unsigned int sz)
619 {
620 	struct thread_stack *ts = thread__stack(thread, cpu);
621 	const size_t bsz = sizeof(struct branch_entry);
622 	struct branch_stack *src;
623 	struct branch_entry *be;
624 	unsigned int nr;
625 
626 	dst->nr = 0;
627 
628 	if (!ts)
629 		return;
630 
631 	src = ts->br_stack_rb;
632 	if (!src->nr)
633 		return;
634 
635 	dst->nr = min((unsigned int)src->nr, sz);
636 
637 	be = &dst->entries[0];
638 	nr = min(ts->br_stack_sz - ts->br_stack_pos, (unsigned int)dst->nr);
639 	memcpy(be, &src->entries[ts->br_stack_pos], bsz * nr);
640 
641 	if (src->nr >= ts->br_stack_sz) {
642 		sz -= nr;
643 		be = &dst->entries[nr];
644 		nr = min(ts->br_stack_pos, sz);
645 		memcpy(be, &src->entries[0], bsz * ts->br_stack_pos);
646 	}
647 }
648 
649 /* Start of user space branch entries */
650 static bool us_start(struct branch_entry *be, u64 kernel_start, bool *start)
651 {
652 	if (!*start)
653 		*start = be->to && be->to < kernel_start;
654 
655 	return *start;
656 }
657 
658 /*
659  * Start of branch entries after the ip fell in between 2 branches, or user
660  * space branch entries.
661  */
662 static bool ks_start(struct branch_entry *be, u64 sample_ip, u64 kernel_start,
663 		     bool *start, struct branch_entry *nb)
664 {
665 	if (!*start) {
666 		*start = (nb && sample_ip >= be->to && sample_ip <= nb->from) ||
667 			 be->from < kernel_start ||
668 			 (be->to && be->to < kernel_start);
669 	}
670 
671 	return *start;
672 }
673 
674 /*
675  * Hardware sample records, created some time after the event occurred, need to
676  * have subsequent addresses removed from the branch stack.
677  */
678 void thread_stack__br_sample_late(struct thread *thread, int cpu,
679 				  struct branch_stack *dst, unsigned int sz,
680 				  u64 ip, u64 kernel_start)
681 {
682 	struct thread_stack *ts = thread__stack(thread, cpu);
683 	struct branch_entry *d, *s, *spos, *ssz;
684 	struct branch_stack *src;
685 	unsigned int nr = 0;
686 	bool start = false;
687 
688 	dst->nr = 0;
689 
690 	if (!ts)
691 		return;
692 
693 	src = ts->br_stack_rb;
694 	if (!src->nr)
695 		return;
696 
697 	spos = &src->entries[ts->br_stack_pos];
698 	ssz  = &src->entries[ts->br_stack_sz];
699 
700 	d = &dst->entries[0];
701 	s = spos;
702 
703 	if (ip < kernel_start) {
704 		/*
705 		 * User space sample: start copying branch entries when the
706 		 * branch is in user space.
707 		 */
708 		for (s = spos; s < ssz && nr < sz; s++) {
709 			if (us_start(s, kernel_start, &start)) {
710 				*d++ = *s;
711 				nr += 1;
712 			}
713 		}
714 
715 		if (src->nr >= ts->br_stack_sz) {
716 			for (s = &src->entries[0]; s < spos && nr < sz; s++) {
717 				if (us_start(s, kernel_start, &start)) {
718 					*d++ = *s;
719 					nr += 1;
720 				}
721 			}
722 		}
723 	} else {
724 		struct branch_entry *nb = NULL;
725 
726 		/*
727 		 * Kernel space sample: start copying branch entries when the ip
728 		 * falls in between 2 branches (or the branch is in user space
729 		 * because then the start must have been missed).
730 		 */
731 		for (s = spos; s < ssz && nr < sz; s++) {
732 			if (ks_start(s, ip, kernel_start, &start, nb)) {
733 				*d++ = *s;
734 				nr += 1;
735 			}
736 			nb = s;
737 		}
738 
739 		if (src->nr >= ts->br_stack_sz) {
740 			for (s = &src->entries[0]; s < spos && nr < sz; s++) {
741 				if (ks_start(s, ip, kernel_start, &start, nb)) {
742 					*d++ = *s;
743 					nr += 1;
744 				}
745 				nb = s;
746 			}
747 		}
748 	}
749 
750 	dst->nr = nr;
751 }
752 
753 struct call_return_processor *
754 call_return_processor__new(int (*process)(struct call_return *cr, u64 *parent_db_id, void *data),
755 			   void *data)
756 {
757 	struct call_return_processor *crp;
758 
759 	crp = zalloc(sizeof(struct call_return_processor));
760 	if (!crp)
761 		return NULL;
762 	crp->cpr = call_path_root__new();
763 	if (!crp->cpr)
764 		goto out_free;
765 	crp->process = process;
766 	crp->data = data;
767 	return crp;
768 
769 out_free:
770 	free(crp);
771 	return NULL;
772 }
773 
774 void call_return_processor__free(struct call_return_processor *crp)
775 {
776 	if (crp) {
777 		call_path_root__free(crp->cpr);
778 		free(crp);
779 	}
780 }
781 
782 static int thread_stack__push_cp(struct thread_stack *ts, u64 ret_addr,
783 				 u64 timestamp, u64 ref, struct call_path *cp,
784 				 bool no_call, bool trace_end)
785 {
786 	struct thread_stack_entry *tse;
787 	int err;
788 
789 	if (!cp)
790 		return -ENOMEM;
791 
792 	if (ts->cnt == ts->sz) {
793 		err = thread_stack__grow(ts);
794 		if (err)
795 			return err;
796 	}
797 
798 	tse = &ts->stack[ts->cnt++];
799 	tse->ret_addr = ret_addr;
800 	tse->timestamp = timestamp;
801 	tse->ref = ref;
802 	tse->branch_count = ts->branch_count;
803 	tse->insn_count = ts->insn_count;
804 	tse->cyc_count = ts->cyc_count;
805 	tse->cp = cp;
806 	tse->no_call = no_call;
807 	tse->trace_end = trace_end;
808 	tse->non_call = false;
809 	tse->db_id = 0;
810 
811 	return 0;
812 }
813 
814 static int thread_stack__pop_cp(struct thread *thread, struct thread_stack *ts,
815 				u64 ret_addr, u64 timestamp, u64 ref,
816 				struct symbol *sym)
817 {
818 	int err;
819 
820 	if (!ts->cnt)
821 		return 1;
822 
823 	if (ts->cnt == 1) {
824 		struct thread_stack_entry *tse = &ts->stack[0];
825 
826 		if (tse->cp->sym == sym)
827 			return thread_stack__call_return(thread, ts, --ts->cnt,
828 							 timestamp, ref, false);
829 	}
830 
831 	if (ts->stack[ts->cnt - 1].ret_addr == ret_addr &&
832 	    !ts->stack[ts->cnt - 1].non_call) {
833 		return thread_stack__call_return(thread, ts, --ts->cnt,
834 						 timestamp, ref, false);
835 	} else {
836 		size_t i = ts->cnt - 1;
837 
838 		while (i--) {
839 			if (ts->stack[i].ret_addr != ret_addr ||
840 			    ts->stack[i].non_call)
841 				continue;
842 			i += 1;
843 			while (ts->cnt > i) {
844 				err = thread_stack__call_return(thread, ts,
845 								--ts->cnt,
846 								timestamp, ref,
847 								true);
848 				if (err)
849 					return err;
850 			}
851 			return thread_stack__call_return(thread, ts, --ts->cnt,
852 							 timestamp, ref, false);
853 		}
854 	}
855 
856 	return 1;
857 }
858 
859 static int thread_stack__bottom(struct thread_stack *ts,
860 				struct perf_sample *sample,
861 				struct addr_location *from_al,
862 				struct addr_location *to_al, u64 ref)
863 {
864 	struct call_path_root *cpr = ts->crp->cpr;
865 	struct call_path *cp;
866 	struct symbol *sym;
867 	u64 ip;
868 
869 	if (sample->ip) {
870 		ip = sample->ip;
871 		sym = from_al->sym;
872 	} else if (sample->addr) {
873 		ip = sample->addr;
874 		sym = to_al->sym;
875 	} else {
876 		return 0;
877 	}
878 
879 	cp = call_path__findnew(cpr, &cpr->call_path, sym, ip,
880 				ts->kernel_start);
881 
882 	return thread_stack__push_cp(ts, ip, sample->time, ref, cp,
883 				     true, false);
884 }
885 
886 static int thread_stack__pop_ks(struct thread *thread, struct thread_stack *ts,
887 				struct perf_sample *sample, u64 ref)
888 {
889 	u64 tm = sample->time;
890 	int err;
891 
892 	/* Return to userspace, so pop all kernel addresses */
893 	while (thread_stack__in_kernel(ts)) {
894 		err = thread_stack__call_return(thread, ts, --ts->cnt,
895 						tm, ref, true);
896 		if (err)
897 			return err;
898 	}
899 
900 	return 0;
901 }
902 
903 static int thread_stack__no_call_return(struct thread *thread,
904 					struct thread_stack *ts,
905 					struct perf_sample *sample,
906 					struct addr_location *from_al,
907 					struct addr_location *to_al, u64 ref)
908 {
909 	struct call_path_root *cpr = ts->crp->cpr;
910 	struct call_path *root = &cpr->call_path;
911 	struct symbol *fsym = from_al->sym;
912 	struct symbol *tsym = to_al->sym;
913 	struct call_path *cp, *parent;
914 	u64 ks = ts->kernel_start;
915 	u64 addr = sample->addr;
916 	u64 tm = sample->time;
917 	u64 ip = sample->ip;
918 	int err;
919 
920 	if (ip >= ks && addr < ks) {
921 		/* Return to userspace, so pop all kernel addresses */
922 		err = thread_stack__pop_ks(thread, ts, sample, ref);
923 		if (err)
924 			return err;
925 
926 		/* If the stack is empty, push the userspace address */
927 		if (!ts->cnt) {
928 			cp = call_path__findnew(cpr, root, tsym, addr, ks);
929 			return thread_stack__push_cp(ts, 0, tm, ref, cp, true,
930 						     false);
931 		}
932 	} else if (thread_stack__in_kernel(ts) && ip < ks) {
933 		/* Return to userspace, so pop all kernel addresses */
934 		err = thread_stack__pop_ks(thread, ts, sample, ref);
935 		if (err)
936 			return err;
937 	}
938 
939 	if (ts->cnt)
940 		parent = ts->stack[ts->cnt - 1].cp;
941 	else
942 		parent = root;
943 
944 	if (parent->sym == from_al->sym) {
945 		/*
946 		 * At the bottom of the stack, assume the missing 'call' was
947 		 * before the trace started. So, pop the current symbol and push
948 		 * the 'to' symbol.
949 		 */
950 		if (ts->cnt == 1) {
951 			err = thread_stack__call_return(thread, ts, --ts->cnt,
952 							tm, ref, false);
953 			if (err)
954 				return err;
955 		}
956 
957 		if (!ts->cnt) {
958 			cp = call_path__findnew(cpr, root, tsym, addr, ks);
959 
960 			return thread_stack__push_cp(ts, addr, tm, ref, cp,
961 						     true, false);
962 		}
963 
964 		/*
965 		 * Otherwise assume the 'return' is being used as a jump (e.g.
966 		 * retpoline) and just push the 'to' symbol.
967 		 */
968 		cp = call_path__findnew(cpr, parent, tsym, addr, ks);
969 
970 		err = thread_stack__push_cp(ts, 0, tm, ref, cp, true, false);
971 		if (!err)
972 			ts->stack[ts->cnt - 1].non_call = true;
973 
974 		return err;
975 	}
976 
977 	/*
978 	 * Assume 'parent' has not yet returned, so push 'to', and then push and
979 	 * pop 'from'.
980 	 */
981 
982 	cp = call_path__findnew(cpr, parent, tsym, addr, ks);
983 
984 	err = thread_stack__push_cp(ts, addr, tm, ref, cp, true, false);
985 	if (err)
986 		return err;
987 
988 	cp = call_path__findnew(cpr, cp, fsym, ip, ks);
989 
990 	err = thread_stack__push_cp(ts, ip, tm, ref, cp, true, false);
991 	if (err)
992 		return err;
993 
994 	return thread_stack__call_return(thread, ts, --ts->cnt, tm, ref, false);
995 }
996 
997 static int thread_stack__trace_begin(struct thread *thread,
998 				     struct thread_stack *ts, u64 timestamp,
999 				     u64 ref)
1000 {
1001 	struct thread_stack_entry *tse;
1002 	int err;
1003 
1004 	if (!ts->cnt)
1005 		return 0;
1006 
1007 	/* Pop trace end */
1008 	tse = &ts->stack[ts->cnt - 1];
1009 	if (tse->trace_end) {
1010 		err = thread_stack__call_return(thread, ts, --ts->cnt,
1011 						timestamp, ref, false);
1012 		if (err)
1013 			return err;
1014 	}
1015 
1016 	return 0;
1017 }
1018 
1019 static int thread_stack__trace_end(struct thread_stack *ts,
1020 				   struct perf_sample *sample, u64 ref)
1021 {
1022 	struct call_path_root *cpr = ts->crp->cpr;
1023 	struct call_path *cp;
1024 	u64 ret_addr;
1025 
1026 	/* No point having 'trace end' on the bottom of the stack */
1027 	if (!ts->cnt || (ts->cnt == 1 && ts->stack[0].ref == ref))
1028 		return 0;
1029 
1030 	cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp, NULL, 0,
1031 				ts->kernel_start);
1032 
1033 	ret_addr = sample->ip + sample->insn_len;
1034 
1035 	return thread_stack__push_cp(ts, ret_addr, sample->time, ref, cp,
1036 				     false, true);
1037 }
1038 
1039 static bool is_x86_retpoline(const char *name)
1040 {
1041 	return strstr(name, "__x86_indirect_thunk_") == name;
1042 }
1043 
1044 /*
1045  * x86 retpoline functions pollute the call graph. This function removes them.
1046  * This does not handle function return thunks, nor is there any improvement
1047  * for the handling of inline thunks or extern thunks.
1048  */
1049 static int thread_stack__x86_retpoline(struct thread_stack *ts,
1050 				       struct perf_sample *sample,
1051 				       struct addr_location *to_al)
1052 {
1053 	struct thread_stack_entry *tse = &ts->stack[ts->cnt - 1];
1054 	struct call_path_root *cpr = ts->crp->cpr;
1055 	struct symbol *sym = tse->cp->sym;
1056 	struct symbol *tsym = to_al->sym;
1057 	struct call_path *cp;
1058 
1059 	if (sym && is_x86_retpoline(sym->name)) {
1060 		/*
1061 		 * This is a x86 retpoline fn. It pollutes the call graph by
1062 		 * showing up everywhere there is an indirect branch, but does
1063 		 * not itself mean anything. Here the top-of-stack is removed,
1064 		 * by decrementing the stack count, and then further down, the
1065 		 * resulting top-of-stack is replaced with the actual target.
1066 		 * The result is that the retpoline functions will no longer
1067 		 * appear in the call graph. Note this only affects the call
1068 		 * graph, since all the original branches are left unchanged.
1069 		 */
1070 		ts->cnt -= 1;
1071 		sym = ts->stack[ts->cnt - 2].cp->sym;
1072 		if (sym && sym == tsym && to_al->addr != tsym->start) {
1073 			/*
1074 			 * Target is back to the middle of the symbol we came
1075 			 * from so assume it is an indirect jmp and forget it
1076 			 * altogether.
1077 			 */
1078 			ts->cnt -= 1;
1079 			return 0;
1080 		}
1081 	} else if (sym && sym == tsym) {
1082 		/*
1083 		 * Target is back to the symbol we came from so assume it is an
1084 		 * indirect jmp and forget it altogether.
1085 		 */
1086 		ts->cnt -= 1;
1087 		return 0;
1088 	}
1089 
1090 	cp = call_path__findnew(cpr, ts->stack[ts->cnt - 2].cp, tsym,
1091 				sample->addr, ts->kernel_start);
1092 	if (!cp)
1093 		return -ENOMEM;
1094 
1095 	/* Replace the top-of-stack with the actual target */
1096 	ts->stack[ts->cnt - 1].cp = cp;
1097 
1098 	return 0;
1099 }
1100 
1101 int thread_stack__process(struct thread *thread, struct comm *comm,
1102 			  struct perf_sample *sample,
1103 			  struct addr_location *from_al,
1104 			  struct addr_location *to_al, u64 ref,
1105 			  struct call_return_processor *crp)
1106 {
1107 	struct thread_stack *ts = thread__stack(thread, sample->cpu);
1108 	enum retpoline_state_t rstate;
1109 	int err = 0;
1110 
1111 	if (ts && !ts->crp) {
1112 		/* Supersede thread_stack__event() */
1113 		thread_stack__reset(thread, ts);
1114 		ts = NULL;
1115 	}
1116 
1117 	if (!ts) {
1118 		ts = thread_stack__new(thread, sample->cpu, crp, true, 0);
1119 		if (!ts)
1120 			return -ENOMEM;
1121 		ts->comm = comm;
1122 	}
1123 
1124 	rstate = ts->rstate;
1125 	if (rstate == X86_RETPOLINE_DETECTED)
1126 		ts->rstate = X86_RETPOLINE_POSSIBLE;
1127 
1128 	/* Flush stack on exec */
1129 	if (ts->comm != comm && thread__pid(thread) == thread__tid(thread)) {
1130 		err = __thread_stack__flush(thread, ts);
1131 		if (err)
1132 			return err;
1133 		ts->comm = comm;
1134 	}
1135 
1136 	/* If the stack is empty, put the current symbol on the stack */
1137 	if (!ts->cnt) {
1138 		err = thread_stack__bottom(ts, sample, from_al, to_al, ref);
1139 		if (err)
1140 			return err;
1141 	}
1142 
1143 	ts->branch_count += 1;
1144 	ts->insn_count += sample->insn_cnt;
1145 	ts->cyc_count += sample->cyc_cnt;
1146 	ts->last_time = sample->time;
1147 
1148 	if (sample->flags & PERF_IP_FLAG_CALL) {
1149 		bool trace_end = sample->flags & PERF_IP_FLAG_TRACE_END;
1150 		struct call_path_root *cpr = ts->crp->cpr;
1151 		struct call_path *cp;
1152 		u64 ret_addr;
1153 
1154 		if (!sample->ip || !sample->addr)
1155 			return 0;
1156 
1157 		ret_addr = sample->ip + sample->insn_len;
1158 		if (ret_addr == sample->addr)
1159 			return 0; /* Zero-length calls are excluded */
1160 
1161 		cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp,
1162 					to_al->sym, sample->addr,
1163 					ts->kernel_start);
1164 		err = thread_stack__push_cp(ts, ret_addr, sample->time, ref,
1165 					    cp, false, trace_end);
1166 
1167 		/*
1168 		 * A call to the same symbol but not the start of the symbol,
1169 		 * may be the start of a x86 retpoline.
1170 		 */
1171 		if (!err && rstate == X86_RETPOLINE_POSSIBLE && to_al->sym &&
1172 		    from_al->sym == to_al->sym &&
1173 		    to_al->addr != to_al->sym->start)
1174 			ts->rstate = X86_RETPOLINE_DETECTED;
1175 
1176 	} else if (sample->flags & PERF_IP_FLAG_RETURN) {
1177 		if (!sample->addr) {
1178 			u32 return_from_kernel = PERF_IP_FLAG_SYSCALLRET |
1179 						 PERF_IP_FLAG_INTERRUPT;
1180 
1181 			if (!(sample->flags & return_from_kernel))
1182 				return 0;
1183 
1184 			/* Pop kernel stack */
1185 			return thread_stack__pop_ks(thread, ts, sample, ref);
1186 		}
1187 
1188 		if (!sample->ip)
1189 			return 0;
1190 
1191 		/* x86 retpoline 'return' doesn't match the stack */
1192 		if (rstate == X86_RETPOLINE_DETECTED && ts->cnt > 2 &&
1193 		    ts->stack[ts->cnt - 1].ret_addr != sample->addr)
1194 			return thread_stack__x86_retpoline(ts, sample, to_al);
1195 
1196 		err = thread_stack__pop_cp(thread, ts, sample->addr,
1197 					   sample->time, ref, from_al->sym);
1198 		if (err) {
1199 			if (err < 0)
1200 				return err;
1201 			err = thread_stack__no_call_return(thread, ts, sample,
1202 							   from_al, to_al, ref);
1203 		}
1204 	} else if (sample->flags & PERF_IP_FLAG_TRACE_BEGIN) {
1205 		err = thread_stack__trace_begin(thread, ts, sample->time, ref);
1206 	} else if (sample->flags & PERF_IP_FLAG_TRACE_END) {
1207 		err = thread_stack__trace_end(ts, sample, ref);
1208 	} else if (sample->flags & PERF_IP_FLAG_BRANCH &&
1209 		   from_al->sym != to_al->sym && to_al->sym &&
1210 		   to_al->addr == to_al->sym->start) {
1211 		struct call_path_root *cpr = ts->crp->cpr;
1212 		struct call_path *cp;
1213 
1214 		/*
1215 		 * The compiler might optimize a call/ret combination by making
1216 		 * it a jmp. Make that visible by recording on the stack a
1217 		 * branch to the start of a different symbol. Note, that means
1218 		 * when a ret pops the stack, all jmps must be popped off first.
1219 		 */
1220 		cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp,
1221 					to_al->sym, sample->addr,
1222 					ts->kernel_start);
1223 		err = thread_stack__push_cp(ts, 0, sample->time, ref, cp, false,
1224 					    false);
1225 		if (!err)
1226 			ts->stack[ts->cnt - 1].non_call = true;
1227 	}
1228 
1229 	return err;
1230 }
1231 
1232 size_t thread_stack__depth(struct thread *thread, int cpu)
1233 {
1234 	struct thread_stack *ts = thread__stack(thread, cpu);
1235 
1236 	if (!ts)
1237 		return 0;
1238 	return ts->cnt;
1239 }
1240