1 // SPDX-License-Identifier: GPL-2.0-only
3 * thread-stack.c: Synthesize a thread's stack using call / return events
4 * Copyright (c) 2014, Intel Corporation.
7 #include <linux/rbtree.h>
8 #include <linux/list.h>
9 #include <linux/log2.h>
19 #include "call-path.h"
20 #include "thread-stack.h"
22 #define STACK_GROWTH 2048
25 * State of retpoline detection.
27 * RETPOLINE_NONE: no retpoline detection
28 * X86_RETPOLINE_POSSIBLE: x86 retpoline possible
29 * X86_RETPOLINE_DETECTED: x86 retpoline detected
31 enum retpoline_state_t {
33 X86_RETPOLINE_POSSIBLE,
34 X86_RETPOLINE_DETECTED,
38 * struct thread_stack_entry - thread stack entry.
39 * @ret_addr: return address
40 * @timestamp: timestamp (if known)
41 * @ref: external reference (e.g. db_id of sample)
42 * @branch_count: the branch count when the entry was created
43 * @insn_count: the instruction count when the entry was created
44 * @cyc_count the cycle count when the entry was created
45 * @db_id: id used for db-export
47 * @no_call: a 'call' was not seen
48 * @trace_end: a 'call' but trace ended
49 * @non_call: a branch but not a 'call' to the start of a different symbol
51 struct thread_stack_entry {
66 * struct thread_stack - thread stack constructed from 'call' and 'return'
68 * @stack: array that holds the stack
69 * @cnt: number of entries in the stack
70 * @sz: current maximum stack size
71 * @trace_nr: current trace number
72 * @branch_count: running branch count
73 * @insn_count: running instruction count
74 * @cyc_count running cycle count
75 * @kernel_start: kernel start address
76 * @last_time: last timestamp
77 * @crp: call/return processor
79 * @arr_sz: size of array if this is the first element of an array
80 * @rstate: used to detect retpolines
83 struct thread_stack_entry *stack;
92 struct call_return_processor *crp;
95 enum retpoline_state_t rstate;
99 * Assume pid == tid == 0 identifies the idle task as defined by
100 * perf_session__register_idle_thread(). The idle task is really 1 task per cpu,
101 * and therefore requires a stack for each cpu.
103 static inline bool thread_stack__per_cpu(struct thread *thread)
105 return !(thread->tid || thread->pid_);
108 static int thread_stack__grow(struct thread_stack *ts)
110 struct thread_stack_entry *new_stack;
113 new_sz = ts->sz + STACK_GROWTH;
114 sz = new_sz * sizeof(struct thread_stack_entry);
116 new_stack = realloc(ts->stack, sz);
120 ts->stack = new_stack;
126 static int thread_stack__init(struct thread_stack *ts, struct thread *thread,
127 struct call_return_processor *crp)
131 err = thread_stack__grow(ts);
135 if (thread->mg && thread->mg->machine) {
136 struct machine *machine = thread->mg->machine;
137 const char *arch = perf_env__arch(machine->env);
139 ts->kernel_start = machine__kernel_start(machine);
140 if (!strcmp(arch, "x86"))
141 ts->rstate = X86_RETPOLINE_POSSIBLE;
143 ts->kernel_start = 1ULL << 63;
150 static struct thread_stack *thread_stack__new(struct thread *thread, int cpu,
151 struct call_return_processor *crp)
153 struct thread_stack *ts = thread->ts, *new_ts;
154 unsigned int old_sz = ts ? ts->arr_sz : 0;
155 unsigned int new_sz = 1;
157 if (thread_stack__per_cpu(thread) && cpu > 0)
158 new_sz = roundup_pow_of_two(cpu + 1);
160 if (!ts || new_sz > old_sz) {
161 new_ts = calloc(new_sz, sizeof(*ts));
165 memcpy(new_ts, ts, old_sz * sizeof(*ts));
166 new_ts->arr_sz = new_sz;
172 if (thread_stack__per_cpu(thread) && cpu > 0 &&
173 (unsigned int)cpu < ts->arr_sz)
177 thread_stack__init(ts, thread, crp))
183 static struct thread_stack *thread__cpu_stack(struct thread *thread, int cpu)
185 struct thread_stack *ts = thread->ts;
190 if (!ts || (unsigned int)cpu >= ts->arr_sz)
201 static inline struct thread_stack *thread__stack(struct thread *thread,
207 if (thread_stack__per_cpu(thread))
208 return thread__cpu_stack(thread, cpu);
213 static int thread_stack__push(struct thread_stack *ts, u64 ret_addr,
218 if (ts->cnt == ts->sz) {
219 err = thread_stack__grow(ts);
221 pr_warning("Out of memory: discarding thread stack\n");
226 ts->stack[ts->cnt].trace_end = trace_end;
227 ts->stack[ts->cnt++].ret_addr = ret_addr;
232 static void thread_stack__pop(struct thread_stack *ts, u64 ret_addr)
237 * In some cases there may be functions which are not seen to return.
238 * For example when setjmp / longjmp has been used. Or the perf context
239 * switch in the kernel which doesn't stop and start tracing in exactly
240 * the same code path. When that happens the return address will be
241 * further down the stack. If the return address is not found at all,
242 * we assume the opposite (i.e. this is a return for a call that wasn't
243 * seen for some reason) and leave the stack alone.
245 for (i = ts->cnt; i; ) {
246 if (ts->stack[--i].ret_addr == ret_addr) {
253 static void thread_stack__pop_trace_end(struct thread_stack *ts)
257 for (i = ts->cnt; i; ) {
258 if (ts->stack[--i].trace_end)
265 static bool thread_stack__in_kernel(struct thread_stack *ts)
270 return ts->stack[ts->cnt - 1].cp->in_kernel;
273 static int thread_stack__call_return(struct thread *thread,
274 struct thread_stack *ts, size_t idx,
275 u64 timestamp, u64 ref, bool no_return)
277 struct call_return_processor *crp = ts->crp;
278 struct thread_stack_entry *tse;
279 struct call_return cr = {
286 tse = &ts->stack[idx];
288 cr.call_time = tse->timestamp;
289 cr.return_time = timestamp;
290 cr.branch_count = ts->branch_count - tse->branch_count;
291 cr.insn_count = ts->insn_count - tse->insn_count;
292 cr.cyc_count = ts->cyc_count - tse->cyc_count;
293 cr.db_id = tse->db_id;
294 cr.call_ref = tse->ref;
297 cr.flags |= CALL_RETURN_NO_CALL;
299 cr.flags |= CALL_RETURN_NO_RETURN;
301 cr.flags |= CALL_RETURN_NON_CALL;
304 * The parent db_id must be assigned before exporting the child. Note
305 * it is not possible to export the parent first because its information
306 * is not yet complete because its 'return' has not yet been processed.
308 parent_db_id = idx ? &(tse - 1)->db_id : NULL;
310 return crp->process(&cr, parent_db_id, crp->data);
313 static int __thread_stack__flush(struct thread *thread, struct thread_stack *ts)
315 struct call_return_processor *crp = ts->crp;
324 err = thread_stack__call_return(thread, ts, --ts->cnt,
325 ts->last_time, 0, true);
327 pr_err("Error flushing thread stack!\n");
336 int thread_stack__flush(struct thread *thread)
338 struct thread_stack *ts = thread->ts;
343 for (pos = 0; pos < ts->arr_sz; pos++) {
344 int ret = __thread_stack__flush(thread, ts + pos);
354 int thread_stack__event(struct thread *thread, int cpu, u32 flags, u64 from_ip,
355 u64 to_ip, u16 insn_len, u64 trace_nr)
357 struct thread_stack *ts = thread__stack(thread, cpu);
363 ts = thread_stack__new(thread, cpu, NULL);
365 pr_warning("Out of memory: no thread stack\n");
368 ts->trace_nr = trace_nr;
372 * When the trace is discontinuous, the trace_nr changes. In that case
373 * the stack might be completely invalid. Better to report nothing than
374 * to report something misleading, so flush the stack.
376 if (trace_nr != ts->trace_nr) {
378 __thread_stack__flush(thread, ts);
379 ts->trace_nr = trace_nr;
382 /* Stop here if thread_stack__process() is in use */
386 if (flags & PERF_IP_FLAG_CALL) {
391 ret_addr = from_ip + insn_len;
392 if (ret_addr == to_ip)
393 return 0; /* Zero-length calls are excluded */
394 return thread_stack__push(ts, ret_addr,
395 flags & PERF_IP_FLAG_TRACE_END);
396 } else if (flags & PERF_IP_FLAG_TRACE_BEGIN) {
398 * If the caller did not change the trace number (which would
399 * have flushed the stack) then try to make sense of the stack.
400 * Possibly, tracing began after returning to the current
401 * address, so try to pop that. Also, do not expect a call made
402 * when the trace ended, to return, so pop that.
404 thread_stack__pop(ts, to_ip);
405 thread_stack__pop_trace_end(ts);
406 } else if ((flags & PERF_IP_FLAG_RETURN) && from_ip) {
407 thread_stack__pop(ts, to_ip);
413 void thread_stack__set_trace_nr(struct thread *thread, int cpu, u64 trace_nr)
415 struct thread_stack *ts = thread__stack(thread, cpu);
420 if (trace_nr != ts->trace_nr) {
422 __thread_stack__flush(thread, ts);
423 ts->trace_nr = trace_nr;
427 static void __thread_stack__free(struct thread *thread, struct thread_stack *ts)
429 __thread_stack__flush(thread, ts);
433 static void thread_stack__reset(struct thread *thread, struct thread_stack *ts)
435 unsigned int arr_sz = ts->arr_sz;
437 __thread_stack__free(thread, ts);
438 memset(ts, 0, sizeof(*ts));
442 void thread_stack__free(struct thread *thread)
444 struct thread_stack *ts = thread->ts;
448 for (pos = 0; pos < ts->arr_sz; pos++)
449 __thread_stack__free(thread, ts + pos);
454 static inline u64 callchain_context(u64 ip, u64 kernel_start)
456 return ip < kernel_start ? PERF_CONTEXT_USER : PERF_CONTEXT_KERNEL;
459 void thread_stack__sample(struct thread *thread, int cpu,
460 struct ip_callchain *chain,
461 size_t sz, u64 ip, u64 kernel_start)
463 struct thread_stack *ts = thread__stack(thread, cpu);
464 u64 context = callchain_context(ip, kernel_start);
473 chain->ips[0] = context;
481 last_context = context;
483 for (i = 2, j = 1; i < sz && j <= ts->cnt; i++, j++) {
484 ip = ts->stack[ts->cnt - j].ret_addr;
485 context = callchain_context(ip, kernel_start);
486 if (context != last_context) {
489 chain->ips[i++] = context;
490 last_context = context;
498 struct call_return_processor *
499 call_return_processor__new(int (*process)(struct call_return *cr, u64 *parent_db_id, void *data),
502 struct call_return_processor *crp;
504 crp = zalloc(sizeof(struct call_return_processor));
507 crp->cpr = call_path_root__new();
510 crp->process = process;
519 void call_return_processor__free(struct call_return_processor *crp)
522 call_path_root__free(crp->cpr);
527 static int thread_stack__push_cp(struct thread_stack *ts, u64 ret_addr,
528 u64 timestamp, u64 ref, struct call_path *cp,
529 bool no_call, bool trace_end)
531 struct thread_stack_entry *tse;
537 if (ts->cnt == ts->sz) {
538 err = thread_stack__grow(ts);
543 tse = &ts->stack[ts->cnt++];
544 tse->ret_addr = ret_addr;
545 tse->timestamp = timestamp;
547 tse->branch_count = ts->branch_count;
548 tse->insn_count = ts->insn_count;
549 tse->cyc_count = ts->cyc_count;
551 tse->no_call = no_call;
552 tse->trace_end = trace_end;
553 tse->non_call = false;
559 static int thread_stack__pop_cp(struct thread *thread, struct thread_stack *ts,
560 u64 ret_addr, u64 timestamp, u64 ref,
569 struct thread_stack_entry *tse = &ts->stack[0];
571 if (tse->cp->sym == sym)
572 return thread_stack__call_return(thread, ts, --ts->cnt,
573 timestamp, ref, false);
576 if (ts->stack[ts->cnt - 1].ret_addr == ret_addr &&
577 !ts->stack[ts->cnt - 1].non_call) {
578 return thread_stack__call_return(thread, ts, --ts->cnt,
579 timestamp, ref, false);
581 size_t i = ts->cnt - 1;
584 if (ts->stack[i].ret_addr != ret_addr ||
585 ts->stack[i].non_call)
588 while (ts->cnt > i) {
589 err = thread_stack__call_return(thread, ts,
596 return thread_stack__call_return(thread, ts, --ts->cnt,
597 timestamp, ref, false);
604 static int thread_stack__bottom(struct thread_stack *ts,
605 struct perf_sample *sample,
606 struct addr_location *from_al,
607 struct addr_location *to_al, u64 ref)
609 struct call_path_root *cpr = ts->crp->cpr;
610 struct call_path *cp;
617 } else if (sample->addr) {
624 cp = call_path__findnew(cpr, &cpr->call_path, sym, ip,
627 return thread_stack__push_cp(ts, ip, sample->time, ref, cp,
631 static int thread_stack__pop_ks(struct thread *thread, struct thread_stack *ts,
632 struct perf_sample *sample, u64 ref)
634 u64 tm = sample->time;
637 /* Return to userspace, so pop all kernel addresses */
638 while (thread_stack__in_kernel(ts)) {
639 err = thread_stack__call_return(thread, ts, --ts->cnt,
648 static int thread_stack__no_call_return(struct thread *thread,
649 struct thread_stack *ts,
650 struct perf_sample *sample,
651 struct addr_location *from_al,
652 struct addr_location *to_al, u64 ref)
654 struct call_path_root *cpr = ts->crp->cpr;
655 struct call_path *root = &cpr->call_path;
656 struct symbol *fsym = from_al->sym;
657 struct symbol *tsym = to_al->sym;
658 struct call_path *cp, *parent;
659 u64 ks = ts->kernel_start;
660 u64 addr = sample->addr;
661 u64 tm = sample->time;
665 if (ip >= ks && addr < ks) {
666 /* Return to userspace, so pop all kernel addresses */
667 err = thread_stack__pop_ks(thread, ts, sample, ref);
671 /* If the stack is empty, push the userspace address */
673 cp = call_path__findnew(cpr, root, tsym, addr, ks);
674 return thread_stack__push_cp(ts, 0, tm, ref, cp, true,
677 } else if (thread_stack__in_kernel(ts) && ip < ks) {
678 /* Return to userspace, so pop all kernel addresses */
679 err = thread_stack__pop_ks(thread, ts, sample, ref);
685 parent = ts->stack[ts->cnt - 1].cp;
689 if (parent->sym == from_al->sym) {
691 * At the bottom of the stack, assume the missing 'call' was
692 * before the trace started. So, pop the current symbol and push
696 err = thread_stack__call_return(thread, ts, --ts->cnt,
703 cp = call_path__findnew(cpr, root, tsym, addr, ks);
705 return thread_stack__push_cp(ts, addr, tm, ref, cp,
710 * Otherwise assume the 'return' is being used as a jump (e.g.
711 * retpoline) and just push the 'to' symbol.
713 cp = call_path__findnew(cpr, parent, tsym, addr, ks);
715 err = thread_stack__push_cp(ts, 0, tm, ref, cp, true, false);
717 ts->stack[ts->cnt - 1].non_call = true;
723 * Assume 'parent' has not yet returned, so push 'to', and then push and
727 cp = call_path__findnew(cpr, parent, tsym, addr, ks);
729 err = thread_stack__push_cp(ts, addr, tm, ref, cp, true, false);
733 cp = call_path__findnew(cpr, cp, fsym, ip, ks);
735 err = thread_stack__push_cp(ts, ip, tm, ref, cp, true, false);
739 return thread_stack__call_return(thread, ts, --ts->cnt, tm, ref, false);
742 static int thread_stack__trace_begin(struct thread *thread,
743 struct thread_stack *ts, u64 timestamp,
746 struct thread_stack_entry *tse;
753 tse = &ts->stack[ts->cnt - 1];
754 if (tse->trace_end) {
755 err = thread_stack__call_return(thread, ts, --ts->cnt,
756 timestamp, ref, false);
764 static int thread_stack__trace_end(struct thread_stack *ts,
765 struct perf_sample *sample, u64 ref)
767 struct call_path_root *cpr = ts->crp->cpr;
768 struct call_path *cp;
771 /* No point having 'trace end' on the bottom of the stack */
772 if (!ts->cnt || (ts->cnt == 1 && ts->stack[0].ref == ref))
775 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp, NULL, 0,
778 ret_addr = sample->ip + sample->insn_len;
780 return thread_stack__push_cp(ts, ret_addr, sample->time, ref, cp,
784 static bool is_x86_retpoline(const char *name)
786 const char *p = strstr(name, "__x86_indirect_thunk_");
788 return p == name || !strcmp(name, "__indirect_thunk_start");
792 * x86 retpoline functions pollute the call graph. This function removes them.
793 * This does not handle function return thunks, nor is there any improvement
794 * for the handling of inline thunks or extern thunks.
796 static int thread_stack__x86_retpoline(struct thread_stack *ts,
797 struct perf_sample *sample,
798 struct addr_location *to_al)
800 struct thread_stack_entry *tse = &ts->stack[ts->cnt - 1];
801 struct call_path_root *cpr = ts->crp->cpr;
802 struct symbol *sym = tse->cp->sym;
803 struct symbol *tsym = to_al->sym;
804 struct call_path *cp;
806 if (sym && is_x86_retpoline(sym->name)) {
808 * This is a x86 retpoline fn. It pollutes the call graph by
809 * showing up everywhere there is an indirect branch, but does
810 * not itself mean anything. Here the top-of-stack is removed,
811 * by decrementing the stack count, and then further down, the
812 * resulting top-of-stack is replaced with the actual target.
813 * The result is that the retpoline functions will no longer
814 * appear in the call graph. Note this only affects the call
815 * graph, since all the original branches are left unchanged.
818 sym = ts->stack[ts->cnt - 2].cp->sym;
819 if (sym && sym == tsym && to_al->addr != tsym->start) {
821 * Target is back to the middle of the symbol we came
822 * from so assume it is an indirect jmp and forget it
828 } else if (sym && sym == tsym) {
830 * Target is back to the symbol we came from so assume it is an
831 * indirect jmp and forget it altogether.
837 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 2].cp, tsym,
838 sample->addr, ts->kernel_start);
842 /* Replace the top-of-stack with the actual target */
843 ts->stack[ts->cnt - 1].cp = cp;
848 int thread_stack__process(struct thread *thread, struct comm *comm,
849 struct perf_sample *sample,
850 struct addr_location *from_al,
851 struct addr_location *to_al, u64 ref,
852 struct call_return_processor *crp)
854 struct thread_stack *ts = thread__stack(thread, sample->cpu);
855 enum retpoline_state_t rstate;
858 if (ts && !ts->crp) {
859 /* Supersede thread_stack__event() */
860 thread_stack__reset(thread, ts);
865 ts = thread_stack__new(thread, sample->cpu, crp);
872 if (rstate == X86_RETPOLINE_DETECTED)
873 ts->rstate = X86_RETPOLINE_POSSIBLE;
875 /* Flush stack on exec */
876 if (ts->comm != comm && thread->pid_ == thread->tid) {
877 err = __thread_stack__flush(thread, ts);
883 /* If the stack is empty, put the current symbol on the stack */
885 err = thread_stack__bottom(ts, sample, from_al, to_al, ref);
890 ts->branch_count += 1;
891 ts->insn_count += sample->insn_cnt;
892 ts->cyc_count += sample->cyc_cnt;
893 ts->last_time = sample->time;
895 if (sample->flags & PERF_IP_FLAG_CALL) {
896 bool trace_end = sample->flags & PERF_IP_FLAG_TRACE_END;
897 struct call_path_root *cpr = ts->crp->cpr;
898 struct call_path *cp;
901 if (!sample->ip || !sample->addr)
904 ret_addr = sample->ip + sample->insn_len;
905 if (ret_addr == sample->addr)
906 return 0; /* Zero-length calls are excluded */
908 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp,
909 to_al->sym, sample->addr,
911 err = thread_stack__push_cp(ts, ret_addr, sample->time, ref,
912 cp, false, trace_end);
915 * A call to the same symbol but not the start of the symbol,
916 * may be the start of a x86 retpoline.
918 if (!err && rstate == X86_RETPOLINE_POSSIBLE && to_al->sym &&
919 from_al->sym == to_al->sym &&
920 to_al->addr != to_al->sym->start)
921 ts->rstate = X86_RETPOLINE_DETECTED;
923 } else if (sample->flags & PERF_IP_FLAG_RETURN) {
925 u32 return_from_kernel = PERF_IP_FLAG_SYSCALLRET |
926 PERF_IP_FLAG_INTERRUPT;
928 if (!(sample->flags & return_from_kernel))
931 /* Pop kernel stack */
932 return thread_stack__pop_ks(thread, ts, sample, ref);
938 /* x86 retpoline 'return' doesn't match the stack */
939 if (rstate == X86_RETPOLINE_DETECTED && ts->cnt > 2 &&
940 ts->stack[ts->cnt - 1].ret_addr != sample->addr)
941 return thread_stack__x86_retpoline(ts, sample, to_al);
943 err = thread_stack__pop_cp(thread, ts, sample->addr,
944 sample->time, ref, from_al->sym);
948 err = thread_stack__no_call_return(thread, ts, sample,
949 from_al, to_al, ref);
951 } else if (sample->flags & PERF_IP_FLAG_TRACE_BEGIN) {
952 err = thread_stack__trace_begin(thread, ts, sample->time, ref);
953 } else if (sample->flags & PERF_IP_FLAG_TRACE_END) {
954 err = thread_stack__trace_end(ts, sample, ref);
955 } else if (sample->flags & PERF_IP_FLAG_BRANCH &&
956 from_al->sym != to_al->sym && to_al->sym &&
957 to_al->addr == to_al->sym->start) {
958 struct call_path_root *cpr = ts->crp->cpr;
959 struct call_path *cp;
962 * The compiler might optimize a call/ret combination by making
963 * it a jmp. Make that visible by recording on the stack a
964 * branch to the start of a different symbol. Note, that means
965 * when a ret pops the stack, all jmps must be popped off first.
967 cp = call_path__findnew(cpr, ts->stack[ts->cnt - 1].cp,
968 to_al->sym, sample->addr,
970 err = thread_stack__push_cp(ts, 0, sample->time, ref, cp, false,
973 ts->stack[ts->cnt - 1].non_call = true;
979 size_t thread_stack__depth(struct thread *thread, int cpu)
981 struct thread_stack *ts = thread__stack(thread, cpu);