Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include <linux/compiler.h>
3#include <linux/types.h>
4#include <linux/zalloc.h>
5#include <inttypes.h>
6#include <limits.h>
7#include <unistd.h>
8#include "tests.h"
9#include "debug.h"
10#include "machine.h"
11#include "event.h"
12#include "../util/unwind.h"
13#include "perf_regs.h"
14#include "map.h"
15#include "symbol.h"
16#include "thread.h"
17#include "callchain.h"
18#include "util/synthetic-events.h"
19
20/* For bsearch. We try to unwind functions in shared object. */
21#include <stdlib.h>
22
23/*
24 * The test will assert frames are on the stack but tail call optimizations lose
25 * the frame of the caller. Clang can disable this optimization on a called
26 * function but GCC currently (11/2020) lacks this attribute. The barrier is
27 * used to inhibit tail calls in these cases.
28 */
29#ifdef __has_attribute
30#if __has_attribute(disable_tail_calls)
31#define NO_TAIL_CALL_ATTRIBUTE __attribute__((disable_tail_calls))
32#define NO_TAIL_CALL_BARRIER
33#endif
34#endif
35#ifndef NO_TAIL_CALL_ATTRIBUTE
36#define NO_TAIL_CALL_ATTRIBUTE
37#define NO_TAIL_CALL_BARRIER __asm__ __volatile__("" : : : "memory");
38#endif
39
40static int mmap_handler(struct perf_tool *tool __maybe_unused,
41 union perf_event *event,
42 struct perf_sample *sample,
43 struct machine *machine)
44{
45 return machine__process_mmap2_event(machine, event, sample);
46}
47
48static int init_live_machine(struct machine *machine)
49{
50 union perf_event event;
51 pid_t pid = getpid();
52
53 memset(&event, 0, sizeof(event));
54 return perf_event__synthesize_mmap_events(NULL, &event, pid, pid,
55 mmap_handler, machine, true);
56}
57
58/*
59 * We need to keep these functions global, despite the
60 * fact that they are used only locally in this object,
61 * in order to keep them around even if the binary is
62 * stripped. If they are gone, the unwind check for
63 * symbol fails.
64 */
65int test_dwarf_unwind__thread(struct thread *thread);
66int test_dwarf_unwind__compare(void *p1, void *p2);
67int test_dwarf_unwind__krava_3(struct thread *thread);
68int test_dwarf_unwind__krava_2(struct thread *thread);
69int test_dwarf_unwind__krava_1(struct thread *thread);
70
71#define MAX_STACK 8
72
73static int unwind_entry(struct unwind_entry *entry, void *arg)
74{
75 unsigned long *cnt = (unsigned long *) arg;
76 char *symbol = entry->ms.sym ? entry->ms.sym->name : NULL;
77 static const char *funcs[MAX_STACK] = {
78 "test__arch_unwind_sample",
79 "test_dwarf_unwind__thread",
80 "test_dwarf_unwind__compare",
81 "bsearch",
82 "test_dwarf_unwind__krava_3",
83 "test_dwarf_unwind__krava_2",
84 "test_dwarf_unwind__krava_1",
85 "test__dwarf_unwind"
86 };
87 /*
88 * The funcs[MAX_STACK] array index, based on the
89 * callchain order setup.
90 */
91 int idx = callchain_param.order == ORDER_CALLER ?
92 MAX_STACK - *cnt - 1 : *cnt;
93
94 if (*cnt >= MAX_STACK) {
95 pr_debug("failed: crossed the max stack value %d\n", MAX_STACK);
96 return -1;
97 }
98
99 if (!symbol) {
100 pr_debug("failed: got unresolved address 0x%" PRIx64 "\n",
101 entry->ip);
102 return -1;
103 }
104
105 (*cnt)++;
106 pr_debug("got: %s 0x%" PRIx64 ", expecting %s\n",
107 symbol, entry->ip, funcs[idx]);
108 return strcmp((const char *) symbol, funcs[idx]);
109}
110
111NO_TAIL_CALL_ATTRIBUTE noinline int test_dwarf_unwind__thread(struct thread *thread)
112{
113 struct perf_sample sample;
114 unsigned long cnt = 0;
115 int err = -1;
116
117 memset(&sample, 0, sizeof(sample));
118
119 if (test__arch_unwind_sample(&sample, thread)) {
120 pr_debug("failed to get unwind sample\n");
121 goto out;
122 }
123
124 err = unwind__get_entries(unwind_entry, &cnt, thread,
125 &sample, MAX_STACK, false);
126 if (err)
127 pr_debug("unwind failed\n");
128 else if (cnt != MAX_STACK) {
129 pr_debug("got wrong number of stack entries %lu != %d\n",
130 cnt, MAX_STACK);
131 err = -1;
132 }
133
134 out:
135 zfree(&sample.user_stack.data);
136 zfree(&sample.user_regs.regs);
137 return err;
138}
139
140static int global_unwind_retval = -INT_MAX;
141
142NO_TAIL_CALL_ATTRIBUTE noinline int test_dwarf_unwind__compare(void *p1, void *p2)
143{
144 /* Any possible value should be 'thread' */
145 struct thread *thread = *(struct thread **)p1;
146
147 if (global_unwind_retval == -INT_MAX) {
148 /* Call unwinder twice for both callchain orders. */
149 callchain_param.order = ORDER_CALLER;
150
151 global_unwind_retval = test_dwarf_unwind__thread(thread);
152 if (!global_unwind_retval) {
153 callchain_param.order = ORDER_CALLEE;
154 global_unwind_retval = test_dwarf_unwind__thread(thread);
155 }
156 }
157
158 return p1 - p2;
159}
160
161NO_TAIL_CALL_ATTRIBUTE noinline int test_dwarf_unwind__krava_3(struct thread *thread)
162{
163 struct thread *array[2] = {thread, thread};
164 void *fp = &bsearch;
165 /*
166 * make _bsearch a volatile function pointer to
167 * prevent potential optimization, which may expand
168 * bsearch and call compare directly from this function,
169 * instead of libc shared object.
170 */
171 void *(*volatile _bsearch)(void *, void *, size_t,
172 size_t, int (*)(void *, void *));
173
174 _bsearch = fp;
175 _bsearch(array, &thread, 2, sizeof(struct thread **),
176 test_dwarf_unwind__compare);
177 return global_unwind_retval;
178}
179
180NO_TAIL_CALL_ATTRIBUTE noinline int test_dwarf_unwind__krava_2(struct thread *thread)
181{
182 int ret;
183
184 ret = test_dwarf_unwind__krava_3(thread);
185 NO_TAIL_CALL_BARRIER;
186 return ret;
187}
188
189NO_TAIL_CALL_ATTRIBUTE noinline int test_dwarf_unwind__krava_1(struct thread *thread)
190{
191 int ret;
192
193 ret = test_dwarf_unwind__krava_2(thread);
194 NO_TAIL_CALL_BARRIER;
195 return ret;
196}
197
198static int test__dwarf_unwind(struct test_suite *test __maybe_unused,
199 int subtest __maybe_unused)
200{
201 struct machine *machine;
202 struct thread *thread;
203 int err = -1;
204
205 machine = machine__new_host();
206 if (!machine) {
207 pr_err("Could not get machine\n");
208 return -1;
209 }
210
211 if (machine__create_kernel_maps(machine)) {
212 pr_err("Failed to create kernel maps\n");
213 return -1;
214 }
215
216 callchain_param.record_mode = CALLCHAIN_DWARF;
217 dwarf_callchain_users = true;
218
219 if (init_live_machine(machine)) {
220 pr_err("Could not init machine\n");
221 goto out;
222 }
223
224 if (verbose > 1)
225 machine__fprintf(machine, stderr);
226
227 thread = machine__find_thread(machine, getpid(), getpid());
228 if (!thread) {
229 pr_err("Could not get thread\n");
230 goto out;
231 }
232
233 err = test_dwarf_unwind__krava_1(thread);
234 thread__put(thread);
235
236 out:
237 machine__delete_threads(machine);
238 machine__delete(machine);
239 return err;
240}
241
242DEFINE_SUITE("Test dwarf unwind", dwarf_unwind);
1#include <linux/compiler.h>
2#include <linux/types.h>
3#include <unistd.h>
4#include "tests.h"
5#include "debug.h"
6#include "machine.h"
7#include "event.h"
8#include "unwind.h"
9#include "perf_regs.h"
10#include "map.h"
11#include "thread.h"
12#include "callchain.h"
13
14#if defined (__x86_64__) || defined (__i386__)
15#include "arch-tests.h"
16#endif
17
18/* For bsearch. We try to unwind functions in shared object. */
19#include <stdlib.h>
20
21static int mmap_handler(struct perf_tool *tool __maybe_unused,
22 union perf_event *event,
23 struct perf_sample *sample,
24 struct machine *machine)
25{
26 return machine__process_mmap2_event(machine, event, sample);
27}
28
29static int init_live_machine(struct machine *machine)
30{
31 union perf_event event;
32 pid_t pid = getpid();
33
34 return perf_event__synthesize_mmap_events(NULL, &event, pid, pid,
35 mmap_handler, machine, true, 500);
36}
37
38#define MAX_STACK 8
39
40static int unwind_entry(struct unwind_entry *entry, void *arg)
41{
42 unsigned long *cnt = (unsigned long *) arg;
43 char *symbol = entry->sym ? entry->sym->name : NULL;
44 static const char *funcs[MAX_STACK] = {
45 "test__arch_unwind_sample",
46 "unwind_thread",
47 "compare",
48 "bsearch",
49 "krava_3",
50 "krava_2",
51 "krava_1",
52 "test__dwarf_unwind"
53 };
54 /*
55 * The funcs[MAX_STACK] array index, based on the
56 * callchain order setup.
57 */
58 int idx = callchain_param.order == ORDER_CALLER ?
59 MAX_STACK - *cnt - 1 : *cnt;
60
61 if (*cnt >= MAX_STACK) {
62 pr_debug("failed: crossed the max stack value %d\n", MAX_STACK);
63 return -1;
64 }
65
66 if (!symbol) {
67 pr_debug("failed: got unresolved address 0x%" PRIx64 "\n",
68 entry->ip);
69 return -1;
70 }
71
72 (*cnt)++;
73 pr_debug("got: %s 0x%" PRIx64 ", expecting %s\n",
74 symbol, entry->ip, funcs[idx]);
75 return strcmp((const char *) symbol, funcs[idx]);
76}
77
78__attribute__ ((noinline))
79static int unwind_thread(struct thread *thread)
80{
81 struct perf_sample sample;
82 unsigned long cnt = 0;
83 int err = -1;
84
85 memset(&sample, 0, sizeof(sample));
86
87 if (test__arch_unwind_sample(&sample, thread)) {
88 pr_debug("failed to get unwind sample\n");
89 goto out;
90 }
91
92 err = unwind__get_entries(unwind_entry, &cnt, thread,
93 &sample, MAX_STACK);
94 if (err)
95 pr_debug("unwind failed\n");
96 else if (cnt != MAX_STACK) {
97 pr_debug("got wrong number of stack entries %lu != %d\n",
98 cnt, MAX_STACK);
99 err = -1;
100 }
101
102 out:
103 free(sample.user_stack.data);
104 free(sample.user_regs.regs);
105 return err;
106}
107
108static int global_unwind_retval = -INT_MAX;
109
110__attribute__ ((noinline))
111static int compare(void *p1, void *p2)
112{
113 /* Any possible value should be 'thread' */
114 struct thread *thread = *(struct thread **)p1;
115
116 if (global_unwind_retval == -INT_MAX) {
117 /* Call unwinder twice for both callchain orders. */
118 callchain_param.order = ORDER_CALLER;
119
120 global_unwind_retval = unwind_thread(thread);
121 if (!global_unwind_retval) {
122 callchain_param.order = ORDER_CALLEE;
123 global_unwind_retval = unwind_thread(thread);
124 }
125 }
126
127 return p1 - p2;
128}
129
130__attribute__ ((noinline))
131static int krava_3(struct thread *thread)
132{
133 struct thread *array[2] = {thread, thread};
134 void *fp = &bsearch;
135 /*
136 * make _bsearch a volatile function pointer to
137 * prevent potential optimization, which may expand
138 * bsearch and call compare directly from this function,
139 * instead of libc shared object.
140 */
141 void *(*volatile _bsearch)(void *, void *, size_t,
142 size_t, int (*)(void *, void *));
143
144 _bsearch = fp;
145 _bsearch(array, &thread, 2, sizeof(struct thread **), compare);
146 return global_unwind_retval;
147}
148
149__attribute__ ((noinline))
150static int krava_2(struct thread *thread)
151{
152 return krava_3(thread);
153}
154
155__attribute__ ((noinline))
156static int krava_1(struct thread *thread)
157{
158 return krava_2(thread);
159}
160
161int test__dwarf_unwind(int subtest __maybe_unused)
162{
163 struct machine *machine;
164 struct thread *thread;
165 int err = -1;
166
167 machine = machine__new_host();
168 if (!machine) {
169 pr_err("Could not get machine\n");
170 return -1;
171 }
172
173 if (machine__create_kernel_maps(machine)) {
174 pr_err("Failed to create kernel maps\n");
175 return -1;
176 }
177
178 callchain_param.record_mode = CALLCHAIN_DWARF;
179
180 if (init_live_machine(machine)) {
181 pr_err("Could not init machine\n");
182 goto out;
183 }
184
185 if (verbose > 1)
186 machine__fprintf(machine, stderr);
187
188 thread = machine__find_thread(machine, getpid(), getpid());
189 if (!thread) {
190 pr_err("Could not get thread\n");
191 goto out;
192 }
193
194 err = krava_1(thread);
195 thread__put(thread);
196
197 out:
198 machine__delete_threads(machine);
199 machine__delete(machine);
200 return err;
201}