xref: /linux/tools/perf/tests/dwarf-unwind.c (revision 4a8033ec560c9fedaaddaf9c8a3e4b6244d8d61a)
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 
40 static 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 
48 static 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  */
65 int test_dwarf_unwind__thread(struct thread *thread);
66 int test_dwarf_unwind__compare(void *p1, void *p2);
67 int test_dwarf_unwind__krava_3(struct thread *thread);
68 int test_dwarf_unwind__krava_2(struct thread *thread);
69 int test_dwarf_unwind__krava_1(struct thread *thread);
70 
71 #define MAX_STACK 8
72 
73 static 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 
111 NO_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);
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 
140 static int global_unwind_retval = -INT_MAX;
141 
142 NO_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 
161 NO_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 
180 NO_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 
189 NO_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 
198 int test__dwarf_unwind(struct test *test __maybe_unused, int subtest __maybe_unused)
199 {
200 	struct machine *machine;
201 	struct thread *thread;
202 	int err = -1;
203 
204 	machine = machine__new_host();
205 	if (!machine) {
206 		pr_err("Could not get machine\n");
207 		return -1;
208 	}
209 
210 	if (machine__create_kernel_maps(machine)) {
211 		pr_err("Failed to create kernel maps\n");
212 		return -1;
213 	}
214 
215 	callchain_param.record_mode = CALLCHAIN_DWARF;
216 	dwarf_callchain_users = true;
217 
218 	if (init_live_machine(machine)) {
219 		pr_err("Could not init machine\n");
220 		goto out;
221 	}
222 
223 	if (verbose > 1)
224 		machine__fprintf(machine, stderr);
225 
226 	thread = machine__find_thread(machine, getpid(), getpid());
227 	if (!thread) {
228 		pr_err("Could not get thread\n");
229 		goto out;
230 	}
231 
232 	err = test_dwarf_unwind__krava_1(thread);
233 	thread__put(thread);
234 
235  out:
236 	machine__delete_threads(machine);
237 	machine__delete(machine);
238 	return err;
239 }
240