1 /* Copyright (c) 2016 Facebook
2 *
3 * This program is free software; you can redistribute it and/or
4 * modify it under the terms of version 2 of the GNU General Public
5 * License as published by the Free Software Foundation.
6 */
7 #include <stdio.h>
8 #include <unistd.h>
9 #include <stdlib.h>
10 #include <stdbool.h>
11 #include <string.h>
12 #include <fcntl.h>
13 #include <poll.h>
14 #include <sys/ioctl.h>
15 #include <linux/perf_event.h>
16 #include <linux/bpf.h>
17 #include <signal.h>
18 #include <assert.h>
19 #include <errno.h>
20 #include <sys/resource.h>
21 #include "libbpf.h"
22 #include "bpf_load.h"
23 #include "perf-sys.h"
24
25 #define SAMPLE_FREQ 50
26
27 static bool sys_read_seen, sys_write_seen;
28
print_ksym(__u64 addr)29 static void print_ksym(__u64 addr)
30 {
31 struct ksym *sym;
32
33 if (!addr)
34 return;
35 sym = ksym_search(addr);
36 printf("%s;", sym->name);
37 if (!strstr(sym->name, "sys_read"))
38 sys_read_seen = true;
39 else if (!strstr(sym->name, "sys_write"))
40 sys_write_seen = true;
41 }
42
print_addr(__u64 addr)43 static void print_addr(__u64 addr)
44 {
45 if (!addr)
46 return;
47 printf("%llx;", addr);
48 }
49
50 #define TASK_COMM_LEN 16
51
52 struct key_t {
53 char comm[TASK_COMM_LEN];
54 __u32 kernstack;
55 __u32 userstack;
56 };
57
print_stack(struct key_t * key,__u64 count)58 static void print_stack(struct key_t *key, __u64 count)
59 {
60 __u64 ip[PERF_MAX_STACK_DEPTH] = {};
61 static bool warned;
62 int i;
63
64 printf("%3lld %s;", count, key->comm);
65 if (bpf_map_lookup_elem(map_fd[1], &key->kernstack, ip) != 0) {
66 printf("---;");
67 } else {
68 for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
69 print_ksym(ip[i]);
70 }
71 printf("-;");
72 if (bpf_map_lookup_elem(map_fd[1], &key->userstack, ip) != 0) {
73 printf("---;");
74 } else {
75 for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
76 print_addr(ip[i]);
77 }
78 if (count < 6)
79 printf("\r");
80 else
81 printf("\n");
82
83 if (key->kernstack == -EEXIST && !warned) {
84 printf("stackmap collisions seen. Consider increasing size\n");
85 warned = true;
86 } else if ((int)key->kernstack < 0 && (int)key->userstack < 0) {
87 printf("err stackid %d %d\n", key->kernstack, key->userstack);
88 }
89 }
90
int_exit(int sig)91 static void int_exit(int sig)
92 {
93 kill(0, SIGKILL);
94 exit(0);
95 }
96
print_stacks(void)97 static void print_stacks(void)
98 {
99 struct key_t key = {}, next_key;
100 __u64 value;
101 __u32 stackid = 0, next_id;
102 int fd = map_fd[0], stack_map = map_fd[1];
103
104 sys_read_seen = sys_write_seen = false;
105 while (bpf_map_get_next_key(fd, &key, &next_key) == 0) {
106 bpf_map_lookup_elem(fd, &next_key, &value);
107 print_stack(&next_key, value);
108 bpf_map_delete_elem(fd, &next_key);
109 key = next_key;
110 }
111 printf("\n");
112 if (!sys_read_seen || !sys_write_seen) {
113 printf("BUG kernel stack doesn't contain sys_read() and sys_write()\n");
114 int_exit(0);
115 }
116
117 /* clear stack map */
118 while (bpf_map_get_next_key(stack_map, &stackid, &next_id) == 0) {
119 bpf_map_delete_elem(stack_map, &next_id);
120 stackid = next_id;
121 }
122 }
123
generate_load(void)124 static inline int generate_load(void)
125 {
126 if (system("dd if=/dev/zero of=/dev/null count=5000k status=none") < 0) {
127 printf("failed to generate some load with dd: %s\n", strerror(errno));
128 return -1;
129 }
130
131 return 0;
132 }
133
test_perf_event_all_cpu(struct perf_event_attr * attr)134 static void test_perf_event_all_cpu(struct perf_event_attr *attr)
135 {
136 int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
137 int *pmu_fd = malloc(nr_cpus * sizeof(int));
138 int i, error = 0;
139
140 /* open perf_event on all cpus */
141 for (i = 0; i < nr_cpus; i++) {
142 pmu_fd[i] = sys_perf_event_open(attr, -1, i, -1, 0);
143 if (pmu_fd[i] < 0) {
144 printf("sys_perf_event_open failed\n");
145 error = 1;
146 goto all_cpu_err;
147 }
148 assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
149 assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_ENABLE) == 0);
150 }
151
152 if (generate_load() < 0) {
153 error = 1;
154 goto all_cpu_err;
155 }
156 print_stacks();
157 all_cpu_err:
158 for (i--; i >= 0; i--) {
159 ioctl(pmu_fd[i], PERF_EVENT_IOC_DISABLE);
160 close(pmu_fd[i]);
161 }
162 free(pmu_fd);
163 if (error)
164 int_exit(0);
165 }
166
test_perf_event_task(struct perf_event_attr * attr)167 static void test_perf_event_task(struct perf_event_attr *attr)
168 {
169 int pmu_fd, error = 0;
170
171 /* open task bound event */
172 pmu_fd = sys_perf_event_open(attr, 0, -1, -1, 0);
173 if (pmu_fd < 0) {
174 printf("sys_perf_event_open failed\n");
175 int_exit(0);
176 }
177 assert(ioctl(pmu_fd, PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
178 assert(ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE) == 0);
179
180 if (generate_load() < 0) {
181 error = 1;
182 goto err;
183 }
184 print_stacks();
185 err:
186 ioctl(pmu_fd, PERF_EVENT_IOC_DISABLE);
187 close(pmu_fd);
188 if (error)
189 int_exit(0);
190 }
191
test_bpf_perf_event(void)192 static void test_bpf_perf_event(void)
193 {
194 struct perf_event_attr attr_type_hw = {
195 .sample_freq = SAMPLE_FREQ,
196 .freq = 1,
197 .type = PERF_TYPE_HARDWARE,
198 .config = PERF_COUNT_HW_CPU_CYCLES,
199 .inherit = 1,
200 };
201 struct perf_event_attr attr_type_sw = {
202 .sample_freq = SAMPLE_FREQ,
203 .freq = 1,
204 .type = PERF_TYPE_SOFTWARE,
205 .config = PERF_COUNT_SW_CPU_CLOCK,
206 .inherit = 1,
207 };
208 struct perf_event_attr attr_hw_cache_l1d = {
209 .sample_freq = SAMPLE_FREQ,
210 .freq = 1,
211 .type = PERF_TYPE_HW_CACHE,
212 .config =
213 PERF_COUNT_HW_CACHE_L1D |
214 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
215 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16),
216 .inherit = 1,
217 };
218 struct perf_event_attr attr_hw_cache_branch_miss = {
219 .sample_freq = SAMPLE_FREQ,
220 .freq = 1,
221 .type = PERF_TYPE_HW_CACHE,
222 .config =
223 PERF_COUNT_HW_CACHE_BPU |
224 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
225 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16),
226 .inherit = 1,
227 };
228 struct perf_event_attr attr_type_raw = {
229 .sample_freq = SAMPLE_FREQ,
230 .freq = 1,
231 .type = PERF_TYPE_RAW,
232 /* Intel Instruction Retired */
233 .config = 0xc0,
234 .inherit = 1,
235 };
236
237 printf("Test HW_CPU_CYCLES\n");
238 test_perf_event_all_cpu(&attr_type_hw);
239 test_perf_event_task(&attr_type_hw);
240
241 printf("Test SW_CPU_CLOCK\n");
242 test_perf_event_all_cpu(&attr_type_sw);
243 test_perf_event_task(&attr_type_sw);
244
245 printf("Test HW_CACHE_L1D\n");
246 test_perf_event_all_cpu(&attr_hw_cache_l1d);
247 test_perf_event_task(&attr_hw_cache_l1d);
248
249 printf("Test HW_CACHE_BPU\n");
250 test_perf_event_all_cpu(&attr_hw_cache_branch_miss);
251 test_perf_event_task(&attr_hw_cache_branch_miss);
252
253 printf("Test Instruction Retired\n");
254 test_perf_event_all_cpu(&attr_type_raw);
255 test_perf_event_task(&attr_type_raw);
256
257 printf("*** PASS ***\n");
258 }
259
260
main(int argc,char ** argv)261 int main(int argc, char **argv)
262 {
263 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
264 char filename[256];
265
266 snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
267 setrlimit(RLIMIT_MEMLOCK, &r);
268
269 signal(SIGINT, int_exit);
270 signal(SIGTERM, int_exit);
271
272 if (load_kallsyms()) {
273 printf("failed to process /proc/kallsyms\n");
274 return 1;
275 }
276
277 if (load_bpf_file(filename)) {
278 printf("%s", bpf_log_buf);
279 return 2;
280 }
281
282 if (fork() == 0) {
283 read_trace_pipe();
284 return 0;
285 }
286 test_bpf_perf_event();
287 int_exit(0);
288 return 0;
289 }
290