1 /* Copyright (c) 2016 Facebook
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.
14 #include <sys/ioctl.h>
15 #include <linux/perf_event.h>
16 #include <linux/bpf.h>
20 #include <sys/resource.h>
25 #define SAMPLE_FREQ 50
27 static bool sys_read_seen, sys_write_seen;
29 static void print_ksym(__u64 addr)
35 sym = ksym_search(addr);
36 printf("%s;", sym->name);
37 if (!strstr(sym->name, "sys_read"))
39 else if (!strstr(sym->name, "sys_write"))
40 sys_write_seen = true;
43 static void print_addr(__u64 addr)
47 printf("%llx;", addr);
50 #define TASK_COMM_LEN 16
53 char comm[TASK_COMM_LEN];
58 static void print_stack(struct key_t *key, __u64 count)
60 __u64 ip[PERF_MAX_STACK_DEPTH] = {};
64 printf("%3lld %s;", count, key->comm);
65 if (bpf_map_lookup_elem(map_fd[1], &key->kernstack, ip) != 0) {
68 for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
72 if (bpf_map_lookup_elem(map_fd[1], &key->userstack, ip) != 0) {
75 for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
83 if (key->kernstack == -EEXIST && !warned) {
84 printf("stackmap collisions seen. Consider increasing size\n");
86 } else if ((int)key->kernstack < 0 && (int)key->userstack < 0) {
87 printf("err stackid %d %d\n", key->kernstack, key->userstack);
91 static void int_exit(int sig)
97 static void print_stacks(void)
99 struct key_t key = {}, next_key;
101 __u32 stackid = 0, next_id;
102 int fd = map_fd[0], stack_map = map_fd[1];
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);
112 if (!sys_read_seen || !sys_write_seen) {
113 printf("BUG kernel stack doesn't contain sys_read() and sys_write()\n");
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);
124 static inline int generate_load(void)
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));
134 static void test_perf_event_all_cpu(struct perf_event_attr *attr)
136 int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
137 int *pmu_fd = malloc(nr_cpus * sizeof(int));
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);
144 printf("sys_perf_event_open failed\n");
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);
152 if (generate_load() < 0) {
158 for (i--; i >= 0; i--) {
159 ioctl(pmu_fd[i], PERF_EVENT_IOC_DISABLE);
167 static void test_perf_event_task(struct perf_event_attr *attr)
169 int pmu_fd, error = 0;
171 /* open task bound event */
172 pmu_fd = sys_perf_event_open(attr, 0, -1, -1, 0);
174 printf("sys_perf_event_open failed\n");
177 assert(ioctl(pmu_fd, PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
178 assert(ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE) == 0);
180 if (generate_load() < 0) {
186 ioctl(pmu_fd, PERF_EVENT_IOC_DISABLE);
192 static void test_bpf_perf_event(void)
194 struct perf_event_attr attr_type_hw = {
195 .sample_freq = SAMPLE_FREQ,
197 .type = PERF_TYPE_HARDWARE,
198 .config = PERF_COUNT_HW_CPU_CYCLES,
201 struct perf_event_attr attr_type_sw = {
202 .sample_freq = SAMPLE_FREQ,
204 .type = PERF_TYPE_SOFTWARE,
205 .config = PERF_COUNT_SW_CPU_CLOCK,
208 struct perf_event_attr attr_hw_cache_l1d = {
209 .sample_freq = SAMPLE_FREQ,
211 .type = PERF_TYPE_HW_CACHE,
213 PERF_COUNT_HW_CACHE_L1D |
214 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
215 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16),
218 struct perf_event_attr attr_hw_cache_branch_miss = {
219 .sample_freq = SAMPLE_FREQ,
221 .type = PERF_TYPE_HW_CACHE,
223 PERF_COUNT_HW_CACHE_BPU |
224 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
225 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16),
228 struct perf_event_attr attr_type_raw = {
229 .sample_freq = SAMPLE_FREQ,
231 .type = PERF_TYPE_RAW,
232 /* Intel Instruction Retired */
237 printf("Test HW_CPU_CYCLES\n");
238 test_perf_event_all_cpu(&attr_type_hw);
239 test_perf_event_task(&attr_type_hw);
241 printf("Test SW_CPU_CLOCK\n");
242 test_perf_event_all_cpu(&attr_type_sw);
243 test_perf_event_task(&attr_type_sw);
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);
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);
253 printf("Test Instruction Retired\n");
254 test_perf_event_all_cpu(&attr_type_raw);
255 test_perf_event_task(&attr_type_raw);
257 printf("*** PASS ***\n");
261 int main(int argc, char **argv)
263 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
266 snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
267 setrlimit(RLIMIT_MEMLOCK, &r);
269 signal(SIGINT, int_exit);
270 signal(SIGTERM, int_exit);
272 if (load_kallsyms()) {
273 printf("failed to process /proc/kallsyms\n");
277 if (load_bpf_file(filename)) {
278 printf("%s", bpf_log_buf);
286 test_bpf_perf_event();