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 #include "trace_helpers.h" 25 26 #define SAMPLE_FREQ 50 27 28 static bool sys_read_seen, sys_write_seen; 29 30 static void print_ksym(__u64 addr) 31 { 32 struct ksym *sym; 33 34 if (!addr) 35 return; 36 sym = ksym_search(addr); 37 if (!sym) { 38 printf("ksym not found. Is kallsyms loaded?\n"); 39 return; 40 } 41 42 printf("%s;", sym->name); 43 if (!strcmp(sym->name, "sys_read")) 44 sys_read_seen = true; 45 else if (!strcmp(sym->name, "sys_write")) 46 sys_write_seen = true; 47 } 48 49 static void print_addr(__u64 addr) 50 { 51 if (!addr) 52 return; 53 printf("%llx;", addr); 54 } 55 56 #define TASK_COMM_LEN 16 57 58 struct key_t { 59 char comm[TASK_COMM_LEN]; 60 __u32 kernstack; 61 __u32 userstack; 62 }; 63 64 static void print_stack(struct key_t *key, __u64 count) 65 { 66 __u64 ip[PERF_MAX_STACK_DEPTH] = {}; 67 static bool warned; 68 int i; 69 70 printf("%3lld %s;", count, key->comm); 71 if (bpf_map_lookup_elem(map_fd[1], &key->kernstack, ip) != 0) { 72 printf("---;"); 73 } else { 74 for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--) 75 print_ksym(ip[i]); 76 } 77 printf("-;"); 78 if (bpf_map_lookup_elem(map_fd[1], &key->userstack, ip) != 0) { 79 printf("---;"); 80 } else { 81 for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--) 82 print_addr(ip[i]); 83 } 84 if (count < 6) 85 printf("\r"); 86 else 87 printf("\n"); 88 89 if (key->kernstack == -EEXIST && !warned) { 90 printf("stackmap collisions seen. Consider increasing size\n"); 91 warned = true; 92 } else if ((int)key->kernstack < 0 && (int)key->userstack < 0) { 93 printf("err stackid %d %d\n", key->kernstack, key->userstack); 94 } 95 } 96 97 static void int_exit(int sig) 98 { 99 kill(0, SIGKILL); 100 exit(0); 101 } 102 103 static void print_stacks(void) 104 { 105 struct key_t key = {}, next_key; 106 __u64 value; 107 __u32 stackid = 0, next_id; 108 int fd = map_fd[0], stack_map = map_fd[1]; 109 110 sys_read_seen = sys_write_seen = false; 111 while (bpf_map_get_next_key(fd, &key, &next_key) == 0) { 112 bpf_map_lookup_elem(fd, &next_key, &value); 113 print_stack(&next_key, value); 114 bpf_map_delete_elem(fd, &next_key); 115 key = next_key; 116 } 117 printf("\n"); 118 if (!sys_read_seen || !sys_write_seen) { 119 printf("BUG kernel stack doesn't contain sys_read() and sys_write()\n"); 120 int_exit(0); 121 } 122 123 /* clear stack map */ 124 while (bpf_map_get_next_key(stack_map, &stackid, &next_id) == 0) { 125 bpf_map_delete_elem(stack_map, &next_id); 126 stackid = next_id; 127 } 128 } 129 130 static inline int generate_load(void) 131 { 132 if (system("dd if=/dev/zero of=/dev/null count=5000k status=none") < 0) { 133 printf("failed to generate some load with dd: %s\n", strerror(errno)); 134 return -1; 135 } 136 137 return 0; 138 } 139 140 static void test_perf_event_all_cpu(struct perf_event_attr *attr) 141 { 142 int nr_cpus = sysconf(_SC_NPROCESSORS_CONF); 143 int *pmu_fd = malloc(nr_cpus * sizeof(int)); 144 int i, error = 0; 145 146 /* system wide perf event, no need to inherit */ 147 attr->inherit = 0; 148 149 /* open perf_event on all cpus */ 150 for (i = 0; i < nr_cpus; i++) { 151 pmu_fd[i] = sys_perf_event_open(attr, -1, i, -1, 0); 152 if (pmu_fd[i] < 0) { 153 printf("sys_perf_event_open failed\n"); 154 error = 1; 155 goto all_cpu_err; 156 } 157 assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0); 158 assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_ENABLE) == 0); 159 } 160 161 if (generate_load() < 0) { 162 error = 1; 163 goto all_cpu_err; 164 } 165 print_stacks(); 166 all_cpu_err: 167 for (i--; i >= 0; i--) { 168 ioctl(pmu_fd[i], PERF_EVENT_IOC_DISABLE); 169 close(pmu_fd[i]); 170 } 171 free(pmu_fd); 172 if (error) 173 int_exit(0); 174 } 175 176 static void test_perf_event_task(struct perf_event_attr *attr) 177 { 178 int pmu_fd, error = 0; 179 180 /* per task perf event, enable inherit so the "dd ..." command can be traced properly. 181 * Enabling inherit will cause bpf_perf_prog_read_time helper failure. 182 */ 183 attr->inherit = 1; 184 185 /* open task bound event */ 186 pmu_fd = sys_perf_event_open(attr, 0, -1, -1, 0); 187 if (pmu_fd < 0) { 188 printf("sys_perf_event_open failed\n"); 189 int_exit(0); 190 } 191 assert(ioctl(pmu_fd, PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0); 192 assert(ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE) == 0); 193 194 if (generate_load() < 0) { 195 error = 1; 196 goto err; 197 } 198 print_stacks(); 199 err: 200 ioctl(pmu_fd, PERF_EVENT_IOC_DISABLE); 201 close(pmu_fd); 202 if (error) 203 int_exit(0); 204 } 205 206 static void test_bpf_perf_event(void) 207 { 208 struct perf_event_attr attr_type_hw = { 209 .sample_freq = SAMPLE_FREQ, 210 .freq = 1, 211 .type = PERF_TYPE_HARDWARE, 212 .config = PERF_COUNT_HW_CPU_CYCLES, 213 }; 214 struct perf_event_attr attr_type_sw = { 215 .sample_freq = SAMPLE_FREQ, 216 .freq = 1, 217 .type = PERF_TYPE_SOFTWARE, 218 .config = PERF_COUNT_SW_CPU_CLOCK, 219 }; 220 struct perf_event_attr attr_hw_cache_l1d = { 221 .sample_freq = SAMPLE_FREQ, 222 .freq = 1, 223 .type = PERF_TYPE_HW_CACHE, 224 .config = 225 PERF_COUNT_HW_CACHE_L1D | 226 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 227 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16), 228 }; 229 struct perf_event_attr attr_hw_cache_branch_miss = { 230 .sample_freq = SAMPLE_FREQ, 231 .freq = 1, 232 .type = PERF_TYPE_HW_CACHE, 233 .config = 234 PERF_COUNT_HW_CACHE_BPU | 235 (PERF_COUNT_HW_CACHE_OP_READ << 8) | 236 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16), 237 }; 238 struct perf_event_attr attr_type_raw = { 239 .sample_freq = SAMPLE_FREQ, 240 .freq = 1, 241 .type = PERF_TYPE_RAW, 242 /* Intel Instruction Retired */ 243 .config = 0xc0, 244 }; 245 struct perf_event_attr attr_type_raw_lock_load = { 246 .sample_freq = SAMPLE_FREQ, 247 .freq = 1, 248 .type = PERF_TYPE_RAW, 249 /* Intel MEM_UOPS_RETIRED.LOCK_LOADS */ 250 .config = 0x21d0, 251 /* Request to record lock address from PEBS */ 252 .sample_type = PERF_SAMPLE_ADDR, 253 /* Record address value requires precise event */ 254 .precise_ip = 2, 255 }; 256 257 printf("Test HW_CPU_CYCLES\n"); 258 test_perf_event_all_cpu(&attr_type_hw); 259 test_perf_event_task(&attr_type_hw); 260 261 printf("Test SW_CPU_CLOCK\n"); 262 test_perf_event_all_cpu(&attr_type_sw); 263 test_perf_event_task(&attr_type_sw); 264 265 printf("Test HW_CACHE_L1D\n"); 266 test_perf_event_all_cpu(&attr_hw_cache_l1d); 267 test_perf_event_task(&attr_hw_cache_l1d); 268 269 printf("Test HW_CACHE_BPU\n"); 270 test_perf_event_all_cpu(&attr_hw_cache_branch_miss); 271 test_perf_event_task(&attr_hw_cache_branch_miss); 272 273 printf("Test Instruction Retired\n"); 274 test_perf_event_all_cpu(&attr_type_raw); 275 test_perf_event_task(&attr_type_raw); 276 277 printf("Test Lock Load\n"); 278 test_perf_event_all_cpu(&attr_type_raw_lock_load); 279 test_perf_event_task(&attr_type_raw_lock_load); 280 281 printf("*** PASS ***\n"); 282 } 283 284 285 int main(int argc, char **argv) 286 { 287 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY}; 288 char filename[256]; 289 290 snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]); 291 setrlimit(RLIMIT_MEMLOCK, &r); 292 293 signal(SIGINT, int_exit); 294 signal(SIGTERM, int_exit); 295 296 if (load_kallsyms()) { 297 printf("failed to process /proc/kallsyms\n"); 298 return 1; 299 } 300 301 if (load_bpf_file(filename)) { 302 printf("%s", bpf_log_buf); 303 return 2; 304 } 305 306 if (fork() == 0) { 307 read_trace_pipe(); 308 return 0; 309 } 310 test_bpf_perf_event(); 311 int_exit(0); 312 return 0; 313 } 314