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 #define _GNU_SOURCE 8 #include <sched.h> 9 #include <stdio.h> 10 #include <sys/types.h> 11 #include <asm/unistd.h> 12 #include <unistd.h> 13 #include <assert.h> 14 #include <sys/wait.h> 15 #include <stdlib.h> 16 #include <signal.h> 17 #include <linux/bpf.h> 18 #include <string.h> 19 #include <time.h> 20 #include <sys/resource.h> 21 #include <arpa/inet.h> 22 #include <errno.h> 23 24 #include <bpf/bpf.h> 25 #include "bpf_load.h" 26 27 #define TEST_BIT(t) (1U << (t)) 28 #define MAX_NR_CPUS 1024 29 30 static __u64 time_get_ns(void) 31 { 32 struct timespec ts; 33 34 clock_gettime(CLOCK_MONOTONIC, &ts); 35 return ts.tv_sec * 1000000000ull + ts.tv_nsec; 36 } 37 38 enum test_type { 39 HASH_PREALLOC, 40 PERCPU_HASH_PREALLOC, 41 HASH_KMALLOC, 42 PERCPU_HASH_KMALLOC, 43 LRU_HASH_PREALLOC, 44 NOCOMMON_LRU_HASH_PREALLOC, 45 LPM_KMALLOC, 46 HASH_LOOKUP, 47 ARRAY_LOOKUP, 48 INNER_LRU_HASH_PREALLOC, 49 LRU_HASH_LOOKUP, 50 NR_TESTS, 51 }; 52 53 const char *test_map_names[NR_TESTS] = { 54 [HASH_PREALLOC] = "hash_map", 55 [PERCPU_HASH_PREALLOC] = "percpu_hash_map", 56 [HASH_KMALLOC] = "hash_map_alloc", 57 [PERCPU_HASH_KMALLOC] = "percpu_hash_map_alloc", 58 [LRU_HASH_PREALLOC] = "lru_hash_map", 59 [NOCOMMON_LRU_HASH_PREALLOC] = "nocommon_lru_hash_map", 60 [LPM_KMALLOC] = "lpm_trie_map_alloc", 61 [HASH_LOOKUP] = "hash_map", 62 [ARRAY_LOOKUP] = "array_map", 63 [INNER_LRU_HASH_PREALLOC] = "inner_lru_hash_map", 64 [LRU_HASH_LOOKUP] = "lru_hash_lookup_map", 65 }; 66 67 static int test_flags = ~0; 68 static uint32_t num_map_entries; 69 static uint32_t inner_lru_hash_size; 70 static int inner_lru_hash_idx = -1; 71 static int array_of_lru_hashs_idx = -1; 72 static int lru_hash_lookup_idx = -1; 73 static int lru_hash_lookup_test_entries = 32; 74 static uint32_t max_cnt = 1000000; 75 76 static int check_test_flags(enum test_type t) 77 { 78 return test_flags & TEST_BIT(t); 79 } 80 81 static void test_hash_prealloc(int cpu) 82 { 83 __u64 start_time; 84 int i; 85 86 start_time = time_get_ns(); 87 for (i = 0; i < max_cnt; i++) 88 syscall(__NR_getuid); 89 printf("%d:hash_map_perf pre-alloc %lld events per sec\n", 90 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 91 } 92 93 static int pre_test_lru_hash_lookup(int tasks) 94 { 95 int fd = map_fd[lru_hash_lookup_idx]; 96 uint32_t key; 97 long val = 1; 98 int ret; 99 100 if (num_map_entries > lru_hash_lookup_test_entries) 101 lru_hash_lookup_test_entries = num_map_entries; 102 103 /* Populate the lru_hash_map for LRU_HASH_LOOKUP perf test. 104 * 105 * It is fine that the user requests for a map with 106 * num_map_entries < 32 and some of the later lru hash lookup 107 * may return not found. For LRU map, we are not interested 108 * in such small map performance. 109 */ 110 for (key = 0; key < lru_hash_lookup_test_entries; key++) { 111 ret = bpf_map_update_elem(fd, &key, &val, BPF_NOEXIST); 112 if (ret) 113 return ret; 114 } 115 116 return 0; 117 } 118 119 static void do_test_lru(enum test_type test, int cpu) 120 { 121 static int inner_lru_map_fds[MAX_NR_CPUS]; 122 123 struct sockaddr_in6 in6 = { .sin6_family = AF_INET6 }; 124 const char *test_name; 125 __u64 start_time; 126 int i, ret; 127 128 if (test == INNER_LRU_HASH_PREALLOC) { 129 int outer_fd = map_fd[array_of_lru_hashs_idx]; 130 unsigned int mycpu, mynode; 131 132 assert(cpu < MAX_NR_CPUS); 133 134 if (cpu) { 135 ret = syscall(__NR_getcpu, &mycpu, &mynode, NULL); 136 assert(!ret); 137 138 inner_lru_map_fds[cpu] = 139 bpf_create_map_node(BPF_MAP_TYPE_LRU_HASH, 140 test_map_names[INNER_LRU_HASH_PREALLOC], 141 sizeof(uint32_t), 142 sizeof(long), 143 inner_lru_hash_size, 0, 144 mynode); 145 if (inner_lru_map_fds[cpu] == -1) { 146 printf("cannot create BPF_MAP_TYPE_LRU_HASH %s(%d)\n", 147 strerror(errno), errno); 148 exit(1); 149 } 150 } else { 151 inner_lru_map_fds[cpu] = map_fd[inner_lru_hash_idx]; 152 } 153 154 ret = bpf_map_update_elem(outer_fd, &cpu, 155 &inner_lru_map_fds[cpu], 156 BPF_ANY); 157 if (ret) { 158 printf("cannot update ARRAY_OF_LRU_HASHS with key:%u. %s(%d)\n", 159 cpu, strerror(errno), errno); 160 exit(1); 161 } 162 } 163 164 in6.sin6_addr.s6_addr16[0] = 0xdead; 165 in6.sin6_addr.s6_addr16[1] = 0xbeef; 166 167 if (test == LRU_HASH_PREALLOC) { 168 test_name = "lru_hash_map_perf"; 169 in6.sin6_addr.s6_addr16[2] = 0; 170 } else if (test == NOCOMMON_LRU_HASH_PREALLOC) { 171 test_name = "nocommon_lru_hash_map_perf"; 172 in6.sin6_addr.s6_addr16[2] = 1; 173 } else if (test == INNER_LRU_HASH_PREALLOC) { 174 test_name = "inner_lru_hash_map_perf"; 175 in6.sin6_addr.s6_addr16[2] = 2; 176 } else if (test == LRU_HASH_LOOKUP) { 177 test_name = "lru_hash_lookup_perf"; 178 in6.sin6_addr.s6_addr16[2] = 3; 179 in6.sin6_addr.s6_addr32[3] = 0; 180 } else { 181 assert(0); 182 } 183 184 start_time = time_get_ns(); 185 for (i = 0; i < max_cnt; i++) { 186 ret = connect(-1, (const struct sockaddr *)&in6, sizeof(in6)); 187 assert(ret == -1 && errno == EBADF); 188 if (in6.sin6_addr.s6_addr32[3] < 189 lru_hash_lookup_test_entries - 32) 190 in6.sin6_addr.s6_addr32[3] += 32; 191 else 192 in6.sin6_addr.s6_addr32[3] = 0; 193 } 194 printf("%d:%s pre-alloc %lld events per sec\n", 195 cpu, test_name, 196 max_cnt * 1000000000ll / (time_get_ns() - start_time)); 197 } 198 199 static void test_lru_hash_prealloc(int cpu) 200 { 201 do_test_lru(LRU_HASH_PREALLOC, cpu); 202 } 203 204 static void test_nocommon_lru_hash_prealloc(int cpu) 205 { 206 do_test_lru(NOCOMMON_LRU_HASH_PREALLOC, cpu); 207 } 208 209 static void test_inner_lru_hash_prealloc(int cpu) 210 { 211 do_test_lru(INNER_LRU_HASH_PREALLOC, cpu); 212 } 213 214 static void test_lru_hash_lookup(int cpu) 215 { 216 do_test_lru(LRU_HASH_LOOKUP, cpu); 217 } 218 219 static void test_percpu_hash_prealloc(int cpu) 220 { 221 __u64 start_time; 222 int i; 223 224 start_time = time_get_ns(); 225 for (i = 0; i < max_cnt; i++) 226 syscall(__NR_geteuid); 227 printf("%d:percpu_hash_map_perf pre-alloc %lld events per sec\n", 228 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 229 } 230 231 static void test_hash_kmalloc(int cpu) 232 { 233 __u64 start_time; 234 int i; 235 236 start_time = time_get_ns(); 237 for (i = 0; i < max_cnt; i++) 238 syscall(__NR_getgid); 239 printf("%d:hash_map_perf kmalloc %lld events per sec\n", 240 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 241 } 242 243 static void test_percpu_hash_kmalloc(int cpu) 244 { 245 __u64 start_time; 246 int i; 247 248 start_time = time_get_ns(); 249 for (i = 0; i < max_cnt; i++) 250 syscall(__NR_getegid); 251 printf("%d:percpu_hash_map_perf kmalloc %lld events per sec\n", 252 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 253 } 254 255 static void test_lpm_kmalloc(int cpu) 256 { 257 __u64 start_time; 258 int i; 259 260 start_time = time_get_ns(); 261 for (i = 0; i < max_cnt; i++) 262 syscall(__NR_gettid); 263 printf("%d:lpm_perf kmalloc %lld events per sec\n", 264 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 265 } 266 267 static void test_hash_lookup(int cpu) 268 { 269 __u64 start_time; 270 int i; 271 272 start_time = time_get_ns(); 273 for (i = 0; i < max_cnt; i++) 274 syscall(__NR_getpgid, 0); 275 printf("%d:hash_lookup %lld lookups per sec\n", 276 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time)); 277 } 278 279 static void test_array_lookup(int cpu) 280 { 281 __u64 start_time; 282 int i; 283 284 start_time = time_get_ns(); 285 for (i = 0; i < max_cnt; i++) 286 syscall(__NR_getppid, 0); 287 printf("%d:array_lookup %lld lookups per sec\n", 288 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time)); 289 } 290 291 typedef int (*pre_test_func)(int tasks); 292 const pre_test_func pre_test_funcs[] = { 293 [LRU_HASH_LOOKUP] = pre_test_lru_hash_lookup, 294 }; 295 296 typedef void (*test_func)(int cpu); 297 const test_func test_funcs[] = { 298 [HASH_PREALLOC] = test_hash_prealloc, 299 [PERCPU_HASH_PREALLOC] = test_percpu_hash_prealloc, 300 [HASH_KMALLOC] = test_hash_kmalloc, 301 [PERCPU_HASH_KMALLOC] = test_percpu_hash_kmalloc, 302 [LRU_HASH_PREALLOC] = test_lru_hash_prealloc, 303 [NOCOMMON_LRU_HASH_PREALLOC] = test_nocommon_lru_hash_prealloc, 304 [LPM_KMALLOC] = test_lpm_kmalloc, 305 [HASH_LOOKUP] = test_hash_lookup, 306 [ARRAY_LOOKUP] = test_array_lookup, 307 [INNER_LRU_HASH_PREALLOC] = test_inner_lru_hash_prealloc, 308 [LRU_HASH_LOOKUP] = test_lru_hash_lookup, 309 }; 310 311 static int pre_test(int tasks) 312 { 313 int i; 314 315 for (i = 0; i < NR_TESTS; i++) { 316 if (pre_test_funcs[i] && check_test_flags(i)) { 317 int ret = pre_test_funcs[i](tasks); 318 319 if (ret) 320 return ret; 321 } 322 } 323 324 return 0; 325 } 326 327 static void loop(int cpu) 328 { 329 cpu_set_t cpuset; 330 int i; 331 332 CPU_ZERO(&cpuset); 333 CPU_SET(cpu, &cpuset); 334 sched_setaffinity(0, sizeof(cpuset), &cpuset); 335 336 for (i = 0; i < NR_TESTS; i++) { 337 if (check_test_flags(i)) 338 test_funcs[i](cpu); 339 } 340 } 341 342 static void run_perf_test(int tasks) 343 { 344 pid_t pid[tasks]; 345 int i; 346 347 assert(!pre_test(tasks)); 348 349 for (i = 0; i < tasks; i++) { 350 pid[i] = fork(); 351 if (pid[i] == 0) { 352 loop(i); 353 exit(0); 354 } else if (pid[i] == -1) { 355 printf("couldn't spawn #%d process\n", i); 356 exit(1); 357 } 358 } 359 for (i = 0; i < tasks; i++) { 360 int status; 361 362 assert(waitpid(pid[i], &status, 0) == pid[i]); 363 assert(status == 0); 364 } 365 } 366 367 static void fill_lpm_trie(void) 368 { 369 struct bpf_lpm_trie_key *key; 370 unsigned long value = 0; 371 unsigned int i; 372 int r; 373 374 key = alloca(sizeof(*key) + 4); 375 key->prefixlen = 32; 376 377 for (i = 0; i < 512; ++i) { 378 key->prefixlen = rand() % 33; 379 key->data[0] = rand() & 0xff; 380 key->data[1] = rand() & 0xff; 381 key->data[2] = rand() & 0xff; 382 key->data[3] = rand() & 0xff; 383 r = bpf_map_update_elem(map_fd[6], key, &value, 0); 384 assert(!r); 385 } 386 387 key->prefixlen = 32; 388 key->data[0] = 192; 389 key->data[1] = 168; 390 key->data[2] = 0; 391 key->data[3] = 1; 392 value = 128; 393 394 r = bpf_map_update_elem(map_fd[6], key, &value, 0); 395 assert(!r); 396 } 397 398 static void fixup_map(struct bpf_map_data *map, int idx) 399 { 400 int i; 401 402 if (!strcmp("inner_lru_hash_map", map->name)) { 403 inner_lru_hash_idx = idx; 404 inner_lru_hash_size = map->def.max_entries; 405 } 406 407 if (!strcmp("array_of_lru_hashs", map->name)) { 408 if (inner_lru_hash_idx == -1) { 409 printf("inner_lru_hash_map must be defined before array_of_lru_hashs\n"); 410 exit(1); 411 } 412 map->def.inner_map_idx = inner_lru_hash_idx; 413 array_of_lru_hashs_idx = idx; 414 } 415 416 if (!strcmp("lru_hash_lookup_map", map->name)) 417 lru_hash_lookup_idx = idx; 418 419 if (num_map_entries <= 0) 420 return; 421 422 inner_lru_hash_size = num_map_entries; 423 424 /* Only change the max_entries for the enabled test(s) */ 425 for (i = 0; i < NR_TESTS; i++) { 426 if (!strcmp(test_map_names[i], map->name) && 427 (check_test_flags(i))) { 428 map->def.max_entries = num_map_entries; 429 } 430 } 431 } 432 433 int main(int argc, char **argv) 434 { 435 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY}; 436 char filename[256]; 437 int num_cpu = 8; 438 439 snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]); 440 setrlimit(RLIMIT_MEMLOCK, &r); 441 442 if (argc > 1) 443 test_flags = atoi(argv[1]) ? : test_flags; 444 445 if (argc > 2) 446 num_cpu = atoi(argv[2]) ? : num_cpu; 447 448 if (argc > 3) 449 num_map_entries = atoi(argv[3]); 450 451 if (argc > 4) 452 max_cnt = atoi(argv[4]); 453 454 if (load_bpf_file_fixup_map(filename, fixup_map)) { 455 printf("%s", bpf_log_buf); 456 return 1; 457 } 458 459 fill_lpm_trie(); 460 461 run_perf_test(num_cpu); 462 463 return 0; 464 } 465