samples/bpf/map_perf_test_user.c

   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 "libbpf.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         NR_TESTS,
  50 };
  51
  52 const char *test_map_names[NR_TESTS] = {
  53         [HASH_PREALLOC] = "hash_map",
  54         [PERCPU_HASH_PREALLOC] = "percpu_hash_map",
  55         [HASH_KMALLOC] = "hash_map_alloc",
  56         [PERCPU_HASH_KMALLOC] = "percpu_hash_map_alloc",
  57         [LRU_HASH_PREALLOC] = "lru_hash_map",
  58         [NOCOMMON_LRU_HASH_PREALLOC] = "nocommon_lru_hash_map",
  59         [LPM_KMALLOC] = "lpm_trie_map_alloc",
  60         [HASH_LOOKUP] = "hash_map",
  61         [ARRAY_LOOKUP] = "array_map",
  62         [INNER_LRU_HASH_PREALLOC] = "inner_lru_hash_map",
  63 };
  64
  65 static int test_flags = ~0;
  66 static uint32_t num_map_entries;
  67 static uint32_t inner_lru_hash_size;
  68 static int inner_lru_hash_idx = -1;
  69 static int array_of_lru_hashs_idx = -1;
  70 static uint32_t max_cnt = 1000000;
  71
  72 static int check_test_flags(enum test_type t)
  73 {
  74         return test_flags & TEST_BIT(t);
  75 }
  76
  77 static void test_hash_prealloc(int cpu)
  78 {
  79         __u64 start_time;
  80         int i;
  81
  82         start_time = time_get_ns();
  83         for (i = 0; i < max_cnt; i++)
  84                 syscall(__NR_getuid);
  85         printf("%d:hash_map_perf pre-alloc %lld events per sec\n",
  86                cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
  87 }
  88
  89 static void do_test_lru(enum test_type test, int cpu)
  90 {
  91         static int inner_lru_map_fds[MAX_NR_CPUS];
  92
  93         struct sockaddr_in6 in6 = { .sin6_family = AF_INET6 };
  94         const char *test_name;
  95         __u64 start_time;
  96         int i, ret;
  97
  98         if (test == INNER_LRU_HASH_PREALLOC) {
  99                 int outer_fd = map_fd[array_of_lru_hashs_idx];
 100
 101                 assert(cpu < MAX_NR_CPUS);
 102
 103                 if (cpu) {
 104                         inner_lru_map_fds[cpu] =
 105                                 bpf_create_map(BPF_MAP_TYPE_LRU_HASH,
 106                                                sizeof(uint32_t), sizeof(long),
 107                                                inner_lru_hash_size, 0);
 108                         if (inner_lru_map_fds[cpu] == -1) {
 109                                 printf("cannot create BPF_MAP_TYPE_LRU_HASH %s(%d)\n",
 110                                        strerror(errno), errno);
 111                                 exit(1);
 112                         }
 113                 } else {
 114                         inner_lru_map_fds[cpu] = map_fd[inner_lru_hash_idx];
 115                 }
 116
 117                 ret = bpf_map_update_elem(outer_fd, &cpu,
 118                                           &inner_lru_map_fds[cpu],
 119                                           BPF_ANY);
 120                 if (ret) {
 121                         printf("cannot update ARRAY_OF_LRU_HASHS with key:%u. %s(%d)\n",
 122                                cpu, strerror(errno), errno);
 123                         exit(1);
 124                 }
 125         }
 126
 127         in6.sin6_addr.s6_addr16[0] = 0xdead;
 128         in6.sin6_addr.s6_addr16[1] = 0xbeef;
 129
 130         if (test == LRU_HASH_PREALLOC) {
 131                 test_name = "lru_hash_map_perf";
 132                 in6.sin6_addr.s6_addr16[7] = 0;
 133         } else if (test == NOCOMMON_LRU_HASH_PREALLOC) {
 134                 test_name = "nocommon_lru_hash_map_perf";
 135                 in6.sin6_addr.s6_addr16[7] = 1;
 136         } else if (test == INNER_LRU_HASH_PREALLOC) {
 137                 test_name = "inner_lru_hash_map_perf";
 138                 in6.sin6_addr.s6_addr16[7] = 2;
 139         } else {
 140                 assert(0);
 141         }
 142
 143         start_time = time_get_ns();
 144         for (i = 0; i < max_cnt; i++) {
 145                 ret = connect(-1, (const struct sockaddr *)&in6, sizeof(in6));
 146                 assert(ret == -1 && errno == EBADF);
 147         }
 148         printf("%d:%s pre-alloc %lld events per sec\n",
 149                cpu, test_name,
 150                max_cnt * 1000000000ll / (time_get_ns() - start_time));
 151 }
 152
 153 static void test_lru_hash_prealloc(int cpu)
 154 {
 155         do_test_lru(LRU_HASH_PREALLOC, cpu);
 156 }
 157
 158 static void test_nocommon_lru_hash_prealloc(int cpu)
 159 {
 160         do_test_lru(NOCOMMON_LRU_HASH_PREALLOC, cpu);
 161 }
 162
 163 static void test_inner_lru_hash_prealloc(int cpu)
 164 {
 165         do_test_lru(INNER_LRU_HASH_PREALLOC, cpu);
 166 }
 167
 168 static void test_percpu_hash_prealloc(int cpu)
 169 {
 170         __u64 start_time;
 171         int i;
 172
 173         start_time = time_get_ns();
 174         for (i = 0; i < max_cnt; i++)
 175                 syscall(__NR_geteuid);
 176         printf("%d:percpu_hash_map_perf pre-alloc %lld events per sec\n",
 177                cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
 178 }
 179
 180 static void test_hash_kmalloc(int cpu)
 181 {
 182         __u64 start_time;
 183         int i;
 184
 185         start_time = time_get_ns();
 186         for (i = 0; i < max_cnt; i++)
 187                 syscall(__NR_getgid);
 188         printf("%d:hash_map_perf kmalloc %lld events per sec\n",
 189                cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
 190 }
 191
 192 static void test_percpu_hash_kmalloc(int cpu)
 193 {
 194         __u64 start_time;
 195         int i;
 196
 197         start_time = time_get_ns();
 198         for (i = 0; i < max_cnt; i++)
 199                 syscall(__NR_getegid);
 200         printf("%d:percpu_hash_map_perf kmalloc %lld events per sec\n",
 201                cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
 202 }
 203
 204 static void test_lpm_kmalloc(int cpu)
 205 {
 206         __u64 start_time;
 207         int i;
 208
 209         start_time = time_get_ns();
 210         for (i = 0; i < max_cnt; i++)
 211                 syscall(__NR_gettid);
 212         printf("%d:lpm_perf kmalloc %lld events per sec\n",
 213                cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
 214 }
 215
 216 static void test_hash_lookup(int cpu)
 217 {
 218         __u64 start_time;
 219         int i;
 220
 221         start_time = time_get_ns();
 222         for (i = 0; i < max_cnt; i++)
 223                 syscall(__NR_getpgid, 0);
 224         printf("%d:hash_lookup %lld lookups per sec\n",
 225                cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time));
 226 }
 227
 228 static void test_array_lookup(int cpu)
 229 {
 230         __u64 start_time;
 231         int i;
 232
 233         start_time = time_get_ns();
 234         for (i = 0; i < max_cnt; i++)
 235                 syscall(__NR_getpgrp, 0);
 236         printf("%d:array_lookup %lld lookups per sec\n",
 237                cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time));
 238 }
 239
 240 typedef void (*test_func)(int cpu);
 241 const test_func test_funcs[] = {
 242         [HASH_PREALLOC] = test_hash_prealloc,
 243         [PERCPU_HASH_PREALLOC] = test_percpu_hash_prealloc,
 244         [HASH_KMALLOC] = test_hash_kmalloc,
 245         [PERCPU_HASH_KMALLOC] = test_percpu_hash_kmalloc,
 246         [LRU_HASH_PREALLOC] = test_lru_hash_prealloc,
 247         [NOCOMMON_LRU_HASH_PREALLOC] = test_nocommon_lru_hash_prealloc,
 248         [LPM_KMALLOC] = test_lpm_kmalloc,
 249         [HASH_LOOKUP] = test_hash_lookup,
 250         [ARRAY_LOOKUP] = test_array_lookup,
 251         [INNER_LRU_HASH_PREALLOC] = test_inner_lru_hash_prealloc,
 252 };
 253
 254 static void loop(int cpu)
 255 {
 256         cpu_set_t cpuset;
 257         int i;
 258
 259         CPU_ZERO(&cpuset);
 260         CPU_SET(cpu, &cpuset);
 261         sched_setaffinity(0, sizeof(cpuset), &cpuset);
 262
 263         for (i = 0; i < NR_TESTS; i++) {
 264                 if (check_test_flags(i))
 265                         test_funcs[i](cpu);
 266         }
 267 }
 268
 269 static void run_perf_test(int tasks)
 270 {
 271         pid_t pid[tasks];
 272         int i;
 273
 274         for (i = 0; i < tasks; i++) {
 275                 pid[i] = fork();
 276                 if (pid[i] == 0) {
 277                         loop(i);
 278                         exit(0);
 279                 } else if (pid[i] == -1) {
 280                         printf("couldn't spawn #%d process\n", i);
 281                         exit(1);
 282                 }
 283         }
 284         for (i = 0; i < tasks; i++) {
 285                 int status;
 286
 287                 assert(waitpid(pid[i], &status, 0) == pid[i]);
 288                 assert(status == 0);
 289         }
 290 }
 291
 292 static void fill_lpm_trie(void)
 293 {
 294         struct bpf_lpm_trie_key *key;
 295         unsigned long value = 0;
 296         unsigned int i;
 297         int r;
 298
 299         key = alloca(sizeof(*key) + 4);
 300         key->prefixlen = 32;
 301
 302         for (i = 0; i < 512; ++i) {
 303                 key->prefixlen = rand() % 33;
 304                 key->data[0] = rand() & 0xff;
 305                 key->data[1] = rand() & 0xff;
 306                 key->data[2] = rand() & 0xff;
 307                 key->data[3] = rand() & 0xff;
 308                 r = bpf_map_update_elem(map_fd[6], key, &value, 0);
 309                 assert(!r);
 310         }
 311
 312         key->prefixlen = 32;
 313         key->data[0] = 192;
 314         key->data[1] = 168;
 315         key->data[2] = 0;
 316         key->data[3] = 1;
 317         value = 128;
 318
 319         r = bpf_map_update_elem(map_fd[6], key, &value, 0);
 320         assert(!r);
 321 }
 322
 323 static void fixup_map(struct bpf_map_data *map, int idx)
 324 {
 325         int i;
 326
 327         if (!strcmp("inner_lru_hash_map", map->name)) {
 328                 inner_lru_hash_idx = idx;
 329                 inner_lru_hash_size = map->def.max_entries;
 330         }
 331
 332         if (!strcmp("array_of_lru_hashs", map->name)) {
 333                 if (inner_lru_hash_idx == -1) {
 334                         printf("inner_lru_hash_map must be defined before array_of_lru_hashs\n");
 335                         exit(1);
 336                 }
 337                 map->def.inner_map_idx = inner_lru_hash_idx;
 338                 array_of_lru_hashs_idx = idx;
 339         }
 340
 341         if (num_map_entries <= 0)
 342                 return;
 343
 344         inner_lru_hash_size = num_map_entries;
 345
 346         /* Only change the max_entries for the enabled test(s) */
 347         for (i = 0; i < NR_TESTS; i++) {
 348                 if (!strcmp(test_map_names[i], map->name) &&
 349                     (check_test_flags(i))) {
 350                         map->def.max_entries = num_map_entries;
 351                 }
 352         }
 353 }
 354
 355 int main(int argc, char **argv)
 356 {
 357         struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
 358         char filename[256];
 359         int num_cpu = 8;
 360
 361         snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
 362         setrlimit(RLIMIT_MEMLOCK, &r);
 363
 364         if (argc > 1)
 365                 test_flags = atoi(argv[1]) ? : test_flags;
 366
 367         if (argc > 2)
 368                 num_cpu = atoi(argv[2]) ? : num_cpu;
 369
 370         if (argc > 3)
 371                 num_map_entries = atoi(argv[3]);
 372
 373         if (argc > 4)
 374                 max_cnt = atoi(argv[4]);
 375
 376         if (load_bpf_file_fixup_map(filename, fixup_map)) {
 377                 printf("%s", bpf_log_buf);
 378                 return 1;
 379         }
 380
 381         fill_lpm_trie();
 382
 383         run_perf_test(num_cpu);
 384
 385         return 0;
 386 }