1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2016 Facebook
3 */
4 #define _GNU_SOURCE
5 #include <sched.h>
6 #include <stdio.h>
7 #include <sys/types.h>
8 #include <asm/unistd.h>
9 #include <unistd.h>
10 #include <assert.h>
11 #include <sys/wait.h>
12 #include <stdlib.h>
13 #include <signal.h>
14 #include <linux/bpf.h>
15 #include <string.h>
16 #include <time.h>
17 #include <sys/resource.h>
18 #include <arpa/inet.h>
19 #include <errno.h>
20
21 #include <bpf/bpf.h>
22 #include "bpf_load.h"
23
24 #define TEST_BIT(t) (1U << (t))
25 #define MAX_NR_CPUS 1024
26
time_get_ns(void)27 static __u64 time_get_ns(void)
28 {
29 struct timespec ts;
30
31 clock_gettime(CLOCK_MONOTONIC, &ts);
32 return ts.tv_sec * 1000000000ull + ts.tv_nsec;
33 }
34
35 enum test_type {
36 HASH_PREALLOC,
37 PERCPU_HASH_PREALLOC,
38 HASH_KMALLOC,
39 PERCPU_HASH_KMALLOC,
40 LRU_HASH_PREALLOC,
41 NOCOMMON_LRU_HASH_PREALLOC,
42 LPM_KMALLOC,
43 HASH_LOOKUP,
44 ARRAY_LOOKUP,
45 INNER_LRU_HASH_PREALLOC,
46 LRU_HASH_LOOKUP,
47 NR_TESTS,
48 };
49
50 const char *test_map_names[NR_TESTS] = {
51 [HASH_PREALLOC] = "hash_map",
52 [PERCPU_HASH_PREALLOC] = "percpu_hash_map",
53 [HASH_KMALLOC] = "hash_map_alloc",
54 [PERCPU_HASH_KMALLOC] = "percpu_hash_map_alloc",
55 [LRU_HASH_PREALLOC] = "lru_hash_map",
56 [NOCOMMON_LRU_HASH_PREALLOC] = "nocommon_lru_hash_map",
57 [LPM_KMALLOC] = "lpm_trie_map_alloc",
58 [HASH_LOOKUP] = "hash_map",
59 [ARRAY_LOOKUP] = "array_map",
60 [INNER_LRU_HASH_PREALLOC] = "inner_lru_hash_map",
61 [LRU_HASH_LOOKUP] = "lru_hash_lookup_map",
62 };
63
64 static int test_flags = ~0;
65 static uint32_t num_map_entries;
66 static uint32_t inner_lru_hash_size;
67 static int inner_lru_hash_idx = -1;
68 static int array_of_lru_hashs_idx = -1;
69 static int lru_hash_lookup_idx = -1;
70 static int lru_hash_lookup_test_entries = 32;
71 static uint32_t max_cnt = 1000000;
72
check_test_flags(enum test_type t)73 static int check_test_flags(enum test_type t)
74 {
75 return test_flags & TEST_BIT(t);
76 }
77
test_hash_prealloc(int cpu)78 static void test_hash_prealloc(int cpu)
79 {
80 __u64 start_time;
81 int i;
82
83 start_time = time_get_ns();
84 for (i = 0; i < max_cnt; i++)
85 syscall(__NR_getuid);
86 printf("%d:hash_map_perf pre-alloc %lld events per sec\n",
87 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
88 }
89
pre_test_lru_hash_lookup(int tasks)90 static int pre_test_lru_hash_lookup(int tasks)
91 {
92 int fd = map_fd[lru_hash_lookup_idx];
93 uint32_t key;
94 long val = 1;
95 int ret;
96
97 if (num_map_entries > lru_hash_lookup_test_entries)
98 lru_hash_lookup_test_entries = num_map_entries;
99
100 /* Populate the lru_hash_map for LRU_HASH_LOOKUP perf test.
101 *
102 * It is fine that the user requests for a map with
103 * num_map_entries < 32 and some of the later lru hash lookup
104 * may return not found. For LRU map, we are not interested
105 * in such small map performance.
106 */
107 for (key = 0; key < lru_hash_lookup_test_entries; key++) {
108 ret = bpf_map_update_elem(fd, &key, &val, BPF_NOEXIST);
109 if (ret)
110 return ret;
111 }
112
113 return 0;
114 }
115
do_test_lru(enum test_type test,int cpu)116 static void do_test_lru(enum test_type test, int cpu)
117 {
118 static int inner_lru_map_fds[MAX_NR_CPUS];
119
120 struct sockaddr_in6 in6 = { .sin6_family = AF_INET6 };
121 const char *test_name;
122 __u64 start_time;
123 int i, ret;
124
125 if (test == INNER_LRU_HASH_PREALLOC) {
126 int outer_fd = map_fd[array_of_lru_hashs_idx];
127 unsigned int mycpu, mynode;
128
129 assert(cpu < MAX_NR_CPUS);
130
131 if (cpu) {
132 ret = syscall(__NR_getcpu, &mycpu, &mynode, NULL);
133 assert(!ret);
134
135 inner_lru_map_fds[cpu] =
136 bpf_create_map_node(BPF_MAP_TYPE_LRU_HASH,
137 test_map_names[INNER_LRU_HASH_PREALLOC],
138 sizeof(uint32_t),
139 sizeof(long),
140 inner_lru_hash_size, 0,
141 mynode);
142 if (inner_lru_map_fds[cpu] == -1) {
143 printf("cannot create BPF_MAP_TYPE_LRU_HASH %s(%d)\n",
144 strerror(errno), errno);
145 exit(1);
146 }
147 } else {
148 inner_lru_map_fds[cpu] = map_fd[inner_lru_hash_idx];
149 }
150
151 ret = bpf_map_update_elem(outer_fd, &cpu,
152 &inner_lru_map_fds[cpu],
153 BPF_ANY);
154 if (ret) {
155 printf("cannot update ARRAY_OF_LRU_HASHS with key:%u. %s(%d)\n",
156 cpu, strerror(errno), errno);
157 exit(1);
158 }
159 }
160
161 in6.sin6_addr.s6_addr16[0] = 0xdead;
162 in6.sin6_addr.s6_addr16[1] = 0xbeef;
163
164 if (test == LRU_HASH_PREALLOC) {
165 test_name = "lru_hash_map_perf";
166 in6.sin6_addr.s6_addr16[2] = 0;
167 } else if (test == NOCOMMON_LRU_HASH_PREALLOC) {
168 test_name = "nocommon_lru_hash_map_perf";
169 in6.sin6_addr.s6_addr16[2] = 1;
170 } else if (test == INNER_LRU_HASH_PREALLOC) {
171 test_name = "inner_lru_hash_map_perf";
172 in6.sin6_addr.s6_addr16[2] = 2;
173 } else if (test == LRU_HASH_LOOKUP) {
174 test_name = "lru_hash_lookup_perf";
175 in6.sin6_addr.s6_addr16[2] = 3;
176 in6.sin6_addr.s6_addr32[3] = 0;
177 } else {
178 assert(0);
179 }
180
181 start_time = time_get_ns();
182 for (i = 0; i < max_cnt; i++) {
183 ret = connect(-1, (const struct sockaddr *)&in6, sizeof(in6));
184 assert(ret == -1 && errno == EBADF);
185 if (in6.sin6_addr.s6_addr32[3] <
186 lru_hash_lookup_test_entries - 32)
187 in6.sin6_addr.s6_addr32[3] += 32;
188 else
189 in6.sin6_addr.s6_addr32[3] = 0;
190 }
191 printf("%d:%s pre-alloc %lld events per sec\n",
192 cpu, test_name,
193 max_cnt * 1000000000ll / (time_get_ns() - start_time));
194 }
195
test_lru_hash_prealloc(int cpu)196 static void test_lru_hash_prealloc(int cpu)
197 {
198 do_test_lru(LRU_HASH_PREALLOC, cpu);
199 }
200
test_nocommon_lru_hash_prealloc(int cpu)201 static void test_nocommon_lru_hash_prealloc(int cpu)
202 {
203 do_test_lru(NOCOMMON_LRU_HASH_PREALLOC, cpu);
204 }
205
test_inner_lru_hash_prealloc(int cpu)206 static void test_inner_lru_hash_prealloc(int cpu)
207 {
208 do_test_lru(INNER_LRU_HASH_PREALLOC, cpu);
209 }
210
test_lru_hash_lookup(int cpu)211 static void test_lru_hash_lookup(int cpu)
212 {
213 do_test_lru(LRU_HASH_LOOKUP, cpu);
214 }
215
test_percpu_hash_prealloc(int cpu)216 static void test_percpu_hash_prealloc(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_geteuid);
224 printf("%d:percpu_hash_map_perf pre-alloc %lld events per sec\n",
225 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
226 }
227
test_hash_kmalloc(int cpu)228 static void test_hash_kmalloc(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_getgid);
236 printf("%d:hash_map_perf kmalloc %lld events per sec\n",
237 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
238 }
239
test_percpu_hash_kmalloc(int cpu)240 static void test_percpu_hash_kmalloc(int cpu)
241 {
242 __u64 start_time;
243 int i;
244
245 start_time = time_get_ns();
246 for (i = 0; i < max_cnt; i++)
247 syscall(__NR_getegid);
248 printf("%d:percpu_hash_map_perf kmalloc %lld events per sec\n",
249 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
250 }
251
test_lpm_kmalloc(int cpu)252 static void test_lpm_kmalloc(int cpu)
253 {
254 __u64 start_time;
255 int i;
256
257 start_time = time_get_ns();
258 for (i = 0; i < max_cnt; i++)
259 syscall(__NR_gettid);
260 printf("%d:lpm_perf kmalloc %lld events per sec\n",
261 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
262 }
263
test_hash_lookup(int cpu)264 static void test_hash_lookup(int cpu)
265 {
266 __u64 start_time;
267 int i;
268
269 start_time = time_get_ns();
270 for (i = 0; i < max_cnt; i++)
271 syscall(__NR_getpgid, 0);
272 printf("%d:hash_lookup %lld lookups per sec\n",
273 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time));
274 }
275
test_array_lookup(int cpu)276 static void test_array_lookup(int cpu)
277 {
278 __u64 start_time;
279 int i;
280
281 start_time = time_get_ns();
282 for (i = 0; i < max_cnt; i++)
283 syscall(__NR_getppid, 0);
284 printf("%d:array_lookup %lld lookups per sec\n",
285 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time));
286 }
287
288 typedef int (*pre_test_func)(int tasks);
289 const pre_test_func pre_test_funcs[] = {
290 [LRU_HASH_LOOKUP] = pre_test_lru_hash_lookup,
291 };
292
293 typedef void (*test_func)(int cpu);
294 const test_func test_funcs[] = {
295 [HASH_PREALLOC] = test_hash_prealloc,
296 [PERCPU_HASH_PREALLOC] = test_percpu_hash_prealloc,
297 [HASH_KMALLOC] = test_hash_kmalloc,
298 [PERCPU_HASH_KMALLOC] = test_percpu_hash_kmalloc,
299 [LRU_HASH_PREALLOC] = test_lru_hash_prealloc,
300 [NOCOMMON_LRU_HASH_PREALLOC] = test_nocommon_lru_hash_prealloc,
301 [LPM_KMALLOC] = test_lpm_kmalloc,
302 [HASH_LOOKUP] = test_hash_lookup,
303 [ARRAY_LOOKUP] = test_array_lookup,
304 [INNER_LRU_HASH_PREALLOC] = test_inner_lru_hash_prealloc,
305 [LRU_HASH_LOOKUP] = test_lru_hash_lookup,
306 };
307
pre_test(int tasks)308 static int pre_test(int tasks)
309 {
310 int i;
311
312 for (i = 0; i < NR_TESTS; i++) {
313 if (pre_test_funcs[i] && check_test_flags(i)) {
314 int ret = pre_test_funcs[i](tasks);
315
316 if (ret)
317 return ret;
318 }
319 }
320
321 return 0;
322 }
323
loop(int cpu)324 static void loop(int cpu)
325 {
326 cpu_set_t cpuset;
327 int i;
328
329 CPU_ZERO(&cpuset);
330 CPU_SET(cpu, &cpuset);
331 sched_setaffinity(0, sizeof(cpuset), &cpuset);
332
333 for (i = 0; i < NR_TESTS; i++) {
334 if (check_test_flags(i))
335 test_funcs[i](cpu);
336 }
337 }
338
run_perf_test(int tasks)339 static void run_perf_test(int tasks)
340 {
341 pid_t pid[tasks];
342 int i;
343
344 assert(!pre_test(tasks));
345
346 for (i = 0; i < tasks; i++) {
347 pid[i] = fork();
348 if (pid[i] == 0) {
349 loop(i);
350 exit(0);
351 } else if (pid[i] == -1) {
352 printf("couldn't spawn #%d process\n", i);
353 exit(1);
354 }
355 }
356 for (i = 0; i < tasks; i++) {
357 int status;
358
359 assert(waitpid(pid[i], &status, 0) == pid[i]);
360 assert(status == 0);
361 }
362 }
363
fill_lpm_trie(void)364 static void fill_lpm_trie(void)
365 {
366 struct bpf_lpm_trie_key *key;
367 unsigned long value = 0;
368 unsigned int i;
369 int r;
370
371 key = alloca(sizeof(*key) + 4);
372 key->prefixlen = 32;
373
374 for (i = 0; i < 512; ++i) {
375 key->prefixlen = rand() % 33;
376 key->data[0] = rand() & 0xff;
377 key->data[1] = rand() & 0xff;
378 key->data[2] = rand() & 0xff;
379 key->data[3] = rand() & 0xff;
380 r = bpf_map_update_elem(map_fd[6], key, &value, 0);
381 assert(!r);
382 }
383
384 key->prefixlen = 32;
385 key->data[0] = 192;
386 key->data[1] = 168;
387 key->data[2] = 0;
388 key->data[3] = 1;
389 value = 128;
390
391 r = bpf_map_update_elem(map_fd[6], key, &value, 0);
392 assert(!r);
393 }
394
fixup_map(struct bpf_map_data * map,int idx)395 static void fixup_map(struct bpf_map_data *map, int idx)
396 {
397 int i;
398
399 if (!strcmp("inner_lru_hash_map", map->name)) {
400 inner_lru_hash_idx = idx;
401 inner_lru_hash_size = map->def.max_entries;
402 }
403
404 if (!strcmp("array_of_lru_hashs", map->name)) {
405 if (inner_lru_hash_idx == -1) {
406 printf("inner_lru_hash_map must be defined before array_of_lru_hashs\n");
407 exit(1);
408 }
409 map->def.inner_map_idx = inner_lru_hash_idx;
410 array_of_lru_hashs_idx = idx;
411 }
412
413 if (!strcmp("lru_hash_lookup_map", map->name))
414 lru_hash_lookup_idx = idx;
415
416 if (num_map_entries <= 0)
417 return;
418
419 inner_lru_hash_size = num_map_entries;
420
421 /* Only change the max_entries for the enabled test(s) */
422 for (i = 0; i < NR_TESTS; i++) {
423 if (!strcmp(test_map_names[i], map->name) &&
424 (check_test_flags(i))) {
425 map->def.max_entries = num_map_entries;
426 }
427 }
428 }
429
main(int argc,char ** argv)430 int main(int argc, char **argv)
431 {
432 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
433 char filename[256];
434 int num_cpu = 8;
435
436 snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
437 setrlimit(RLIMIT_MEMLOCK, &r);
438
439 if (argc > 1)
440 test_flags = atoi(argv[1]) ? : test_flags;
441
442 if (argc > 2)
443 num_cpu = atoi(argv[2]) ? : num_cpu;
444
445 if (argc > 3)
446 num_map_entries = atoi(argv[3]);
447
448 if (argc > 4)
449 max_cnt = atoi(argv[4]);
450
451 if (load_bpf_file_fixup_map(filename, fixup_map)) {
452 printf("%s", bpf_log_buf);
453 return 1;
454 }
455
456 fill_lpm_trie();
457
458 run_perf_test(num_cpu);
459
460 return 0;
461 }
462