Loading...
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 <string.h>
15#include <time.h>
16#include <sys/resource.h>
17#include <arpa/inet.h>
18#include <errno.h>
19
20#include <bpf/bpf.h>
21#include <bpf/libbpf.h>
22
23#define TEST_BIT(t) (1U << (t))
24#define MAX_NR_CPUS 1024
25
26static __u64 time_get_ns(void)
27{
28 struct timespec ts;
29
30 clock_gettime(CLOCK_MONOTONIC, &ts);
31 return ts.tv_sec * 1000000000ull + ts.tv_nsec;
32}
33
34enum test_type {
35 HASH_PREALLOC,
36 PERCPU_HASH_PREALLOC,
37 HASH_KMALLOC,
38 PERCPU_HASH_KMALLOC,
39 LRU_HASH_PREALLOC,
40 NOCOMMON_LRU_HASH_PREALLOC,
41 LPM_KMALLOC,
42 HASH_LOOKUP,
43 ARRAY_LOOKUP,
44 INNER_LRU_HASH_PREALLOC,
45 LRU_HASH_LOOKUP,
46 NR_TESTS,
47};
48
49const char *test_map_names[NR_TESTS] = {
50 [HASH_PREALLOC] = "hash_map",
51 [PERCPU_HASH_PREALLOC] = "percpu_hash_map",
52 [HASH_KMALLOC] = "hash_map_alloc",
53 [PERCPU_HASH_KMALLOC] = "percpu_hash_map_alloc",
54 [LRU_HASH_PREALLOC] = "lru_hash_map",
55 [NOCOMMON_LRU_HASH_PREALLOC] = "nocommon_lru_hash_map",
56 [LPM_KMALLOC] = "lpm_trie_map_alloc",
57 [HASH_LOOKUP] = "hash_map",
58 [ARRAY_LOOKUP] = "array_map",
59 [INNER_LRU_HASH_PREALLOC] = "inner_lru_hash_map",
60 [LRU_HASH_LOOKUP] = "lru_hash_lookup_map",
61};
62
63enum map_idx {
64 array_of_lru_hashs_idx,
65 hash_map_alloc_idx,
66 lru_hash_lookup_idx,
67 NR_IDXES,
68};
69
70static int map_fd[NR_IDXES];
71
72static int test_flags = ~0;
73static uint32_t num_map_entries;
74static uint32_t inner_lru_hash_size;
75static int lru_hash_lookup_test_entries = 32;
76static uint32_t max_cnt = 1000000;
77
78static int check_test_flags(enum test_type t)
79{
80 return test_flags & TEST_BIT(t);
81}
82
83static void test_hash_prealloc(int cpu)
84{
85 __u64 start_time;
86 int i;
87
88 start_time = time_get_ns();
89 for (i = 0; i < max_cnt; i++)
90 syscall(__NR_getuid);
91 printf("%d:hash_map_perf pre-alloc %lld events per sec\n",
92 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
93}
94
95static int pre_test_lru_hash_lookup(int tasks)
96{
97 int fd = map_fd[lru_hash_lookup_idx];
98 uint32_t key;
99 long val = 1;
100 int ret;
101
102 if (num_map_entries > lru_hash_lookup_test_entries)
103 lru_hash_lookup_test_entries = num_map_entries;
104
105 /* Populate the lru_hash_map for LRU_HASH_LOOKUP perf test.
106 *
107 * It is fine that the user requests for a map with
108 * num_map_entries < 32 and some of the later lru hash lookup
109 * may return not found. For LRU map, we are not interested
110 * in such small map performance.
111 */
112 for (key = 0; key < lru_hash_lookup_test_entries; key++) {
113 ret = bpf_map_update_elem(fd, &key, &val, BPF_NOEXIST);
114 if (ret)
115 return ret;
116 }
117
118 return 0;
119}
120
121static void do_test_lru(enum test_type test, int cpu)
122{
123 static int inner_lru_map_fds[MAX_NR_CPUS];
124
125 struct sockaddr_in6 in6 = { .sin6_family = AF_INET6 };
126 const char *test_name;
127 __u64 start_time;
128 int i, ret;
129
130 if (test == INNER_LRU_HASH_PREALLOC && cpu) {
131 /* If CPU is not 0, create inner_lru hash map and insert the fd
132 * value into the array_of_lru_hash map. In case of CPU 0,
133 * 'inner_lru_hash_map' was statically inserted on the map init
134 */
135 int outer_fd = map_fd[array_of_lru_hashs_idx];
136 unsigned int mycpu, mynode;
137
138 assert(cpu < MAX_NR_CPUS);
139
140 ret = syscall(__NR_getcpu, &mycpu, &mynode, NULL);
141 assert(!ret);
142
143 inner_lru_map_fds[cpu] =
144 bpf_create_map_node(BPF_MAP_TYPE_LRU_HASH,
145 test_map_names[INNER_LRU_HASH_PREALLOC],
146 sizeof(uint32_t),
147 sizeof(long),
148 inner_lru_hash_size, 0,
149 mynode);
150 if (inner_lru_map_fds[cpu] == -1) {
151 printf("cannot create BPF_MAP_TYPE_LRU_HASH %s(%d)\n",
152 strerror(errno), errno);
153 exit(1);
154 }
155
156 ret = bpf_map_update_elem(outer_fd, &cpu,
157 &inner_lru_map_fds[cpu],
158 BPF_ANY);
159 if (ret) {
160 printf("cannot update ARRAY_OF_LRU_HASHS with key:%u. %s(%d)\n",
161 cpu, strerror(errno), errno);
162 exit(1);
163 }
164 }
165
166 in6.sin6_addr.s6_addr16[0] = 0xdead;
167 in6.sin6_addr.s6_addr16[1] = 0xbeef;
168
169 if (test == LRU_HASH_PREALLOC) {
170 test_name = "lru_hash_map_perf";
171 in6.sin6_addr.s6_addr16[2] = 0;
172 } else if (test == NOCOMMON_LRU_HASH_PREALLOC) {
173 test_name = "nocommon_lru_hash_map_perf";
174 in6.sin6_addr.s6_addr16[2] = 1;
175 } else if (test == INNER_LRU_HASH_PREALLOC) {
176 test_name = "inner_lru_hash_map_perf";
177 in6.sin6_addr.s6_addr16[2] = 2;
178 } else if (test == LRU_HASH_LOOKUP) {
179 test_name = "lru_hash_lookup_perf";
180 in6.sin6_addr.s6_addr16[2] = 3;
181 in6.sin6_addr.s6_addr32[3] = 0;
182 } else {
183 assert(0);
184 }
185
186 start_time = time_get_ns();
187 for (i = 0; i < max_cnt; i++) {
188 ret = connect(-1, (const struct sockaddr *)&in6, sizeof(in6));
189 assert(ret == -1 && errno == EBADF);
190 if (in6.sin6_addr.s6_addr32[3] <
191 lru_hash_lookup_test_entries - 32)
192 in6.sin6_addr.s6_addr32[3] += 32;
193 else
194 in6.sin6_addr.s6_addr32[3] = 0;
195 }
196 printf("%d:%s pre-alloc %lld events per sec\n",
197 cpu, test_name,
198 max_cnt * 1000000000ll / (time_get_ns() - start_time));
199}
200
201static void test_lru_hash_prealloc(int cpu)
202{
203 do_test_lru(LRU_HASH_PREALLOC, cpu);
204}
205
206static void test_nocommon_lru_hash_prealloc(int cpu)
207{
208 do_test_lru(NOCOMMON_LRU_HASH_PREALLOC, cpu);
209}
210
211static void test_inner_lru_hash_prealloc(int cpu)
212{
213 do_test_lru(INNER_LRU_HASH_PREALLOC, cpu);
214}
215
216static void test_lru_hash_lookup(int cpu)
217{
218 do_test_lru(LRU_HASH_LOOKUP, cpu);
219}
220
221static void test_percpu_hash_prealloc(int cpu)
222{
223 __u64 start_time;
224 int i;
225
226 start_time = time_get_ns();
227 for (i = 0; i < max_cnt; i++)
228 syscall(__NR_geteuid);
229 printf("%d:percpu_hash_map_perf pre-alloc %lld events per sec\n",
230 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
231}
232
233static void test_hash_kmalloc(int cpu)
234{
235 __u64 start_time;
236 int i;
237
238 start_time = time_get_ns();
239 for (i = 0; i < max_cnt; i++)
240 syscall(__NR_getgid);
241 printf("%d:hash_map_perf kmalloc %lld events per sec\n",
242 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
243}
244
245static void test_percpu_hash_kmalloc(int cpu)
246{
247 __u64 start_time;
248 int i;
249
250 start_time = time_get_ns();
251 for (i = 0; i < max_cnt; i++)
252 syscall(__NR_getegid);
253 printf("%d:percpu_hash_map_perf kmalloc %lld events per sec\n",
254 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
255}
256
257static void test_lpm_kmalloc(int cpu)
258{
259 __u64 start_time;
260 int i;
261
262 start_time = time_get_ns();
263 for (i = 0; i < max_cnt; i++)
264 syscall(__NR_gettid);
265 printf("%d:lpm_perf kmalloc %lld events per sec\n",
266 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
267}
268
269static void test_hash_lookup(int cpu)
270{
271 __u64 start_time;
272 int i;
273
274 start_time = time_get_ns();
275 for (i = 0; i < max_cnt; i++)
276 syscall(__NR_getpgid, 0);
277 printf("%d:hash_lookup %lld lookups per sec\n",
278 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time));
279}
280
281static void test_array_lookup(int cpu)
282{
283 __u64 start_time;
284 int i;
285
286 start_time = time_get_ns();
287 for (i = 0; i < max_cnt; i++)
288 syscall(__NR_getppid, 0);
289 printf("%d:array_lookup %lld lookups per sec\n",
290 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time));
291}
292
293typedef int (*pre_test_func)(int tasks);
294const pre_test_func pre_test_funcs[] = {
295 [LRU_HASH_LOOKUP] = pre_test_lru_hash_lookup,
296};
297
298typedef void (*test_func)(int cpu);
299const test_func test_funcs[] = {
300 [HASH_PREALLOC] = test_hash_prealloc,
301 [PERCPU_HASH_PREALLOC] = test_percpu_hash_prealloc,
302 [HASH_KMALLOC] = test_hash_kmalloc,
303 [PERCPU_HASH_KMALLOC] = test_percpu_hash_kmalloc,
304 [LRU_HASH_PREALLOC] = test_lru_hash_prealloc,
305 [NOCOMMON_LRU_HASH_PREALLOC] = test_nocommon_lru_hash_prealloc,
306 [LPM_KMALLOC] = test_lpm_kmalloc,
307 [HASH_LOOKUP] = test_hash_lookup,
308 [ARRAY_LOOKUP] = test_array_lookup,
309 [INNER_LRU_HASH_PREALLOC] = test_inner_lru_hash_prealloc,
310 [LRU_HASH_LOOKUP] = test_lru_hash_lookup,
311};
312
313static int pre_test(int tasks)
314{
315 int i;
316
317 for (i = 0; i < NR_TESTS; i++) {
318 if (pre_test_funcs[i] && check_test_flags(i)) {
319 int ret = pre_test_funcs[i](tasks);
320
321 if (ret)
322 return ret;
323 }
324 }
325
326 return 0;
327}
328
329static void loop(int cpu)
330{
331 cpu_set_t cpuset;
332 int i;
333
334 CPU_ZERO(&cpuset);
335 CPU_SET(cpu, &cpuset);
336 sched_setaffinity(0, sizeof(cpuset), &cpuset);
337
338 for (i = 0; i < NR_TESTS; i++) {
339 if (check_test_flags(i))
340 test_funcs[i](cpu);
341 }
342}
343
344static void run_perf_test(int tasks)
345{
346 pid_t pid[tasks];
347 int i;
348
349 assert(!pre_test(tasks));
350
351 for (i = 0; i < tasks; i++) {
352 pid[i] = fork();
353 if (pid[i] == 0) {
354 loop(i);
355 exit(0);
356 } else if (pid[i] == -1) {
357 printf("couldn't spawn #%d process\n", i);
358 exit(1);
359 }
360 }
361 for (i = 0; i < tasks; i++) {
362 int status;
363
364 assert(waitpid(pid[i], &status, 0) == pid[i]);
365 assert(status == 0);
366 }
367}
368
369static void fill_lpm_trie(void)
370{
371 struct bpf_lpm_trie_key *key;
372 unsigned long value = 0;
373 unsigned int i;
374 int r;
375
376 key = alloca(sizeof(*key) + 4);
377 key->prefixlen = 32;
378
379 for (i = 0; i < 512; ++i) {
380 key->prefixlen = rand() % 33;
381 key->data[0] = rand() & 0xff;
382 key->data[1] = rand() & 0xff;
383 key->data[2] = rand() & 0xff;
384 key->data[3] = rand() & 0xff;
385 r = bpf_map_update_elem(map_fd[hash_map_alloc_idx],
386 key, &value, 0);
387 assert(!r);
388 }
389
390 key->prefixlen = 32;
391 key->data[0] = 192;
392 key->data[1] = 168;
393 key->data[2] = 0;
394 key->data[3] = 1;
395 value = 128;
396
397 r = bpf_map_update_elem(map_fd[hash_map_alloc_idx], key, &value, 0);
398 assert(!r);
399}
400
401static void fixup_map(struct bpf_object *obj)
402{
403 struct bpf_map *map;
404 int i;
405
406 bpf_object__for_each_map(map, obj) {
407 const char *name = bpf_map__name(map);
408
409 /* Only change the max_entries for the enabled test(s) */
410 for (i = 0; i < NR_TESTS; i++) {
411 if (!strcmp(test_map_names[i], name) &&
412 (check_test_flags(i))) {
413 bpf_map__resize(map, num_map_entries);
414 continue;
415 }
416 }
417 }
418
419 inner_lru_hash_size = num_map_entries;
420}
421
422int main(int argc, char **argv)
423{
424 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
425 int nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
426 struct bpf_link *links[8];
427 struct bpf_program *prog;
428 struct bpf_object *obj;
429 struct bpf_map *map;
430 char filename[256];
431 int i = 0;
432
433 if (setrlimit(RLIMIT_MEMLOCK, &r)) {
434 perror("setrlimit(RLIMIT_MEMLOCK)");
435 return 1;
436 }
437
438 if (argc > 1)
439 test_flags = atoi(argv[1]) ? : test_flags;
440
441 if (argc > 2)
442 nr_cpus = atoi(argv[2]) ? : nr_cpus;
443
444 if (argc > 3)
445 num_map_entries = atoi(argv[3]);
446
447 if (argc > 4)
448 max_cnt = atoi(argv[4]);
449
450 snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
451 obj = bpf_object__open_file(filename, NULL);
452 if (libbpf_get_error(obj)) {
453 fprintf(stderr, "ERROR: opening BPF object file failed\n");
454 return 0;
455 }
456
457 map = bpf_object__find_map_by_name(obj, "inner_lru_hash_map");
458 if (libbpf_get_error(map)) {
459 fprintf(stderr, "ERROR: finding a map in obj file failed\n");
460 goto cleanup;
461 }
462
463 inner_lru_hash_size = bpf_map__max_entries(map);
464 if (!inner_lru_hash_size) {
465 fprintf(stderr, "ERROR: failed to get map attribute\n");
466 goto cleanup;
467 }
468
469 /* resize BPF map prior to loading */
470 if (num_map_entries > 0)
471 fixup_map(obj);
472
473 /* load BPF program */
474 if (bpf_object__load(obj)) {
475 fprintf(stderr, "ERROR: loading BPF object file failed\n");
476 goto cleanup;
477 }
478
479 map_fd[0] = bpf_object__find_map_fd_by_name(obj, "array_of_lru_hashs");
480 map_fd[1] = bpf_object__find_map_fd_by_name(obj, "hash_map_alloc");
481 map_fd[2] = bpf_object__find_map_fd_by_name(obj, "lru_hash_lookup_map");
482 if (map_fd[0] < 0 || map_fd[1] < 0 || map_fd[2] < 0) {
483 fprintf(stderr, "ERROR: finding a map in obj file failed\n");
484 goto cleanup;
485 }
486
487 bpf_object__for_each_program(prog, obj) {
488 links[i] = bpf_program__attach(prog);
489 if (libbpf_get_error(links[i])) {
490 fprintf(stderr, "ERROR: bpf_program__attach failed\n");
491 links[i] = NULL;
492 goto cleanup;
493 }
494 i++;
495 }
496
497 fill_lpm_trie();
498
499 run_perf_test(nr_cpus);
500
501cleanup:
502 for (i--; i >= 0; i--)
503 bpf_link__destroy(links[i]);
504
505 bpf_object__close(obj);
506 return 0;
507}
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
27static __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
35enum 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
50const 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
64static int test_flags = ~0;
65static uint32_t num_map_entries;
66static uint32_t inner_lru_hash_size;
67static int inner_lru_hash_idx = -1;
68static int array_of_lru_hashs_idx = -1;
69static int lru_hash_lookup_idx = -1;
70static int lru_hash_lookup_test_entries = 32;
71static uint32_t max_cnt = 1000000;
72
73static int check_test_flags(enum test_type t)
74{
75 return test_flags & TEST_BIT(t);
76}
77
78static 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
90static 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
116static 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
196static void test_lru_hash_prealloc(int cpu)
197{
198 do_test_lru(LRU_HASH_PREALLOC, cpu);
199}
200
201static void test_nocommon_lru_hash_prealloc(int cpu)
202{
203 do_test_lru(NOCOMMON_LRU_HASH_PREALLOC, cpu);
204}
205
206static void test_inner_lru_hash_prealloc(int cpu)
207{
208 do_test_lru(INNER_LRU_HASH_PREALLOC, cpu);
209}
210
211static void test_lru_hash_lookup(int cpu)
212{
213 do_test_lru(LRU_HASH_LOOKUP, cpu);
214}
215
216static 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
228static 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
240static 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
252static 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
264static 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
276static 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
288typedef int (*pre_test_func)(int tasks);
289const pre_test_func pre_test_funcs[] = {
290 [LRU_HASH_LOOKUP] = pre_test_lru_hash_lookup,
291};
292
293typedef void (*test_func)(int cpu);
294const 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
308static 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
324static 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
339static 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
364static 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
395static 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
430int 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}