1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2016 Facebook
8 #include <asm/unistd.h>
16 #include <sys/resource.h>
17 #include <arpa/inet.h>
21 #include <bpf/libbpf.h>
23 #define TEST_BIT(t) (1U << (t))
24 #define MAX_NR_CPUS 1024
26 static __u64 time_get_ns(void)
30 clock_gettime(CLOCK_MONOTONIC, &ts);
31 return ts.tv_sec * 1000000000ull + ts.tv_nsec;
40 NOCOMMON_LRU_HASH_PREALLOC,
44 INNER_LRU_HASH_PREALLOC,
49 const 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",
64 array_of_lru_hashs_idx,
70 static int map_fd[NR_IDXES];
72 static int test_flags = ~0;
73 static uint32_t num_map_entries;
74 static uint32_t inner_lru_hash_size;
75 static int lru_hash_lookup_test_entries = 32;
76 static uint32_t max_cnt = 1000000;
78 static int check_test_flags(enum test_type t)
80 return test_flags & TEST_BIT(t);
83 static void test_hash_prealloc(int cpu)
88 start_time = time_get_ns();
89 for (i = 0; i < max_cnt; i++)
91 printf("%d:hash_map_perf pre-alloc %lld events per sec\n",
92 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
95 static int pre_test_lru_hash_lookup(int tasks)
97 int fd = map_fd[lru_hash_lookup_idx];
102 if (num_map_entries > lru_hash_lookup_test_entries)
103 lru_hash_lookup_test_entries = num_map_entries;
105 /* Populate the lru_hash_map for LRU_HASH_LOOKUP perf test.
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.
112 for (key = 0; key < lru_hash_lookup_test_entries; key++) {
113 ret = bpf_map_update_elem(fd, &key, &val, BPF_NOEXIST);
121 static void do_test_lru(enum test_type test, int cpu)
123 static int inner_lru_map_fds[MAX_NR_CPUS];
125 struct sockaddr_in6 in6 = { .sin6_family = AF_INET6 };
126 const char *test_name;
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
135 int outer_fd = map_fd[array_of_lru_hashs_idx];
136 unsigned int mycpu, mynode;
137 LIBBPF_OPTS(bpf_map_create_opts, opts,
138 .map_flags = BPF_F_NUMA_NODE,
141 assert(cpu < MAX_NR_CPUS);
143 ret = syscall(__NR_getcpu, &mycpu, &mynode, NULL);
146 opts.numa_node = mynode;
147 inner_lru_map_fds[cpu] =
148 bpf_map_create(BPF_MAP_TYPE_LRU_HASH,
149 test_map_names[INNER_LRU_HASH_PREALLOC],
152 inner_lru_hash_size, &opts);
153 if (inner_lru_map_fds[cpu] == -1) {
154 printf("cannot create BPF_MAP_TYPE_LRU_HASH %s(%d)\n",
155 strerror(errno), errno);
159 ret = bpf_map_update_elem(outer_fd, &cpu,
160 &inner_lru_map_fds[cpu],
163 printf("cannot update ARRAY_OF_LRU_HASHS with key:%u. %s(%d)\n",
164 cpu, strerror(errno), errno);
169 in6.sin6_addr.s6_addr16[0] = 0xdead;
170 in6.sin6_addr.s6_addr16[1] = 0xbeef;
172 if (test == LRU_HASH_PREALLOC) {
173 test_name = "lru_hash_map_perf";
174 in6.sin6_addr.s6_addr16[2] = 0;
175 } else if (test == NOCOMMON_LRU_HASH_PREALLOC) {
176 test_name = "nocommon_lru_hash_map_perf";
177 in6.sin6_addr.s6_addr16[2] = 1;
178 } else if (test == INNER_LRU_HASH_PREALLOC) {
179 test_name = "inner_lru_hash_map_perf";
180 in6.sin6_addr.s6_addr16[2] = 2;
181 } else if (test == LRU_HASH_LOOKUP) {
182 test_name = "lru_hash_lookup_perf";
183 in6.sin6_addr.s6_addr16[2] = 3;
184 in6.sin6_addr.s6_addr32[3] = 0;
189 start_time = time_get_ns();
190 for (i = 0; i < max_cnt; i++) {
191 ret = connect(-1, (const struct sockaddr *)&in6, sizeof(in6));
192 assert(ret == -1 && errno == EBADF);
193 if (in6.sin6_addr.s6_addr32[3] <
194 lru_hash_lookup_test_entries - 32)
195 in6.sin6_addr.s6_addr32[3] += 32;
197 in6.sin6_addr.s6_addr32[3] = 0;
199 printf("%d:%s pre-alloc %lld events per sec\n",
201 max_cnt * 1000000000ll / (time_get_ns() - start_time));
204 static void test_lru_hash_prealloc(int cpu)
206 do_test_lru(LRU_HASH_PREALLOC, cpu);
209 static void test_nocommon_lru_hash_prealloc(int cpu)
211 do_test_lru(NOCOMMON_LRU_HASH_PREALLOC, cpu);
214 static void test_inner_lru_hash_prealloc(int cpu)
216 do_test_lru(INNER_LRU_HASH_PREALLOC, cpu);
219 static void test_lru_hash_lookup(int cpu)
221 do_test_lru(LRU_HASH_LOOKUP, cpu);
224 static void test_percpu_hash_prealloc(int cpu)
229 start_time = time_get_ns();
230 for (i = 0; i < max_cnt; i++)
231 syscall(__NR_geteuid);
232 printf("%d:percpu_hash_map_perf pre-alloc %lld events per sec\n",
233 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
236 static void test_hash_kmalloc(int cpu)
241 start_time = time_get_ns();
242 for (i = 0; i < max_cnt; i++)
243 syscall(__NR_getgid);
244 printf("%d:hash_map_perf kmalloc %lld events per sec\n",
245 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
248 static void test_percpu_hash_kmalloc(int cpu)
253 start_time = time_get_ns();
254 for (i = 0; i < max_cnt; i++)
255 syscall(__NR_getegid);
256 printf("%d:percpu_hash_map_perf kmalloc %lld events per sec\n",
257 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
260 static void test_lpm_kmalloc(int cpu)
265 start_time = time_get_ns();
266 for (i = 0; i < max_cnt; i++)
267 syscall(__NR_gettid);
268 printf("%d:lpm_perf kmalloc %lld events per sec\n",
269 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time));
272 static void test_hash_lookup(int cpu)
277 start_time = time_get_ns();
278 for (i = 0; i < max_cnt; i++)
279 syscall(__NR_getpgid, 0);
280 printf("%d:hash_lookup %lld lookups per sec\n",
281 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time));
284 static void test_array_lookup(int cpu)
289 start_time = time_get_ns();
290 for (i = 0; i < max_cnt; i++)
291 syscall(__NR_getppid, 0);
292 printf("%d:array_lookup %lld lookups per sec\n",
293 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time));
296 typedef int (*pre_test_func)(int tasks);
297 const pre_test_func pre_test_funcs[] = {
298 [LRU_HASH_LOOKUP] = pre_test_lru_hash_lookup,
301 typedef void (*test_func)(int cpu);
302 const test_func test_funcs[] = {
303 [HASH_PREALLOC] = test_hash_prealloc,
304 [PERCPU_HASH_PREALLOC] = test_percpu_hash_prealloc,
305 [HASH_KMALLOC] = test_hash_kmalloc,
306 [PERCPU_HASH_KMALLOC] = test_percpu_hash_kmalloc,
307 [LRU_HASH_PREALLOC] = test_lru_hash_prealloc,
308 [NOCOMMON_LRU_HASH_PREALLOC] = test_nocommon_lru_hash_prealloc,
309 [LPM_KMALLOC] = test_lpm_kmalloc,
310 [HASH_LOOKUP] = test_hash_lookup,
311 [ARRAY_LOOKUP] = test_array_lookup,
312 [INNER_LRU_HASH_PREALLOC] = test_inner_lru_hash_prealloc,
313 [LRU_HASH_LOOKUP] = test_lru_hash_lookup,
316 static int pre_test(int tasks)
320 for (i = 0; i < NR_TESTS; i++) {
321 if (pre_test_funcs[i] && check_test_flags(i)) {
322 int ret = pre_test_funcs[i](tasks);
332 static void loop(int cpu)
338 CPU_SET(cpu, &cpuset);
339 sched_setaffinity(0, sizeof(cpuset), &cpuset);
341 for (i = 0; i < NR_TESTS; i++) {
342 if (check_test_flags(i))
347 static void run_perf_test(int tasks)
352 assert(!pre_test(tasks));
354 for (i = 0; i < tasks; i++) {
359 } else if (pid[i] == -1) {
360 printf("couldn't spawn #%d process\n", i);
364 for (i = 0; i < tasks; i++) {
367 assert(waitpid(pid[i], &status, 0) == pid[i]);
372 static void fill_lpm_trie(void)
374 struct bpf_lpm_trie_key *key;
375 unsigned long value = 0;
379 key = alloca(sizeof(*key) + 4);
382 for (i = 0; i < 512; ++i) {
383 key->prefixlen = rand() % 33;
384 key->data[0] = rand() & 0xff;
385 key->data[1] = rand() & 0xff;
386 key->data[2] = rand() & 0xff;
387 key->data[3] = rand() & 0xff;
388 r = bpf_map_update_elem(map_fd[hash_map_alloc_idx],
400 r = bpf_map_update_elem(map_fd[hash_map_alloc_idx], key, &value, 0);
404 static void fixup_map(struct bpf_object *obj)
409 bpf_object__for_each_map(map, obj) {
410 const char *name = bpf_map__name(map);
412 /* Only change the max_entries for the enabled test(s) */
413 for (i = 0; i < NR_TESTS; i++) {
414 if (!strcmp(test_map_names[i], name) &&
415 (check_test_flags(i))) {
416 bpf_map__resize(map, num_map_entries);
422 inner_lru_hash_size = num_map_entries;
425 int main(int argc, char **argv)
427 int nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
428 struct bpf_link *links[8];
429 struct bpf_program *prog;
430 struct bpf_object *obj;
436 test_flags = atoi(argv[1]) ? : test_flags;
439 nr_cpus = atoi(argv[2]) ? : nr_cpus;
442 num_map_entries = atoi(argv[3]);
445 max_cnt = atoi(argv[4]);
447 snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
448 obj = bpf_object__open_file(filename, NULL);
449 if (libbpf_get_error(obj)) {
450 fprintf(stderr, "ERROR: opening BPF object file failed\n");
454 map = bpf_object__find_map_by_name(obj, "inner_lru_hash_map");
455 if (libbpf_get_error(map)) {
456 fprintf(stderr, "ERROR: finding a map in obj file failed\n");
460 inner_lru_hash_size = bpf_map__max_entries(map);
461 if (!inner_lru_hash_size) {
462 fprintf(stderr, "ERROR: failed to get map attribute\n");
466 /* resize BPF map prior to loading */
467 if (num_map_entries > 0)
470 /* load BPF program */
471 if (bpf_object__load(obj)) {
472 fprintf(stderr, "ERROR: loading BPF object file failed\n");
476 map_fd[0] = bpf_object__find_map_fd_by_name(obj, "array_of_lru_hashs");
477 map_fd[1] = bpf_object__find_map_fd_by_name(obj, "hash_map_alloc");
478 map_fd[2] = bpf_object__find_map_fd_by_name(obj, "lru_hash_lookup_map");
479 if (map_fd[0] < 0 || map_fd[1] < 0 || map_fd[2] < 0) {
480 fprintf(stderr, "ERROR: finding a map in obj file failed\n");
484 bpf_object__for_each_program(prog, obj) {
485 links[i] = bpf_program__attach(prog);
486 if (libbpf_get_error(links[i])) {
487 fprintf(stderr, "ERROR: bpf_program__attach failed\n");
496 run_perf_test(nr_cpus);
499 for (i--; i >= 0; i--)
500 bpf_link__destroy(links[i]);
502 bpf_object__close(obj);