1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (c) 2016 Facebook
6 #include <linux/types.h>
16 #include <sys/resource.h>
24 #define min(a, b) ((a) < (b) ? (a) : (b))
26 # define offsetof(TYPE, MEMBER) ((size_t)&((TYPE *)0)->MEMBER)
28 #define container_of(ptr, type, member) ({ \
29 const typeof( ((type *)0)->member ) *__mptr = (ptr); \
30 (type *)( (char *)__mptr - offsetof(type,member) );})
33 static unsigned long long *dist_keys;
34 static unsigned int dist_key_counts;
37 struct list_head *next, *prev;
40 static inline void INIT_LIST_HEAD(struct list_head *list)
46 static inline int list_empty(const struct list_head *head)
48 return head->next == head;
51 static inline void __list_add(struct list_head *new,
52 struct list_head *prev,
53 struct list_head *next)
61 static inline void list_add(struct list_head *new, struct list_head *head)
63 __list_add(new, head, head->next);
66 static inline void __list_del(struct list_head *prev, struct list_head *next)
72 static inline void __list_del_entry(struct list_head *entry)
74 __list_del(entry->prev, entry->next);
77 static inline void list_move(struct list_head *list, struct list_head *head)
79 __list_del_entry(list);
83 #define list_entry(ptr, type, member) \
84 container_of(ptr, type, member)
86 #define list_last_entry(ptr, type, member) \
87 list_entry((ptr)->prev, type, member)
89 struct pfect_lru_node {
90 struct list_head list;
91 unsigned long long key;
95 struct list_head list;
96 struct pfect_lru_node *free_nodes;
97 unsigned int cur_size;
98 unsigned int lru_size;
99 unsigned int nr_unique;
100 unsigned int nr_misses;
105 static void pfect_lru_init(struct pfect_lru *lru, unsigned int lru_size,
106 unsigned int nr_possible_elems)
108 lru->map_fd = bpf_create_map(BPF_MAP_TYPE_HASH,
109 sizeof(unsigned long long),
110 sizeof(struct pfect_lru_node *),
111 nr_possible_elems, 0);
112 assert(lru->map_fd != -1);
114 lru->free_nodes = malloc(lru_size * sizeof(struct pfect_lru_node));
115 assert(lru->free_nodes);
117 INIT_LIST_HEAD(&lru->list);
119 lru->lru_size = lru_size;
120 lru->nr_unique = lru->nr_misses = lru->total = 0;
123 static void pfect_lru_destroy(struct pfect_lru *lru)
126 free(lru->free_nodes);
129 static int pfect_lru_lookup_or_insert(struct pfect_lru *lru,
130 unsigned long long key)
132 struct pfect_lru_node *node = NULL;
136 if (!bpf_map_lookup_elem(lru->map_fd, &key, &node)) {
138 list_move(&node->list, &lru->list);
144 if (lru->cur_size < lru->lru_size) {
145 node = &lru->free_nodes[lru->cur_size++];
146 INIT_LIST_HEAD(&node->list);
148 struct pfect_lru_node *null_node = NULL;
150 node = list_last_entry(&lru->list,
151 struct pfect_lru_node,
153 bpf_map_update_elem(lru->map_fd, &node->key, &null_node, BPF_EXIST);
157 list_move(&node->list, &lru->list);
161 assert(!bpf_map_update_elem(lru->map_fd, &key, &node, BPF_EXIST));
164 assert(!bpf_map_update_elem(lru->map_fd, &key, &node, BPF_NOEXIST));
170 static unsigned int read_keys(const char *dist_file,
171 unsigned long long **keys)
174 unsigned long long *retkeys;
175 unsigned int counts = 0;
180 dist_fd = open(dist_file, 0);
181 assert(dist_fd != -1);
183 assert(fstat(dist_fd, &fst) == 0);
184 b = malloc(fst.st_size);
187 assert(read(dist_fd, b, fst.st_size) == fst.st_size);
189 for (i = 0; i < fst.st_size; i++) {
193 counts++; /* in case the last line has no \n */
195 retkeys = malloc(counts * sizeof(unsigned long long));
199 for (l = strtok(b, "\n"); l; l = strtok(NULL, "\n"))
200 retkeys[counts++] = strtoull(l, NULL, 10);
208 static int create_map(int map_type, int map_flags, unsigned int size)
212 map_fd = bpf_create_map(map_type, sizeof(unsigned long long),
213 sizeof(unsigned long long), size, map_flags);
216 perror("bpf_create_map");
221 static int sched_next_online(int pid, int next_to_try)
225 if (next_to_try == nr_cpus)
228 while (next_to_try < nr_cpus) {
230 CPU_SET(next_to_try++, &cpuset);
231 if (!sched_setaffinity(pid, sizeof(cpuset), &cpuset))
238 static void run_parallel(unsigned int tasks, void (*fn)(int i, void *data),
241 int next_sched_cpu = 0;
245 for (i = 0; i < tasks; i++) {
248 next_sched_cpu = sched_next_online(0, next_sched_cpu);
251 } else if (pid[i] == -1) {
252 printf("couldn't spawn #%d process\n", i);
255 /* It is mostly redundant and just allow the parent
256 * process to update next_shced_cpu for the next child
259 next_sched_cpu = sched_next_online(pid[i], next_sched_cpu);
261 for (i = 0; i < tasks; i++) {
264 assert(waitpid(pid[i], &status, 0) == pid[i]);
269 static void do_test_lru_dist(int task, void *data)
271 unsigned int nr_misses = 0;
272 struct pfect_lru pfect_lru;
273 unsigned long long key, value = 1234;
276 unsigned int lru_map_fd = ((unsigned int *)data)[0];
277 unsigned int lru_size = ((unsigned int *)data)[1];
278 unsigned long long key_offset = task * dist_key_counts;
280 pfect_lru_init(&pfect_lru, lru_size, dist_key_counts);
282 for (i = 0; i < dist_key_counts; i++) {
283 key = dist_keys[i] + key_offset;
285 pfect_lru_lookup_or_insert(&pfect_lru, key);
287 if (!bpf_map_lookup_elem(lru_map_fd, &key, &value))
290 if (bpf_map_update_elem(lru_map_fd, &key, &value, BPF_NOEXIST)) {
291 printf("bpf_map_update_elem(lru_map_fd, %llu): errno:%d\n",
299 printf(" task:%d BPF LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
300 task, pfect_lru.nr_unique, dist_key_counts, nr_misses,
302 printf(" task:%d Perfect LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
303 task, pfect_lru.nr_unique, pfect_lru.total,
304 pfect_lru.nr_misses, pfect_lru.total);
306 pfect_lru_destroy(&pfect_lru);
310 static void test_parallel_lru_dist(int map_type, int map_flags,
311 int nr_tasks, unsigned int lru_size)
316 printf("%s (map_type:%d map_flags:0x%X):\n", __func__, map_type,
319 if (map_flags & BPF_F_NO_COMMON_LRU)
320 lru_map_fd = create_map(map_type, map_flags,
323 lru_map_fd = create_map(map_type, map_flags,
324 nr_tasks * lru_size);
325 assert(lru_map_fd != -1);
327 child_data[0] = lru_map_fd;
328 child_data[1] = lru_size;
330 run_parallel(nr_tasks, do_test_lru_dist, child_data);
335 static void test_lru_loss0(int map_type, int map_flags)
337 unsigned long long key, value[nr_cpus];
338 unsigned int old_unused_losses = 0;
339 unsigned int new_unused_losses = 0;
340 unsigned int used_losses = 0;
343 printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
346 assert(sched_next_online(0, 0) != -1);
348 if (map_flags & BPF_F_NO_COMMON_LRU)
349 map_fd = create_map(map_type, map_flags, 900 * nr_cpus);
351 map_fd = create_map(map_type, map_flags, 900);
353 assert(map_fd != -1);
357 for (key = 1; key <= 1000; key++) {
358 int start_key, end_key;
360 assert(bpf_map_update_elem(map_fd, &key, value, BPF_NOEXIST) == 0);
363 end_key = min(key, 900);
365 while (start_key <= end_key) {
366 bpf_map_lookup_elem(map_fd, &start_key, value);
371 for (key = 1; key <= 1000; key++) {
372 if (bpf_map_lookup_elem(map_fd, &key, value)) {
384 printf("older-elem-losses:%d(/100) active-elem-losses:%d(/800) "
385 "newer-elem-losses:%d(/100)\n",
386 old_unused_losses, used_losses, new_unused_losses);
389 static void test_lru_loss1(int map_type, int map_flags)
391 unsigned long long key, value[nr_cpus];
393 unsigned int nr_losses = 0;
395 printf("%s (map_type:%d map_flags:0x%X): ", __func__, map_type,
398 assert(sched_next_online(0, 0) != -1);
400 if (map_flags & BPF_F_NO_COMMON_LRU)
401 map_fd = create_map(map_type, map_flags, 1000 * nr_cpus);
403 map_fd = create_map(map_type, map_flags, 1000);
405 assert(map_fd != -1);
409 for (key = 1; key <= 1000; key++)
410 assert(!bpf_map_update_elem(map_fd, &key, value, BPF_NOEXIST));
412 for (key = 1; key <= 1000; key++) {
413 if (bpf_map_lookup_elem(map_fd, &key, value))
419 printf("nr_losses:%d(/1000)\n", nr_losses);
422 static void do_test_parallel_lru_loss(int task, void *data)
424 const unsigned int nr_stable_elems = 1000;
425 const unsigned int nr_repeats = 100000;
427 int map_fd = *(int *)data;
428 unsigned long long stable_base;
429 unsigned long long key, value[nr_cpus];
430 unsigned long long next_ins_key;
431 unsigned int nr_losses = 0;
434 stable_base = task * nr_repeats * 2 + 1;
435 next_ins_key = stable_base;
437 for (i = 0; i < nr_stable_elems; i++) {
438 assert(bpf_map_update_elem(map_fd, &next_ins_key, value,
443 for (i = 0; i < nr_repeats; i++) {
449 key = rn % nr_stable_elems + stable_base;
450 bpf_map_lookup_elem(map_fd, &key, value);
452 bpf_map_update_elem(map_fd, &next_ins_key, value,
459 for (i = 0; i < nr_stable_elems; i++) {
460 if (bpf_map_lookup_elem(map_fd, &key, value))
465 printf(" task:%d nr_losses:%u\n", task, nr_losses);
468 static void test_parallel_lru_loss(int map_type, int map_flags, int nr_tasks)
472 printf("%s (map_type:%d map_flags:0x%X):\n", __func__, map_type,
475 /* Give 20% more than the active working set */
476 if (map_flags & BPF_F_NO_COMMON_LRU)
477 map_fd = create_map(map_type, map_flags,
478 nr_cpus * (1000 + 200));
480 map_fd = create_map(map_type, map_flags,
481 nr_tasks * (1000 + 200));
483 assert(map_fd != -1);
485 run_parallel(nr_tasks, do_test_parallel_lru_loss, &map_fd);
490 int main(int argc, char **argv)
492 int map_flags[] = {0, BPF_F_NO_COMMON_LRU};
493 const char *dist_file;
499 printf("Usage: %s <dist-file> <lru-size> <nr-tasks>\n",
505 lru_size = atoi(argv[2]);
506 nr_tasks = atoi(argv[3]);
508 setbuf(stdout, NULL);
512 nr_cpus = bpf_num_possible_cpus();
513 assert(nr_cpus != -1);
514 printf("nr_cpus:%d\n\n", nr_cpus);
516 nr_tasks = min(nr_tasks, nr_cpus);
518 dist_key_counts = read_keys(dist_file, &dist_keys);
519 if (!dist_key_counts) {
520 printf("%s has no key\n", dist_file);
524 for (f = 0; f < sizeof(map_flags) / sizeof(*map_flags); f++) {
525 test_lru_loss0(BPF_MAP_TYPE_LRU_HASH, map_flags[f]);
526 test_lru_loss1(BPF_MAP_TYPE_LRU_HASH, map_flags[f]);
527 test_parallel_lru_loss(BPF_MAP_TYPE_LRU_HASH, map_flags[f],
529 test_parallel_lru_dist(BPF_MAP_TYPE_LRU_HASH, map_flags[f],