2 * Copyright (c) 2016 Facebook
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
9 #include <linux/types.h>
12 #include <linux/bpf.h>
19 #include <sys/resource.h>
27 #define min(a, b) ((a) < (b) ? (a) : (b))
29 # define offsetof(TYPE, MEMBER) ((size_t)&((TYPE *)0)->MEMBER)
31 #define container_of(ptr, type, member) ({ \
32 const typeof( ((type *)0)->member ) *__mptr = (ptr); \
33 (type *)( (char *)__mptr - offsetof(type,member) );})
36 static unsigned long long *dist_keys
;
37 static unsigned int dist_key_counts
;
40 struct list_head
*next
, *prev
;
43 static inline void INIT_LIST_HEAD(struct list_head
*list
)
49 static inline int list_empty(const struct list_head
*head
)
51 return head
->next
== head
;
54 static inline void __list_add(struct list_head
*new,
55 struct list_head
*prev
,
56 struct list_head
*next
)
64 static inline void list_add(struct list_head
*new, struct list_head
*head
)
66 __list_add(new, head
, head
->next
);
69 static inline void __list_del(struct list_head
*prev
, struct list_head
*next
)
75 static inline void __list_del_entry(struct list_head
*entry
)
77 __list_del(entry
->prev
, entry
->next
);
80 static inline void list_move(struct list_head
*list
, struct list_head
*head
)
82 __list_del_entry(list
);
86 #define list_entry(ptr, type, member) \
87 container_of(ptr, type, member)
89 #define list_last_entry(ptr, type, member) \
90 list_entry((ptr)->prev, type, member)
92 struct pfect_lru_node
{
93 struct list_head list
;
94 unsigned long long key
;
98 struct list_head list
;
99 struct pfect_lru_node
*free_nodes
;
100 unsigned int cur_size
;
101 unsigned int lru_size
;
102 unsigned int nr_unique
;
103 unsigned int nr_misses
;
108 static void pfect_lru_init(struct pfect_lru
*lru
, unsigned int lru_size
,
109 unsigned int nr_possible_elems
)
111 lru
->map_fd
= bpf_create_map(BPF_MAP_TYPE_HASH
,
112 sizeof(unsigned long long),
113 sizeof(struct pfect_lru_node
*),
114 nr_possible_elems
, 0);
115 assert(lru
->map_fd
!= -1);
117 lru
->free_nodes
= malloc(lru_size
* sizeof(struct pfect_lru_node
));
118 assert(lru
->free_nodes
);
120 INIT_LIST_HEAD(&lru
->list
);
122 lru
->lru_size
= lru_size
;
123 lru
->nr_unique
= lru
->nr_misses
= lru
->total
= 0;
126 static void pfect_lru_destroy(struct pfect_lru
*lru
)
129 free(lru
->free_nodes
);
132 static int pfect_lru_lookup_or_insert(struct pfect_lru
*lru
,
133 unsigned long long key
)
135 struct pfect_lru_node
*node
= NULL
;
139 if (!bpf_map_lookup_elem(lru
->map_fd
, &key
, &node
)) {
141 list_move(&node
->list
, &lru
->list
);
147 if (lru
->cur_size
< lru
->lru_size
) {
148 node
= &lru
->free_nodes
[lru
->cur_size
++];
149 INIT_LIST_HEAD(&node
->list
);
151 struct pfect_lru_node
*null_node
= NULL
;
153 node
= list_last_entry(&lru
->list
,
154 struct pfect_lru_node
,
156 bpf_map_update_elem(lru
->map_fd
, &node
->key
, &null_node
, BPF_EXIST
);
160 list_move(&node
->list
, &lru
->list
);
164 assert(!bpf_map_update_elem(lru
->map_fd
, &key
, &node
, BPF_EXIST
));
167 assert(!bpf_map_update_elem(lru
->map_fd
, &key
, &node
, BPF_NOEXIST
));
173 static unsigned int read_keys(const char *dist_file
,
174 unsigned long long **keys
)
177 unsigned long long *retkeys
;
178 unsigned int counts
= 0;
183 dist_fd
= open(dist_file
, 0);
184 assert(dist_fd
!= -1);
186 assert(fstat(dist_fd
, &fst
) == 0);
187 b
= malloc(fst
.st_size
);
190 assert(read(dist_fd
, b
, fst
.st_size
) == fst
.st_size
);
192 for (i
= 0; i
< fst
.st_size
; i
++) {
196 counts
++; /* in case the last line has no \n */
198 retkeys
= malloc(counts
* sizeof(unsigned long long));
202 for (l
= strtok(b
, "\n"); l
; l
= strtok(NULL
, "\n"))
203 retkeys
[counts
++] = strtoull(l
, NULL
, 10);
211 static int create_map(int map_type
, int map_flags
, unsigned int size
)
215 map_fd
= bpf_create_map(map_type
, sizeof(unsigned long long),
216 sizeof(unsigned long long), size
, map_flags
);
219 perror("bpf_create_map");
224 static int sched_next_online(int pid
, int next_to_try
)
228 if (next_to_try
== nr_cpus
)
231 while (next_to_try
< nr_cpus
) {
233 CPU_SET(next_to_try
++, &cpuset
);
234 if (!sched_setaffinity(pid
, sizeof(cpuset
), &cpuset
))
241 static void run_parallel(unsigned int tasks
, void (*fn
)(int i
, void *data
),
244 int next_sched_cpu
= 0;
248 for (i
= 0; i
< tasks
; i
++) {
251 next_sched_cpu
= sched_next_online(0, next_sched_cpu
);
254 } else if (pid
[i
] == -1) {
255 printf("couldn't spawn #%d process\n", i
);
258 /* It is mostly redundant and just allow the parent
259 * process to update next_shced_cpu for the next child
262 next_sched_cpu
= sched_next_online(pid
[i
], next_sched_cpu
);
264 for (i
= 0; i
< tasks
; i
++) {
267 assert(waitpid(pid
[i
], &status
, 0) == pid
[i
]);
272 static void do_test_lru_dist(int task
, void *data
)
274 unsigned int nr_misses
= 0;
275 struct pfect_lru pfect_lru
;
276 unsigned long long key
, value
= 1234;
279 unsigned int lru_map_fd
= ((unsigned int *)data
)[0];
280 unsigned int lru_size
= ((unsigned int *)data
)[1];
281 unsigned long long key_offset
= task
* dist_key_counts
;
283 pfect_lru_init(&pfect_lru
, lru_size
, dist_key_counts
);
285 for (i
= 0; i
< dist_key_counts
; i
++) {
286 key
= dist_keys
[i
] + key_offset
;
288 pfect_lru_lookup_or_insert(&pfect_lru
, key
);
290 if (!bpf_map_lookup_elem(lru_map_fd
, &key
, &value
))
293 if (bpf_map_update_elem(lru_map_fd
, &key
, &value
, BPF_NOEXIST
)) {
294 printf("bpf_map_update_elem(lru_map_fd, %llu): errno:%d\n",
302 printf(" task:%d BPF LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
303 task
, pfect_lru
.nr_unique
, dist_key_counts
, nr_misses
,
305 printf(" task:%d Perfect LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
306 task
, pfect_lru
.nr_unique
, pfect_lru
.total
,
307 pfect_lru
.nr_misses
, pfect_lru
.total
);
309 pfect_lru_destroy(&pfect_lru
);
313 static void test_parallel_lru_dist(int map_type
, int map_flags
,
314 int nr_tasks
, unsigned int lru_size
)
319 printf("%s (map_type:%d map_flags:0x%X):\n", __func__
, map_type
,
322 if (map_flags
& BPF_F_NO_COMMON_LRU
)
323 lru_map_fd
= create_map(map_type
, map_flags
,
326 lru_map_fd
= create_map(map_type
, map_flags
,
327 nr_tasks
* lru_size
);
328 assert(lru_map_fd
!= -1);
330 child_data
[0] = lru_map_fd
;
331 child_data
[1] = lru_size
;
333 run_parallel(nr_tasks
, do_test_lru_dist
, child_data
);
338 static void test_lru_loss0(int map_type
, int map_flags
)
340 unsigned long long key
, value
[nr_cpus
];
341 unsigned int old_unused_losses
= 0;
342 unsigned int new_unused_losses
= 0;
343 unsigned int used_losses
= 0;
346 printf("%s (map_type:%d map_flags:0x%X): ", __func__
, map_type
,
349 assert(sched_next_online(0, 0) != -1);
351 if (map_flags
& BPF_F_NO_COMMON_LRU
)
352 map_fd
= create_map(map_type
, map_flags
, 900 * nr_cpus
);
354 map_fd
= create_map(map_type
, map_flags
, 900);
356 assert(map_fd
!= -1);
360 for (key
= 1; key
<= 1000; key
++) {
361 int start_key
, end_key
;
363 assert(bpf_map_update_elem(map_fd
, &key
, value
, BPF_NOEXIST
) == 0);
366 end_key
= min(key
, 900);
368 while (start_key
<= end_key
) {
369 bpf_map_lookup_elem(map_fd
, &start_key
, value
);
374 for (key
= 1; key
<= 1000; key
++) {
375 if (bpf_map_lookup_elem(map_fd
, &key
, value
)) {
387 printf("older-elem-losses:%d(/100) active-elem-losses:%d(/800) "
388 "newer-elem-losses:%d(/100)\n",
389 old_unused_losses
, used_losses
, new_unused_losses
);
392 static void test_lru_loss1(int map_type
, int map_flags
)
394 unsigned long long key
, value
[nr_cpus
];
396 unsigned int nr_losses
= 0;
398 printf("%s (map_type:%d map_flags:0x%X): ", __func__
, map_type
,
401 assert(sched_next_online(0, 0) != -1);
403 if (map_flags
& BPF_F_NO_COMMON_LRU
)
404 map_fd
= create_map(map_type
, map_flags
, 1000 * nr_cpus
);
406 map_fd
= create_map(map_type
, map_flags
, 1000);
408 assert(map_fd
!= -1);
412 for (key
= 1; key
<= 1000; key
++)
413 assert(!bpf_map_update_elem(map_fd
, &key
, value
, BPF_NOEXIST
));
415 for (key
= 1; key
<= 1000; key
++) {
416 if (bpf_map_lookup_elem(map_fd
, &key
, value
))
422 printf("nr_losses:%d(/1000)\n", nr_losses
);
425 static void do_test_parallel_lru_loss(int task
, void *data
)
427 const unsigned int nr_stable_elems
= 1000;
428 const unsigned int nr_repeats
= 100000;
430 int map_fd
= *(int *)data
;
431 unsigned long long stable_base
;
432 unsigned long long key
, value
[nr_cpus
];
433 unsigned long long next_ins_key
;
434 unsigned int nr_losses
= 0;
437 stable_base
= task
* nr_repeats
* 2 + 1;
438 next_ins_key
= stable_base
;
440 for (i
= 0; i
< nr_stable_elems
; i
++) {
441 assert(bpf_map_update_elem(map_fd
, &next_ins_key
, value
,
446 for (i
= 0; i
< nr_repeats
; i
++) {
452 key
= rn
% nr_stable_elems
+ stable_base
;
453 bpf_map_lookup_elem(map_fd
, &key
, value
);
455 bpf_map_update_elem(map_fd
, &next_ins_key
, value
,
462 for (i
= 0; i
< nr_stable_elems
; i
++) {
463 if (bpf_map_lookup_elem(map_fd
, &key
, value
))
468 printf(" task:%d nr_losses:%u\n", task
, nr_losses
);
471 static void test_parallel_lru_loss(int map_type
, int map_flags
, int nr_tasks
)
475 printf("%s (map_type:%d map_flags:0x%X):\n", __func__
, map_type
,
478 /* Give 20% more than the active working set */
479 if (map_flags
& BPF_F_NO_COMMON_LRU
)
480 map_fd
= create_map(map_type
, map_flags
,
481 nr_cpus
* (1000 + 200));
483 map_fd
= create_map(map_type
, map_flags
,
484 nr_tasks
* (1000 + 200));
486 assert(map_fd
!= -1);
488 run_parallel(nr_tasks
, do_test_parallel_lru_loss
, &map_fd
);
493 int main(int argc
, char **argv
)
495 struct rlimit r
= {RLIM_INFINITY
, RLIM_INFINITY
};
496 int map_flags
[] = {0, BPF_F_NO_COMMON_LRU
};
497 const char *dist_file
;
503 printf("Usage: %s <dist-file> <lru-size> <nr-tasks>\n",
509 lru_size
= atoi(argv
[2]);
510 nr_tasks
= atoi(argv
[3]);
512 setbuf(stdout
, NULL
);
514 assert(!setrlimit(RLIMIT_MEMLOCK
, &r
));
518 nr_cpus
= bpf_num_possible_cpus();
519 assert(nr_cpus
!= -1);
520 printf("nr_cpus:%d\n\n", nr_cpus
);
522 nr_tasks
= min(nr_tasks
, nr_cpus
);
524 dist_key_counts
= read_keys(dist_file
, &dist_keys
);
525 if (!dist_key_counts
) {
526 printf("%s has no key\n", dist_file
);
530 for (f
= 0; f
< sizeof(map_flags
) / sizeof(*map_flags
); f
++) {
531 test_lru_loss0(BPF_MAP_TYPE_LRU_HASH
, map_flags
[f
]);
532 test_lru_loss1(BPF_MAP_TYPE_LRU_HASH
, map_flags
[f
]);
533 test_parallel_lru_loss(BPF_MAP_TYPE_LRU_HASH
, map_flags
[f
],
535 test_parallel_lru_dist(BPF_MAP_TYPE_LRU_HASH
, map_flags
[f
],