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))
28 #define offsetof(TYPE, MEMBER) ((size_t)&((TYPE *)0)->MEMBER)
29 #define container_of(ptr, type, member) ({ \
30 const typeof( ((type *)0)->member ) *__mptr = (ptr); \
31 (type *)( (char *)__mptr - offsetof(type,member) );})
34 static unsigned long long *dist_keys
;
35 static unsigned int dist_key_counts
;
38 struct list_head
*next
, *prev
;
41 static inline void INIT_LIST_HEAD(struct list_head
*list
)
47 static inline int list_empty(const struct list_head
*head
)
49 return head
->next
== head
;
52 static inline void __list_add(struct list_head
*new,
53 struct list_head
*prev
,
54 struct list_head
*next
)
62 static inline void list_add(struct list_head
*new, struct list_head
*head
)
64 __list_add(new, head
, head
->next
);
67 static inline void __list_del(struct list_head
*prev
, struct list_head
*next
)
73 static inline void __list_del_entry(struct list_head
*entry
)
75 __list_del(entry
->prev
, entry
->next
);
78 static inline void list_move(struct list_head
*list
, struct list_head
*head
)
80 __list_del_entry(list
);
84 #define list_entry(ptr, type, member) \
85 container_of(ptr, type, member)
87 #define list_last_entry(ptr, type, member) \
88 list_entry((ptr)->prev, type, member)
90 struct pfect_lru_node
{
91 struct list_head list
;
92 unsigned long long key
;
96 struct list_head list
;
97 struct pfect_lru_node
*free_nodes
;
98 unsigned int cur_size
;
99 unsigned int lru_size
;
100 unsigned int nr_unique
;
101 unsigned int nr_misses
;
106 static void pfect_lru_init(struct pfect_lru
*lru
, unsigned int lru_size
,
107 unsigned int nr_possible_elems
)
109 lru
->map_fd
= bpf_create_map(BPF_MAP_TYPE_HASH
,
110 sizeof(unsigned long long),
111 sizeof(struct pfect_lru_node
*),
112 nr_possible_elems
, 0);
113 assert(lru
->map_fd
!= -1);
115 lru
->free_nodes
= malloc(lru_size
* sizeof(struct pfect_lru_node
));
116 assert(lru
->free_nodes
);
118 INIT_LIST_HEAD(&lru
->list
);
120 lru
->lru_size
= lru_size
;
121 lru
->nr_unique
= lru
->nr_misses
= lru
->total
= 0;
124 static void pfect_lru_destroy(struct pfect_lru
*lru
)
127 free(lru
->free_nodes
);
130 static int pfect_lru_lookup_or_insert(struct pfect_lru
*lru
,
131 unsigned long long key
)
133 struct pfect_lru_node
*node
= NULL
;
137 if (!bpf_map_lookup_elem(lru
->map_fd
, &key
, &node
)) {
139 list_move(&node
->list
, &lru
->list
);
145 if (lru
->cur_size
< lru
->lru_size
) {
146 node
= &lru
->free_nodes
[lru
->cur_size
++];
147 INIT_LIST_HEAD(&node
->list
);
149 struct pfect_lru_node
*null_node
= NULL
;
151 node
= list_last_entry(&lru
->list
,
152 struct pfect_lru_node
,
154 bpf_map_update_elem(lru
->map_fd
, &node
->key
, &null_node
, BPF_EXIST
);
158 list_move(&node
->list
, &lru
->list
);
162 assert(!bpf_map_update_elem(lru
->map_fd
, &key
, &node
, BPF_EXIST
));
165 assert(!bpf_map_update_elem(lru
->map_fd
, &key
, &node
, BPF_NOEXIST
));
171 static unsigned int read_keys(const char *dist_file
,
172 unsigned long long **keys
)
175 unsigned long long *retkeys
;
176 unsigned int counts
= 0;
181 dist_fd
= open(dist_file
, 0);
182 assert(dist_fd
!= -1);
184 assert(fstat(dist_fd
, &fst
) == 0);
185 b
= malloc(fst
.st_size
);
188 assert(read(dist_fd
, b
, fst
.st_size
) == fst
.st_size
);
190 for (i
= 0; i
< fst
.st_size
; i
++) {
194 counts
++; /* in case the last line has no \n */
196 retkeys
= malloc(counts
* sizeof(unsigned long long));
200 for (l
= strtok(b
, "\n"); l
; l
= strtok(NULL
, "\n"))
201 retkeys
[counts
++] = strtoull(l
, NULL
, 10);
209 static int create_map(int map_type
, int map_flags
, unsigned int size
)
213 map_fd
= bpf_create_map(map_type
, sizeof(unsigned long long),
214 sizeof(unsigned long long), size
, map_flags
);
217 perror("bpf_create_map");
222 static int sched_next_online(int pid
, int next_to_try
)
226 if (next_to_try
== nr_cpus
)
229 while (next_to_try
< nr_cpus
) {
231 CPU_SET(next_to_try
++, &cpuset
);
232 if (!sched_setaffinity(pid
, sizeof(cpuset
), &cpuset
))
239 static void run_parallel(unsigned int tasks
, void (*fn
)(int i
, void *data
),
242 int next_sched_cpu
= 0;
246 for (i
= 0; i
< tasks
; i
++) {
249 next_sched_cpu
= sched_next_online(0, next_sched_cpu
);
252 } else if (pid
[i
] == -1) {
253 printf("couldn't spawn #%d process\n", i
);
256 /* It is mostly redundant and just allow the parent
257 * process to update next_shced_cpu for the next child
260 next_sched_cpu
= sched_next_online(pid
[i
], next_sched_cpu
);
262 for (i
= 0; i
< tasks
; i
++) {
265 assert(waitpid(pid
[i
], &status
, 0) == pid
[i
]);
270 static void do_test_lru_dist(int task
, void *data
)
272 unsigned int nr_misses
= 0;
273 struct pfect_lru pfect_lru
;
274 unsigned long long key
, value
= 1234;
277 unsigned int lru_map_fd
= ((unsigned int *)data
)[0];
278 unsigned int lru_size
= ((unsigned int *)data
)[1];
279 unsigned long long key_offset
= task
* dist_key_counts
;
281 pfect_lru_init(&pfect_lru
, lru_size
, dist_key_counts
);
283 for (i
= 0; i
< dist_key_counts
; i
++) {
284 key
= dist_keys
[i
] + key_offset
;
286 pfect_lru_lookup_or_insert(&pfect_lru
, key
);
288 if (!bpf_map_lookup_elem(lru_map_fd
, &key
, &value
))
291 if (bpf_map_update_elem(lru_map_fd
, &key
, &value
, BPF_NOEXIST
)) {
292 printf("bpf_map_update_elem(lru_map_fd, %llu): errno:%d\n",
300 printf(" task:%d BPF LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
301 task
, pfect_lru
.nr_unique
, dist_key_counts
, nr_misses
,
303 printf(" task:%d Perfect LRU: nr_unique:%u(/%u) nr_misses:%u(/%u)\n",
304 task
, pfect_lru
.nr_unique
, pfect_lru
.total
,
305 pfect_lru
.nr_misses
, pfect_lru
.total
);
307 pfect_lru_destroy(&pfect_lru
);
311 static void test_parallel_lru_dist(int map_type
, int map_flags
,
312 int nr_tasks
, unsigned int lru_size
)
317 printf("%s (map_type:%d map_flags:0x%X):\n", __func__
, map_type
,
320 if (map_flags
& BPF_F_NO_COMMON_LRU
)
321 lru_map_fd
= create_map(map_type
, map_flags
,
324 lru_map_fd
= create_map(map_type
, map_flags
,
325 nr_tasks
* lru_size
);
326 assert(lru_map_fd
!= -1);
328 child_data
[0] = lru_map_fd
;
329 child_data
[1] = lru_size
;
331 run_parallel(nr_tasks
, do_test_lru_dist
, child_data
);
336 static void test_lru_loss0(int map_type
, int map_flags
)
338 unsigned long long key
, value
[nr_cpus
];
339 unsigned int old_unused_losses
= 0;
340 unsigned int new_unused_losses
= 0;
341 unsigned int used_losses
= 0;
344 printf("%s (map_type:%d map_flags:0x%X): ", __func__
, map_type
,
347 assert(sched_next_online(0, 0) != -1);
349 if (map_flags
& BPF_F_NO_COMMON_LRU
)
350 map_fd
= create_map(map_type
, map_flags
, 900 * nr_cpus
);
352 map_fd
= create_map(map_type
, map_flags
, 900);
354 assert(map_fd
!= -1);
358 for (key
= 1; key
<= 1000; key
++) {
359 int start_key
, end_key
;
361 assert(bpf_map_update_elem(map_fd
, &key
, value
, BPF_NOEXIST
) == 0);
364 end_key
= min(key
, 900);
366 while (start_key
<= end_key
) {
367 bpf_map_lookup_elem(map_fd
, &start_key
, value
);
372 for (key
= 1; key
<= 1000; key
++) {
373 if (bpf_map_lookup_elem(map_fd
, &key
, value
)) {
385 printf("older-elem-losses:%d(/100) active-elem-losses:%d(/800) "
386 "newer-elem-losses:%d(/100)\n",
387 old_unused_losses
, used_losses
, new_unused_losses
);
390 static void test_lru_loss1(int map_type
, int map_flags
)
392 unsigned long long key
, value
[nr_cpus
];
394 unsigned int nr_losses
= 0;
396 printf("%s (map_type:%d map_flags:0x%X): ", __func__
, map_type
,
399 assert(sched_next_online(0, 0) != -1);
401 if (map_flags
& BPF_F_NO_COMMON_LRU
)
402 map_fd
= create_map(map_type
, map_flags
, 1000 * nr_cpus
);
404 map_fd
= create_map(map_type
, map_flags
, 1000);
406 assert(map_fd
!= -1);
410 for (key
= 1; key
<= 1000; key
++)
411 assert(!bpf_map_update_elem(map_fd
, &key
, value
, BPF_NOEXIST
));
413 for (key
= 1; key
<= 1000; key
++) {
414 if (bpf_map_lookup_elem(map_fd
, &key
, value
))
420 printf("nr_losses:%d(/1000)\n", nr_losses
);
423 static void do_test_parallel_lru_loss(int task
, void *data
)
425 const unsigned int nr_stable_elems
= 1000;
426 const unsigned int nr_repeats
= 100000;
428 int map_fd
= *(int *)data
;
429 unsigned long long stable_base
;
430 unsigned long long key
, value
[nr_cpus
];
431 unsigned long long next_ins_key
;
432 unsigned int nr_losses
= 0;
435 stable_base
= task
* nr_repeats
* 2 + 1;
436 next_ins_key
= stable_base
;
438 for (i
= 0; i
< nr_stable_elems
; i
++) {
439 assert(bpf_map_update_elem(map_fd
, &next_ins_key
, value
,
444 for (i
= 0; i
< nr_repeats
; i
++) {
450 key
= rn
% nr_stable_elems
+ stable_base
;
451 bpf_map_lookup_elem(map_fd
, &key
, value
);
453 bpf_map_update_elem(map_fd
, &next_ins_key
, value
,
460 for (i
= 0; i
< nr_stable_elems
; i
++) {
461 if (bpf_map_lookup_elem(map_fd
, &key
, value
))
466 printf(" task:%d nr_losses:%u\n", task
, nr_losses
);
469 static void test_parallel_lru_loss(int map_type
, int map_flags
, int nr_tasks
)
473 printf("%s (map_type:%d map_flags:0x%X):\n", __func__
, map_type
,
476 /* Give 20% more than the active working set */
477 if (map_flags
& BPF_F_NO_COMMON_LRU
)
478 map_fd
= create_map(map_type
, map_flags
,
479 nr_cpus
* (1000 + 200));
481 map_fd
= create_map(map_type
, map_flags
,
482 nr_tasks
* (1000 + 200));
484 assert(map_fd
!= -1);
486 run_parallel(nr_tasks
, do_test_parallel_lru_loss
, &map_fd
);
491 int main(int argc
, char **argv
)
493 struct rlimit r
= {RLIM_INFINITY
, RLIM_INFINITY
};
494 int map_flags
[] = {0, BPF_F_NO_COMMON_LRU
};
495 const char *dist_file
;
501 printf("Usage: %s <dist-file> <lru-size> <nr-tasks>\n",
507 lru_size
= atoi(argv
[2]);
508 nr_tasks
= atoi(argv
[3]);
510 setbuf(stdout
, NULL
);
512 assert(!setrlimit(RLIMIT_MEMLOCK
, &r
));
516 nr_cpus
= bpf_num_possible_cpus();
517 assert(nr_cpus
!= -1);
518 printf("nr_cpus:%d\n\n", nr_cpus
);
520 nr_tasks
= min(nr_tasks
, nr_cpus
);
522 dist_key_counts
= read_keys(dist_file
, &dist_keys
);
523 if (!dist_key_counts
) {
524 printf("%s has no key\n", dist_file
);
528 for (f
= 0; f
< sizeof(map_flags
) / sizeof(*map_flags
); f
++) {
529 test_lru_loss0(BPF_MAP_TYPE_LRU_HASH
, map_flags
[f
]);
530 test_lru_loss1(BPF_MAP_TYPE_LRU_HASH
, map_flags
[f
]);
531 test_parallel_lru_loss(BPF_MAP_TYPE_LRU_HASH
, map_flags
[f
],
533 test_parallel_lru_dist(BPF_MAP_TYPE_LRU_HASH
, map_flags
[f
],