1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
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.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 #include <linux/bpf.h>
14 #include <linux/jhash.h>
15 #include <linux/filter.h>
16 #include <linux/rculist_nulls.h>
17 #include "percpu_freelist.h"
18 #include "bpf_lru_list.h"
19 #include "map_in_map.h"
21 #define HTAB_CREATE_FLAG_MASK \
22 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE)
25 struct hlist_nulls_head head
;
31 struct bucket
*buckets
;
34 struct pcpu_freelist freelist
;
37 struct htab_elem
*__percpu
*extra_elems
;
38 atomic_t count
; /* number of elements in this hashtable */
39 u32 n_buckets
; /* number of hash buckets */
40 u32 elem_size
; /* size of each element in bytes */
43 /* each htab element is struct htab_elem + key + value */
46 struct hlist_nulls_node hash_node
;
50 struct bpf_htab
*htab
;
51 struct pcpu_freelist_node fnode
;
57 struct bpf_lru_node lru_node
;
60 char key
[0] __aligned(8);
63 static bool htab_lru_map_delete_node(void *arg
, struct bpf_lru_node
*node
);
65 static bool htab_is_lru(const struct bpf_htab
*htab
)
67 return htab
->map
.map_type
== BPF_MAP_TYPE_LRU_HASH
||
68 htab
->map
.map_type
== BPF_MAP_TYPE_LRU_PERCPU_HASH
;
71 static bool htab_is_percpu(const struct bpf_htab
*htab
)
73 return htab
->map
.map_type
== BPF_MAP_TYPE_PERCPU_HASH
||
74 htab
->map
.map_type
== BPF_MAP_TYPE_LRU_PERCPU_HASH
;
77 static bool htab_is_prealloc(const struct bpf_htab
*htab
)
79 return !(htab
->map
.map_flags
& BPF_F_NO_PREALLOC
);
82 static inline void htab_elem_set_ptr(struct htab_elem
*l
, u32 key_size
,
85 *(void __percpu
**)(l
->key
+ key_size
) = pptr
;
88 static inline void __percpu
*htab_elem_get_ptr(struct htab_elem
*l
, u32 key_size
)
90 return *(void __percpu
**)(l
->key
+ key_size
);
93 static void *fd_htab_map_get_ptr(const struct bpf_map
*map
, struct htab_elem
*l
)
95 return *(void **)(l
->key
+ roundup(map
->key_size
, 8));
98 static struct htab_elem
*get_htab_elem(struct bpf_htab
*htab
, int i
)
100 return (struct htab_elem
*) (htab
->elems
+ i
* htab
->elem_size
);
103 static void htab_free_elems(struct bpf_htab
*htab
)
107 if (!htab_is_percpu(htab
))
110 for (i
= 0; i
< htab
->map
.max_entries
; i
++) {
113 pptr
= htab_elem_get_ptr(get_htab_elem(htab
, i
),
118 bpf_map_area_free(htab
->elems
);
121 static struct htab_elem
*prealloc_lru_pop(struct bpf_htab
*htab
, void *key
,
124 struct bpf_lru_node
*node
= bpf_lru_pop_free(&htab
->lru
, hash
);
128 l
= container_of(node
, struct htab_elem
, lru_node
);
129 memcpy(l
->key
, key
, htab
->map
.key_size
);
136 static int prealloc_init(struct bpf_htab
*htab
)
138 u32 num_entries
= htab
->map
.max_entries
;
139 int err
= -ENOMEM
, i
;
141 if (!htab_is_percpu(htab
) && !htab_is_lru(htab
))
142 num_entries
+= num_possible_cpus();
144 htab
->elems
= bpf_map_area_alloc(htab
->elem_size
* num_entries
,
145 htab
->map
.numa_node
);
149 if (!htab_is_percpu(htab
))
150 goto skip_percpu_elems
;
152 for (i
= 0; i
< num_entries
; i
++) {
153 u32 size
= round_up(htab
->map
.value_size
, 8);
156 pptr
= __alloc_percpu_gfp(size
, 8, GFP_USER
| __GFP_NOWARN
);
159 htab_elem_set_ptr(get_htab_elem(htab
, i
), htab
->map
.key_size
,
164 if (htab_is_lru(htab
))
165 err
= bpf_lru_init(&htab
->lru
,
166 htab
->map
.map_flags
& BPF_F_NO_COMMON_LRU
,
167 offsetof(struct htab_elem
, hash
) -
168 offsetof(struct htab_elem
, lru_node
),
169 htab_lru_map_delete_node
,
172 err
= pcpu_freelist_init(&htab
->freelist
);
177 if (htab_is_lru(htab
))
178 bpf_lru_populate(&htab
->lru
, htab
->elems
,
179 offsetof(struct htab_elem
, lru_node
),
180 htab
->elem_size
, num_entries
);
182 pcpu_freelist_populate(&htab
->freelist
,
183 htab
->elems
+ offsetof(struct htab_elem
, fnode
),
184 htab
->elem_size
, num_entries
);
189 htab_free_elems(htab
);
193 static void prealloc_destroy(struct bpf_htab
*htab
)
195 htab_free_elems(htab
);
197 if (htab_is_lru(htab
))
198 bpf_lru_destroy(&htab
->lru
);
200 pcpu_freelist_destroy(&htab
->freelist
);
203 static int alloc_extra_elems(struct bpf_htab
*htab
)
205 struct htab_elem
*__percpu
*pptr
, *l_new
;
206 struct pcpu_freelist_node
*l
;
209 pptr
= __alloc_percpu_gfp(sizeof(struct htab_elem
*), 8,
210 GFP_USER
| __GFP_NOWARN
);
214 for_each_possible_cpu(cpu
) {
215 l
= pcpu_freelist_pop(&htab
->freelist
);
216 /* pop will succeed, since prealloc_init()
217 * preallocated extra num_possible_cpus elements
219 l_new
= container_of(l
, struct htab_elem
, fnode
);
220 *per_cpu_ptr(pptr
, cpu
) = l_new
;
222 htab
->extra_elems
= pptr
;
226 /* Called from syscall */
227 static struct bpf_map
*htab_map_alloc(union bpf_attr
*attr
)
229 bool percpu
= (attr
->map_type
== BPF_MAP_TYPE_PERCPU_HASH
||
230 attr
->map_type
== BPF_MAP_TYPE_LRU_PERCPU_HASH
);
231 bool lru
= (attr
->map_type
== BPF_MAP_TYPE_LRU_HASH
||
232 attr
->map_type
== BPF_MAP_TYPE_LRU_PERCPU_HASH
);
233 /* percpu_lru means each cpu has its own LRU list.
234 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
235 * the map's value itself is percpu. percpu_lru has
236 * nothing to do with the map's value.
238 bool percpu_lru
= (attr
->map_flags
& BPF_F_NO_COMMON_LRU
);
239 bool prealloc
= !(attr
->map_flags
& BPF_F_NO_PREALLOC
);
240 int numa_node
= bpf_map_attr_numa_node(attr
);
241 struct bpf_htab
*htab
;
245 BUILD_BUG_ON(offsetof(struct htab_elem
, htab
) !=
246 offsetof(struct htab_elem
, hash_node
.pprev
));
247 BUILD_BUG_ON(offsetof(struct htab_elem
, fnode
.next
) !=
248 offsetof(struct htab_elem
, hash_node
.pprev
));
250 if (lru
&& !capable(CAP_SYS_ADMIN
))
251 /* LRU implementation is much complicated than other
252 * maps. Hence, limit to CAP_SYS_ADMIN for now.
254 return ERR_PTR(-EPERM
);
256 if (attr
->map_flags
& ~HTAB_CREATE_FLAG_MASK
)
257 /* reserved bits should not be used */
258 return ERR_PTR(-EINVAL
);
260 if (!lru
&& percpu_lru
)
261 return ERR_PTR(-EINVAL
);
263 if (lru
&& !prealloc
)
264 return ERR_PTR(-ENOTSUPP
);
266 if (numa_node
!= NUMA_NO_NODE
&& (percpu
|| percpu_lru
))
267 return ERR_PTR(-EINVAL
);
269 htab
= kzalloc(sizeof(*htab
), GFP_USER
);
271 return ERR_PTR(-ENOMEM
);
273 /* mandatory map attributes */
274 htab
->map
.map_type
= attr
->map_type
;
275 htab
->map
.key_size
= attr
->key_size
;
276 htab
->map
.value_size
= attr
->value_size
;
277 htab
->map
.max_entries
= attr
->max_entries
;
278 htab
->map
.map_flags
= attr
->map_flags
;
279 htab
->map
.numa_node
= numa_node
;
281 /* check sanity of attributes.
282 * value_size == 0 may be allowed in the future to use map as a set
285 if (htab
->map
.max_entries
== 0 || htab
->map
.key_size
== 0 ||
286 htab
->map
.value_size
== 0)
290 /* ensure each CPU's lru list has >=1 elements.
291 * since we are at it, make each lru list has the same
292 * number of elements.
294 htab
->map
.max_entries
= roundup(attr
->max_entries
,
295 num_possible_cpus());
296 if (htab
->map
.max_entries
< attr
->max_entries
)
297 htab
->map
.max_entries
= rounddown(attr
->max_entries
,
298 num_possible_cpus());
301 /* hash table size must be power of 2 */
302 htab
->n_buckets
= roundup_pow_of_two(htab
->map
.max_entries
);
305 if (htab
->map
.key_size
> MAX_BPF_STACK
)
306 /* eBPF programs initialize keys on stack, so they cannot be
307 * larger than max stack size
311 if (htab
->map
.value_size
>= KMALLOC_MAX_SIZE
-
312 MAX_BPF_STACK
- sizeof(struct htab_elem
))
313 /* if value_size is bigger, the user space won't be able to
314 * access the elements via bpf syscall. This check also makes
315 * sure that the elem_size doesn't overflow and it's
316 * kmalloc-able later in htab_map_update_elem()
320 if (percpu
&& round_up(htab
->map
.value_size
, 8) > PCPU_MIN_UNIT_SIZE
)
321 /* make sure the size for pcpu_alloc() is reasonable */
324 htab
->elem_size
= sizeof(struct htab_elem
) +
325 round_up(htab
->map
.key_size
, 8);
327 htab
->elem_size
+= sizeof(void *);
329 htab
->elem_size
+= round_up(htab
->map
.value_size
, 8);
331 /* prevent zero size kmalloc and check for u32 overflow */
332 if (htab
->n_buckets
== 0 ||
333 htab
->n_buckets
> U32_MAX
/ sizeof(struct bucket
))
336 cost
= (u64
) htab
->n_buckets
* sizeof(struct bucket
) +
337 (u64
) htab
->elem_size
* htab
->map
.max_entries
;
340 cost
+= (u64
) round_up(htab
->map
.value_size
, 8) *
341 num_possible_cpus() * htab
->map
.max_entries
;
343 cost
+= (u64
) htab
->elem_size
* num_possible_cpus();
345 if (cost
>= U32_MAX
- PAGE_SIZE
)
346 /* make sure page count doesn't overflow */
349 htab
->map
.pages
= round_up(cost
, PAGE_SIZE
) >> PAGE_SHIFT
;
351 /* if map size is larger than memlock limit, reject it early */
352 err
= bpf_map_precharge_memlock(htab
->map
.pages
);
357 htab
->buckets
= bpf_map_area_alloc(htab
->n_buckets
*
358 sizeof(struct bucket
),
359 htab
->map
.numa_node
);
363 for (i
= 0; i
< htab
->n_buckets
; i
++) {
364 INIT_HLIST_NULLS_HEAD(&htab
->buckets
[i
].head
, i
);
365 raw_spin_lock_init(&htab
->buckets
[i
].lock
);
369 err
= prealloc_init(htab
);
373 if (!percpu
&& !lru
) {
374 /* lru itself can remove the least used element, so
375 * there is no need for an extra elem during map_update.
377 err
= alloc_extra_elems(htab
);
386 prealloc_destroy(htab
);
388 bpf_map_area_free(htab
->buckets
);
394 static inline u32
htab_map_hash(const void *key
, u32 key_len
)
396 return jhash(key
, key_len
, 0);
399 static inline struct bucket
*__select_bucket(struct bpf_htab
*htab
, u32 hash
)
401 return &htab
->buckets
[hash
& (htab
->n_buckets
- 1)];
404 static inline struct hlist_nulls_head
*select_bucket(struct bpf_htab
*htab
, u32 hash
)
406 return &__select_bucket(htab
, hash
)->head
;
409 /* this lookup function can only be called with bucket lock taken */
410 static struct htab_elem
*lookup_elem_raw(struct hlist_nulls_head
*head
, u32 hash
,
411 void *key
, u32 key_size
)
413 struct hlist_nulls_node
*n
;
416 hlist_nulls_for_each_entry_rcu(l
, n
, head
, hash_node
)
417 if (l
->hash
== hash
&& !memcmp(&l
->key
, key
, key_size
))
423 /* can be called without bucket lock. it will repeat the loop in
424 * the unlikely event when elements moved from one bucket into another
425 * while link list is being walked
427 static struct htab_elem
*lookup_nulls_elem_raw(struct hlist_nulls_head
*head
,
429 u32 key_size
, u32 n_buckets
)
431 struct hlist_nulls_node
*n
;
435 hlist_nulls_for_each_entry_rcu(l
, n
, head
, hash_node
)
436 if (l
->hash
== hash
&& !memcmp(&l
->key
, key
, key_size
))
439 if (unlikely(get_nulls_value(n
) != (hash
& (n_buckets
- 1))))
445 /* Called from syscall or from eBPF program directly, so
446 * arguments have to match bpf_map_lookup_elem() exactly.
447 * The return value is adjusted by BPF instructions
448 * in htab_map_gen_lookup().
450 static void *__htab_map_lookup_elem(struct bpf_map
*map
, void *key
)
452 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
453 struct hlist_nulls_head
*head
;
457 /* Must be called with rcu_read_lock. */
458 WARN_ON_ONCE(!rcu_read_lock_held());
460 key_size
= map
->key_size
;
462 hash
= htab_map_hash(key
, key_size
);
464 head
= select_bucket(htab
, hash
);
466 l
= lookup_nulls_elem_raw(head
, hash
, key
, key_size
, htab
->n_buckets
);
471 static void *htab_map_lookup_elem(struct bpf_map
*map
, void *key
)
473 struct htab_elem
*l
= __htab_map_lookup_elem(map
, key
);
476 return l
->key
+ round_up(map
->key_size
, 8);
481 /* inline bpf_map_lookup_elem() call.
484 * bpf_map_lookup_elem
485 * map->ops->map_lookup_elem
486 * htab_map_lookup_elem
487 * __htab_map_lookup_elem
490 * __htab_map_lookup_elem
492 static u32
htab_map_gen_lookup(struct bpf_map
*map
, struct bpf_insn
*insn_buf
)
494 struct bpf_insn
*insn
= insn_buf
;
495 const int ret
= BPF_REG_0
;
497 *insn
++ = BPF_EMIT_CALL((u64 (*)(u64
, u64
, u64
, u64
, u64
))__htab_map_lookup_elem
);
498 *insn
++ = BPF_JMP_IMM(BPF_JEQ
, ret
, 0, 1);
499 *insn
++ = BPF_ALU64_IMM(BPF_ADD
, ret
,
500 offsetof(struct htab_elem
, key
) +
501 round_up(map
->key_size
, 8));
502 return insn
- insn_buf
;
505 static void *htab_lru_map_lookup_elem(struct bpf_map
*map
, void *key
)
507 struct htab_elem
*l
= __htab_map_lookup_elem(map
, key
);
510 bpf_lru_node_set_ref(&l
->lru_node
);
511 return l
->key
+ round_up(map
->key_size
, 8);
517 static u32
htab_lru_map_gen_lookup(struct bpf_map
*map
,
518 struct bpf_insn
*insn_buf
)
520 struct bpf_insn
*insn
= insn_buf
;
521 const int ret
= BPF_REG_0
;
522 const int ref_reg
= BPF_REG_1
;
524 *insn
++ = BPF_EMIT_CALL((u64 (*)(u64
, u64
, u64
, u64
, u64
))__htab_map_lookup_elem
);
525 *insn
++ = BPF_JMP_IMM(BPF_JEQ
, ret
, 0, 4);
526 *insn
++ = BPF_LDX_MEM(BPF_B
, ref_reg
, ret
,
527 offsetof(struct htab_elem
, lru_node
) +
528 offsetof(struct bpf_lru_node
, ref
));
529 *insn
++ = BPF_JMP_IMM(BPF_JNE
, ref_reg
, 0, 1);
530 *insn
++ = BPF_ST_MEM(BPF_B
, ret
,
531 offsetof(struct htab_elem
, lru_node
) +
532 offsetof(struct bpf_lru_node
, ref
),
534 *insn
++ = BPF_ALU64_IMM(BPF_ADD
, ret
,
535 offsetof(struct htab_elem
, key
) +
536 round_up(map
->key_size
, 8));
537 return insn
- insn_buf
;
540 /* It is called from the bpf_lru_list when the LRU needs to delete
541 * older elements from the htab.
543 static bool htab_lru_map_delete_node(void *arg
, struct bpf_lru_node
*node
)
545 struct bpf_htab
*htab
= (struct bpf_htab
*)arg
;
546 struct htab_elem
*l
= NULL
, *tgt_l
;
547 struct hlist_nulls_head
*head
;
548 struct hlist_nulls_node
*n
;
552 tgt_l
= container_of(node
, struct htab_elem
, lru_node
);
553 b
= __select_bucket(htab
, tgt_l
->hash
);
556 raw_spin_lock_irqsave(&b
->lock
, flags
);
558 hlist_nulls_for_each_entry_rcu(l
, n
, head
, hash_node
)
560 hlist_nulls_del_rcu(&l
->hash_node
);
564 raw_spin_unlock_irqrestore(&b
->lock
, flags
);
569 /* Called from syscall */
570 static int htab_map_get_next_key(struct bpf_map
*map
, void *key
, void *next_key
)
572 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
573 struct hlist_nulls_head
*head
;
574 struct htab_elem
*l
, *next_l
;
578 WARN_ON_ONCE(!rcu_read_lock_held());
580 key_size
= map
->key_size
;
583 goto find_first_elem
;
585 hash
= htab_map_hash(key
, key_size
);
587 head
= select_bucket(htab
, hash
);
590 l
= lookup_nulls_elem_raw(head
, hash
, key
, key_size
, htab
->n_buckets
);
593 goto find_first_elem
;
595 /* key was found, get next key in the same bucket */
596 next_l
= hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l
->hash_node
)),
597 struct htab_elem
, hash_node
);
600 /* if next elem in this hash list is non-zero, just return it */
601 memcpy(next_key
, next_l
->key
, key_size
);
605 /* no more elements in this hash list, go to the next bucket */
606 i
= hash
& (htab
->n_buckets
- 1);
610 /* iterate over buckets */
611 for (; i
< htab
->n_buckets
; i
++) {
612 head
= select_bucket(htab
, i
);
614 /* pick first element in the bucket */
615 next_l
= hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head
)),
616 struct htab_elem
, hash_node
);
618 /* if it's not empty, just return it */
619 memcpy(next_key
, next_l
->key
, key_size
);
624 /* iterated over all buckets and all elements */
628 static void htab_elem_free(struct bpf_htab
*htab
, struct htab_elem
*l
)
630 if (htab
->map
.map_type
== BPF_MAP_TYPE_PERCPU_HASH
)
631 free_percpu(htab_elem_get_ptr(l
, htab
->map
.key_size
));
635 static void htab_elem_free_rcu(struct rcu_head
*head
)
637 struct htab_elem
*l
= container_of(head
, struct htab_elem
, rcu
);
638 struct bpf_htab
*htab
= l
->htab
;
640 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
641 * we're calling kfree, otherwise deadlock is possible if kprobes
642 * are placed somewhere inside of slub
645 __this_cpu_inc(bpf_prog_active
);
646 htab_elem_free(htab
, l
);
647 __this_cpu_dec(bpf_prog_active
);
651 static void free_htab_elem(struct bpf_htab
*htab
, struct htab_elem
*l
)
653 struct bpf_map
*map
= &htab
->map
;
655 if (map
->ops
->map_fd_put_ptr
) {
656 void *ptr
= fd_htab_map_get_ptr(map
, l
);
658 map
->ops
->map_fd_put_ptr(ptr
);
661 if (htab_is_prealloc(htab
)) {
662 pcpu_freelist_push(&htab
->freelist
, &l
->fnode
);
664 atomic_dec(&htab
->count
);
666 call_rcu(&l
->rcu
, htab_elem_free_rcu
);
670 static void pcpu_copy_value(struct bpf_htab
*htab
, void __percpu
*pptr
,
671 void *value
, bool onallcpus
)
674 /* copy true value_size bytes */
675 memcpy(this_cpu_ptr(pptr
), value
, htab
->map
.value_size
);
677 u32 size
= round_up(htab
->map
.value_size
, 8);
680 for_each_possible_cpu(cpu
) {
681 bpf_long_memcpy(per_cpu_ptr(pptr
, cpu
),
688 static bool fd_htab_map_needs_adjust(const struct bpf_htab
*htab
)
690 return htab
->map
.map_type
== BPF_MAP_TYPE_HASH_OF_MAPS
&&
694 static u32
htab_size_value(const struct bpf_htab
*htab
, bool percpu
)
696 u32 size
= htab
->map
.value_size
;
698 if (percpu
|| fd_htab_map_needs_adjust(htab
))
699 size
= round_up(size
, 8);
703 static struct htab_elem
*alloc_htab_elem(struct bpf_htab
*htab
, void *key
,
704 void *value
, u32 key_size
, u32 hash
,
705 bool percpu
, bool onallcpus
,
706 struct htab_elem
*old_elem
)
708 u32 size
= htab_size_value(htab
, percpu
);
709 bool prealloc
= htab_is_prealloc(htab
);
710 struct htab_elem
*l_new
, **pl_new
;
715 /* if we're updating the existing element,
716 * use per-cpu extra elems to avoid freelist_pop/push
718 pl_new
= this_cpu_ptr(htab
->extra_elems
);
722 struct pcpu_freelist_node
*l
;
724 l
= pcpu_freelist_pop(&htab
->freelist
);
726 return ERR_PTR(-E2BIG
);
727 l_new
= container_of(l
, struct htab_elem
, fnode
);
730 if (atomic_inc_return(&htab
->count
) > htab
->map
.max_entries
)
732 /* when map is full and update() is replacing
733 * old element, it's ok to allocate, since
734 * old element will be freed immediately.
735 * Otherwise return an error
737 atomic_dec(&htab
->count
);
738 return ERR_PTR(-E2BIG
);
740 l_new
= kmalloc_node(htab
->elem_size
, GFP_ATOMIC
| __GFP_NOWARN
,
741 htab
->map
.numa_node
);
743 return ERR_PTR(-ENOMEM
);
746 memcpy(l_new
->key
, key
, key_size
);
749 pptr
= htab_elem_get_ptr(l_new
, key_size
);
751 /* alloc_percpu zero-fills */
752 pptr
= __alloc_percpu_gfp(size
, 8,
753 GFP_ATOMIC
| __GFP_NOWARN
);
756 return ERR_PTR(-ENOMEM
);
760 pcpu_copy_value(htab
, pptr
, value
, onallcpus
);
763 htab_elem_set_ptr(l_new
, key_size
, pptr
);
765 memcpy(l_new
->key
+ round_up(key_size
, 8), value
, size
);
772 static int check_flags(struct bpf_htab
*htab
, struct htab_elem
*l_old
,
775 if (l_old
&& map_flags
== BPF_NOEXIST
)
776 /* elem already exists */
779 if (!l_old
&& map_flags
== BPF_EXIST
)
780 /* elem doesn't exist, cannot update it */
786 /* Called from syscall or from eBPF program */
787 static int htab_map_update_elem(struct bpf_map
*map
, void *key
, void *value
,
790 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
791 struct htab_elem
*l_new
= NULL
, *l_old
;
792 struct hlist_nulls_head
*head
;
798 if (unlikely(map_flags
> BPF_EXIST
))
802 WARN_ON_ONCE(!rcu_read_lock_held());
804 key_size
= map
->key_size
;
806 hash
= htab_map_hash(key
, key_size
);
808 b
= __select_bucket(htab
, hash
);
811 /* bpf_map_update_elem() can be called in_irq() */
812 raw_spin_lock_irqsave(&b
->lock
, flags
);
814 l_old
= lookup_elem_raw(head
, hash
, key
, key_size
);
816 ret
= check_flags(htab
, l_old
, map_flags
);
820 l_new
= alloc_htab_elem(htab
, key
, value
, key_size
, hash
, false, false,
823 /* all pre-allocated elements are in use or memory exhausted */
824 ret
= PTR_ERR(l_new
);
828 /* add new element to the head of the list, so that
829 * concurrent search will find it before old elem
831 hlist_nulls_add_head_rcu(&l_new
->hash_node
, head
);
833 hlist_nulls_del_rcu(&l_old
->hash_node
);
834 if (!htab_is_prealloc(htab
))
835 free_htab_elem(htab
, l_old
);
839 raw_spin_unlock_irqrestore(&b
->lock
, flags
);
843 static int htab_lru_map_update_elem(struct bpf_map
*map
, void *key
, void *value
,
846 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
847 struct htab_elem
*l_new
, *l_old
= NULL
;
848 struct hlist_nulls_head
*head
;
854 if (unlikely(map_flags
> BPF_EXIST
))
858 WARN_ON_ONCE(!rcu_read_lock_held());
860 key_size
= map
->key_size
;
862 hash
= htab_map_hash(key
, key_size
);
864 b
= __select_bucket(htab
, hash
);
867 /* For LRU, we need to alloc before taking bucket's
868 * spinlock because getting free nodes from LRU may need
869 * to remove older elements from htab and this removal
870 * operation will need a bucket lock.
872 l_new
= prealloc_lru_pop(htab
, key
, hash
);
875 memcpy(l_new
->key
+ round_up(map
->key_size
, 8), value
, map
->value_size
);
877 /* bpf_map_update_elem() can be called in_irq() */
878 raw_spin_lock_irqsave(&b
->lock
, flags
);
880 l_old
= lookup_elem_raw(head
, hash
, key
, key_size
);
882 ret
= check_flags(htab
, l_old
, map_flags
);
886 /* add new element to the head of the list, so that
887 * concurrent search will find it before old elem
889 hlist_nulls_add_head_rcu(&l_new
->hash_node
, head
);
891 bpf_lru_node_set_ref(&l_new
->lru_node
);
892 hlist_nulls_del_rcu(&l_old
->hash_node
);
897 raw_spin_unlock_irqrestore(&b
->lock
, flags
);
900 bpf_lru_push_free(&htab
->lru
, &l_new
->lru_node
);
902 bpf_lru_push_free(&htab
->lru
, &l_old
->lru_node
);
907 static int __htab_percpu_map_update_elem(struct bpf_map
*map
, void *key
,
908 void *value
, u64 map_flags
,
911 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
912 struct htab_elem
*l_new
= NULL
, *l_old
;
913 struct hlist_nulls_head
*head
;
919 if (unlikely(map_flags
> BPF_EXIST
))
923 WARN_ON_ONCE(!rcu_read_lock_held());
925 key_size
= map
->key_size
;
927 hash
= htab_map_hash(key
, key_size
);
929 b
= __select_bucket(htab
, hash
);
932 /* bpf_map_update_elem() can be called in_irq() */
933 raw_spin_lock_irqsave(&b
->lock
, flags
);
935 l_old
= lookup_elem_raw(head
, hash
, key
, key_size
);
937 ret
= check_flags(htab
, l_old
, map_flags
);
942 /* per-cpu hash map can update value in-place */
943 pcpu_copy_value(htab
, htab_elem_get_ptr(l_old
, key_size
),
946 l_new
= alloc_htab_elem(htab
, key
, value
, key_size
,
947 hash
, true, onallcpus
, NULL
);
949 ret
= PTR_ERR(l_new
);
952 hlist_nulls_add_head_rcu(&l_new
->hash_node
, head
);
956 raw_spin_unlock_irqrestore(&b
->lock
, flags
);
960 static int __htab_lru_percpu_map_update_elem(struct bpf_map
*map
, void *key
,
961 void *value
, u64 map_flags
,
964 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
965 struct htab_elem
*l_new
= NULL
, *l_old
;
966 struct hlist_nulls_head
*head
;
972 if (unlikely(map_flags
> BPF_EXIST
))
976 WARN_ON_ONCE(!rcu_read_lock_held());
978 key_size
= map
->key_size
;
980 hash
= htab_map_hash(key
, key_size
);
982 b
= __select_bucket(htab
, hash
);
985 /* For LRU, we need to alloc before taking bucket's
986 * spinlock because LRU's elem alloc may need
987 * to remove older elem from htab and this removal
988 * operation will need a bucket lock.
990 if (map_flags
!= BPF_EXIST
) {
991 l_new
= prealloc_lru_pop(htab
, key
, hash
);
996 /* bpf_map_update_elem() can be called in_irq() */
997 raw_spin_lock_irqsave(&b
->lock
, flags
);
999 l_old
= lookup_elem_raw(head
, hash
, key
, key_size
);
1001 ret
= check_flags(htab
, l_old
, map_flags
);
1006 bpf_lru_node_set_ref(&l_old
->lru_node
);
1008 /* per-cpu hash map can update value in-place */
1009 pcpu_copy_value(htab
, htab_elem_get_ptr(l_old
, key_size
),
1012 pcpu_copy_value(htab
, htab_elem_get_ptr(l_new
, key_size
),
1014 hlist_nulls_add_head_rcu(&l_new
->hash_node
, head
);
1019 raw_spin_unlock_irqrestore(&b
->lock
, flags
);
1021 bpf_lru_push_free(&htab
->lru
, &l_new
->lru_node
);
1025 static int htab_percpu_map_update_elem(struct bpf_map
*map
, void *key
,
1026 void *value
, u64 map_flags
)
1028 return __htab_percpu_map_update_elem(map
, key
, value
, map_flags
, false);
1031 static int htab_lru_percpu_map_update_elem(struct bpf_map
*map
, void *key
,
1032 void *value
, u64 map_flags
)
1034 return __htab_lru_percpu_map_update_elem(map
, key
, value
, map_flags
,
1038 /* Called from syscall or from eBPF program */
1039 static int htab_map_delete_elem(struct bpf_map
*map
, void *key
)
1041 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
1042 struct hlist_nulls_head
*head
;
1044 struct htab_elem
*l
;
1045 unsigned long flags
;
1049 WARN_ON_ONCE(!rcu_read_lock_held());
1051 key_size
= map
->key_size
;
1053 hash
= htab_map_hash(key
, key_size
);
1054 b
= __select_bucket(htab
, hash
);
1057 raw_spin_lock_irqsave(&b
->lock
, flags
);
1059 l
= lookup_elem_raw(head
, hash
, key
, key_size
);
1062 hlist_nulls_del_rcu(&l
->hash_node
);
1063 free_htab_elem(htab
, l
);
1067 raw_spin_unlock_irqrestore(&b
->lock
, flags
);
1071 static int htab_lru_map_delete_elem(struct bpf_map
*map
, void *key
)
1073 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
1074 struct hlist_nulls_head
*head
;
1076 struct htab_elem
*l
;
1077 unsigned long flags
;
1081 WARN_ON_ONCE(!rcu_read_lock_held());
1083 key_size
= map
->key_size
;
1085 hash
= htab_map_hash(key
, key_size
);
1086 b
= __select_bucket(htab
, hash
);
1089 raw_spin_lock_irqsave(&b
->lock
, flags
);
1091 l
= lookup_elem_raw(head
, hash
, key
, key_size
);
1094 hlist_nulls_del_rcu(&l
->hash_node
);
1098 raw_spin_unlock_irqrestore(&b
->lock
, flags
);
1100 bpf_lru_push_free(&htab
->lru
, &l
->lru_node
);
1104 static void delete_all_elements(struct bpf_htab
*htab
)
1108 for (i
= 0; i
< htab
->n_buckets
; i
++) {
1109 struct hlist_nulls_head
*head
= select_bucket(htab
, i
);
1110 struct hlist_nulls_node
*n
;
1111 struct htab_elem
*l
;
1113 hlist_nulls_for_each_entry_safe(l
, n
, head
, hash_node
) {
1114 hlist_nulls_del_rcu(&l
->hash_node
);
1115 htab_elem_free(htab
, l
);
1120 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1121 static void htab_map_free(struct bpf_map
*map
)
1123 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
1125 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1126 * so the programs (can be more than one that used this map) were
1127 * disconnected from events. Wait for outstanding critical sections in
1128 * these programs to complete
1132 /* some of free_htab_elem() callbacks for elements of this map may
1133 * not have executed. Wait for them.
1136 if (!htab_is_prealloc(htab
))
1137 delete_all_elements(htab
);
1139 prealloc_destroy(htab
);
1141 free_percpu(htab
->extra_elems
);
1142 bpf_map_area_free(htab
->buckets
);
1146 const struct bpf_map_ops htab_map_ops
= {
1147 .map_alloc
= htab_map_alloc
,
1148 .map_free
= htab_map_free
,
1149 .map_get_next_key
= htab_map_get_next_key
,
1150 .map_lookup_elem
= htab_map_lookup_elem
,
1151 .map_update_elem
= htab_map_update_elem
,
1152 .map_delete_elem
= htab_map_delete_elem
,
1153 .map_gen_lookup
= htab_map_gen_lookup
,
1156 const struct bpf_map_ops htab_lru_map_ops
= {
1157 .map_alloc
= htab_map_alloc
,
1158 .map_free
= htab_map_free
,
1159 .map_get_next_key
= htab_map_get_next_key
,
1160 .map_lookup_elem
= htab_lru_map_lookup_elem
,
1161 .map_update_elem
= htab_lru_map_update_elem
,
1162 .map_delete_elem
= htab_lru_map_delete_elem
,
1163 .map_gen_lookup
= htab_lru_map_gen_lookup
,
1166 /* Called from eBPF program */
1167 static void *htab_percpu_map_lookup_elem(struct bpf_map
*map
, void *key
)
1169 struct htab_elem
*l
= __htab_map_lookup_elem(map
, key
);
1172 return this_cpu_ptr(htab_elem_get_ptr(l
, map
->key_size
));
1177 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map
*map
, void *key
)
1179 struct htab_elem
*l
= __htab_map_lookup_elem(map
, key
);
1182 bpf_lru_node_set_ref(&l
->lru_node
);
1183 return this_cpu_ptr(htab_elem_get_ptr(l
, map
->key_size
));
1189 int bpf_percpu_hash_copy(struct bpf_map
*map
, void *key
, void *value
)
1191 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
1192 struct htab_elem
*l
;
1193 void __percpu
*pptr
;
1198 /* per_cpu areas are zero-filled and bpf programs can only
1199 * access 'value_size' of them, so copying rounded areas
1200 * will not leak any kernel data
1202 size
= round_up(map
->value_size
, 8);
1204 l
= __htab_map_lookup_elem(map
, key
);
1207 if (htab_is_lru(htab
))
1208 bpf_lru_node_set_ref(&l
->lru_node
);
1209 pptr
= htab_elem_get_ptr(l
, map
->key_size
);
1210 for_each_possible_cpu(cpu
) {
1211 bpf_long_memcpy(value
+ off
,
1212 per_cpu_ptr(pptr
, cpu
), size
);
1221 int bpf_percpu_hash_update(struct bpf_map
*map
, void *key
, void *value
,
1224 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
1228 if (htab_is_lru(htab
))
1229 ret
= __htab_lru_percpu_map_update_elem(map
, key
, value
,
1232 ret
= __htab_percpu_map_update_elem(map
, key
, value
, map_flags
,
1239 const struct bpf_map_ops htab_percpu_map_ops
= {
1240 .map_alloc
= htab_map_alloc
,
1241 .map_free
= htab_map_free
,
1242 .map_get_next_key
= htab_map_get_next_key
,
1243 .map_lookup_elem
= htab_percpu_map_lookup_elem
,
1244 .map_update_elem
= htab_percpu_map_update_elem
,
1245 .map_delete_elem
= htab_map_delete_elem
,
1248 const struct bpf_map_ops htab_lru_percpu_map_ops
= {
1249 .map_alloc
= htab_map_alloc
,
1250 .map_free
= htab_map_free
,
1251 .map_get_next_key
= htab_map_get_next_key
,
1252 .map_lookup_elem
= htab_lru_percpu_map_lookup_elem
,
1253 .map_update_elem
= htab_lru_percpu_map_update_elem
,
1254 .map_delete_elem
= htab_lru_map_delete_elem
,
1257 static struct bpf_map
*fd_htab_map_alloc(union bpf_attr
*attr
)
1259 if (attr
->value_size
!= sizeof(u32
))
1260 return ERR_PTR(-EINVAL
);
1261 return htab_map_alloc(attr
);
1264 static void fd_htab_map_free(struct bpf_map
*map
)
1266 struct bpf_htab
*htab
= container_of(map
, struct bpf_htab
, map
);
1267 struct hlist_nulls_node
*n
;
1268 struct hlist_nulls_head
*head
;
1269 struct htab_elem
*l
;
1272 for (i
= 0; i
< htab
->n_buckets
; i
++) {
1273 head
= select_bucket(htab
, i
);
1275 hlist_nulls_for_each_entry_safe(l
, n
, head
, hash_node
) {
1276 void *ptr
= fd_htab_map_get_ptr(map
, l
);
1278 map
->ops
->map_fd_put_ptr(ptr
);
1285 /* only called from syscall */
1286 int bpf_fd_htab_map_lookup_elem(struct bpf_map
*map
, void *key
, u32
*value
)
1291 if (!map
->ops
->map_fd_sys_lookup_elem
)
1295 ptr
= htab_map_lookup_elem(map
, key
);
1297 *value
= map
->ops
->map_fd_sys_lookup_elem(READ_ONCE(*ptr
));
1305 /* only called from syscall */
1306 int bpf_fd_htab_map_update_elem(struct bpf_map
*map
, struct file
*map_file
,
1307 void *key
, void *value
, u64 map_flags
)
1311 u32 ufd
= *(u32
*)value
;
1313 ptr
= map
->ops
->map_fd_get_ptr(map
, map_file
, ufd
);
1315 return PTR_ERR(ptr
);
1317 ret
= htab_map_update_elem(map
, key
, &ptr
, map_flags
);
1319 map
->ops
->map_fd_put_ptr(ptr
);
1324 static struct bpf_map
*htab_of_map_alloc(union bpf_attr
*attr
)
1326 struct bpf_map
*map
, *inner_map_meta
;
1328 inner_map_meta
= bpf_map_meta_alloc(attr
->inner_map_fd
);
1329 if (IS_ERR(inner_map_meta
))
1330 return inner_map_meta
;
1332 map
= fd_htab_map_alloc(attr
);
1334 bpf_map_meta_free(inner_map_meta
);
1338 map
->inner_map_meta
= inner_map_meta
;
1343 static void *htab_of_map_lookup_elem(struct bpf_map
*map
, void *key
)
1345 struct bpf_map
**inner_map
= htab_map_lookup_elem(map
, key
);
1350 return READ_ONCE(*inner_map
);
1353 static u32
htab_of_map_gen_lookup(struct bpf_map
*map
,
1354 struct bpf_insn
*insn_buf
)
1356 struct bpf_insn
*insn
= insn_buf
;
1357 const int ret
= BPF_REG_0
;
1359 *insn
++ = BPF_EMIT_CALL((u64 (*)(u64
, u64
, u64
, u64
, u64
))__htab_map_lookup_elem
);
1360 *insn
++ = BPF_JMP_IMM(BPF_JEQ
, ret
, 0, 2);
1361 *insn
++ = BPF_ALU64_IMM(BPF_ADD
, ret
,
1362 offsetof(struct htab_elem
, key
) +
1363 round_up(map
->key_size
, 8));
1364 *insn
++ = BPF_LDX_MEM(BPF_DW
, ret
, ret
, 0);
1366 return insn
- insn_buf
;
1369 static void htab_of_map_free(struct bpf_map
*map
)
1371 bpf_map_meta_free(map
->inner_map_meta
);
1372 fd_htab_map_free(map
);
1375 const struct bpf_map_ops htab_of_maps_map_ops
= {
1376 .map_alloc
= htab_of_map_alloc
,
1377 .map_free
= htab_of_map_free
,
1378 .map_get_next_key
= htab_map_get_next_key
,
1379 .map_lookup_elem
= htab_of_map_lookup_elem
,
1380 .map_delete_elem
= htab_map_delete_elem
,
1381 .map_fd_get_ptr
= bpf_map_fd_get_ptr
,
1382 .map_fd_put_ptr
= bpf_map_fd_put_ptr
,
1383 .map_fd_sys_lookup_elem
= bpf_map_fd_sys_lookup_elem
,
1384 .map_gen_lookup
= htab_of_map_gen_lookup
,