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 | \
23 BPF_F_RDONLY | BPF_F_WRONLY)
26 struct hlist_nulls_head head;
32 struct bucket *buckets;
35 struct pcpu_freelist freelist;
38 struct htab_elem *__percpu *extra_elems;
39 atomic_t count; /* number of elements in this hashtable */
40 u32 n_buckets; /* number of hash buckets */
41 u32 elem_size; /* size of each element in bytes */
44 /* each htab element is struct htab_elem + key + value */
47 struct hlist_nulls_node hash_node;
51 struct bpf_htab *htab;
52 struct pcpu_freelist_node fnode;
58 struct bpf_lru_node lru_node;
61 char key[0] __aligned(8);
64 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
66 static bool htab_is_lru(const struct bpf_htab *htab)
68 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
69 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
72 static bool htab_is_percpu(const struct bpf_htab *htab)
74 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
75 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
78 static bool htab_is_prealloc(const struct bpf_htab *htab)
80 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
83 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
86 *(void __percpu **)(l->key + key_size) = pptr;
89 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
91 return *(void __percpu **)(l->key + key_size);
94 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
96 return *(void **)(l->key + roundup(map->key_size, 8));
99 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
101 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
104 static void htab_free_elems(struct bpf_htab *htab)
108 if (!htab_is_percpu(htab))
111 for (i = 0; i < htab->map.max_entries; i++) {
114 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
120 bpf_map_area_free(htab->elems);
123 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
126 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
130 l = container_of(node, struct htab_elem, lru_node);
131 memcpy(l->key, key, htab->map.key_size);
138 static int prealloc_init(struct bpf_htab *htab)
140 u32 num_entries = htab->map.max_entries;
141 int err = -ENOMEM, i;
143 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
144 num_entries += num_possible_cpus();
146 htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
147 htab->map.numa_node);
151 if (!htab_is_percpu(htab))
152 goto skip_percpu_elems;
154 for (i = 0; i < num_entries; i++) {
155 u32 size = round_up(htab->map.value_size, 8);
158 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
161 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
167 if (htab_is_lru(htab))
168 err = bpf_lru_init(&htab->lru,
169 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
170 offsetof(struct htab_elem, hash) -
171 offsetof(struct htab_elem, lru_node),
172 htab_lru_map_delete_node,
175 err = pcpu_freelist_init(&htab->freelist);
180 if (htab_is_lru(htab))
181 bpf_lru_populate(&htab->lru, htab->elems,
182 offsetof(struct htab_elem, lru_node),
183 htab->elem_size, num_entries);
185 pcpu_freelist_populate(&htab->freelist,
186 htab->elems + offsetof(struct htab_elem, fnode),
187 htab->elem_size, num_entries);
192 htab_free_elems(htab);
196 static void prealloc_destroy(struct bpf_htab *htab)
198 htab_free_elems(htab);
200 if (htab_is_lru(htab))
201 bpf_lru_destroy(&htab->lru);
203 pcpu_freelist_destroy(&htab->freelist);
206 static int alloc_extra_elems(struct bpf_htab *htab)
208 struct htab_elem *__percpu *pptr, *l_new;
209 struct pcpu_freelist_node *l;
212 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
213 GFP_USER | __GFP_NOWARN);
217 for_each_possible_cpu(cpu) {
218 l = pcpu_freelist_pop(&htab->freelist);
219 /* pop will succeed, since prealloc_init()
220 * preallocated extra num_possible_cpus elements
222 l_new = container_of(l, struct htab_elem, fnode);
223 *per_cpu_ptr(pptr, cpu) = l_new;
225 htab->extra_elems = pptr;
229 /* Called from syscall */
230 static int htab_map_alloc_check(union bpf_attr *attr)
232 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
233 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
234 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
235 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
236 /* percpu_lru means each cpu has its own LRU list.
237 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
238 * the map's value itself is percpu. percpu_lru has
239 * nothing to do with the map's value.
241 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
242 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
243 int numa_node = bpf_map_attr_numa_node(attr);
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.
256 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
257 /* reserved bits should not be used */
260 if (!lru && percpu_lru)
263 if (lru && !prealloc)
266 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
269 /* check sanity of attributes.
270 * value_size == 0 may be allowed in the future to use map as a set
272 if (attr->max_entries == 0 || attr->key_size == 0 ||
273 attr->value_size == 0)
276 if (attr->key_size > MAX_BPF_STACK)
277 /* eBPF programs initialize keys on stack, so they cannot be
278 * larger than max stack size
282 if (attr->value_size >= KMALLOC_MAX_SIZE -
283 MAX_BPF_STACK - sizeof(struct htab_elem))
284 /* if value_size is bigger, the user space won't be able to
285 * access the elements via bpf syscall. This check also makes
286 * sure that the elem_size doesn't overflow and it's
287 * kmalloc-able later in htab_map_update_elem()
294 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
296 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
297 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
298 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
299 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
300 /* percpu_lru means each cpu has its own LRU list.
301 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
302 * the map's value itself is percpu. percpu_lru has
303 * nothing to do with the map's value.
305 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
306 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
307 struct bpf_htab *htab;
311 htab = kzalloc(sizeof(*htab), GFP_USER);
313 return ERR_PTR(-ENOMEM);
315 bpf_map_init_from_attr(&htab->map, attr);
318 /* ensure each CPU's lru list has >=1 elements.
319 * since we are at it, make each lru list has the same
320 * number of elements.
322 htab->map.max_entries = roundup(attr->max_entries,
323 num_possible_cpus());
324 if (htab->map.max_entries < attr->max_entries)
325 htab->map.max_entries = rounddown(attr->max_entries,
326 num_possible_cpus());
329 /* hash table size must be power of 2 */
330 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
332 htab->elem_size = sizeof(struct htab_elem) +
333 round_up(htab->map.key_size, 8);
335 htab->elem_size += sizeof(void *);
337 htab->elem_size += round_up(htab->map.value_size, 8);
340 /* prevent zero size kmalloc and check for u32 overflow */
341 if (htab->n_buckets == 0 ||
342 htab->n_buckets > U32_MAX / sizeof(struct bucket))
345 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
346 (u64) htab->elem_size * htab->map.max_entries;
349 cost += (u64) round_up(htab->map.value_size, 8) *
350 num_possible_cpus() * htab->map.max_entries;
352 cost += (u64) htab->elem_size * num_possible_cpus();
354 if (cost >= U32_MAX - PAGE_SIZE)
355 /* make sure page count doesn't overflow */
358 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
360 /* if map size is larger than memlock limit, reject it early */
361 err = bpf_map_precharge_memlock(htab->map.pages);
366 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
367 sizeof(struct bucket),
368 htab->map.numa_node);
372 for (i = 0; i < htab->n_buckets; i++) {
373 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
374 raw_spin_lock_init(&htab->buckets[i].lock);
378 err = prealloc_init(htab);
382 if (!percpu && !lru) {
383 /* lru itself can remove the least used element, so
384 * there is no need for an extra elem during map_update.
386 err = alloc_extra_elems(htab);
395 prealloc_destroy(htab);
397 bpf_map_area_free(htab->buckets);
403 static inline u32 htab_map_hash(const void *key, u32 key_len)
405 return jhash(key, key_len, 0);
408 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
410 return &htab->buckets[hash & (htab->n_buckets - 1)];
413 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
415 return &__select_bucket(htab, hash)->head;
418 /* this lookup function can only be called with bucket lock taken */
419 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
420 void *key, u32 key_size)
422 struct hlist_nulls_node *n;
425 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
426 if (l->hash == hash && !memcmp(&l->key, key, key_size))
432 /* can be called without bucket lock. it will repeat the loop in
433 * the unlikely event when elements moved from one bucket into another
434 * while link list is being walked
436 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
438 u32 key_size, u32 n_buckets)
440 struct hlist_nulls_node *n;
444 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
445 if (l->hash == hash && !memcmp(&l->key, key, key_size))
448 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
454 /* Called from syscall or from eBPF program directly, so
455 * arguments have to match bpf_map_lookup_elem() exactly.
456 * The return value is adjusted by BPF instructions
457 * in htab_map_gen_lookup().
459 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
461 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
462 struct hlist_nulls_head *head;
466 /* Must be called with rcu_read_lock. */
467 WARN_ON_ONCE(!rcu_read_lock_held());
469 key_size = map->key_size;
471 hash = htab_map_hash(key, key_size);
473 head = select_bucket(htab, hash);
475 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
480 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
482 struct htab_elem *l = __htab_map_lookup_elem(map, key);
485 return l->key + round_up(map->key_size, 8);
490 /* inline bpf_map_lookup_elem() call.
493 * bpf_map_lookup_elem
494 * map->ops->map_lookup_elem
495 * htab_map_lookup_elem
496 * __htab_map_lookup_elem
499 * __htab_map_lookup_elem
501 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
503 struct bpf_insn *insn = insn_buf;
504 const int ret = BPF_REG_0;
506 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
507 (void *(*)(struct bpf_map *map, void *key))NULL));
508 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
509 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
510 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
511 offsetof(struct htab_elem, key) +
512 round_up(map->key_size, 8));
513 return insn - insn_buf;
516 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
518 struct htab_elem *l = __htab_map_lookup_elem(map, key);
521 bpf_lru_node_set_ref(&l->lru_node);
522 return l->key + round_up(map->key_size, 8);
528 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
529 struct bpf_insn *insn_buf)
531 struct bpf_insn *insn = insn_buf;
532 const int ret = BPF_REG_0;
533 const int ref_reg = BPF_REG_1;
535 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
536 (void *(*)(struct bpf_map *map, void *key))NULL));
537 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
538 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
539 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
540 offsetof(struct htab_elem, lru_node) +
541 offsetof(struct bpf_lru_node, ref));
542 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
543 *insn++ = BPF_ST_MEM(BPF_B, ret,
544 offsetof(struct htab_elem, lru_node) +
545 offsetof(struct bpf_lru_node, ref),
547 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
548 offsetof(struct htab_elem, key) +
549 round_up(map->key_size, 8));
550 return insn - insn_buf;
553 /* It is called from the bpf_lru_list when the LRU needs to delete
554 * older elements from the htab.
556 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
558 struct bpf_htab *htab = (struct bpf_htab *)arg;
559 struct htab_elem *l = NULL, *tgt_l;
560 struct hlist_nulls_head *head;
561 struct hlist_nulls_node *n;
565 tgt_l = container_of(node, struct htab_elem, lru_node);
566 b = __select_bucket(htab, tgt_l->hash);
569 raw_spin_lock_irqsave(&b->lock, flags);
571 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
573 hlist_nulls_del_rcu(&l->hash_node);
577 raw_spin_unlock_irqrestore(&b->lock, flags);
582 /* Called from syscall */
583 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
585 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
586 struct hlist_nulls_head *head;
587 struct htab_elem *l, *next_l;
591 WARN_ON_ONCE(!rcu_read_lock_held());
593 key_size = map->key_size;
596 goto find_first_elem;
598 hash = htab_map_hash(key, key_size);
600 head = select_bucket(htab, hash);
603 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
606 goto find_first_elem;
608 /* key was found, get next key in the same bucket */
609 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
610 struct htab_elem, hash_node);
613 /* if next elem in this hash list is non-zero, just return it */
614 memcpy(next_key, next_l->key, key_size);
618 /* no more elements in this hash list, go to the next bucket */
619 i = hash & (htab->n_buckets - 1);
623 /* iterate over buckets */
624 for (; i < htab->n_buckets; i++) {
625 head = select_bucket(htab, i);
627 /* pick first element in the bucket */
628 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
629 struct htab_elem, hash_node);
631 /* if it's not empty, just return it */
632 memcpy(next_key, next_l->key, key_size);
637 /* iterated over all buckets and all elements */
641 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
643 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
644 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
648 static void htab_elem_free_rcu(struct rcu_head *head)
650 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
651 struct bpf_htab *htab = l->htab;
653 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
654 * we're calling kfree, otherwise deadlock is possible if kprobes
655 * are placed somewhere inside of slub
658 __this_cpu_inc(bpf_prog_active);
659 htab_elem_free(htab, l);
660 __this_cpu_dec(bpf_prog_active);
664 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
666 struct bpf_map *map = &htab->map;
668 if (map->ops->map_fd_put_ptr) {
669 void *ptr = fd_htab_map_get_ptr(map, l);
671 map->ops->map_fd_put_ptr(ptr);
674 if (htab_is_prealloc(htab)) {
675 pcpu_freelist_push(&htab->freelist, &l->fnode);
677 atomic_dec(&htab->count);
679 call_rcu(&l->rcu, htab_elem_free_rcu);
683 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
684 void *value, bool onallcpus)
687 /* copy true value_size bytes */
688 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
690 u32 size = round_up(htab->map.value_size, 8);
693 for_each_possible_cpu(cpu) {
694 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
701 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
703 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
707 static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
709 u32 size = htab->map.value_size;
711 if (percpu || fd_htab_map_needs_adjust(htab))
712 size = round_up(size, 8);
716 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
717 void *value, u32 key_size, u32 hash,
718 bool percpu, bool onallcpus,
719 struct htab_elem *old_elem)
721 u32 size = htab_size_value(htab, percpu);
722 bool prealloc = htab_is_prealloc(htab);
723 struct htab_elem *l_new, **pl_new;
728 /* if we're updating the existing element,
729 * use per-cpu extra elems to avoid freelist_pop/push
731 pl_new = this_cpu_ptr(htab->extra_elems);
735 struct pcpu_freelist_node *l;
737 l = pcpu_freelist_pop(&htab->freelist);
739 return ERR_PTR(-E2BIG);
740 l_new = container_of(l, struct htab_elem, fnode);
743 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
745 /* when map is full and update() is replacing
746 * old element, it's ok to allocate, since
747 * old element will be freed immediately.
748 * Otherwise return an error
750 atomic_dec(&htab->count);
751 return ERR_PTR(-E2BIG);
753 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
754 htab->map.numa_node);
756 return ERR_PTR(-ENOMEM);
759 memcpy(l_new->key, key, key_size);
762 pptr = htab_elem_get_ptr(l_new, key_size);
764 /* alloc_percpu zero-fills */
765 pptr = __alloc_percpu_gfp(size, 8,
766 GFP_ATOMIC | __GFP_NOWARN);
769 return ERR_PTR(-ENOMEM);
773 pcpu_copy_value(htab, pptr, value, onallcpus);
776 htab_elem_set_ptr(l_new, key_size, pptr);
778 memcpy(l_new->key + round_up(key_size, 8), value, size);
785 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
788 if (l_old && map_flags == BPF_NOEXIST)
789 /* elem already exists */
792 if (!l_old && map_flags == BPF_EXIST)
793 /* elem doesn't exist, cannot update it */
799 /* Called from syscall or from eBPF program */
800 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
803 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
804 struct htab_elem *l_new = NULL, *l_old;
805 struct hlist_nulls_head *head;
811 if (unlikely(map_flags > BPF_EXIST))
815 WARN_ON_ONCE(!rcu_read_lock_held());
817 key_size = map->key_size;
819 hash = htab_map_hash(key, key_size);
821 b = __select_bucket(htab, hash);
824 /* bpf_map_update_elem() can be called in_irq() */
825 raw_spin_lock_irqsave(&b->lock, flags);
827 l_old = lookup_elem_raw(head, hash, key, key_size);
829 ret = check_flags(htab, l_old, map_flags);
833 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
836 /* all pre-allocated elements are in use or memory exhausted */
837 ret = PTR_ERR(l_new);
841 /* add new element to the head of the list, so that
842 * concurrent search will find it before old elem
844 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
846 hlist_nulls_del_rcu(&l_old->hash_node);
847 if (!htab_is_prealloc(htab))
848 free_htab_elem(htab, l_old);
852 raw_spin_unlock_irqrestore(&b->lock, flags);
856 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
859 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
860 struct htab_elem *l_new, *l_old = NULL;
861 struct hlist_nulls_head *head;
867 if (unlikely(map_flags > BPF_EXIST))
871 WARN_ON_ONCE(!rcu_read_lock_held());
873 key_size = map->key_size;
875 hash = htab_map_hash(key, key_size);
877 b = __select_bucket(htab, hash);
880 /* For LRU, we need to alloc before taking bucket's
881 * spinlock because getting free nodes from LRU may need
882 * to remove older elements from htab and this removal
883 * operation will need a bucket lock.
885 l_new = prealloc_lru_pop(htab, key, hash);
888 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
890 /* bpf_map_update_elem() can be called in_irq() */
891 raw_spin_lock_irqsave(&b->lock, flags);
893 l_old = lookup_elem_raw(head, hash, key, key_size);
895 ret = check_flags(htab, l_old, map_flags);
899 /* add new element to the head of the list, so that
900 * concurrent search will find it before old elem
902 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
904 bpf_lru_node_set_ref(&l_new->lru_node);
905 hlist_nulls_del_rcu(&l_old->hash_node);
910 raw_spin_unlock_irqrestore(&b->lock, flags);
913 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
915 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
920 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
921 void *value, u64 map_flags,
924 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
925 struct htab_elem *l_new = NULL, *l_old;
926 struct hlist_nulls_head *head;
932 if (unlikely(map_flags > BPF_EXIST))
936 WARN_ON_ONCE(!rcu_read_lock_held());
938 key_size = map->key_size;
940 hash = htab_map_hash(key, key_size);
942 b = __select_bucket(htab, hash);
945 /* bpf_map_update_elem() can be called in_irq() */
946 raw_spin_lock_irqsave(&b->lock, flags);
948 l_old = lookup_elem_raw(head, hash, key, key_size);
950 ret = check_flags(htab, l_old, map_flags);
955 /* per-cpu hash map can update value in-place */
956 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
959 l_new = alloc_htab_elem(htab, key, value, key_size,
960 hash, true, onallcpus, NULL);
962 ret = PTR_ERR(l_new);
965 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
969 raw_spin_unlock_irqrestore(&b->lock, flags);
973 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
974 void *value, u64 map_flags,
977 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
978 struct htab_elem *l_new = NULL, *l_old;
979 struct hlist_nulls_head *head;
985 if (unlikely(map_flags > BPF_EXIST))
989 WARN_ON_ONCE(!rcu_read_lock_held());
991 key_size = map->key_size;
993 hash = htab_map_hash(key, key_size);
995 b = __select_bucket(htab, hash);
998 /* For LRU, we need to alloc before taking bucket's
999 * spinlock because LRU's elem alloc may need
1000 * to remove older elem from htab and this removal
1001 * operation will need a bucket lock.
1003 if (map_flags != BPF_EXIST) {
1004 l_new = prealloc_lru_pop(htab, key, hash);
1009 /* bpf_map_update_elem() can be called in_irq() */
1010 raw_spin_lock_irqsave(&b->lock, flags);
1012 l_old = lookup_elem_raw(head, hash, key, key_size);
1014 ret = check_flags(htab, l_old, map_flags);
1019 bpf_lru_node_set_ref(&l_old->lru_node);
1021 /* per-cpu hash map can update value in-place */
1022 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1025 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1027 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1032 raw_spin_unlock_irqrestore(&b->lock, flags);
1034 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1038 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1039 void *value, u64 map_flags)
1041 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1044 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1045 void *value, u64 map_flags)
1047 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1051 /* Called from syscall or from eBPF program */
1052 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1054 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1055 struct hlist_nulls_head *head;
1057 struct htab_elem *l;
1058 unsigned long flags;
1062 WARN_ON_ONCE(!rcu_read_lock_held());
1064 key_size = map->key_size;
1066 hash = htab_map_hash(key, key_size);
1067 b = __select_bucket(htab, hash);
1070 raw_spin_lock_irqsave(&b->lock, flags);
1072 l = lookup_elem_raw(head, hash, key, key_size);
1075 hlist_nulls_del_rcu(&l->hash_node);
1076 free_htab_elem(htab, l);
1080 raw_spin_unlock_irqrestore(&b->lock, flags);
1084 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1086 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1087 struct hlist_nulls_head *head;
1089 struct htab_elem *l;
1090 unsigned long flags;
1094 WARN_ON_ONCE(!rcu_read_lock_held());
1096 key_size = map->key_size;
1098 hash = htab_map_hash(key, key_size);
1099 b = __select_bucket(htab, hash);
1102 raw_spin_lock_irqsave(&b->lock, flags);
1104 l = lookup_elem_raw(head, hash, key, key_size);
1107 hlist_nulls_del_rcu(&l->hash_node);
1111 raw_spin_unlock_irqrestore(&b->lock, flags);
1113 bpf_lru_push_free(&htab->lru, &l->lru_node);
1117 static void delete_all_elements(struct bpf_htab *htab)
1121 for (i = 0; i < htab->n_buckets; i++) {
1122 struct hlist_nulls_head *head = select_bucket(htab, i);
1123 struct hlist_nulls_node *n;
1124 struct htab_elem *l;
1126 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1127 hlist_nulls_del_rcu(&l->hash_node);
1128 htab_elem_free(htab, l);
1133 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1134 static void htab_map_free(struct bpf_map *map)
1136 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1138 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1139 * so the programs (can be more than one that used this map) were
1140 * disconnected from events. Wait for outstanding critical sections in
1141 * these programs to complete
1145 /* some of free_htab_elem() callbacks for elements of this map may
1146 * not have executed. Wait for them.
1149 if (!htab_is_prealloc(htab))
1150 delete_all_elements(htab);
1152 prealloc_destroy(htab);
1154 free_percpu(htab->extra_elems);
1155 bpf_map_area_free(htab->buckets);
1159 const struct bpf_map_ops htab_map_ops = {
1160 .map_alloc_check = htab_map_alloc_check,
1161 .map_alloc = htab_map_alloc,
1162 .map_free = htab_map_free,
1163 .map_get_next_key = htab_map_get_next_key,
1164 .map_lookup_elem = htab_map_lookup_elem,
1165 .map_update_elem = htab_map_update_elem,
1166 .map_delete_elem = htab_map_delete_elem,
1167 .map_gen_lookup = htab_map_gen_lookup,
1170 const struct bpf_map_ops htab_lru_map_ops = {
1171 .map_alloc_check = htab_map_alloc_check,
1172 .map_alloc = htab_map_alloc,
1173 .map_free = htab_map_free,
1174 .map_get_next_key = htab_map_get_next_key,
1175 .map_lookup_elem = htab_lru_map_lookup_elem,
1176 .map_update_elem = htab_lru_map_update_elem,
1177 .map_delete_elem = htab_lru_map_delete_elem,
1178 .map_gen_lookup = htab_lru_map_gen_lookup,
1181 /* Called from eBPF program */
1182 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1184 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1187 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1192 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1194 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1197 bpf_lru_node_set_ref(&l->lru_node);
1198 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1204 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1206 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1207 struct htab_elem *l;
1208 void __percpu *pptr;
1213 /* per_cpu areas are zero-filled and bpf programs can only
1214 * access 'value_size' of them, so copying rounded areas
1215 * will not leak any kernel data
1217 size = round_up(map->value_size, 8);
1219 l = __htab_map_lookup_elem(map, key);
1222 if (htab_is_lru(htab))
1223 bpf_lru_node_set_ref(&l->lru_node);
1224 pptr = htab_elem_get_ptr(l, map->key_size);
1225 for_each_possible_cpu(cpu) {
1226 bpf_long_memcpy(value + off,
1227 per_cpu_ptr(pptr, cpu), size);
1236 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1239 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1243 if (htab_is_lru(htab))
1244 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1247 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1254 const struct bpf_map_ops htab_percpu_map_ops = {
1255 .map_alloc_check = htab_map_alloc_check,
1256 .map_alloc = htab_map_alloc,
1257 .map_free = htab_map_free,
1258 .map_get_next_key = htab_map_get_next_key,
1259 .map_lookup_elem = htab_percpu_map_lookup_elem,
1260 .map_update_elem = htab_percpu_map_update_elem,
1261 .map_delete_elem = htab_map_delete_elem,
1264 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1265 .map_alloc_check = htab_map_alloc_check,
1266 .map_alloc = htab_map_alloc,
1267 .map_free = htab_map_free,
1268 .map_get_next_key = htab_map_get_next_key,
1269 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1270 .map_update_elem = htab_lru_percpu_map_update_elem,
1271 .map_delete_elem = htab_lru_map_delete_elem,
1274 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1276 if (attr->value_size != sizeof(u32))
1278 return htab_map_alloc_check(attr);
1281 static void fd_htab_map_free(struct bpf_map *map)
1283 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1284 struct hlist_nulls_node *n;
1285 struct hlist_nulls_head *head;
1286 struct htab_elem *l;
1289 for (i = 0; i < htab->n_buckets; i++) {
1290 head = select_bucket(htab, i);
1292 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1293 void *ptr = fd_htab_map_get_ptr(map, l);
1295 map->ops->map_fd_put_ptr(ptr);
1302 /* only called from syscall */
1303 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1308 if (!map->ops->map_fd_sys_lookup_elem)
1312 ptr = htab_map_lookup_elem(map, key);
1314 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1322 /* only called from syscall */
1323 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1324 void *key, void *value, u64 map_flags)
1328 u32 ufd = *(u32 *)value;
1330 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1332 return PTR_ERR(ptr);
1334 ret = htab_map_update_elem(map, key, &ptr, map_flags);
1336 map->ops->map_fd_put_ptr(ptr);
1341 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1343 struct bpf_map *map, *inner_map_meta;
1345 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1346 if (IS_ERR(inner_map_meta))
1347 return inner_map_meta;
1349 map = htab_map_alloc(attr);
1351 bpf_map_meta_free(inner_map_meta);
1355 map->inner_map_meta = inner_map_meta;
1360 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1362 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
1367 return READ_ONCE(*inner_map);
1370 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1371 struct bpf_insn *insn_buf)
1373 struct bpf_insn *insn = insn_buf;
1374 const int ret = BPF_REG_0;
1376 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1377 (void *(*)(struct bpf_map *map, void *key))NULL));
1378 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1379 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1380 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1381 offsetof(struct htab_elem, key) +
1382 round_up(map->key_size, 8));
1383 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1385 return insn - insn_buf;
1388 static void htab_of_map_free(struct bpf_map *map)
1390 bpf_map_meta_free(map->inner_map_meta);
1391 fd_htab_map_free(map);
1394 const struct bpf_map_ops htab_of_maps_map_ops = {
1395 .map_alloc_check = fd_htab_map_alloc_check,
1396 .map_alloc = htab_of_map_alloc,
1397 .map_free = htab_of_map_free,
1398 .map_get_next_key = htab_map_get_next_key,
1399 .map_lookup_elem = htab_of_map_lookup_elem,
1400 .map_delete_elem = htab_map_delete_elem,
1401 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1402 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1403 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1404 .map_gen_lookup = htab_of_map_gen_lookup,