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"
22 struct hlist_nulls_head head;
28 struct bucket *buckets;
31 struct pcpu_freelist freelist;
34 struct htab_elem *__percpu *extra_elems;
35 atomic_t count; /* number of elements in this hashtable */
36 u32 n_buckets; /* number of hash buckets */
37 u32 elem_size; /* size of each element in bytes */
40 /* each htab element is struct htab_elem + key + value */
43 struct hlist_nulls_node hash_node;
47 struct bpf_htab *htab;
48 struct pcpu_freelist_node fnode;
54 struct bpf_lru_node lru_node;
57 char key[0] __aligned(8);
60 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
62 static bool htab_is_lru(const struct bpf_htab *htab)
64 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
65 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
68 static bool htab_is_percpu(const struct bpf_htab *htab)
70 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
71 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
74 static bool htab_is_prealloc(const struct bpf_htab *htab)
76 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
79 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
82 *(void __percpu **)(l->key + key_size) = pptr;
85 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
87 return *(void __percpu **)(l->key + key_size);
90 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
92 return *(void **)(l->key + roundup(map->key_size, 8));
95 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
97 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
100 static void htab_free_elems(struct bpf_htab *htab)
104 if (!htab_is_percpu(htab))
107 for (i = 0; i < htab->map.max_entries; i++) {
110 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
115 bpf_map_area_free(htab->elems);
118 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
121 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
125 l = container_of(node, struct htab_elem, lru_node);
126 memcpy(l->key, key, htab->map.key_size);
133 static int prealloc_init(struct bpf_htab *htab)
135 u32 num_entries = htab->map.max_entries;
136 int err = -ENOMEM, i;
138 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
139 num_entries += num_possible_cpus();
141 htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries);
145 if (!htab_is_percpu(htab))
146 goto skip_percpu_elems;
148 for (i = 0; i < num_entries; i++) {
149 u32 size = round_up(htab->map.value_size, 8);
152 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
155 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
160 if (htab_is_lru(htab))
161 err = bpf_lru_init(&htab->lru,
162 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
163 offsetof(struct htab_elem, hash) -
164 offsetof(struct htab_elem, lru_node),
165 htab_lru_map_delete_node,
168 err = pcpu_freelist_init(&htab->freelist);
173 if (htab_is_lru(htab))
174 bpf_lru_populate(&htab->lru, htab->elems,
175 offsetof(struct htab_elem, lru_node),
176 htab->elem_size, num_entries);
178 pcpu_freelist_populate(&htab->freelist,
179 htab->elems + offsetof(struct htab_elem, fnode),
180 htab->elem_size, num_entries);
185 htab_free_elems(htab);
189 static void prealloc_destroy(struct bpf_htab *htab)
191 htab_free_elems(htab);
193 if (htab_is_lru(htab))
194 bpf_lru_destroy(&htab->lru);
196 pcpu_freelist_destroy(&htab->freelist);
199 static int alloc_extra_elems(struct bpf_htab *htab)
201 struct htab_elem *__percpu *pptr, *l_new;
202 struct pcpu_freelist_node *l;
205 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
206 GFP_USER | __GFP_NOWARN);
210 for_each_possible_cpu(cpu) {
211 l = pcpu_freelist_pop(&htab->freelist);
212 /* pop will succeed, since prealloc_init()
213 * preallocated extra num_possible_cpus elements
215 l_new = container_of(l, struct htab_elem, fnode);
216 *per_cpu_ptr(pptr, cpu) = l_new;
218 htab->extra_elems = pptr;
222 /* Called from syscall */
223 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
225 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
226 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
227 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
228 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
229 /* percpu_lru means each cpu has its own LRU list.
230 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
231 * the map's value itself is percpu. percpu_lru has
232 * nothing to do with the map's value.
234 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
235 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
236 struct bpf_htab *htab;
240 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
241 offsetof(struct htab_elem, hash_node.pprev));
242 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
243 offsetof(struct htab_elem, hash_node.pprev));
245 if (lru && !capable(CAP_SYS_ADMIN))
246 /* LRU implementation is much complicated than other
247 * maps. Hence, limit to CAP_SYS_ADMIN for now.
249 return ERR_PTR(-EPERM);
251 if (attr->map_flags & ~(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU))
252 /* reserved bits should not be used */
253 return ERR_PTR(-EINVAL);
255 if (!lru && percpu_lru)
256 return ERR_PTR(-EINVAL);
258 if (lru && !prealloc)
259 return ERR_PTR(-ENOTSUPP);
261 htab = kzalloc(sizeof(*htab), GFP_USER);
263 return ERR_PTR(-ENOMEM);
265 /* mandatory map attributes */
266 htab->map.map_type = attr->map_type;
267 htab->map.key_size = attr->key_size;
268 htab->map.value_size = attr->value_size;
269 htab->map.max_entries = attr->max_entries;
270 htab->map.map_flags = attr->map_flags;
272 /* check sanity of attributes.
273 * value_size == 0 may be allowed in the future to use map as a set
276 if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
277 htab->map.value_size == 0)
281 /* ensure each CPU's lru list has >=1 elements.
282 * since we are at it, make each lru list has the same
283 * number of elements.
285 htab->map.max_entries = roundup(attr->max_entries,
286 num_possible_cpus());
287 if (htab->map.max_entries < attr->max_entries)
288 htab->map.max_entries = rounddown(attr->max_entries,
289 num_possible_cpus());
292 /* hash table size must be power of 2 */
293 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
296 if (htab->map.key_size > MAX_BPF_STACK)
297 /* eBPF programs initialize keys on stack, so they cannot be
298 * larger than max stack size
302 if (htab->map.value_size >= KMALLOC_MAX_SIZE -
303 MAX_BPF_STACK - sizeof(struct htab_elem))
304 /* if value_size is bigger, the user space won't be able to
305 * access the elements via bpf syscall. This check also makes
306 * sure that the elem_size doesn't overflow and it's
307 * kmalloc-able later in htab_map_update_elem()
311 if (percpu && round_up(htab->map.value_size, 8) > PCPU_MIN_UNIT_SIZE)
312 /* make sure the size for pcpu_alloc() is reasonable */
315 htab->elem_size = sizeof(struct htab_elem) +
316 round_up(htab->map.key_size, 8);
318 htab->elem_size += sizeof(void *);
320 htab->elem_size += round_up(htab->map.value_size, 8);
322 /* prevent zero size kmalloc and check for u32 overflow */
323 if (htab->n_buckets == 0 ||
324 htab->n_buckets > U32_MAX / sizeof(struct bucket))
327 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
328 (u64) htab->elem_size * htab->map.max_entries;
331 cost += (u64) round_up(htab->map.value_size, 8) *
332 num_possible_cpus() * htab->map.max_entries;
334 cost += (u64) htab->elem_size * num_possible_cpus();
336 if (cost >= U32_MAX - PAGE_SIZE)
337 /* make sure page count doesn't overflow */
340 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
342 /* if map size is larger than memlock limit, reject it early */
343 err = bpf_map_precharge_memlock(htab->map.pages);
348 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
349 sizeof(struct bucket));
353 for (i = 0; i < htab->n_buckets; i++) {
354 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
355 raw_spin_lock_init(&htab->buckets[i].lock);
359 err = prealloc_init(htab);
363 if (!percpu && !lru) {
364 /* lru itself can remove the least used element, so
365 * there is no need for an extra elem during map_update.
367 err = alloc_extra_elems(htab);
376 prealloc_destroy(htab);
378 bpf_map_area_free(htab->buckets);
384 static inline u32 htab_map_hash(const void *key, u32 key_len)
386 return jhash(key, key_len, 0);
389 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
391 return &htab->buckets[hash & (htab->n_buckets - 1)];
394 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
396 return &__select_bucket(htab, hash)->head;
399 /* this lookup function can only be called with bucket lock taken */
400 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
401 void *key, u32 key_size)
403 struct hlist_nulls_node *n;
406 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
407 if (l->hash == hash && !memcmp(&l->key, key, key_size))
413 /* can be called without bucket lock. it will repeat the loop in
414 * the unlikely event when elements moved from one bucket into another
415 * while link list is being walked
417 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
419 u32 key_size, u32 n_buckets)
421 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))
429 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
435 /* Called from syscall or from eBPF program directly, so
436 * arguments have to match bpf_map_lookup_elem() exactly.
437 * The return value is adjusted by BPF instructions
438 * in htab_map_gen_lookup().
440 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
442 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
443 struct hlist_nulls_head *head;
447 /* Must be called with rcu_read_lock. */
448 WARN_ON_ONCE(!rcu_read_lock_held());
450 key_size = map->key_size;
452 hash = htab_map_hash(key, key_size);
454 head = select_bucket(htab, hash);
456 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
461 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
463 struct htab_elem *l = __htab_map_lookup_elem(map, key);
466 return l->key + round_up(map->key_size, 8);
471 /* inline bpf_map_lookup_elem() call.
474 * bpf_map_lookup_elem
475 * map->ops->map_lookup_elem
476 * htab_map_lookup_elem
477 * __htab_map_lookup_elem
480 * __htab_map_lookup_elem
482 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
484 struct bpf_insn *insn = insn_buf;
485 const int ret = BPF_REG_0;
487 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
488 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
489 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
490 offsetof(struct htab_elem, key) +
491 round_up(map->key_size, 8));
492 return insn - insn_buf;
495 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
497 struct htab_elem *l = __htab_map_lookup_elem(map, key);
500 bpf_lru_node_set_ref(&l->lru_node);
501 return l->key + round_up(map->key_size, 8);
507 /* It is called from the bpf_lru_list when the LRU needs to delete
508 * older elements from the htab.
510 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
512 struct bpf_htab *htab = (struct bpf_htab *)arg;
513 struct htab_elem *l = NULL, *tgt_l;
514 struct hlist_nulls_head *head;
515 struct hlist_nulls_node *n;
519 tgt_l = container_of(node, struct htab_elem, lru_node);
520 b = __select_bucket(htab, tgt_l->hash);
523 raw_spin_lock_irqsave(&b->lock, flags);
525 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
527 hlist_nulls_del_rcu(&l->hash_node);
531 raw_spin_unlock_irqrestore(&b->lock, flags);
536 /* Called from syscall */
537 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
539 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
540 struct hlist_nulls_head *head;
541 struct htab_elem *l, *next_l;
545 WARN_ON_ONCE(!rcu_read_lock_held());
547 key_size = map->key_size;
550 goto find_first_elem;
552 hash = htab_map_hash(key, key_size);
554 head = select_bucket(htab, hash);
557 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
560 goto find_first_elem;
562 /* key was found, get next key in the same bucket */
563 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
564 struct htab_elem, hash_node);
567 /* if next elem in this hash list is non-zero, just return it */
568 memcpy(next_key, next_l->key, key_size);
572 /* no more elements in this hash list, go to the next bucket */
573 i = hash & (htab->n_buckets - 1);
577 /* iterate over buckets */
578 for (; i < htab->n_buckets; i++) {
579 head = select_bucket(htab, i);
581 /* pick first element in the bucket */
582 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
583 struct htab_elem, hash_node);
585 /* if it's not empty, just return it */
586 memcpy(next_key, next_l->key, key_size);
591 /* iterated over all buckets and all elements */
595 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
597 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
598 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
602 static void htab_elem_free_rcu(struct rcu_head *head)
604 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
605 struct bpf_htab *htab = l->htab;
607 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
608 * we're calling kfree, otherwise deadlock is possible if kprobes
609 * are placed somewhere inside of slub
612 __this_cpu_inc(bpf_prog_active);
613 htab_elem_free(htab, l);
614 __this_cpu_dec(bpf_prog_active);
618 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
620 struct bpf_map *map = &htab->map;
622 if (map->ops->map_fd_put_ptr) {
623 void *ptr = fd_htab_map_get_ptr(map, l);
625 map->ops->map_fd_put_ptr(ptr);
628 if (htab_is_prealloc(htab)) {
629 pcpu_freelist_push(&htab->freelist, &l->fnode);
631 atomic_dec(&htab->count);
633 call_rcu(&l->rcu, htab_elem_free_rcu);
637 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
638 void *value, bool onallcpus)
641 /* copy true value_size bytes */
642 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
644 u32 size = round_up(htab->map.value_size, 8);
647 for_each_possible_cpu(cpu) {
648 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
655 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
656 void *value, u32 key_size, u32 hash,
657 bool percpu, bool onallcpus,
658 struct htab_elem *old_elem)
660 u32 size = htab->map.value_size;
661 bool prealloc = htab_is_prealloc(htab);
662 struct htab_elem *l_new, **pl_new;
667 /* if we're updating the existing element,
668 * use per-cpu extra elems to avoid freelist_pop/push
670 pl_new = this_cpu_ptr(htab->extra_elems);
674 struct pcpu_freelist_node *l;
676 l = pcpu_freelist_pop(&htab->freelist);
678 return ERR_PTR(-E2BIG);
679 l_new = container_of(l, struct htab_elem, fnode);
682 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
684 /* when map is full and update() is replacing
685 * old element, it's ok to allocate, since
686 * old element will be freed immediately.
687 * Otherwise return an error
689 atomic_dec(&htab->count);
690 return ERR_PTR(-E2BIG);
692 l_new = kmalloc(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN);
694 return ERR_PTR(-ENOMEM);
697 memcpy(l_new->key, key, key_size);
699 /* round up value_size to 8 bytes */
700 size = round_up(size, 8);
703 pptr = htab_elem_get_ptr(l_new, key_size);
705 /* alloc_percpu zero-fills */
706 pptr = __alloc_percpu_gfp(size, 8,
707 GFP_ATOMIC | __GFP_NOWARN);
710 return ERR_PTR(-ENOMEM);
714 pcpu_copy_value(htab, pptr, value, onallcpus);
717 htab_elem_set_ptr(l_new, key_size, pptr);
719 memcpy(l_new->key + round_up(key_size, 8), value, size);
726 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
729 if (l_old && map_flags == BPF_NOEXIST)
730 /* elem already exists */
733 if (!l_old && map_flags == BPF_EXIST)
734 /* elem doesn't exist, cannot update it */
740 /* Called from syscall or from eBPF program */
741 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
744 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
745 struct htab_elem *l_new = NULL, *l_old;
746 struct hlist_nulls_head *head;
752 if (unlikely(map_flags > BPF_EXIST))
756 WARN_ON_ONCE(!rcu_read_lock_held());
758 key_size = map->key_size;
760 hash = htab_map_hash(key, key_size);
762 b = __select_bucket(htab, hash);
765 /* bpf_map_update_elem() can be called in_irq() */
766 raw_spin_lock_irqsave(&b->lock, flags);
768 l_old = lookup_elem_raw(head, hash, key, key_size);
770 ret = check_flags(htab, l_old, map_flags);
774 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
777 /* all pre-allocated elements are in use or memory exhausted */
778 ret = PTR_ERR(l_new);
782 /* add new element to the head of the list, so that
783 * concurrent search will find it before old elem
785 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
787 hlist_nulls_del_rcu(&l_old->hash_node);
788 if (!htab_is_prealloc(htab))
789 free_htab_elem(htab, l_old);
793 raw_spin_unlock_irqrestore(&b->lock, flags);
797 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
800 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
801 struct htab_elem *l_new, *l_old = NULL;
802 struct hlist_nulls_head *head;
808 if (unlikely(map_flags > BPF_EXIST))
812 WARN_ON_ONCE(!rcu_read_lock_held());
814 key_size = map->key_size;
816 hash = htab_map_hash(key, key_size);
818 b = __select_bucket(htab, hash);
821 /* For LRU, we need to alloc before taking bucket's
822 * spinlock because getting free nodes from LRU may need
823 * to remove older elements from htab and this removal
824 * operation will need a bucket lock.
826 l_new = prealloc_lru_pop(htab, key, hash);
829 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
831 /* bpf_map_update_elem() can be called in_irq() */
832 raw_spin_lock_irqsave(&b->lock, flags);
834 l_old = lookup_elem_raw(head, hash, key, key_size);
836 ret = check_flags(htab, l_old, map_flags);
840 /* add new element to the head of the list, so that
841 * concurrent search will find it before old elem
843 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
845 bpf_lru_node_set_ref(&l_new->lru_node);
846 hlist_nulls_del_rcu(&l_old->hash_node);
851 raw_spin_unlock_irqrestore(&b->lock, flags);
854 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
856 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
861 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
862 void *value, u64 map_flags,
865 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
866 struct htab_elem *l_new = NULL, *l_old;
867 struct hlist_nulls_head *head;
873 if (unlikely(map_flags > BPF_EXIST))
877 WARN_ON_ONCE(!rcu_read_lock_held());
879 key_size = map->key_size;
881 hash = htab_map_hash(key, key_size);
883 b = __select_bucket(htab, hash);
886 /* bpf_map_update_elem() can be called in_irq() */
887 raw_spin_lock_irqsave(&b->lock, flags);
889 l_old = lookup_elem_raw(head, hash, key, key_size);
891 ret = check_flags(htab, l_old, map_flags);
896 /* per-cpu hash map can update value in-place */
897 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
900 l_new = alloc_htab_elem(htab, key, value, key_size,
901 hash, true, onallcpus, NULL);
903 ret = PTR_ERR(l_new);
906 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
910 raw_spin_unlock_irqrestore(&b->lock, flags);
914 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
915 void *value, u64 map_flags,
918 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
919 struct htab_elem *l_new = NULL, *l_old;
920 struct hlist_nulls_head *head;
926 if (unlikely(map_flags > BPF_EXIST))
930 WARN_ON_ONCE(!rcu_read_lock_held());
932 key_size = map->key_size;
934 hash = htab_map_hash(key, key_size);
936 b = __select_bucket(htab, hash);
939 /* For LRU, we need to alloc before taking bucket's
940 * spinlock because LRU's elem alloc may need
941 * to remove older elem from htab and this removal
942 * operation will need a bucket lock.
944 if (map_flags != BPF_EXIST) {
945 l_new = prealloc_lru_pop(htab, key, hash);
950 /* bpf_map_update_elem() can be called in_irq() */
951 raw_spin_lock_irqsave(&b->lock, flags);
953 l_old = lookup_elem_raw(head, hash, key, key_size);
955 ret = check_flags(htab, l_old, map_flags);
960 bpf_lru_node_set_ref(&l_old->lru_node);
962 /* per-cpu hash map can update value in-place */
963 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
966 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
968 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
973 raw_spin_unlock_irqrestore(&b->lock, flags);
975 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
979 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
980 void *value, u64 map_flags)
982 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
985 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
986 void *value, u64 map_flags)
988 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
992 /* Called from syscall or from eBPF program */
993 static int htab_map_delete_elem(struct bpf_map *map, void *key)
995 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
996 struct hlist_nulls_head *head;
1003 WARN_ON_ONCE(!rcu_read_lock_held());
1005 key_size = map->key_size;
1007 hash = htab_map_hash(key, key_size);
1008 b = __select_bucket(htab, hash);
1011 raw_spin_lock_irqsave(&b->lock, flags);
1013 l = lookup_elem_raw(head, hash, key, key_size);
1016 hlist_nulls_del_rcu(&l->hash_node);
1017 free_htab_elem(htab, l);
1021 raw_spin_unlock_irqrestore(&b->lock, flags);
1025 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1027 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1028 struct hlist_nulls_head *head;
1030 struct htab_elem *l;
1031 unsigned long flags;
1035 WARN_ON_ONCE(!rcu_read_lock_held());
1037 key_size = map->key_size;
1039 hash = htab_map_hash(key, key_size);
1040 b = __select_bucket(htab, hash);
1043 raw_spin_lock_irqsave(&b->lock, flags);
1045 l = lookup_elem_raw(head, hash, key, key_size);
1048 hlist_nulls_del_rcu(&l->hash_node);
1052 raw_spin_unlock_irqrestore(&b->lock, flags);
1054 bpf_lru_push_free(&htab->lru, &l->lru_node);
1058 static void delete_all_elements(struct bpf_htab *htab)
1062 for (i = 0; i < htab->n_buckets; i++) {
1063 struct hlist_nulls_head *head = select_bucket(htab, i);
1064 struct hlist_nulls_node *n;
1065 struct htab_elem *l;
1067 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1068 hlist_nulls_del_rcu(&l->hash_node);
1069 htab_elem_free(htab, l);
1074 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1075 static void htab_map_free(struct bpf_map *map)
1077 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1079 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1080 * so the programs (can be more than one that used this map) were
1081 * disconnected from events. Wait for outstanding critical sections in
1082 * these programs to complete
1086 /* some of free_htab_elem() callbacks for elements of this map may
1087 * not have executed. Wait for them.
1090 if (!htab_is_prealloc(htab))
1091 delete_all_elements(htab);
1093 prealloc_destroy(htab);
1095 free_percpu(htab->extra_elems);
1096 bpf_map_area_free(htab->buckets);
1100 const struct bpf_map_ops htab_map_ops = {
1101 .map_alloc = htab_map_alloc,
1102 .map_free = htab_map_free,
1103 .map_get_next_key = htab_map_get_next_key,
1104 .map_lookup_elem = htab_map_lookup_elem,
1105 .map_update_elem = htab_map_update_elem,
1106 .map_delete_elem = htab_map_delete_elem,
1107 .map_gen_lookup = htab_map_gen_lookup,
1110 const struct bpf_map_ops htab_lru_map_ops = {
1111 .map_alloc = htab_map_alloc,
1112 .map_free = htab_map_free,
1113 .map_get_next_key = htab_map_get_next_key,
1114 .map_lookup_elem = htab_lru_map_lookup_elem,
1115 .map_update_elem = htab_lru_map_update_elem,
1116 .map_delete_elem = htab_lru_map_delete_elem,
1119 /* Called from eBPF program */
1120 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1122 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1125 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1130 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1132 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1135 bpf_lru_node_set_ref(&l->lru_node);
1136 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1142 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1144 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1145 struct htab_elem *l;
1146 void __percpu *pptr;
1151 /* per_cpu areas are zero-filled and bpf programs can only
1152 * access 'value_size' of them, so copying rounded areas
1153 * will not leak any kernel data
1155 size = round_up(map->value_size, 8);
1157 l = __htab_map_lookup_elem(map, key);
1160 if (htab_is_lru(htab))
1161 bpf_lru_node_set_ref(&l->lru_node);
1162 pptr = htab_elem_get_ptr(l, map->key_size);
1163 for_each_possible_cpu(cpu) {
1164 bpf_long_memcpy(value + off,
1165 per_cpu_ptr(pptr, cpu), size);
1174 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1177 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1181 if (htab_is_lru(htab))
1182 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1185 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1192 const struct bpf_map_ops htab_percpu_map_ops = {
1193 .map_alloc = htab_map_alloc,
1194 .map_free = htab_map_free,
1195 .map_get_next_key = htab_map_get_next_key,
1196 .map_lookup_elem = htab_percpu_map_lookup_elem,
1197 .map_update_elem = htab_percpu_map_update_elem,
1198 .map_delete_elem = htab_map_delete_elem,
1201 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1202 .map_alloc = htab_map_alloc,
1203 .map_free = htab_map_free,
1204 .map_get_next_key = htab_map_get_next_key,
1205 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1206 .map_update_elem = htab_lru_percpu_map_update_elem,
1207 .map_delete_elem = htab_lru_map_delete_elem,
1210 static struct bpf_map *fd_htab_map_alloc(union bpf_attr *attr)
1212 struct bpf_map *map;
1214 if (attr->value_size != sizeof(u32))
1215 return ERR_PTR(-EINVAL);
1217 /* pointer is stored internally */
1218 attr->value_size = sizeof(void *);
1219 map = htab_map_alloc(attr);
1220 attr->value_size = sizeof(u32);
1225 static void fd_htab_map_free(struct bpf_map *map)
1227 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1228 struct hlist_nulls_node *n;
1229 struct hlist_nulls_head *head;
1230 struct htab_elem *l;
1233 for (i = 0; i < htab->n_buckets; i++) {
1234 head = select_bucket(htab, i);
1236 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1237 void *ptr = fd_htab_map_get_ptr(map, l);
1239 map->ops->map_fd_put_ptr(ptr);
1246 /* only called from syscall */
1247 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1248 void *key, void *value, u64 map_flags)
1252 u32 ufd = *(u32 *)value;
1254 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1256 return PTR_ERR(ptr);
1258 ret = htab_map_update_elem(map, key, &ptr, map_flags);
1260 map->ops->map_fd_put_ptr(ptr);
1265 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1267 struct bpf_map *map, *inner_map_meta;
1269 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1270 if (IS_ERR(inner_map_meta))
1271 return inner_map_meta;
1273 map = fd_htab_map_alloc(attr);
1275 bpf_map_meta_free(inner_map_meta);
1279 map->inner_map_meta = inner_map_meta;
1284 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1286 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
1291 return READ_ONCE(*inner_map);
1294 static void htab_of_map_free(struct bpf_map *map)
1296 bpf_map_meta_free(map->inner_map_meta);
1297 fd_htab_map_free(map);
1300 const struct bpf_map_ops htab_of_maps_map_ops = {
1301 .map_alloc = htab_of_map_alloc,
1302 .map_free = htab_of_map_free,
1303 .map_get_next_key = htab_map_get_next_key,
1304 .map_lookup_elem = htab_of_map_lookup_elem,
1305 .map_delete_elem = htab_map_delete_elem,
1306 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1307 .map_fd_put_ptr = bpf_map_fd_put_ptr,