1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
7 #include <linux/jhash.h>
8 #include <linux/filter.h>
9 #include <linux/rculist_nulls.h>
10 #include <linux/random.h>
11 #include <uapi/linux/btf.h>
12 #include "percpu_freelist.h"
13 #include "bpf_lru_list.h"
14 #include "map_in_map.h"
16 #define HTAB_CREATE_FLAG_MASK \
17 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
18 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
21 struct hlist_nulls_head head;
27 struct bucket *buckets;
30 struct pcpu_freelist freelist;
33 struct htab_elem *__percpu *extra_elems;
34 atomic_t count; /* number of elements in this hashtable */
35 u32 n_buckets; /* number of hash buckets */
36 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),
116 bpf_map_area_free(htab->elems);
119 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
122 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
126 l = container_of(node, struct htab_elem, lru_node);
127 memcpy(l->key, key, htab->map.key_size);
134 static int prealloc_init(struct bpf_htab *htab)
136 u32 num_entries = htab->map.max_entries;
137 int err = -ENOMEM, i;
139 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
140 num_entries += num_possible_cpus();
142 htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
143 htab->map.numa_node);
147 if (!htab_is_percpu(htab))
148 goto skip_percpu_elems;
150 for (i = 0; i < num_entries; i++) {
151 u32 size = round_up(htab->map.value_size, 8);
154 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
157 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
163 if (htab_is_lru(htab))
164 err = bpf_lru_init(&htab->lru,
165 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
166 offsetof(struct htab_elem, hash) -
167 offsetof(struct htab_elem, lru_node),
168 htab_lru_map_delete_node,
171 err = pcpu_freelist_init(&htab->freelist);
176 if (htab_is_lru(htab))
177 bpf_lru_populate(&htab->lru, htab->elems,
178 offsetof(struct htab_elem, lru_node),
179 htab->elem_size, num_entries);
181 pcpu_freelist_populate(&htab->freelist,
182 htab->elems + offsetof(struct htab_elem, fnode),
183 htab->elem_size, num_entries);
188 htab_free_elems(htab);
192 static void prealloc_destroy(struct bpf_htab *htab)
194 htab_free_elems(htab);
196 if (htab_is_lru(htab))
197 bpf_lru_destroy(&htab->lru);
199 pcpu_freelist_destroy(&htab->freelist);
202 static int alloc_extra_elems(struct bpf_htab *htab)
204 struct htab_elem *__percpu *pptr, *l_new;
205 struct pcpu_freelist_node *l;
208 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
209 GFP_USER | __GFP_NOWARN);
213 for_each_possible_cpu(cpu) {
214 l = pcpu_freelist_pop(&htab->freelist);
215 /* pop will succeed, since prealloc_init()
216 * preallocated extra num_possible_cpus elements
218 l_new = container_of(l, struct htab_elem, fnode);
219 *per_cpu_ptr(pptr, cpu) = l_new;
221 htab->extra_elems = pptr;
225 /* Called from syscall */
226 static int htab_map_alloc_check(union bpf_attr *attr)
228 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
229 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
230 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
231 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
232 /* percpu_lru means each cpu has its own LRU list.
233 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
234 * the map's value itself is percpu. percpu_lru has
235 * nothing to do with the map's value.
237 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
238 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
239 bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
240 int numa_node = bpf_map_attr_numa_node(attr);
242 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
243 offsetof(struct htab_elem, hash_node.pprev));
244 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
245 offsetof(struct htab_elem, hash_node.pprev));
247 if (lru && !capable(CAP_SYS_ADMIN))
248 /* LRU implementation is much complicated than other
249 * maps. Hence, limit to CAP_SYS_ADMIN for now.
253 if (zero_seed && !capable(CAP_SYS_ADMIN))
254 /* Guard against local DoS, and discourage production use. */
257 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
258 !bpf_map_flags_access_ok(attr->map_flags))
261 if (!lru && percpu_lru)
264 if (lru && !prealloc)
267 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
270 /* check sanity of attributes.
271 * value_size == 0 may be allowed in the future to use map as a set
273 if (attr->max_entries == 0 || attr->key_size == 0 ||
274 attr->value_size == 0)
277 if (attr->key_size > MAX_BPF_STACK)
278 /* eBPF programs initialize keys on stack, so they cannot be
279 * larger than max stack size
283 if (attr->value_size >= KMALLOC_MAX_SIZE -
284 MAX_BPF_STACK - sizeof(struct htab_elem))
285 /* if value_size is bigger, the user space won't be able to
286 * access the elements via bpf syscall. This check also makes
287 * sure that the elem_size doesn't overflow and it's
288 * kmalloc-able later in htab_map_update_elem()
295 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
297 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
298 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
299 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
300 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
301 /* percpu_lru means each cpu has its own LRU list.
302 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
303 * the map's value itself is percpu. percpu_lru has
304 * nothing to do with the map's value.
306 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
307 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
308 struct bpf_htab *htab;
312 htab = kzalloc(sizeof(*htab), GFP_USER);
314 return ERR_PTR(-ENOMEM);
316 bpf_map_init_from_attr(&htab->map, attr);
319 /* ensure each CPU's lru list has >=1 elements.
320 * since we are at it, make each lru list has the same
321 * number of elements.
323 htab->map.max_entries = roundup(attr->max_entries,
324 num_possible_cpus());
325 if (htab->map.max_entries < attr->max_entries)
326 htab->map.max_entries = rounddown(attr->max_entries,
327 num_possible_cpus());
330 /* hash table size must be power of 2 */
331 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
333 htab->elem_size = sizeof(struct htab_elem) +
334 round_up(htab->map.key_size, 8);
336 htab->elem_size += sizeof(void *);
338 htab->elem_size += round_up(htab->map.value_size, 8);
341 /* prevent zero size kmalloc and check for u32 overflow */
342 if (htab->n_buckets == 0 ||
343 htab->n_buckets > U32_MAX / sizeof(struct bucket))
346 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
347 (u64) htab->elem_size * htab->map.max_entries;
350 cost += (u64) round_up(htab->map.value_size, 8) *
351 num_possible_cpus() * htab->map.max_entries;
353 cost += (u64) htab->elem_size * num_possible_cpus();
355 /* if map size is larger than memlock limit, reject it */
356 err = bpf_map_charge_init(&htab->map.memory, cost);
361 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
362 sizeof(struct bucket),
363 htab->map.numa_node);
367 if (htab->map.map_flags & BPF_F_ZERO_SEED)
370 htab->hashrnd = get_random_int();
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);
399 bpf_map_charge_finish(&htab->map.memory);
405 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
407 return jhash(key, key_len, hashrnd);
410 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
412 return &htab->buckets[hash & (htab->n_buckets - 1)];
415 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
417 return &__select_bucket(htab, hash)->head;
420 /* this lookup function can only be called with bucket lock taken */
421 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
422 void *key, u32 key_size)
424 struct hlist_nulls_node *n;
427 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
428 if (l->hash == hash && !memcmp(&l->key, key, key_size))
434 /* can be called without bucket lock. it will repeat the loop in
435 * the unlikely event when elements moved from one bucket into another
436 * while link list is being walked
438 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
440 u32 key_size, u32 n_buckets)
442 struct hlist_nulls_node *n;
446 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
447 if (l->hash == hash && !memcmp(&l->key, key, key_size))
450 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
456 /* Called from syscall or from eBPF program directly, so
457 * arguments have to match bpf_map_lookup_elem() exactly.
458 * The return value is adjusted by BPF instructions
459 * in htab_map_gen_lookup().
461 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
463 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
464 struct hlist_nulls_head *head;
468 /* Must be called with rcu_read_lock. */
469 WARN_ON_ONCE(!rcu_read_lock_held());
471 key_size = map->key_size;
473 hash = htab_map_hash(key, key_size, htab->hashrnd);
475 head = select_bucket(htab, hash);
477 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
482 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
484 struct htab_elem *l = __htab_map_lookup_elem(map, key);
487 return l->key + round_up(map->key_size, 8);
492 /* inline bpf_map_lookup_elem() call.
495 * bpf_map_lookup_elem
496 * map->ops->map_lookup_elem
497 * htab_map_lookup_elem
498 * __htab_map_lookup_elem
501 * __htab_map_lookup_elem
503 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
505 struct bpf_insn *insn = insn_buf;
506 const int ret = BPF_REG_0;
508 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
509 (void *(*)(struct bpf_map *map, void *key))NULL));
510 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
511 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
512 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
513 offsetof(struct htab_elem, key) +
514 round_up(map->key_size, 8));
515 return insn - insn_buf;
518 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
519 void *key, const bool mark)
521 struct htab_elem *l = __htab_map_lookup_elem(map, key);
525 bpf_lru_node_set_ref(&l->lru_node);
526 return l->key + round_up(map->key_size, 8);
532 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
534 return __htab_lru_map_lookup_elem(map, key, true);
537 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
539 return __htab_lru_map_lookup_elem(map, key, false);
542 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
543 struct bpf_insn *insn_buf)
545 struct bpf_insn *insn = insn_buf;
546 const int ret = BPF_REG_0;
547 const int ref_reg = BPF_REG_1;
549 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
550 (void *(*)(struct bpf_map *map, void *key))NULL));
551 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
552 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
553 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
554 offsetof(struct htab_elem, lru_node) +
555 offsetof(struct bpf_lru_node, ref));
556 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
557 *insn++ = BPF_ST_MEM(BPF_B, ret,
558 offsetof(struct htab_elem, lru_node) +
559 offsetof(struct bpf_lru_node, ref),
561 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
562 offsetof(struct htab_elem, key) +
563 round_up(map->key_size, 8));
564 return insn - insn_buf;
567 /* It is called from the bpf_lru_list when the LRU needs to delete
568 * older elements from the htab.
570 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
572 struct bpf_htab *htab = (struct bpf_htab *)arg;
573 struct htab_elem *l = NULL, *tgt_l;
574 struct hlist_nulls_head *head;
575 struct hlist_nulls_node *n;
579 tgt_l = container_of(node, struct htab_elem, lru_node);
580 b = __select_bucket(htab, tgt_l->hash);
583 raw_spin_lock_irqsave(&b->lock, flags);
585 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
587 hlist_nulls_del_rcu(&l->hash_node);
591 raw_spin_unlock_irqrestore(&b->lock, flags);
596 /* Called from syscall */
597 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
599 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
600 struct hlist_nulls_head *head;
601 struct htab_elem *l, *next_l;
605 WARN_ON_ONCE(!rcu_read_lock_held());
607 key_size = map->key_size;
610 goto find_first_elem;
612 hash = htab_map_hash(key, key_size, htab->hashrnd);
614 head = select_bucket(htab, hash);
617 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
620 goto find_first_elem;
622 /* key was found, get next key in the same bucket */
623 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
624 struct htab_elem, hash_node);
627 /* if next elem in this hash list is non-zero, just return it */
628 memcpy(next_key, next_l->key, key_size);
632 /* no more elements in this hash list, go to the next bucket */
633 i = hash & (htab->n_buckets - 1);
637 /* iterate over buckets */
638 for (; i < htab->n_buckets; i++) {
639 head = select_bucket(htab, i);
641 /* pick first element in the bucket */
642 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
643 struct htab_elem, hash_node);
645 /* if it's not empty, just return it */
646 memcpy(next_key, next_l->key, key_size);
651 /* iterated over all buckets and all elements */
655 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
657 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
658 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
662 static void htab_elem_free_rcu(struct rcu_head *head)
664 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
665 struct bpf_htab *htab = l->htab;
667 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
668 * we're calling kfree, otherwise deadlock is possible if kprobes
669 * are placed somewhere inside of slub
672 __this_cpu_inc(bpf_prog_active);
673 htab_elem_free(htab, l);
674 __this_cpu_dec(bpf_prog_active);
678 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
680 struct bpf_map *map = &htab->map;
682 if (map->ops->map_fd_put_ptr) {
683 void *ptr = fd_htab_map_get_ptr(map, l);
685 map->ops->map_fd_put_ptr(ptr);
688 if (htab_is_prealloc(htab)) {
689 __pcpu_freelist_push(&htab->freelist, &l->fnode);
691 atomic_dec(&htab->count);
693 call_rcu(&l->rcu, htab_elem_free_rcu);
697 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
698 void *value, bool onallcpus)
701 /* copy true value_size bytes */
702 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
704 u32 size = round_up(htab->map.value_size, 8);
707 for_each_possible_cpu(cpu) {
708 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
715 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
717 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
721 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
722 void *value, u32 key_size, u32 hash,
723 bool percpu, bool onallcpus,
724 struct htab_elem *old_elem)
726 u32 size = htab->map.value_size;
727 bool prealloc = htab_is_prealloc(htab);
728 struct htab_elem *l_new, **pl_new;
733 /* if we're updating the existing element,
734 * use per-cpu extra elems to avoid freelist_pop/push
736 pl_new = this_cpu_ptr(htab->extra_elems);
740 struct pcpu_freelist_node *l;
742 l = __pcpu_freelist_pop(&htab->freelist);
744 return ERR_PTR(-E2BIG);
745 l_new = container_of(l, struct htab_elem, fnode);
748 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
750 /* when map is full and update() is replacing
751 * old element, it's ok to allocate, since
752 * old element will be freed immediately.
753 * Otherwise return an error
755 l_new = ERR_PTR(-E2BIG);
758 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
759 htab->map.numa_node);
761 l_new = ERR_PTR(-ENOMEM);
764 check_and_init_map_lock(&htab->map,
765 l_new->key + round_up(key_size, 8));
768 memcpy(l_new->key, key, key_size);
770 size = round_up(size, 8);
772 pptr = htab_elem_get_ptr(l_new, key_size);
774 /* alloc_percpu zero-fills */
775 pptr = __alloc_percpu_gfp(size, 8,
776 GFP_ATOMIC | __GFP_NOWARN);
779 l_new = ERR_PTR(-ENOMEM);
784 pcpu_copy_value(htab, pptr, value, onallcpus);
787 htab_elem_set_ptr(l_new, key_size, pptr);
788 } else if (fd_htab_map_needs_adjust(htab)) {
789 size = round_up(size, 8);
790 memcpy(l_new->key + round_up(key_size, 8), value, size);
792 copy_map_value(&htab->map,
793 l_new->key + round_up(key_size, 8),
800 atomic_dec(&htab->count);
804 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
807 if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
808 /* elem already exists */
811 if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
812 /* elem doesn't exist, cannot update it */
818 /* Called from syscall or from eBPF program */
819 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
822 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
823 struct htab_elem *l_new = NULL, *l_old;
824 struct hlist_nulls_head *head;
830 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
834 WARN_ON_ONCE(!rcu_read_lock_held());
836 key_size = map->key_size;
838 hash = htab_map_hash(key, key_size, htab->hashrnd);
840 b = __select_bucket(htab, hash);
843 if (unlikely(map_flags & BPF_F_LOCK)) {
844 if (unlikely(!map_value_has_spin_lock(map)))
846 /* find an element without taking the bucket lock */
847 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
849 ret = check_flags(htab, l_old, map_flags);
853 /* grab the element lock and update value in place */
854 copy_map_value_locked(map,
855 l_old->key + round_up(key_size, 8),
859 /* fall through, grab the bucket lock and lookup again.
860 * 99.9% chance that the element won't be found,
861 * but second lookup under lock has to be done.
865 /* bpf_map_update_elem() can be called in_irq() */
866 raw_spin_lock_irqsave(&b->lock, flags);
868 l_old = lookup_elem_raw(head, hash, key, key_size);
870 ret = check_flags(htab, l_old, map_flags);
874 if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
875 /* first lookup without the bucket lock didn't find the element,
876 * but second lookup with the bucket lock found it.
877 * This case is highly unlikely, but has to be dealt with:
878 * grab the element lock in addition to the bucket lock
879 * and update element in place
881 copy_map_value_locked(map,
882 l_old->key + round_up(key_size, 8),
888 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
891 /* all pre-allocated elements are in use or memory exhausted */
892 ret = PTR_ERR(l_new);
896 /* add new element to the head of the list, so that
897 * concurrent search will find it before old elem
899 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
901 hlist_nulls_del_rcu(&l_old->hash_node);
902 if (!htab_is_prealloc(htab))
903 free_htab_elem(htab, l_old);
907 raw_spin_unlock_irqrestore(&b->lock, flags);
911 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
914 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
915 struct htab_elem *l_new, *l_old = NULL;
916 struct hlist_nulls_head *head;
922 if (unlikely(map_flags > BPF_EXIST))
926 WARN_ON_ONCE(!rcu_read_lock_held());
928 key_size = map->key_size;
930 hash = htab_map_hash(key, key_size, htab->hashrnd);
932 b = __select_bucket(htab, hash);
935 /* For LRU, we need to alloc before taking bucket's
936 * spinlock because getting free nodes from LRU may need
937 * to remove older elements from htab and this removal
938 * operation will need a bucket lock.
940 l_new = prealloc_lru_pop(htab, key, hash);
943 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
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);
954 /* add new element to the head of the list, so that
955 * concurrent search will find it before old elem
957 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
959 bpf_lru_node_set_ref(&l_new->lru_node);
960 hlist_nulls_del_rcu(&l_old->hash_node);
965 raw_spin_unlock_irqrestore(&b->lock, flags);
968 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
970 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
975 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
976 void *value, u64 map_flags,
979 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
980 struct htab_elem *l_new = NULL, *l_old;
981 struct hlist_nulls_head *head;
987 if (unlikely(map_flags > BPF_EXIST))
991 WARN_ON_ONCE(!rcu_read_lock_held());
993 key_size = map->key_size;
995 hash = htab_map_hash(key, key_size, htab->hashrnd);
997 b = __select_bucket(htab, hash);
1000 /* bpf_map_update_elem() can be called in_irq() */
1001 raw_spin_lock_irqsave(&b->lock, flags);
1003 l_old = lookup_elem_raw(head, hash, key, key_size);
1005 ret = check_flags(htab, l_old, map_flags);
1010 /* per-cpu hash map can update value in-place */
1011 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1014 l_new = alloc_htab_elem(htab, key, value, key_size,
1015 hash, true, onallcpus, NULL);
1016 if (IS_ERR(l_new)) {
1017 ret = PTR_ERR(l_new);
1020 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1024 raw_spin_unlock_irqrestore(&b->lock, flags);
1028 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1029 void *value, u64 map_flags,
1032 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1033 struct htab_elem *l_new = NULL, *l_old;
1034 struct hlist_nulls_head *head;
1035 unsigned long flags;
1040 if (unlikely(map_flags > BPF_EXIST))
1044 WARN_ON_ONCE(!rcu_read_lock_held());
1046 key_size = map->key_size;
1048 hash = htab_map_hash(key, key_size, htab->hashrnd);
1050 b = __select_bucket(htab, hash);
1053 /* For LRU, we need to alloc before taking bucket's
1054 * spinlock because LRU's elem alloc may need
1055 * to remove older elem from htab and this removal
1056 * operation will need a bucket lock.
1058 if (map_flags != BPF_EXIST) {
1059 l_new = prealloc_lru_pop(htab, key, hash);
1064 /* bpf_map_update_elem() can be called in_irq() */
1065 raw_spin_lock_irqsave(&b->lock, flags);
1067 l_old = lookup_elem_raw(head, hash, key, key_size);
1069 ret = check_flags(htab, l_old, map_flags);
1074 bpf_lru_node_set_ref(&l_old->lru_node);
1076 /* per-cpu hash map can update value in-place */
1077 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1080 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1082 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1087 raw_spin_unlock_irqrestore(&b->lock, flags);
1089 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1093 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1094 void *value, u64 map_flags)
1096 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1099 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1100 void *value, u64 map_flags)
1102 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1106 /* Called from syscall or from eBPF program */
1107 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1109 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1110 struct hlist_nulls_head *head;
1112 struct htab_elem *l;
1113 unsigned long flags;
1117 WARN_ON_ONCE(!rcu_read_lock_held());
1119 key_size = map->key_size;
1121 hash = htab_map_hash(key, key_size, htab->hashrnd);
1122 b = __select_bucket(htab, hash);
1125 raw_spin_lock_irqsave(&b->lock, flags);
1127 l = lookup_elem_raw(head, hash, key, key_size);
1130 hlist_nulls_del_rcu(&l->hash_node);
1131 free_htab_elem(htab, l);
1135 raw_spin_unlock_irqrestore(&b->lock, flags);
1139 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1141 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1142 struct hlist_nulls_head *head;
1144 struct htab_elem *l;
1145 unsigned long flags;
1149 WARN_ON_ONCE(!rcu_read_lock_held());
1151 key_size = map->key_size;
1153 hash = htab_map_hash(key, key_size, htab->hashrnd);
1154 b = __select_bucket(htab, hash);
1157 raw_spin_lock_irqsave(&b->lock, flags);
1159 l = lookup_elem_raw(head, hash, key, key_size);
1162 hlist_nulls_del_rcu(&l->hash_node);
1166 raw_spin_unlock_irqrestore(&b->lock, flags);
1168 bpf_lru_push_free(&htab->lru, &l->lru_node);
1172 static void delete_all_elements(struct bpf_htab *htab)
1176 for (i = 0; i < htab->n_buckets; i++) {
1177 struct hlist_nulls_head *head = select_bucket(htab, i);
1178 struct hlist_nulls_node *n;
1179 struct htab_elem *l;
1181 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1182 hlist_nulls_del_rcu(&l->hash_node);
1183 htab_elem_free(htab, l);
1188 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1189 static void htab_map_free(struct bpf_map *map)
1191 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1193 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1194 * so the programs (can be more than one that used this map) were
1195 * disconnected from events. Wait for outstanding critical sections in
1196 * these programs to complete
1200 /* some of free_htab_elem() callbacks for elements of this map may
1201 * not have executed. Wait for them.
1204 if (!htab_is_prealloc(htab))
1205 delete_all_elements(htab);
1207 prealloc_destroy(htab);
1209 free_percpu(htab->extra_elems);
1210 bpf_map_area_free(htab->buckets);
1214 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1221 value = htab_map_lookup_elem(map, key);
1227 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1229 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1235 const struct bpf_map_ops htab_map_ops = {
1236 .map_alloc_check = htab_map_alloc_check,
1237 .map_alloc = htab_map_alloc,
1238 .map_free = htab_map_free,
1239 .map_get_next_key = htab_map_get_next_key,
1240 .map_lookup_elem = htab_map_lookup_elem,
1241 .map_update_elem = htab_map_update_elem,
1242 .map_delete_elem = htab_map_delete_elem,
1243 .map_gen_lookup = htab_map_gen_lookup,
1244 .map_seq_show_elem = htab_map_seq_show_elem,
1247 const struct bpf_map_ops htab_lru_map_ops = {
1248 .map_alloc_check = htab_map_alloc_check,
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_map_lookup_elem,
1253 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1254 .map_update_elem = htab_lru_map_update_elem,
1255 .map_delete_elem = htab_lru_map_delete_elem,
1256 .map_gen_lookup = htab_lru_map_gen_lookup,
1257 .map_seq_show_elem = htab_map_seq_show_elem,
1260 /* Called from eBPF program */
1261 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1263 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1266 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1271 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1273 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1276 bpf_lru_node_set_ref(&l->lru_node);
1277 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1283 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1285 struct htab_elem *l;
1286 void __percpu *pptr;
1291 /* per_cpu areas are zero-filled and bpf programs can only
1292 * access 'value_size' of them, so copying rounded areas
1293 * will not leak any kernel data
1295 size = round_up(map->value_size, 8);
1297 l = __htab_map_lookup_elem(map, key);
1300 /* We do not mark LRU map element here in order to not mess up
1301 * eviction heuristics when user space does a map walk.
1303 pptr = htab_elem_get_ptr(l, map->key_size);
1304 for_each_possible_cpu(cpu) {
1305 bpf_long_memcpy(value + off,
1306 per_cpu_ptr(pptr, cpu), size);
1315 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1318 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1322 if (htab_is_lru(htab))
1323 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1326 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1333 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1336 struct htab_elem *l;
1337 void __percpu *pptr;
1342 l = __htab_map_lookup_elem(map, key);
1348 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1349 seq_puts(m, ": {\n");
1350 pptr = htab_elem_get_ptr(l, map->key_size);
1351 for_each_possible_cpu(cpu) {
1352 seq_printf(m, "\tcpu%d: ", cpu);
1353 btf_type_seq_show(map->btf, map->btf_value_type_id,
1354 per_cpu_ptr(pptr, cpu), m);
1362 const struct bpf_map_ops htab_percpu_map_ops = {
1363 .map_alloc_check = htab_map_alloc_check,
1364 .map_alloc = htab_map_alloc,
1365 .map_free = htab_map_free,
1366 .map_get_next_key = htab_map_get_next_key,
1367 .map_lookup_elem = htab_percpu_map_lookup_elem,
1368 .map_update_elem = htab_percpu_map_update_elem,
1369 .map_delete_elem = htab_map_delete_elem,
1370 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1373 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1374 .map_alloc_check = htab_map_alloc_check,
1375 .map_alloc = htab_map_alloc,
1376 .map_free = htab_map_free,
1377 .map_get_next_key = htab_map_get_next_key,
1378 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1379 .map_update_elem = htab_lru_percpu_map_update_elem,
1380 .map_delete_elem = htab_lru_map_delete_elem,
1381 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1384 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1386 if (attr->value_size != sizeof(u32))
1388 return htab_map_alloc_check(attr);
1391 static void fd_htab_map_free(struct bpf_map *map)
1393 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1394 struct hlist_nulls_node *n;
1395 struct hlist_nulls_head *head;
1396 struct htab_elem *l;
1399 for (i = 0; i < htab->n_buckets; i++) {
1400 head = select_bucket(htab, i);
1402 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1403 void *ptr = fd_htab_map_get_ptr(map, l);
1405 map->ops->map_fd_put_ptr(ptr);
1412 /* only called from syscall */
1413 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1418 if (!map->ops->map_fd_sys_lookup_elem)
1422 ptr = htab_map_lookup_elem(map, key);
1424 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1432 /* only called from syscall */
1433 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1434 void *key, void *value, u64 map_flags)
1438 u32 ufd = *(u32 *)value;
1440 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1442 return PTR_ERR(ptr);
1444 ret = htab_map_update_elem(map, key, &ptr, map_flags);
1446 map->ops->map_fd_put_ptr(ptr);
1451 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1453 struct bpf_map *map, *inner_map_meta;
1455 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1456 if (IS_ERR(inner_map_meta))
1457 return inner_map_meta;
1459 map = htab_map_alloc(attr);
1461 bpf_map_meta_free(inner_map_meta);
1465 map->inner_map_meta = inner_map_meta;
1470 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1472 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
1477 return READ_ONCE(*inner_map);
1480 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1481 struct bpf_insn *insn_buf)
1483 struct bpf_insn *insn = insn_buf;
1484 const int ret = BPF_REG_0;
1486 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1487 (void *(*)(struct bpf_map *map, void *key))NULL));
1488 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1489 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1490 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1491 offsetof(struct htab_elem, key) +
1492 round_up(map->key_size, 8));
1493 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1495 return insn - insn_buf;
1498 static void htab_of_map_free(struct bpf_map *map)
1500 bpf_map_meta_free(map->inner_map_meta);
1501 fd_htab_map_free(map);
1504 const struct bpf_map_ops htab_of_maps_map_ops = {
1505 .map_alloc_check = fd_htab_map_alloc_check,
1506 .map_alloc = htab_of_map_alloc,
1507 .map_free = htab_of_map_free,
1508 .map_get_next_key = htab_map_get_next_key,
1509 .map_lookup_elem = htab_of_map_lookup_elem,
1510 .map_delete_elem = htab_map_delete_elem,
1511 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1512 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1513 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1514 .map_gen_lookup = htab_of_map_gen_lookup,
1515 .map_check_btf = map_check_no_btf,