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/btf.h>
15 #include <linux/jhash.h>
16 #include <linux/filter.h>
17 #include <linux/rculist_nulls.h>
18 #include <linux/random.h>
19 #include <uapi/linux/btf.h>
20 #include "percpu_freelist.h"
21 #include "bpf_lru_list.h"
22 #include "map_in_map.h"
24 #define HTAB_CREATE_FLAG_MASK \
25 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
26 BPF_F_RDONLY | BPF_F_WRONLY)
29 struct hlist_nulls_head head;
35 struct bucket *buckets;
38 struct pcpu_freelist freelist;
41 struct htab_elem *__percpu *extra_elems;
42 atomic_t count; /* number of elements in this hashtable */
43 u32 n_buckets; /* number of hash buckets */
44 u32 elem_size; /* size of each element in bytes */
48 /* each htab element is struct htab_elem + key + value */
51 struct hlist_nulls_node hash_node;
55 struct bpf_htab *htab;
56 struct pcpu_freelist_node fnode;
62 struct bpf_lru_node lru_node;
65 char key[0] __aligned(8);
68 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
70 static bool htab_is_lru(const struct bpf_htab *htab)
72 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
73 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
76 static bool htab_is_percpu(const struct bpf_htab *htab)
78 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
79 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
82 static bool htab_is_prealloc(const struct bpf_htab *htab)
84 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
87 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
90 *(void __percpu **)(l->key + key_size) = pptr;
93 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
95 return *(void __percpu **)(l->key + key_size);
98 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
100 return *(void **)(l->key + roundup(map->key_size, 8));
103 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
105 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
108 static void htab_free_elems(struct bpf_htab *htab)
112 if (!htab_is_percpu(htab))
115 for (i = 0; i < htab->map.max_entries; i++) {
118 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
124 bpf_map_area_free(htab->elems);
127 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
130 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
134 l = container_of(node, struct htab_elem, lru_node);
135 memcpy(l->key, key, htab->map.key_size);
142 static int prealloc_init(struct bpf_htab *htab)
144 u32 num_entries = htab->map.max_entries;
145 int err = -ENOMEM, i;
147 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
148 num_entries += num_possible_cpus();
150 htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
151 htab->map.numa_node);
155 if (!htab_is_percpu(htab))
156 goto skip_percpu_elems;
158 for (i = 0; i < num_entries; i++) {
159 u32 size = round_up(htab->map.value_size, 8);
162 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
165 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
171 if (htab_is_lru(htab))
172 err = bpf_lru_init(&htab->lru,
173 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
174 offsetof(struct htab_elem, hash) -
175 offsetof(struct htab_elem, lru_node),
176 htab_lru_map_delete_node,
179 err = pcpu_freelist_init(&htab->freelist);
184 if (htab_is_lru(htab))
185 bpf_lru_populate(&htab->lru, htab->elems,
186 offsetof(struct htab_elem, lru_node),
187 htab->elem_size, num_entries);
189 pcpu_freelist_populate(&htab->freelist,
190 htab->elems + offsetof(struct htab_elem, fnode),
191 htab->elem_size, num_entries);
196 htab_free_elems(htab);
200 static void prealloc_destroy(struct bpf_htab *htab)
202 htab_free_elems(htab);
204 if (htab_is_lru(htab))
205 bpf_lru_destroy(&htab->lru);
207 pcpu_freelist_destroy(&htab->freelist);
210 static int alloc_extra_elems(struct bpf_htab *htab)
212 struct htab_elem *__percpu *pptr, *l_new;
213 struct pcpu_freelist_node *l;
216 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
217 GFP_USER | __GFP_NOWARN);
221 for_each_possible_cpu(cpu) {
222 l = pcpu_freelist_pop(&htab->freelist);
223 /* pop will succeed, since prealloc_init()
224 * preallocated extra num_possible_cpus elements
226 l_new = container_of(l, struct htab_elem, fnode);
227 *per_cpu_ptr(pptr, cpu) = l_new;
229 htab->extra_elems = pptr;
233 /* Called from syscall */
234 static int htab_map_alloc_check(union bpf_attr *attr)
236 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
237 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
238 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
239 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
240 /* percpu_lru means each cpu has its own LRU list.
241 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
242 * the map's value itself is percpu. percpu_lru has
243 * nothing to do with the map's value.
245 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
246 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
247 int numa_node = bpf_map_attr_numa_node(attr);
249 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
250 offsetof(struct htab_elem, hash_node.pprev));
251 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
252 offsetof(struct htab_elem, hash_node.pprev));
254 if (lru && !capable(CAP_SYS_ADMIN))
255 /* LRU implementation is much complicated than other
256 * maps. Hence, limit to CAP_SYS_ADMIN for now.
260 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
261 /* reserved bits should not be used */
264 if (!lru && percpu_lru)
267 if (lru && !prealloc)
270 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
273 /* check sanity of attributes.
274 * value_size == 0 may be allowed in the future to use map as a set
276 if (attr->max_entries == 0 || attr->key_size == 0 ||
277 attr->value_size == 0)
280 if (attr->key_size > MAX_BPF_STACK)
281 /* eBPF programs initialize keys on stack, so they cannot be
282 * larger than max stack size
286 if (attr->value_size >= KMALLOC_MAX_SIZE -
287 MAX_BPF_STACK - sizeof(struct htab_elem))
288 /* if value_size is bigger, the user space won't be able to
289 * access the elements via bpf syscall. This check also makes
290 * sure that the elem_size doesn't overflow and it's
291 * kmalloc-able later in htab_map_update_elem()
298 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
300 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
301 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
302 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
303 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
304 /* percpu_lru means each cpu has its own LRU list.
305 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
306 * the map's value itself is percpu. percpu_lru has
307 * nothing to do with the map's value.
309 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
310 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
311 struct bpf_htab *htab;
315 htab = kzalloc(sizeof(*htab), GFP_USER);
317 return ERR_PTR(-ENOMEM);
319 bpf_map_init_from_attr(&htab->map, attr);
322 /* ensure each CPU's lru list has >=1 elements.
323 * since we are at it, make each lru list has the same
324 * number of elements.
326 htab->map.max_entries = roundup(attr->max_entries,
327 num_possible_cpus());
328 if (htab->map.max_entries < attr->max_entries)
329 htab->map.max_entries = rounddown(attr->max_entries,
330 num_possible_cpus());
333 /* hash table size must be power of 2 */
334 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
336 htab->elem_size = sizeof(struct htab_elem) +
337 round_up(htab->map.key_size, 8);
339 htab->elem_size += sizeof(void *);
341 htab->elem_size += round_up(htab->map.value_size, 8);
344 /* prevent zero size kmalloc and check for u32 overflow */
345 if (htab->n_buckets == 0 ||
346 htab->n_buckets > U32_MAX / sizeof(struct bucket))
349 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
350 (u64) htab->elem_size * htab->map.max_entries;
353 cost += (u64) round_up(htab->map.value_size, 8) *
354 num_possible_cpus() * htab->map.max_entries;
356 cost += (u64) htab->elem_size * num_possible_cpus();
358 if (cost >= U32_MAX - PAGE_SIZE)
359 /* make sure page count doesn't overflow */
362 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
364 /* if map size is larger than memlock limit, reject it early */
365 err = bpf_map_precharge_memlock(htab->map.pages);
370 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
371 sizeof(struct bucket),
372 htab->map.numa_node);
376 htab->hashrnd = get_random_int();
377 for (i = 0; i < htab->n_buckets; i++) {
378 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
379 raw_spin_lock_init(&htab->buckets[i].lock);
383 err = prealloc_init(htab);
387 if (!percpu && !lru) {
388 /* lru itself can remove the least used element, so
389 * there is no need for an extra elem during map_update.
391 err = alloc_extra_elems(htab);
400 prealloc_destroy(htab);
402 bpf_map_area_free(htab->buckets);
408 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
410 return jhash(key, key_len, hashrnd);
413 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
415 return &htab->buckets[hash & (htab->n_buckets - 1)];
418 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
420 return &__select_bucket(htab, hash)->head;
423 /* this lookup function can only be called with bucket lock taken */
424 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
425 void *key, u32 key_size)
427 struct hlist_nulls_node *n;
430 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
431 if (l->hash == hash && !memcmp(&l->key, key, key_size))
437 /* can be called without bucket lock. it will repeat the loop in
438 * the unlikely event when elements moved from one bucket into another
439 * while link list is being walked
441 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
443 u32 key_size, u32 n_buckets)
445 struct hlist_nulls_node *n;
449 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
450 if (l->hash == hash && !memcmp(&l->key, key, key_size))
453 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
459 /* Called from syscall or from eBPF program directly, so
460 * arguments have to match bpf_map_lookup_elem() exactly.
461 * The return value is adjusted by BPF instructions
462 * in htab_map_gen_lookup().
464 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
466 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
467 struct hlist_nulls_head *head;
471 /* Must be called with rcu_read_lock. */
472 WARN_ON_ONCE(!rcu_read_lock_held());
474 key_size = map->key_size;
476 hash = htab_map_hash(key, key_size, htab->hashrnd);
478 head = select_bucket(htab, hash);
480 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
485 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
487 struct htab_elem *l = __htab_map_lookup_elem(map, key);
490 return l->key + round_up(map->key_size, 8);
495 /* inline bpf_map_lookup_elem() call.
498 * bpf_map_lookup_elem
499 * map->ops->map_lookup_elem
500 * htab_map_lookup_elem
501 * __htab_map_lookup_elem
504 * __htab_map_lookup_elem
506 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
508 struct bpf_insn *insn = insn_buf;
509 const int ret = BPF_REG_0;
511 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
512 (void *(*)(struct bpf_map *map, void *key))NULL));
513 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
514 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
515 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
516 offsetof(struct htab_elem, key) +
517 round_up(map->key_size, 8));
518 return insn - insn_buf;
521 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
522 void *key, const bool mark)
524 struct htab_elem *l = __htab_map_lookup_elem(map, key);
528 bpf_lru_node_set_ref(&l->lru_node);
529 return l->key + round_up(map->key_size, 8);
535 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
537 return __htab_lru_map_lookup_elem(map, key, true);
540 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
542 return __htab_lru_map_lookup_elem(map, key, false);
545 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
546 struct bpf_insn *insn_buf)
548 struct bpf_insn *insn = insn_buf;
549 const int ret = BPF_REG_0;
550 const int ref_reg = BPF_REG_1;
552 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
553 (void *(*)(struct bpf_map *map, void *key))NULL));
554 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
555 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
556 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
557 offsetof(struct htab_elem, lru_node) +
558 offsetof(struct bpf_lru_node, ref));
559 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
560 *insn++ = BPF_ST_MEM(BPF_B, ret,
561 offsetof(struct htab_elem, lru_node) +
562 offsetof(struct bpf_lru_node, ref),
564 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
565 offsetof(struct htab_elem, key) +
566 round_up(map->key_size, 8));
567 return insn - insn_buf;
570 /* It is called from the bpf_lru_list when the LRU needs to delete
571 * older elements from the htab.
573 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
575 struct bpf_htab *htab = (struct bpf_htab *)arg;
576 struct htab_elem *l = NULL, *tgt_l;
577 struct hlist_nulls_head *head;
578 struct hlist_nulls_node *n;
582 tgt_l = container_of(node, struct htab_elem, lru_node);
583 b = __select_bucket(htab, tgt_l->hash);
586 raw_spin_lock_irqsave(&b->lock, flags);
588 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
590 hlist_nulls_del_rcu(&l->hash_node);
594 raw_spin_unlock_irqrestore(&b->lock, flags);
599 /* Called from syscall */
600 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
602 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
603 struct hlist_nulls_head *head;
604 struct htab_elem *l, *next_l;
608 WARN_ON_ONCE(!rcu_read_lock_held());
610 key_size = map->key_size;
613 goto find_first_elem;
615 hash = htab_map_hash(key, key_size, htab->hashrnd);
617 head = select_bucket(htab, hash);
620 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
623 goto find_first_elem;
625 /* key was found, get next key in the same bucket */
626 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
627 struct htab_elem, hash_node);
630 /* if next elem in this hash list is non-zero, just return it */
631 memcpy(next_key, next_l->key, key_size);
635 /* no more elements in this hash list, go to the next bucket */
636 i = hash & (htab->n_buckets - 1);
640 /* iterate over buckets */
641 for (; i < htab->n_buckets; i++) {
642 head = select_bucket(htab, i);
644 /* pick first element in the bucket */
645 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
646 struct htab_elem, hash_node);
648 /* if it's not empty, just return it */
649 memcpy(next_key, next_l->key, key_size);
654 /* iterated over all buckets and all elements */
658 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
660 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
661 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
665 static void htab_elem_free_rcu(struct rcu_head *head)
667 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
668 struct bpf_htab *htab = l->htab;
670 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
671 * we're calling kfree, otherwise deadlock is possible if kprobes
672 * are placed somewhere inside of slub
675 __this_cpu_inc(bpf_prog_active);
676 htab_elem_free(htab, l);
677 __this_cpu_dec(bpf_prog_active);
681 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
683 struct bpf_map *map = &htab->map;
685 if (map->ops->map_fd_put_ptr) {
686 void *ptr = fd_htab_map_get_ptr(map, l);
688 map->ops->map_fd_put_ptr(ptr);
691 if (htab_is_prealloc(htab)) {
692 __pcpu_freelist_push(&htab->freelist, &l->fnode);
694 atomic_dec(&htab->count);
696 call_rcu(&l->rcu, htab_elem_free_rcu);
700 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
701 void *value, bool onallcpus)
704 /* copy true value_size bytes */
705 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
707 u32 size = round_up(htab->map.value_size, 8);
710 for_each_possible_cpu(cpu) {
711 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
718 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
720 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
724 static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
726 u32 size = htab->map.value_size;
728 if (percpu || fd_htab_map_needs_adjust(htab))
729 size = round_up(size, 8);
733 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
734 void *value, u32 key_size, u32 hash,
735 bool percpu, bool onallcpus,
736 struct htab_elem *old_elem)
738 u32 size = htab_size_value(htab, percpu);
739 bool prealloc = htab_is_prealloc(htab);
740 struct htab_elem *l_new, **pl_new;
745 /* if we're updating the existing element,
746 * use per-cpu extra elems to avoid freelist_pop/push
748 pl_new = this_cpu_ptr(htab->extra_elems);
752 struct pcpu_freelist_node *l;
754 l = __pcpu_freelist_pop(&htab->freelist);
756 return ERR_PTR(-E2BIG);
757 l_new = container_of(l, struct htab_elem, fnode);
760 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
762 /* when map is full and update() is replacing
763 * old element, it's ok to allocate, since
764 * old element will be freed immediately.
765 * Otherwise return an error
767 l_new = ERR_PTR(-E2BIG);
770 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
771 htab->map.numa_node);
773 l_new = ERR_PTR(-ENOMEM);
778 memcpy(l_new->key, key, key_size);
781 pptr = htab_elem_get_ptr(l_new, key_size);
783 /* alloc_percpu zero-fills */
784 pptr = __alloc_percpu_gfp(size, 8,
785 GFP_ATOMIC | __GFP_NOWARN);
788 l_new = ERR_PTR(-ENOMEM);
793 pcpu_copy_value(htab, pptr, value, onallcpus);
796 htab_elem_set_ptr(l_new, key_size, pptr);
798 memcpy(l_new->key + round_up(key_size, 8), value, size);
804 atomic_dec(&htab->count);
808 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
811 if (l_old && map_flags == BPF_NOEXIST)
812 /* elem already exists */
815 if (!l_old && map_flags == BPF_EXIST)
816 /* elem doesn't exist, cannot update it */
822 /* Called from syscall or from eBPF program */
823 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
826 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
827 struct htab_elem *l_new = NULL, *l_old;
828 struct hlist_nulls_head *head;
834 if (unlikely(map_flags > BPF_EXIST))
838 WARN_ON_ONCE(!rcu_read_lock_held());
840 key_size = map->key_size;
842 hash = htab_map_hash(key, key_size, htab->hashrnd);
844 b = __select_bucket(htab, hash);
847 /* bpf_map_update_elem() can be called in_irq() */
848 raw_spin_lock_irqsave(&b->lock, flags);
850 l_old = lookup_elem_raw(head, hash, key, key_size);
852 ret = check_flags(htab, l_old, map_flags);
856 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
859 /* all pre-allocated elements are in use or memory exhausted */
860 ret = PTR_ERR(l_new);
864 /* add new element to the head of the list, so that
865 * concurrent search will find it before old elem
867 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
869 hlist_nulls_del_rcu(&l_old->hash_node);
870 if (!htab_is_prealloc(htab))
871 free_htab_elem(htab, l_old);
875 raw_spin_unlock_irqrestore(&b->lock, flags);
879 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
882 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
883 struct htab_elem *l_new, *l_old = NULL;
884 struct hlist_nulls_head *head;
890 if (unlikely(map_flags > BPF_EXIST))
894 WARN_ON_ONCE(!rcu_read_lock_held());
896 key_size = map->key_size;
898 hash = htab_map_hash(key, key_size, htab->hashrnd);
900 b = __select_bucket(htab, hash);
903 /* For LRU, we need to alloc before taking bucket's
904 * spinlock because getting free nodes from LRU may need
905 * to remove older elements from htab and this removal
906 * operation will need a bucket lock.
908 l_new = prealloc_lru_pop(htab, key, hash);
911 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
913 /* bpf_map_update_elem() can be called in_irq() */
914 raw_spin_lock_irqsave(&b->lock, flags);
916 l_old = lookup_elem_raw(head, hash, key, key_size);
918 ret = check_flags(htab, l_old, map_flags);
922 /* add new element to the head of the list, so that
923 * concurrent search will find it before old elem
925 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
927 bpf_lru_node_set_ref(&l_new->lru_node);
928 hlist_nulls_del_rcu(&l_old->hash_node);
933 raw_spin_unlock_irqrestore(&b->lock, flags);
936 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
938 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
943 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
944 void *value, u64 map_flags,
947 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
948 struct htab_elem *l_new = NULL, *l_old;
949 struct hlist_nulls_head *head;
955 if (unlikely(map_flags > BPF_EXIST))
959 WARN_ON_ONCE(!rcu_read_lock_held());
961 key_size = map->key_size;
963 hash = htab_map_hash(key, key_size, htab->hashrnd);
965 b = __select_bucket(htab, hash);
968 /* bpf_map_update_elem() can be called in_irq() */
969 raw_spin_lock_irqsave(&b->lock, flags);
971 l_old = lookup_elem_raw(head, hash, key, key_size);
973 ret = check_flags(htab, l_old, map_flags);
978 /* per-cpu hash map can update value in-place */
979 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
982 l_new = alloc_htab_elem(htab, key, value, key_size,
983 hash, true, onallcpus, NULL);
985 ret = PTR_ERR(l_new);
988 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
992 raw_spin_unlock_irqrestore(&b->lock, flags);
996 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
997 void *value, u64 map_flags,
1000 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1001 struct htab_elem *l_new = NULL, *l_old;
1002 struct hlist_nulls_head *head;
1003 unsigned long flags;
1008 if (unlikely(map_flags > BPF_EXIST))
1012 WARN_ON_ONCE(!rcu_read_lock_held());
1014 key_size = map->key_size;
1016 hash = htab_map_hash(key, key_size, htab->hashrnd);
1018 b = __select_bucket(htab, hash);
1021 /* For LRU, we need to alloc before taking bucket's
1022 * spinlock because LRU's elem alloc may need
1023 * to remove older elem from htab and this removal
1024 * operation will need a bucket lock.
1026 if (map_flags != BPF_EXIST) {
1027 l_new = prealloc_lru_pop(htab, key, hash);
1032 /* bpf_map_update_elem() can be called in_irq() */
1033 raw_spin_lock_irqsave(&b->lock, flags);
1035 l_old = lookup_elem_raw(head, hash, key, key_size);
1037 ret = check_flags(htab, l_old, map_flags);
1042 bpf_lru_node_set_ref(&l_old->lru_node);
1044 /* per-cpu hash map can update value in-place */
1045 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1048 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1050 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1055 raw_spin_unlock_irqrestore(&b->lock, flags);
1057 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1061 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1062 void *value, u64 map_flags)
1064 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1067 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1068 void *value, u64 map_flags)
1070 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1074 /* Called from syscall or from eBPF program */
1075 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1077 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1078 struct hlist_nulls_head *head;
1080 struct htab_elem *l;
1081 unsigned long flags;
1085 WARN_ON_ONCE(!rcu_read_lock_held());
1087 key_size = map->key_size;
1089 hash = htab_map_hash(key, key_size, htab->hashrnd);
1090 b = __select_bucket(htab, hash);
1093 raw_spin_lock_irqsave(&b->lock, flags);
1095 l = lookup_elem_raw(head, hash, key, key_size);
1098 hlist_nulls_del_rcu(&l->hash_node);
1099 free_htab_elem(htab, l);
1103 raw_spin_unlock_irqrestore(&b->lock, flags);
1107 static int htab_lru_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);
1134 raw_spin_unlock_irqrestore(&b->lock, flags);
1136 bpf_lru_push_free(&htab->lru, &l->lru_node);
1140 static void delete_all_elements(struct bpf_htab *htab)
1144 for (i = 0; i < htab->n_buckets; i++) {
1145 struct hlist_nulls_head *head = select_bucket(htab, i);
1146 struct hlist_nulls_node *n;
1147 struct htab_elem *l;
1149 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1150 hlist_nulls_del_rcu(&l->hash_node);
1151 htab_elem_free(htab, l);
1156 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1157 static void htab_map_free(struct bpf_map *map)
1159 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1161 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1162 * so the programs (can be more than one that used this map) were
1163 * disconnected from events. Wait for outstanding critical sections in
1164 * these programs to complete
1168 /* some of free_htab_elem() callbacks for elements of this map may
1169 * not have executed. Wait for them.
1172 if (!htab_is_prealloc(htab))
1173 delete_all_elements(htab);
1175 prealloc_destroy(htab);
1177 free_percpu(htab->extra_elems);
1178 bpf_map_area_free(htab->buckets);
1182 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1189 value = htab_map_lookup_elem(map, key);
1195 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1197 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1203 const struct bpf_map_ops htab_map_ops = {
1204 .map_alloc_check = htab_map_alloc_check,
1205 .map_alloc = htab_map_alloc,
1206 .map_free = htab_map_free,
1207 .map_get_next_key = htab_map_get_next_key,
1208 .map_lookup_elem = htab_map_lookup_elem,
1209 .map_update_elem = htab_map_update_elem,
1210 .map_delete_elem = htab_map_delete_elem,
1211 .map_gen_lookup = htab_map_gen_lookup,
1212 .map_seq_show_elem = htab_map_seq_show_elem,
1215 const struct bpf_map_ops htab_lru_map_ops = {
1216 .map_alloc_check = htab_map_alloc_check,
1217 .map_alloc = htab_map_alloc,
1218 .map_free = htab_map_free,
1219 .map_get_next_key = htab_map_get_next_key,
1220 .map_lookup_elem = htab_lru_map_lookup_elem,
1221 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1222 .map_update_elem = htab_lru_map_update_elem,
1223 .map_delete_elem = htab_lru_map_delete_elem,
1224 .map_gen_lookup = htab_lru_map_gen_lookup,
1225 .map_seq_show_elem = htab_map_seq_show_elem,
1228 /* Called from eBPF program */
1229 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1231 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1234 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1239 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1241 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1244 bpf_lru_node_set_ref(&l->lru_node);
1245 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1251 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1253 struct htab_elem *l;
1254 void __percpu *pptr;
1259 /* per_cpu areas are zero-filled and bpf programs can only
1260 * access 'value_size' of them, so copying rounded areas
1261 * will not leak any kernel data
1263 size = round_up(map->value_size, 8);
1265 l = __htab_map_lookup_elem(map, key);
1268 /* We do not mark LRU map element here in order to not mess up
1269 * eviction heuristics when user space does a map walk.
1271 pptr = htab_elem_get_ptr(l, map->key_size);
1272 for_each_possible_cpu(cpu) {
1273 bpf_long_memcpy(value + off,
1274 per_cpu_ptr(pptr, cpu), size);
1283 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1286 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1290 if (htab_is_lru(htab))
1291 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1294 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1301 const struct bpf_map_ops htab_percpu_map_ops = {
1302 .map_alloc_check = htab_map_alloc_check,
1303 .map_alloc = htab_map_alloc,
1304 .map_free = htab_map_free,
1305 .map_get_next_key = htab_map_get_next_key,
1306 .map_lookup_elem = htab_percpu_map_lookup_elem,
1307 .map_update_elem = htab_percpu_map_update_elem,
1308 .map_delete_elem = htab_map_delete_elem,
1311 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1312 .map_alloc_check = htab_map_alloc_check,
1313 .map_alloc = htab_map_alloc,
1314 .map_free = htab_map_free,
1315 .map_get_next_key = htab_map_get_next_key,
1316 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1317 .map_update_elem = htab_lru_percpu_map_update_elem,
1318 .map_delete_elem = htab_lru_map_delete_elem,
1321 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1323 if (attr->value_size != sizeof(u32))
1325 return htab_map_alloc_check(attr);
1328 static void fd_htab_map_free(struct bpf_map *map)
1330 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1331 struct hlist_nulls_node *n;
1332 struct hlist_nulls_head *head;
1333 struct htab_elem *l;
1336 for (i = 0; i < htab->n_buckets; i++) {
1337 head = select_bucket(htab, i);
1339 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1340 void *ptr = fd_htab_map_get_ptr(map, l);
1342 map->ops->map_fd_put_ptr(ptr);
1349 /* only called from syscall */
1350 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1355 if (!map->ops->map_fd_sys_lookup_elem)
1359 ptr = htab_map_lookup_elem(map, key);
1361 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1369 /* only called from syscall */
1370 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1371 void *key, void *value, u64 map_flags)
1375 u32 ufd = *(u32 *)value;
1377 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1379 return PTR_ERR(ptr);
1381 ret = htab_map_update_elem(map, key, &ptr, map_flags);
1383 map->ops->map_fd_put_ptr(ptr);
1388 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1390 struct bpf_map *map, *inner_map_meta;
1392 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1393 if (IS_ERR(inner_map_meta))
1394 return inner_map_meta;
1396 map = htab_map_alloc(attr);
1398 bpf_map_meta_free(inner_map_meta);
1402 map->inner_map_meta = inner_map_meta;
1407 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1409 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
1414 return READ_ONCE(*inner_map);
1417 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1418 struct bpf_insn *insn_buf)
1420 struct bpf_insn *insn = insn_buf;
1421 const int ret = BPF_REG_0;
1423 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1424 (void *(*)(struct bpf_map *map, void *key))NULL));
1425 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1426 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1427 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1428 offsetof(struct htab_elem, key) +
1429 round_up(map->key_size, 8));
1430 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1432 return insn - insn_buf;
1435 static void htab_of_map_free(struct bpf_map *map)
1437 bpf_map_meta_free(map->inner_map_meta);
1438 fd_htab_map_free(map);
1441 const struct bpf_map_ops htab_of_maps_map_ops = {
1442 .map_alloc_check = fd_htab_map_alloc_check,
1443 .map_alloc = htab_of_map_alloc,
1444 .map_free = htab_of_map_free,
1445 .map_get_next_key = htab_map_get_next_key,
1446 .map_lookup_elem = htab_of_map_lookup_elem,
1447 .map_delete_elem = htab_map_delete_elem,
1448 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1449 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1450 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1451 .map_gen_lookup = htab_of_map_gen_lookup,
1452 .map_check_btf = map_check_no_btf,