1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 #include <linux/bpf.h>
14 #include <linux/jhash.h>
15 #include <linux/filter.h>
16 #include <linux/rculist_nulls.h>
17 #include "percpu_freelist.h"
18 #include "bpf_lru_list.h"
19 #include "map_in_map.h"
21 #define HTAB_CREATE_FLAG_MASK \
22 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE)
25 struct hlist_nulls_head head;
31 struct bucket *buckets;
34 struct pcpu_freelist freelist;
37 struct htab_elem *__percpu *extra_elems;
38 atomic_t count; /* number of elements in this hashtable */
39 u32 n_buckets; /* number of hash buckets */
40 u32 elem_size; /* size of each element in bytes */
43 /* each htab element is struct htab_elem + key + value */
46 struct hlist_nulls_node hash_node;
50 struct bpf_htab *htab;
51 struct pcpu_freelist_node fnode;
57 struct bpf_lru_node lru_node;
60 char key[0] __aligned(8);
63 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
65 static bool htab_is_lru(const struct bpf_htab *htab)
67 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
68 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
71 static bool htab_is_percpu(const struct bpf_htab *htab)
73 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
74 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
77 static bool htab_is_prealloc(const struct bpf_htab *htab)
79 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
82 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
85 *(void __percpu **)(l->key + key_size) = pptr;
88 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
90 return *(void __percpu **)(l->key + key_size);
93 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
95 return *(void **)(l->key + roundup(map->key_size, 8));
98 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
100 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
103 static void htab_free_elems(struct bpf_htab *htab)
107 if (!htab_is_percpu(htab))
110 for (i = 0; i < htab->map.max_entries; i++) {
113 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
118 bpf_map_area_free(htab->elems);
121 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
124 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
128 l = container_of(node, struct htab_elem, lru_node);
129 memcpy(l->key, key, htab->map.key_size);
136 static int prealloc_init(struct bpf_htab *htab)
138 u32 num_entries = htab->map.max_entries;
139 int err = -ENOMEM, i;
141 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
142 num_entries += num_possible_cpus();
144 htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
145 htab->map.numa_node);
149 if (!htab_is_percpu(htab))
150 goto skip_percpu_elems;
152 for (i = 0; i < num_entries; i++) {
153 u32 size = round_up(htab->map.value_size, 8);
156 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
159 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
164 if (htab_is_lru(htab))
165 err = bpf_lru_init(&htab->lru,
166 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
167 offsetof(struct htab_elem, hash) -
168 offsetof(struct htab_elem, lru_node),
169 htab_lru_map_delete_node,
172 err = pcpu_freelist_init(&htab->freelist);
177 if (htab_is_lru(htab))
178 bpf_lru_populate(&htab->lru, htab->elems,
179 offsetof(struct htab_elem, lru_node),
180 htab->elem_size, num_entries);
182 pcpu_freelist_populate(&htab->freelist,
183 htab->elems + offsetof(struct htab_elem, fnode),
184 htab->elem_size, num_entries);
189 htab_free_elems(htab);
193 static void prealloc_destroy(struct bpf_htab *htab)
195 htab_free_elems(htab);
197 if (htab_is_lru(htab))
198 bpf_lru_destroy(&htab->lru);
200 pcpu_freelist_destroy(&htab->freelist);
203 static int alloc_extra_elems(struct bpf_htab *htab)
205 struct htab_elem *__percpu *pptr, *l_new;
206 struct pcpu_freelist_node *l;
209 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
210 GFP_USER | __GFP_NOWARN);
214 for_each_possible_cpu(cpu) {
215 l = pcpu_freelist_pop(&htab->freelist);
216 /* pop will succeed, since prealloc_init()
217 * preallocated extra num_possible_cpus elements
219 l_new = container_of(l, struct htab_elem, fnode);
220 *per_cpu_ptr(pptr, cpu) = l_new;
222 htab->extra_elems = pptr;
226 /* Called from syscall */
227 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
229 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
230 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
231 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
232 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
233 /* percpu_lru means each cpu has its own LRU list.
234 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
235 * the map's value itself is percpu. percpu_lru has
236 * nothing to do with the map's value.
238 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
239 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
240 int numa_node = bpf_map_attr_numa_node(attr);
241 struct bpf_htab *htab;
245 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
246 offsetof(struct htab_elem, hash_node.pprev));
247 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
248 offsetof(struct htab_elem, hash_node.pprev));
250 if (lru && !capable(CAP_SYS_ADMIN))
251 /* LRU implementation is much complicated than other
252 * maps. Hence, limit to CAP_SYS_ADMIN for now.
254 return ERR_PTR(-EPERM);
256 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
257 /* reserved bits should not be used */
258 return ERR_PTR(-EINVAL);
260 if (!lru && percpu_lru)
261 return ERR_PTR(-EINVAL);
263 if (lru && !prealloc)
264 return ERR_PTR(-ENOTSUPP);
266 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
267 return ERR_PTR(-EINVAL);
269 htab = kzalloc(sizeof(*htab), GFP_USER);
271 return ERR_PTR(-ENOMEM);
273 /* mandatory map attributes */
274 htab->map.map_type = attr->map_type;
275 htab->map.key_size = attr->key_size;
276 htab->map.value_size = attr->value_size;
277 htab->map.max_entries = attr->max_entries;
278 htab->map.map_flags = attr->map_flags;
279 htab->map.numa_node = numa_node;
281 /* check sanity of attributes.
282 * value_size == 0 may be allowed in the future to use map as a set
285 if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
286 htab->map.value_size == 0)
290 /* ensure each CPU's lru list has >=1 elements.
291 * since we are at it, make each lru list has the same
292 * number of elements.
294 htab->map.max_entries = roundup(attr->max_entries,
295 num_possible_cpus());
296 if (htab->map.max_entries < attr->max_entries)
297 htab->map.max_entries = rounddown(attr->max_entries,
298 num_possible_cpus());
301 /* hash table size must be power of 2 */
302 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
305 if (htab->map.key_size > MAX_BPF_STACK)
306 /* eBPF programs initialize keys on stack, so they cannot be
307 * larger than max stack size
311 if (htab->map.value_size >= KMALLOC_MAX_SIZE -
312 MAX_BPF_STACK - sizeof(struct htab_elem))
313 /* if value_size is bigger, the user space won't be able to
314 * access the elements via bpf syscall. This check also makes
315 * sure that the elem_size doesn't overflow and it's
316 * kmalloc-able later in htab_map_update_elem()
320 htab->elem_size = sizeof(struct htab_elem) +
321 round_up(htab->map.key_size, 8);
323 htab->elem_size += sizeof(void *);
325 htab->elem_size += round_up(htab->map.value_size, 8);
327 /* prevent zero size kmalloc and check for u32 overflow */
328 if (htab->n_buckets == 0 ||
329 htab->n_buckets > U32_MAX / sizeof(struct bucket))
332 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
333 (u64) htab->elem_size * htab->map.max_entries;
336 cost += (u64) round_up(htab->map.value_size, 8) *
337 num_possible_cpus() * htab->map.max_entries;
339 cost += (u64) htab->elem_size * num_possible_cpus();
341 if (cost >= U32_MAX - PAGE_SIZE)
342 /* make sure page count doesn't overflow */
345 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
347 /* if map size is larger than memlock limit, reject it early */
348 err = bpf_map_precharge_memlock(htab->map.pages);
353 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
354 sizeof(struct bucket),
355 htab->map.numa_node);
359 for (i = 0; i < htab->n_buckets; i++) {
360 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
361 raw_spin_lock_init(&htab->buckets[i].lock);
365 err = prealloc_init(htab);
369 if (!percpu && !lru) {
370 /* lru itself can remove the least used element, so
371 * there is no need for an extra elem during map_update.
373 err = alloc_extra_elems(htab);
382 prealloc_destroy(htab);
384 bpf_map_area_free(htab->buckets);
390 static inline u32 htab_map_hash(const void *key, u32 key_len)
392 return jhash(key, key_len, 0);
395 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
397 return &htab->buckets[hash & (htab->n_buckets - 1)];
400 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
402 return &__select_bucket(htab, hash)->head;
405 /* this lookup function can only be called with bucket lock taken */
406 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
407 void *key, u32 key_size)
409 struct hlist_nulls_node *n;
412 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
413 if (l->hash == hash && !memcmp(&l->key, key, key_size))
419 /* can be called without bucket lock. it will repeat the loop in
420 * the unlikely event when elements moved from one bucket into another
421 * while link list is being walked
423 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
425 u32 key_size, u32 n_buckets)
427 struct hlist_nulls_node *n;
431 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
432 if (l->hash == hash && !memcmp(&l->key, key, key_size))
435 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
441 /* Called from syscall or from eBPF program directly, so
442 * arguments have to match bpf_map_lookup_elem() exactly.
443 * The return value is adjusted by BPF instructions
444 * in htab_map_gen_lookup().
446 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
448 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
449 struct hlist_nulls_head *head;
453 /* Must be called with rcu_read_lock. */
454 WARN_ON_ONCE(!rcu_read_lock_held());
456 key_size = map->key_size;
458 hash = htab_map_hash(key, key_size);
460 head = select_bucket(htab, hash);
462 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
467 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
469 struct htab_elem *l = __htab_map_lookup_elem(map, key);
472 return l->key + round_up(map->key_size, 8);
477 /* inline bpf_map_lookup_elem() call.
480 * bpf_map_lookup_elem
481 * map->ops->map_lookup_elem
482 * htab_map_lookup_elem
483 * __htab_map_lookup_elem
486 * __htab_map_lookup_elem
488 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
490 struct bpf_insn *insn = insn_buf;
491 const int ret = BPF_REG_0;
493 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
494 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
495 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
496 offsetof(struct htab_elem, key) +
497 round_up(map->key_size, 8));
498 return insn - insn_buf;
501 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
502 void *key, const bool mark)
504 struct htab_elem *l = __htab_map_lookup_elem(map, key);
508 bpf_lru_node_set_ref(&l->lru_node);
509 return l->key + round_up(map->key_size, 8);
515 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
517 return __htab_lru_map_lookup_elem(map, key, true);
520 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
522 return __htab_lru_map_lookup_elem(map, key, false);
525 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
526 struct bpf_insn *insn_buf)
528 struct bpf_insn *insn = insn_buf;
529 const int ret = BPF_REG_0;
530 const int ref_reg = BPF_REG_1;
532 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
533 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
534 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
535 offsetof(struct htab_elem, lru_node) +
536 offsetof(struct bpf_lru_node, ref));
537 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
538 *insn++ = BPF_ST_MEM(BPF_B, ret,
539 offsetof(struct htab_elem, lru_node) +
540 offsetof(struct bpf_lru_node, ref),
542 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
543 offsetof(struct htab_elem, key) +
544 round_up(map->key_size, 8));
545 return insn - insn_buf;
548 /* It is called from the bpf_lru_list when the LRU needs to delete
549 * older elements from the htab.
551 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
553 struct bpf_htab *htab = (struct bpf_htab *)arg;
554 struct htab_elem *l = NULL, *tgt_l;
555 struct hlist_nulls_head *head;
556 struct hlist_nulls_node *n;
560 tgt_l = container_of(node, struct htab_elem, lru_node);
561 b = __select_bucket(htab, tgt_l->hash);
564 raw_spin_lock_irqsave(&b->lock, flags);
566 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
568 hlist_nulls_del_rcu(&l->hash_node);
572 raw_spin_unlock_irqrestore(&b->lock, flags);
577 /* Called from syscall */
578 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
580 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
581 struct hlist_nulls_head *head;
582 struct htab_elem *l, *next_l;
586 WARN_ON_ONCE(!rcu_read_lock_held());
588 key_size = map->key_size;
591 goto find_first_elem;
593 hash = htab_map_hash(key, key_size);
595 head = select_bucket(htab, hash);
598 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
601 goto find_first_elem;
603 /* key was found, get next key in the same bucket */
604 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
605 struct htab_elem, hash_node);
608 /* if next elem in this hash list is non-zero, just return it */
609 memcpy(next_key, next_l->key, key_size);
613 /* no more elements in this hash list, go to the next bucket */
614 i = hash & (htab->n_buckets - 1);
618 /* iterate over buckets */
619 for (; i < htab->n_buckets; i++) {
620 head = select_bucket(htab, i);
622 /* pick first element in the bucket */
623 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
624 struct htab_elem, hash_node);
626 /* if it's not empty, just return it */
627 memcpy(next_key, next_l->key, key_size);
632 /* iterated over all buckets and all elements */
636 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
638 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
639 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
643 static void htab_elem_free_rcu(struct rcu_head *head)
645 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
646 struct bpf_htab *htab = l->htab;
648 htab_elem_free(htab, l);
651 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
653 struct bpf_map *map = &htab->map;
656 if (map->ops->map_fd_put_ptr) {
657 ptr = fd_htab_map_get_ptr(map, l);
658 map->ops->map_fd_put_ptr(ptr);
662 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
664 htab_put_fd_value(htab, l);
666 if (htab_is_prealloc(htab)) {
667 __pcpu_freelist_push(&htab->freelist, &l->fnode);
669 atomic_dec(&htab->count);
671 call_rcu(&l->rcu, htab_elem_free_rcu);
675 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
676 void *value, bool onallcpus)
679 /* copy true value_size bytes */
680 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
682 u32 size = round_up(htab->map.value_size, 8);
685 for_each_possible_cpu(cpu) {
686 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
693 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
695 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
699 static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
701 u32 size = htab->map.value_size;
703 if (percpu || fd_htab_map_needs_adjust(htab))
704 size = round_up(size, 8);
708 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
709 void *value, u32 key_size, u32 hash,
710 bool percpu, bool onallcpus,
711 struct htab_elem *old_elem)
713 u32 size = htab_size_value(htab, percpu);
714 bool prealloc = htab_is_prealloc(htab);
715 struct htab_elem *l_new, **pl_new;
720 /* if we're updating the existing element,
721 * use per-cpu extra elems to avoid freelist_pop/push
723 pl_new = this_cpu_ptr(htab->extra_elems);
725 htab_put_fd_value(htab, old_elem);
728 struct pcpu_freelist_node *l;
730 l = __pcpu_freelist_pop(&htab->freelist);
732 return ERR_PTR(-E2BIG);
733 l_new = container_of(l, struct htab_elem, fnode);
736 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
738 /* when map is full and update() is replacing
739 * old element, it's ok to allocate, since
740 * old element will be freed immediately.
741 * Otherwise return an error
743 l_new = ERR_PTR(-E2BIG);
746 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
747 htab->map.numa_node);
749 l_new = ERR_PTR(-ENOMEM);
754 memcpy(l_new->key, key, key_size);
757 pptr = htab_elem_get_ptr(l_new, key_size);
759 /* alloc_percpu zero-fills */
760 pptr = __alloc_percpu_gfp(size, 8,
761 GFP_ATOMIC | __GFP_NOWARN);
764 l_new = ERR_PTR(-ENOMEM);
769 pcpu_copy_value(htab, pptr, value, onallcpus);
772 htab_elem_set_ptr(l_new, key_size, pptr);
774 memcpy(l_new->key + round_up(key_size, 8), value, size);
780 atomic_dec(&htab->count);
784 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
787 if (l_old && map_flags == BPF_NOEXIST)
788 /* elem already exists */
791 if (!l_old && map_flags == BPF_EXIST)
792 /* elem doesn't exist, cannot update it */
798 /* Called from syscall or from eBPF program */
799 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
802 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
803 struct htab_elem *l_new = NULL, *l_old;
804 struct hlist_nulls_head *head;
810 if (unlikely(map_flags > BPF_EXIST))
814 WARN_ON_ONCE(!rcu_read_lock_held());
816 key_size = map->key_size;
818 hash = htab_map_hash(key, key_size);
820 b = __select_bucket(htab, hash);
823 /* bpf_map_update_elem() can be called in_irq() */
824 raw_spin_lock_irqsave(&b->lock, flags);
826 l_old = lookup_elem_raw(head, hash, key, key_size);
828 ret = check_flags(htab, l_old, map_flags);
832 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
835 /* all pre-allocated elements are in use or memory exhausted */
836 ret = PTR_ERR(l_new);
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 hlist_nulls_del_rcu(&l_old->hash_node);
846 if (!htab_is_prealloc(htab))
847 free_htab_elem(htab, l_old);
851 raw_spin_unlock_irqrestore(&b->lock, flags);
855 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
858 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
859 struct htab_elem *l_new, *l_old = NULL;
860 struct hlist_nulls_head *head;
866 if (unlikely(map_flags > BPF_EXIST))
870 WARN_ON_ONCE(!rcu_read_lock_held());
872 key_size = map->key_size;
874 hash = htab_map_hash(key, key_size);
876 b = __select_bucket(htab, hash);
879 /* For LRU, we need to alloc before taking bucket's
880 * spinlock because getting free nodes from LRU may need
881 * to remove older elements from htab and this removal
882 * operation will need a bucket lock.
884 l_new = prealloc_lru_pop(htab, key, hash);
887 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
889 /* bpf_map_update_elem() can be called in_irq() */
890 raw_spin_lock_irqsave(&b->lock, flags);
892 l_old = lookup_elem_raw(head, hash, key, key_size);
894 ret = check_flags(htab, l_old, map_flags);
898 /* add new element to the head of the list, so that
899 * concurrent search will find it before old elem
901 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
903 bpf_lru_node_set_ref(&l_new->lru_node);
904 hlist_nulls_del_rcu(&l_old->hash_node);
909 raw_spin_unlock_irqrestore(&b->lock, flags);
912 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
914 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
919 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
920 void *value, u64 map_flags,
923 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
924 struct htab_elem *l_new = NULL, *l_old;
925 struct hlist_nulls_head *head;
931 if (unlikely(map_flags > BPF_EXIST))
935 WARN_ON_ONCE(!rcu_read_lock_held());
937 key_size = map->key_size;
939 hash = htab_map_hash(key, key_size);
941 b = __select_bucket(htab, hash);
944 /* bpf_map_update_elem() can be called in_irq() */
945 raw_spin_lock_irqsave(&b->lock, flags);
947 l_old = lookup_elem_raw(head, hash, key, key_size);
949 ret = check_flags(htab, l_old, map_flags);
954 /* per-cpu hash map can update value in-place */
955 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
958 l_new = alloc_htab_elem(htab, key, value, key_size,
959 hash, true, onallcpus, NULL);
961 ret = PTR_ERR(l_new);
964 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
968 raw_spin_unlock_irqrestore(&b->lock, flags);
972 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
973 void *value, u64 map_flags,
976 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
977 struct htab_elem *l_new = NULL, *l_old;
978 struct hlist_nulls_head *head;
984 if (unlikely(map_flags > BPF_EXIST))
988 WARN_ON_ONCE(!rcu_read_lock_held());
990 key_size = map->key_size;
992 hash = htab_map_hash(key, key_size);
994 b = __select_bucket(htab, hash);
997 /* For LRU, we need to alloc before taking bucket's
998 * spinlock because LRU's elem alloc may need
999 * to remove older elem from htab and this removal
1000 * operation will need a bucket lock.
1002 if (map_flags != BPF_EXIST) {
1003 l_new = prealloc_lru_pop(htab, key, hash);
1008 /* bpf_map_update_elem() can be called in_irq() */
1009 raw_spin_lock_irqsave(&b->lock, flags);
1011 l_old = lookup_elem_raw(head, hash, key, key_size);
1013 ret = check_flags(htab, l_old, map_flags);
1018 bpf_lru_node_set_ref(&l_old->lru_node);
1020 /* per-cpu hash map can update value in-place */
1021 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1024 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1026 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1031 raw_spin_unlock_irqrestore(&b->lock, flags);
1033 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1037 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1038 void *value, u64 map_flags)
1040 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1043 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1044 void *value, u64 map_flags)
1046 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1050 /* Called from syscall or from eBPF program */
1051 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1053 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1054 struct hlist_nulls_head *head;
1056 struct htab_elem *l;
1057 unsigned long flags;
1061 WARN_ON_ONCE(!rcu_read_lock_held());
1063 key_size = map->key_size;
1065 hash = htab_map_hash(key, key_size);
1066 b = __select_bucket(htab, hash);
1069 raw_spin_lock_irqsave(&b->lock, flags);
1071 l = lookup_elem_raw(head, hash, key, key_size);
1074 hlist_nulls_del_rcu(&l->hash_node);
1075 free_htab_elem(htab, l);
1079 raw_spin_unlock_irqrestore(&b->lock, flags);
1083 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1085 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1086 struct hlist_nulls_head *head;
1088 struct htab_elem *l;
1089 unsigned long flags;
1093 WARN_ON_ONCE(!rcu_read_lock_held());
1095 key_size = map->key_size;
1097 hash = htab_map_hash(key, key_size);
1098 b = __select_bucket(htab, hash);
1101 raw_spin_lock_irqsave(&b->lock, flags);
1103 l = lookup_elem_raw(head, hash, key, key_size);
1106 hlist_nulls_del_rcu(&l->hash_node);
1110 raw_spin_unlock_irqrestore(&b->lock, flags);
1112 bpf_lru_push_free(&htab->lru, &l->lru_node);
1116 static void delete_all_elements(struct bpf_htab *htab)
1120 for (i = 0; i < htab->n_buckets; i++) {
1121 struct hlist_nulls_head *head = select_bucket(htab, i);
1122 struct hlist_nulls_node *n;
1123 struct htab_elem *l;
1125 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1126 hlist_nulls_del_rcu(&l->hash_node);
1127 htab_elem_free(htab, l);
1132 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1133 static void htab_map_free(struct bpf_map *map)
1135 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1137 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1138 * so the programs (can be more than one that used this map) were
1139 * disconnected from events. Wait for outstanding critical sections in
1140 * these programs to complete
1144 /* some of free_htab_elem() callbacks for elements of this map may
1145 * not have executed. Wait for them.
1148 if (!htab_is_prealloc(htab))
1149 delete_all_elements(htab);
1151 prealloc_destroy(htab);
1153 free_percpu(htab->extra_elems);
1154 bpf_map_area_free(htab->buckets);
1158 const struct bpf_map_ops htab_map_ops = {
1159 .map_alloc = htab_map_alloc,
1160 .map_free = htab_map_free,
1161 .map_get_next_key = htab_map_get_next_key,
1162 .map_lookup_elem = htab_map_lookup_elem,
1163 .map_update_elem = htab_map_update_elem,
1164 .map_delete_elem = htab_map_delete_elem,
1165 .map_gen_lookup = htab_map_gen_lookup,
1168 const struct bpf_map_ops htab_lru_map_ops = {
1169 .map_alloc = htab_map_alloc,
1170 .map_free = htab_map_free,
1171 .map_get_next_key = htab_map_get_next_key,
1172 .map_lookup_elem = htab_lru_map_lookup_elem,
1173 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1174 .map_update_elem = htab_lru_map_update_elem,
1175 .map_delete_elem = htab_lru_map_delete_elem,
1176 .map_gen_lookup = htab_lru_map_gen_lookup,
1179 /* Called from eBPF program */
1180 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1182 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1185 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1190 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1192 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1195 bpf_lru_node_set_ref(&l->lru_node);
1196 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1202 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1204 struct htab_elem *l;
1205 void __percpu *pptr;
1210 /* per_cpu areas are zero-filled and bpf programs can only
1211 * access 'value_size' of them, so copying rounded areas
1212 * will not leak any kernel data
1214 size = round_up(map->value_size, 8);
1216 l = __htab_map_lookup_elem(map, key);
1219 /* We do not mark LRU map element here in order to not mess up
1220 * eviction heuristics when user space does a map walk.
1222 pptr = htab_elem_get_ptr(l, map->key_size);
1223 for_each_possible_cpu(cpu) {
1224 bpf_long_memcpy(value + off,
1225 per_cpu_ptr(pptr, cpu), size);
1234 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1237 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1241 if (htab_is_lru(htab))
1242 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1245 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1252 const struct bpf_map_ops htab_percpu_map_ops = {
1253 .map_alloc = htab_map_alloc,
1254 .map_free = htab_map_free,
1255 .map_get_next_key = htab_map_get_next_key,
1256 .map_lookup_elem = htab_percpu_map_lookup_elem,
1257 .map_update_elem = htab_percpu_map_update_elem,
1258 .map_delete_elem = htab_map_delete_elem,
1261 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1262 .map_alloc = htab_map_alloc,
1263 .map_free = htab_map_free,
1264 .map_get_next_key = htab_map_get_next_key,
1265 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1266 .map_update_elem = htab_lru_percpu_map_update_elem,
1267 .map_delete_elem = htab_lru_map_delete_elem,
1270 static struct bpf_map *fd_htab_map_alloc(union bpf_attr *attr)
1272 if (attr->value_size != sizeof(u32))
1273 return ERR_PTR(-EINVAL);
1274 return htab_map_alloc(attr);
1277 static void fd_htab_map_free(struct bpf_map *map)
1279 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1280 struct hlist_nulls_node *n;
1281 struct hlist_nulls_head *head;
1282 struct htab_elem *l;
1285 for (i = 0; i < htab->n_buckets; i++) {
1286 head = select_bucket(htab, i);
1288 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1289 void *ptr = fd_htab_map_get_ptr(map, l);
1291 map->ops->map_fd_put_ptr(ptr);
1298 /* only called from syscall */
1299 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1304 if (!map->ops->map_fd_sys_lookup_elem)
1308 ptr = htab_map_lookup_elem(map, key);
1310 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1318 /* only called from syscall */
1319 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1320 void *key, void *value, u64 map_flags)
1324 u32 ufd = *(u32 *)value;
1326 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1328 return PTR_ERR(ptr);
1330 ret = htab_map_update_elem(map, key, &ptr, map_flags);
1332 map->ops->map_fd_put_ptr(ptr);
1337 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1339 struct bpf_map *map, *inner_map_meta;
1341 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1342 if (IS_ERR(inner_map_meta))
1343 return inner_map_meta;
1345 map = fd_htab_map_alloc(attr);
1347 bpf_map_meta_free(inner_map_meta);
1351 map->inner_map_meta = inner_map_meta;
1356 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1358 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
1363 return READ_ONCE(*inner_map);
1366 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1367 struct bpf_insn *insn_buf)
1369 struct bpf_insn *insn = insn_buf;
1370 const int ret = BPF_REG_0;
1372 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
1373 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1374 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1375 offsetof(struct htab_elem, key) +
1376 round_up(map->key_size, 8));
1377 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1379 return insn - insn_buf;
1382 static void htab_of_map_free(struct bpf_map *map)
1384 bpf_map_meta_free(map->inner_map_meta);
1385 fd_htab_map_free(map);
1388 const struct bpf_map_ops htab_of_maps_map_ops = {
1389 .map_alloc = htab_of_map_alloc,
1390 .map_free = htab_of_map_free,
1391 .map_get_next_key = htab_map_get_next_key,
1392 .map_lookup_elem = htab_of_map_lookup_elem,
1393 .map_delete_elem = htab_map_delete_elem,
1394 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1395 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1396 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1397 .map_gen_lookup = htab_of_map_gen_lookup,