1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 #include <linux/bpf.h>
14 #include <linux/jhash.h>
15 #include <linux/filter.h>
16 #include <linux/rculist_nulls.h>
17 #include "percpu_freelist.h"
18 #include "bpf_lru_list.h"
19 #include "map_in_map.h"
21 #define HTAB_CREATE_FLAG_MASK \
22 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
23 BPF_F_RDONLY | BPF_F_WRONLY)
26 struct hlist_nulls_head head;
32 struct bucket *buckets;
35 struct pcpu_freelist freelist;
38 struct htab_elem *__percpu *extra_elems;
39 atomic_t count; /* number of elements in this hashtable */
40 u32 n_buckets; /* number of hash buckets */
41 u32 elem_size; /* size of each element in bytes */
44 /* each htab element is struct htab_elem + key + value */
47 struct hlist_nulls_node hash_node;
51 struct bpf_htab *htab;
52 struct pcpu_freelist_node fnode;
58 struct bpf_lru_node lru_node;
61 char key[0] __aligned(8);
64 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
66 static bool htab_is_lru(const struct bpf_htab *htab)
68 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
69 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
72 static bool htab_is_percpu(const struct bpf_htab *htab)
74 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
75 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
78 static bool htab_is_prealloc(const struct bpf_htab *htab)
80 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
83 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
86 *(void __percpu **)(l->key + key_size) = pptr;
89 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
91 return *(void __percpu **)(l->key + key_size);
94 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
96 return *(void **)(l->key + roundup(map->key_size, 8));
99 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
101 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
104 static void htab_free_elems(struct bpf_htab *htab)
108 if (!htab_is_percpu(htab))
111 for (i = 0; i < htab->map.max_entries; i++) {
114 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
119 bpf_map_area_free(htab->elems);
122 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
125 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
129 l = container_of(node, struct htab_elem, lru_node);
130 memcpy(l->key, key, htab->map.key_size);
137 static int prealloc_init(struct bpf_htab *htab)
139 u32 num_entries = htab->map.max_entries;
140 int err = -ENOMEM, i;
142 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
143 num_entries += num_possible_cpus();
145 htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
146 htab->map.numa_node);
150 if (!htab_is_percpu(htab))
151 goto skip_percpu_elems;
153 for (i = 0; i < num_entries; i++) {
154 u32 size = round_up(htab->map.value_size, 8);
157 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
160 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
165 if (htab_is_lru(htab))
166 err = bpf_lru_init(&htab->lru,
167 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
168 offsetof(struct htab_elem, hash) -
169 offsetof(struct htab_elem, lru_node),
170 htab_lru_map_delete_node,
173 err = pcpu_freelist_init(&htab->freelist);
178 if (htab_is_lru(htab))
179 bpf_lru_populate(&htab->lru, htab->elems,
180 offsetof(struct htab_elem, lru_node),
181 htab->elem_size, num_entries);
183 pcpu_freelist_populate(&htab->freelist,
184 htab->elems + offsetof(struct htab_elem, fnode),
185 htab->elem_size, num_entries);
190 htab_free_elems(htab);
194 static void prealloc_destroy(struct bpf_htab *htab)
196 htab_free_elems(htab);
198 if (htab_is_lru(htab))
199 bpf_lru_destroy(&htab->lru);
201 pcpu_freelist_destroy(&htab->freelist);
204 static int alloc_extra_elems(struct bpf_htab *htab)
206 struct htab_elem *__percpu *pptr, *l_new;
207 struct pcpu_freelist_node *l;
210 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
211 GFP_USER | __GFP_NOWARN);
215 for_each_possible_cpu(cpu) {
216 l = pcpu_freelist_pop(&htab->freelist);
217 /* pop will succeed, since prealloc_init()
218 * preallocated extra num_possible_cpus elements
220 l_new = container_of(l, struct htab_elem, fnode);
221 *per_cpu_ptr(pptr, cpu) = l_new;
223 htab->extra_elems = pptr;
227 /* Called from syscall */
228 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
230 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
231 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
232 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
233 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
234 /* percpu_lru means each cpu has its own LRU list.
235 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
236 * the map's value itself is percpu. percpu_lru has
237 * nothing to do with the map's value.
239 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
240 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
241 int numa_node = bpf_map_attr_numa_node(attr);
242 struct bpf_htab *htab;
246 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
247 offsetof(struct htab_elem, hash_node.pprev));
248 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
249 offsetof(struct htab_elem, hash_node.pprev));
251 if (lru && !capable(CAP_SYS_ADMIN))
252 /* LRU implementation is much complicated than other
253 * maps. Hence, limit to CAP_SYS_ADMIN for now.
255 return ERR_PTR(-EPERM);
257 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
258 /* reserved bits should not be used */
259 return ERR_PTR(-EINVAL);
261 if (!lru && percpu_lru)
262 return ERR_PTR(-EINVAL);
264 if (lru && !prealloc)
265 return ERR_PTR(-ENOTSUPP);
267 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
268 return ERR_PTR(-EINVAL);
270 htab = kzalloc(sizeof(*htab), GFP_USER);
272 return ERR_PTR(-ENOMEM);
274 /* mandatory map attributes */
275 htab->map.map_type = attr->map_type;
276 htab->map.key_size = attr->key_size;
277 htab->map.value_size = attr->value_size;
278 htab->map.max_entries = attr->max_entries;
279 htab->map.map_flags = attr->map_flags;
280 htab->map.numa_node = numa_node;
282 /* check sanity of attributes.
283 * value_size == 0 may be allowed in the future to use map as a set
286 if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
287 htab->map.value_size == 0)
291 /* ensure each CPU's lru list has >=1 elements.
292 * since we are at it, make each lru list has the same
293 * number of elements.
295 htab->map.max_entries = roundup(attr->max_entries,
296 num_possible_cpus());
297 if (htab->map.max_entries < attr->max_entries)
298 htab->map.max_entries = rounddown(attr->max_entries,
299 num_possible_cpus());
302 /* hash table size must be power of 2 */
303 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
306 if (htab->map.key_size > MAX_BPF_STACK)
307 /* eBPF programs initialize keys on stack, so they cannot be
308 * larger than max stack size
312 if (htab->map.value_size >= KMALLOC_MAX_SIZE -
313 MAX_BPF_STACK - sizeof(struct htab_elem))
314 /* if value_size is bigger, the user space won't be able to
315 * access the elements via bpf syscall. This check also makes
316 * sure that the elem_size doesn't overflow and it's
317 * kmalloc-able later in htab_map_update_elem()
321 htab->elem_size = sizeof(struct htab_elem) +
322 round_up(htab->map.key_size, 8);
324 htab->elem_size += sizeof(void *);
326 htab->elem_size += round_up(htab->map.value_size, 8);
328 /* prevent zero size kmalloc and check for u32 overflow */
329 if (htab->n_buckets == 0 ||
330 htab->n_buckets > U32_MAX / sizeof(struct bucket))
333 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
334 (u64) htab->elem_size * htab->map.max_entries;
337 cost += (u64) round_up(htab->map.value_size, 8) *
338 num_possible_cpus() * htab->map.max_entries;
340 cost += (u64) htab->elem_size * num_possible_cpus();
342 if (cost >= U32_MAX - PAGE_SIZE)
343 /* make sure page count doesn't overflow */
346 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
348 /* if map size is larger than memlock limit, reject it early */
349 err = bpf_map_precharge_memlock(htab->map.pages);
354 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
355 sizeof(struct bucket),
356 htab->map.numa_node);
360 for (i = 0; i < htab->n_buckets; i++) {
361 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
362 raw_spin_lock_init(&htab->buckets[i].lock);
366 err = prealloc_init(htab);
370 if (!percpu && !lru) {
371 /* lru itself can remove the least used element, so
372 * there is no need for an extra elem during map_update.
374 err = alloc_extra_elems(htab);
383 prealloc_destroy(htab);
385 bpf_map_area_free(htab->buckets);
391 static inline u32 htab_map_hash(const void *key, u32 key_len)
393 return jhash(key, key_len, 0);
396 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
398 return &htab->buckets[hash & (htab->n_buckets - 1)];
401 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
403 return &__select_bucket(htab, hash)->head;
406 /* this lookup function can only be called with bucket lock taken */
407 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
408 void *key, u32 key_size)
410 struct hlist_nulls_node *n;
413 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
414 if (l->hash == hash && !memcmp(&l->key, key, key_size))
420 /* can be called without bucket lock. it will repeat the loop in
421 * the unlikely event when elements moved from one bucket into another
422 * while link list is being walked
424 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
426 u32 key_size, u32 n_buckets)
428 struct hlist_nulls_node *n;
432 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
433 if (l->hash == hash && !memcmp(&l->key, key, key_size))
436 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
442 /* Called from syscall or from eBPF program directly, so
443 * arguments have to match bpf_map_lookup_elem() exactly.
444 * The return value is adjusted by BPF instructions
445 * in htab_map_gen_lookup().
447 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
449 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
450 struct hlist_nulls_head *head;
454 /* Must be called with rcu_read_lock. */
455 WARN_ON_ONCE(!rcu_read_lock_held());
457 key_size = map->key_size;
459 hash = htab_map_hash(key, key_size);
461 head = select_bucket(htab, hash);
463 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
468 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
470 struct htab_elem *l = __htab_map_lookup_elem(map, key);
473 return l->key + round_up(map->key_size, 8);
478 /* inline bpf_map_lookup_elem() call.
481 * bpf_map_lookup_elem
482 * map->ops->map_lookup_elem
483 * htab_map_lookup_elem
484 * __htab_map_lookup_elem
487 * __htab_map_lookup_elem
489 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
491 struct bpf_insn *insn = insn_buf;
492 const int ret = BPF_REG_0;
494 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
495 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
496 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
497 offsetof(struct htab_elem, key) +
498 round_up(map->key_size, 8));
499 return insn - insn_buf;
502 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
504 struct htab_elem *l = __htab_map_lookup_elem(map, key);
507 bpf_lru_node_set_ref(&l->lru_node);
508 return l->key + round_up(map->key_size, 8);
514 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
515 struct bpf_insn *insn_buf)
517 struct bpf_insn *insn = insn_buf;
518 const int ret = BPF_REG_0;
519 const int ref_reg = BPF_REG_1;
521 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
522 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
523 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
524 offsetof(struct htab_elem, lru_node) +
525 offsetof(struct bpf_lru_node, ref));
526 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
527 *insn++ = BPF_ST_MEM(BPF_B, ret,
528 offsetof(struct htab_elem, lru_node) +
529 offsetof(struct bpf_lru_node, ref),
531 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
532 offsetof(struct htab_elem, key) +
533 round_up(map->key_size, 8));
534 return insn - insn_buf;
537 /* It is called from the bpf_lru_list when the LRU needs to delete
538 * older elements from the htab.
540 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
542 struct bpf_htab *htab = (struct bpf_htab *)arg;
543 struct htab_elem *l = NULL, *tgt_l;
544 struct hlist_nulls_head *head;
545 struct hlist_nulls_node *n;
549 tgt_l = container_of(node, struct htab_elem, lru_node);
550 b = __select_bucket(htab, tgt_l->hash);
553 raw_spin_lock_irqsave(&b->lock, flags);
555 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
557 hlist_nulls_del_rcu(&l->hash_node);
561 raw_spin_unlock_irqrestore(&b->lock, flags);
566 /* Called from syscall */
567 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
569 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
570 struct hlist_nulls_head *head;
571 struct htab_elem *l, *next_l;
575 WARN_ON_ONCE(!rcu_read_lock_held());
577 key_size = map->key_size;
580 goto find_first_elem;
582 hash = htab_map_hash(key, key_size);
584 head = select_bucket(htab, hash);
587 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
590 goto find_first_elem;
592 /* key was found, get next key in the same bucket */
593 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
594 struct htab_elem, hash_node);
597 /* if next elem in this hash list is non-zero, just return it */
598 memcpy(next_key, next_l->key, key_size);
602 /* no more elements in this hash list, go to the next bucket */
603 i = hash & (htab->n_buckets - 1);
607 /* iterate over buckets */
608 for (; i < htab->n_buckets; i++) {
609 head = select_bucket(htab, i);
611 /* pick first element in the bucket */
612 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
613 struct htab_elem, hash_node);
615 /* if it's not empty, just return it */
616 memcpy(next_key, next_l->key, key_size);
621 /* iterated over all buckets and all elements */
625 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
627 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
628 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
632 static void htab_elem_free_rcu(struct rcu_head *head)
634 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
635 struct bpf_htab *htab = l->htab;
637 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
638 * we're calling kfree, otherwise deadlock is possible if kprobes
639 * are placed somewhere inside of slub
642 __this_cpu_inc(bpf_prog_active);
643 htab_elem_free(htab, l);
644 __this_cpu_dec(bpf_prog_active);
648 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
650 struct bpf_map *map = &htab->map;
652 if (map->ops->map_fd_put_ptr) {
653 void *ptr = fd_htab_map_get_ptr(map, l);
655 map->ops->map_fd_put_ptr(ptr);
658 if (htab_is_prealloc(htab)) {
659 pcpu_freelist_push(&htab->freelist, &l->fnode);
661 atomic_dec(&htab->count);
663 call_rcu(&l->rcu, htab_elem_free_rcu);
667 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
668 void *value, bool onallcpus)
671 /* copy true value_size bytes */
672 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
674 u32 size = round_up(htab->map.value_size, 8);
677 for_each_possible_cpu(cpu) {
678 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
685 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
687 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
691 static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
693 u32 size = htab->map.value_size;
695 if (percpu || fd_htab_map_needs_adjust(htab))
696 size = round_up(size, 8);
700 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
701 void *value, u32 key_size, u32 hash,
702 bool percpu, bool onallcpus,
703 struct htab_elem *old_elem)
705 u32 size = htab_size_value(htab, percpu);
706 bool prealloc = htab_is_prealloc(htab);
707 struct htab_elem *l_new, **pl_new;
712 /* if we're updating the existing element,
713 * use per-cpu extra elems to avoid freelist_pop/push
715 pl_new = this_cpu_ptr(htab->extra_elems);
719 struct pcpu_freelist_node *l;
721 l = pcpu_freelist_pop(&htab->freelist);
723 return ERR_PTR(-E2BIG);
724 l_new = container_of(l, struct htab_elem, fnode);
727 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
729 /* when map is full and update() is replacing
730 * old element, it's ok to allocate, since
731 * old element will be freed immediately.
732 * Otherwise return an error
734 atomic_dec(&htab->count);
735 return ERR_PTR(-E2BIG);
737 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
738 htab->map.numa_node);
740 return ERR_PTR(-ENOMEM);
743 memcpy(l_new->key, key, key_size);
746 pptr = htab_elem_get_ptr(l_new, key_size);
748 /* alloc_percpu zero-fills */
749 pptr = __alloc_percpu_gfp(size, 8,
750 GFP_ATOMIC | __GFP_NOWARN);
753 return ERR_PTR(-ENOMEM);
757 pcpu_copy_value(htab, pptr, value, onallcpus);
760 htab_elem_set_ptr(l_new, key_size, pptr);
762 memcpy(l_new->key + round_up(key_size, 8), value, size);
769 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
772 if (l_old && map_flags == BPF_NOEXIST)
773 /* elem already exists */
776 if (!l_old && map_flags == BPF_EXIST)
777 /* elem doesn't exist, cannot update it */
783 /* Called from syscall or from eBPF program */
784 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
787 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
788 struct htab_elem *l_new = NULL, *l_old;
789 struct hlist_nulls_head *head;
795 if (unlikely(map_flags > BPF_EXIST))
799 WARN_ON_ONCE(!rcu_read_lock_held());
801 key_size = map->key_size;
803 hash = htab_map_hash(key, key_size);
805 b = __select_bucket(htab, hash);
808 /* bpf_map_update_elem() can be called in_irq() */
809 raw_spin_lock_irqsave(&b->lock, flags);
811 l_old = lookup_elem_raw(head, hash, key, key_size);
813 ret = check_flags(htab, l_old, map_flags);
817 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
820 /* all pre-allocated elements are in use or memory exhausted */
821 ret = PTR_ERR(l_new);
825 /* add new element to the head of the list, so that
826 * concurrent search will find it before old elem
828 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
830 hlist_nulls_del_rcu(&l_old->hash_node);
831 if (!htab_is_prealloc(htab))
832 free_htab_elem(htab, l_old);
836 raw_spin_unlock_irqrestore(&b->lock, flags);
840 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
843 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
844 struct htab_elem *l_new, *l_old = NULL;
845 struct hlist_nulls_head *head;
851 if (unlikely(map_flags > BPF_EXIST))
855 WARN_ON_ONCE(!rcu_read_lock_held());
857 key_size = map->key_size;
859 hash = htab_map_hash(key, key_size);
861 b = __select_bucket(htab, hash);
864 /* For LRU, we need to alloc before taking bucket's
865 * spinlock because getting free nodes from LRU may need
866 * to remove older elements from htab and this removal
867 * operation will need a bucket lock.
869 l_new = prealloc_lru_pop(htab, key, hash);
872 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
874 /* bpf_map_update_elem() can be called in_irq() */
875 raw_spin_lock_irqsave(&b->lock, flags);
877 l_old = lookup_elem_raw(head, hash, key, key_size);
879 ret = check_flags(htab, l_old, map_flags);
883 /* add new element to the head of the list, so that
884 * concurrent search will find it before old elem
886 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
888 bpf_lru_node_set_ref(&l_new->lru_node);
889 hlist_nulls_del_rcu(&l_old->hash_node);
894 raw_spin_unlock_irqrestore(&b->lock, flags);
897 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
899 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
904 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
905 void *value, u64 map_flags,
908 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
909 struct htab_elem *l_new = NULL, *l_old;
910 struct hlist_nulls_head *head;
916 if (unlikely(map_flags > BPF_EXIST))
920 WARN_ON_ONCE(!rcu_read_lock_held());
922 key_size = map->key_size;
924 hash = htab_map_hash(key, key_size);
926 b = __select_bucket(htab, hash);
929 /* bpf_map_update_elem() can be called in_irq() */
930 raw_spin_lock_irqsave(&b->lock, flags);
932 l_old = lookup_elem_raw(head, hash, key, key_size);
934 ret = check_flags(htab, l_old, map_flags);
939 /* per-cpu hash map can update value in-place */
940 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
943 l_new = alloc_htab_elem(htab, key, value, key_size,
944 hash, true, onallcpus, NULL);
946 ret = PTR_ERR(l_new);
949 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
953 raw_spin_unlock_irqrestore(&b->lock, flags);
957 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
958 void *value, u64 map_flags,
961 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
962 struct htab_elem *l_new = NULL, *l_old;
963 struct hlist_nulls_head *head;
969 if (unlikely(map_flags > BPF_EXIST))
973 WARN_ON_ONCE(!rcu_read_lock_held());
975 key_size = map->key_size;
977 hash = htab_map_hash(key, key_size);
979 b = __select_bucket(htab, hash);
982 /* For LRU, we need to alloc before taking bucket's
983 * spinlock because LRU's elem alloc may need
984 * to remove older elem from htab and this removal
985 * operation will need a bucket lock.
987 if (map_flags != BPF_EXIST) {
988 l_new = prealloc_lru_pop(htab, key, hash);
993 /* bpf_map_update_elem() can be called in_irq() */
994 raw_spin_lock_irqsave(&b->lock, flags);
996 l_old = lookup_elem_raw(head, hash, key, key_size);
998 ret = check_flags(htab, l_old, map_flags);
1003 bpf_lru_node_set_ref(&l_old->lru_node);
1005 /* per-cpu hash map can update value in-place */
1006 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1009 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1011 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1016 raw_spin_unlock_irqrestore(&b->lock, flags);
1018 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1022 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1023 void *value, u64 map_flags)
1025 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1028 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1029 void *value, u64 map_flags)
1031 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1035 /* Called from syscall or from eBPF program */
1036 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1038 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1039 struct hlist_nulls_head *head;
1041 struct htab_elem *l;
1042 unsigned long flags;
1046 WARN_ON_ONCE(!rcu_read_lock_held());
1048 key_size = map->key_size;
1050 hash = htab_map_hash(key, key_size);
1051 b = __select_bucket(htab, hash);
1054 raw_spin_lock_irqsave(&b->lock, flags);
1056 l = lookup_elem_raw(head, hash, key, key_size);
1059 hlist_nulls_del_rcu(&l->hash_node);
1060 free_htab_elem(htab, l);
1064 raw_spin_unlock_irqrestore(&b->lock, flags);
1068 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1070 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1071 struct hlist_nulls_head *head;
1073 struct htab_elem *l;
1074 unsigned long flags;
1078 WARN_ON_ONCE(!rcu_read_lock_held());
1080 key_size = map->key_size;
1082 hash = htab_map_hash(key, key_size);
1083 b = __select_bucket(htab, hash);
1086 raw_spin_lock_irqsave(&b->lock, flags);
1088 l = lookup_elem_raw(head, hash, key, key_size);
1091 hlist_nulls_del_rcu(&l->hash_node);
1095 raw_spin_unlock_irqrestore(&b->lock, flags);
1097 bpf_lru_push_free(&htab->lru, &l->lru_node);
1101 static void delete_all_elements(struct bpf_htab *htab)
1105 for (i = 0; i < htab->n_buckets; i++) {
1106 struct hlist_nulls_head *head = select_bucket(htab, i);
1107 struct hlist_nulls_node *n;
1108 struct htab_elem *l;
1110 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1111 hlist_nulls_del_rcu(&l->hash_node);
1112 htab_elem_free(htab, l);
1117 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1118 static void htab_map_free(struct bpf_map *map)
1120 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1122 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1123 * so the programs (can be more than one that used this map) were
1124 * disconnected from events. Wait for outstanding critical sections in
1125 * these programs to complete
1129 /* some of free_htab_elem() callbacks for elements of this map may
1130 * not have executed. Wait for them.
1133 if (!htab_is_prealloc(htab))
1134 delete_all_elements(htab);
1136 prealloc_destroy(htab);
1138 free_percpu(htab->extra_elems);
1139 bpf_map_area_free(htab->buckets);
1143 const struct bpf_map_ops htab_map_ops = {
1144 .map_alloc = htab_map_alloc,
1145 .map_free = htab_map_free,
1146 .map_get_next_key = htab_map_get_next_key,
1147 .map_lookup_elem = htab_map_lookup_elem,
1148 .map_update_elem = htab_map_update_elem,
1149 .map_delete_elem = htab_map_delete_elem,
1150 .map_gen_lookup = htab_map_gen_lookup,
1153 const struct bpf_map_ops htab_lru_map_ops = {
1154 .map_alloc = htab_map_alloc,
1155 .map_free = htab_map_free,
1156 .map_get_next_key = htab_map_get_next_key,
1157 .map_lookup_elem = htab_lru_map_lookup_elem,
1158 .map_update_elem = htab_lru_map_update_elem,
1159 .map_delete_elem = htab_lru_map_delete_elem,
1160 .map_gen_lookup = htab_lru_map_gen_lookup,
1163 /* Called from eBPF program */
1164 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1166 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1169 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1174 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1176 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1179 bpf_lru_node_set_ref(&l->lru_node);
1180 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1186 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1188 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1189 struct htab_elem *l;
1190 void __percpu *pptr;
1195 /* per_cpu areas are zero-filled and bpf programs can only
1196 * access 'value_size' of them, so copying rounded areas
1197 * will not leak any kernel data
1199 size = round_up(map->value_size, 8);
1201 l = __htab_map_lookup_elem(map, key);
1204 if (htab_is_lru(htab))
1205 bpf_lru_node_set_ref(&l->lru_node);
1206 pptr = htab_elem_get_ptr(l, map->key_size);
1207 for_each_possible_cpu(cpu) {
1208 bpf_long_memcpy(value + off,
1209 per_cpu_ptr(pptr, cpu), size);
1218 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1221 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1225 if (htab_is_lru(htab))
1226 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1229 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1236 const struct bpf_map_ops htab_percpu_map_ops = {
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_percpu_map_lookup_elem,
1241 .map_update_elem = htab_percpu_map_update_elem,
1242 .map_delete_elem = htab_map_delete_elem,
1245 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1246 .map_alloc = htab_map_alloc,
1247 .map_free = htab_map_free,
1248 .map_get_next_key = htab_map_get_next_key,
1249 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1250 .map_update_elem = htab_lru_percpu_map_update_elem,
1251 .map_delete_elem = htab_lru_map_delete_elem,
1254 static struct bpf_map *fd_htab_map_alloc(union bpf_attr *attr)
1256 if (attr->value_size != sizeof(u32))
1257 return ERR_PTR(-EINVAL);
1258 return htab_map_alloc(attr);
1261 static void fd_htab_map_free(struct bpf_map *map)
1263 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1264 struct hlist_nulls_node *n;
1265 struct hlist_nulls_head *head;
1266 struct htab_elem *l;
1269 for (i = 0; i < htab->n_buckets; i++) {
1270 head = select_bucket(htab, i);
1272 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1273 void *ptr = fd_htab_map_get_ptr(map, l);
1275 map->ops->map_fd_put_ptr(ptr);
1282 /* only called from syscall */
1283 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1288 if (!map->ops->map_fd_sys_lookup_elem)
1292 ptr = htab_map_lookup_elem(map, key);
1294 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1302 /* only called from syscall */
1303 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1304 void *key, void *value, u64 map_flags)
1308 u32 ufd = *(u32 *)value;
1310 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1312 return PTR_ERR(ptr);
1314 ret = htab_map_update_elem(map, key, &ptr, map_flags);
1316 map->ops->map_fd_put_ptr(ptr);
1321 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1323 struct bpf_map *map, *inner_map_meta;
1325 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1326 if (IS_ERR(inner_map_meta))
1327 return inner_map_meta;
1329 map = fd_htab_map_alloc(attr);
1331 bpf_map_meta_free(inner_map_meta);
1335 map->inner_map_meta = inner_map_meta;
1340 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1342 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
1347 return READ_ONCE(*inner_map);
1350 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1351 struct bpf_insn *insn_buf)
1353 struct bpf_insn *insn = insn_buf;
1354 const int ret = BPF_REG_0;
1356 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
1357 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1358 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1359 offsetof(struct htab_elem, key) +
1360 round_up(map->key_size, 8));
1361 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1363 return insn - insn_buf;
1366 static void htab_of_map_free(struct bpf_map *map)
1368 bpf_map_meta_free(map->inner_map_meta);
1369 fd_htab_map_free(map);
1372 const struct bpf_map_ops htab_of_maps_map_ops = {
1373 .map_alloc = htab_of_map_alloc,
1374 .map_free = htab_of_map_free,
1375 .map_get_next_key = htab_map_get_next_key,
1376 .map_lookup_elem = htab_of_map_lookup_elem,
1377 .map_delete_elem = htab_map_delete_elem,
1378 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1379 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1380 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1381 .map_gen_lookup = htab_of_map_gen_lookup,