1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
7 #include <linux/jhash.h>
8 #include <linux/filter.h>
9 #include <linux/rculist_nulls.h>
10 #include <linux/random.h>
11 #include <uapi/linux/btf.h>
12 #include <linux/rcupdate_trace.h>
13 #include "percpu_freelist.h"
14 #include "bpf_lru_list.h"
15 #include "map_in_map.h"
17 #define HTAB_CREATE_FLAG_MASK \
18 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
19 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
21 #define BATCH_OPS(_name) \
23 _name##_map_lookup_batch, \
24 .map_lookup_and_delete_batch = \
25 _name##_map_lookup_and_delete_batch, \
27 generic_map_update_batch, \
29 generic_map_delete_batch
32 * The bucket lock has two protection scopes:
34 * 1) Serializing concurrent operations from BPF programs on differrent
37 * 2) Serializing concurrent operations from BPF programs and sys_bpf()
39 * BPF programs can execute in any context including perf, kprobes and
40 * tracing. As there are almost no limits where perf, kprobes and tracing
41 * can be invoked from the lock operations need to be protected against
42 * deadlocks. Deadlocks can be caused by recursion and by an invocation in
43 * the lock held section when functions which acquire this lock are invoked
44 * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
45 * variable bpf_prog_active, which prevents BPF programs attached to perf
46 * events, kprobes and tracing to be invoked before the prior invocation
47 * from one of these contexts completed. sys_bpf() uses the same mechanism
48 * by pinning the task to the current CPU and incrementing the recursion
49 * protection accross the map operation.
51 * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
52 * operations like memory allocations (even with GFP_ATOMIC) from atomic
53 * contexts. This is required because even with GFP_ATOMIC the memory
54 * allocator calls into code pathes which acquire locks with long held lock
55 * sections. To ensure the deterministic behaviour these locks are regular
56 * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
57 * true atomic contexts on an RT kernel are the low level hardware
58 * handling, scheduling, low level interrupt handling, NMIs etc. None of
59 * these contexts should ever do memory allocations.
61 * As regular device interrupt handlers and soft interrupts are forced into
62 * thread context, the existing code which does
63 * spin_lock*(); alloc(GPF_ATOMIC); spin_unlock*();
66 * In theory the BPF locks could be converted to regular spinlocks as well,
67 * but the bucket locks and percpu_freelist locks can be taken from
68 * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
69 * atomic contexts even on RT. These mechanisms require preallocated maps,
70 * so there is no need to invoke memory allocations within the lock held
73 * BPF maps which need dynamic allocation are only used from (forced)
74 * thread context on RT and can therefore use regular spinlocks which in
75 * turn allows to invoke memory allocations from the lock held section.
77 * On a non RT kernel this distinction is neither possible nor required.
78 * spinlock maps to raw_spinlock and the extra code is optimized out by the
82 struct hlist_nulls_head head;
84 raw_spinlock_t raw_lock;
89 #define HASHTAB_MAP_LOCK_COUNT 8
90 #define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
94 struct bucket *buckets;
97 struct pcpu_freelist freelist;
100 struct htab_elem *__percpu *extra_elems;
101 atomic_t count; /* number of elements in this hashtable */
102 u32 n_buckets; /* number of hash buckets */
103 u32 elem_size; /* size of each element in bytes */
105 struct lock_class_key lockdep_key;
106 int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
109 /* each htab element is struct htab_elem + key + value */
112 struct hlist_nulls_node hash_node;
116 struct bpf_htab *htab;
117 struct pcpu_freelist_node fnode;
118 struct htab_elem *batch_flink;
124 struct bpf_lru_node lru_node;
127 char key[] __aligned(8);
130 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
132 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
135 static inline bool htab_use_raw_lock(const struct bpf_htab *htab)
137 return (!IS_ENABLED(CONFIG_PREEMPT_RT) || htab_is_prealloc(htab));
140 static void htab_init_buckets(struct bpf_htab *htab)
144 for (i = 0; i < htab->n_buckets; i++) {
145 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
146 if (htab_use_raw_lock(htab)) {
147 raw_spin_lock_init(&htab->buckets[i].raw_lock);
148 lockdep_set_class(&htab->buckets[i].raw_lock,
151 spin_lock_init(&htab->buckets[i].lock);
152 lockdep_set_class(&htab->buckets[i].lock,
158 static inline int htab_lock_bucket(const struct bpf_htab *htab,
159 struct bucket *b, u32 hash,
160 unsigned long *pflags)
164 hash = hash & HASHTAB_MAP_LOCK_MASK;
167 if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
168 __this_cpu_dec(*(htab->map_locked[hash]));
173 if (htab_use_raw_lock(htab))
174 raw_spin_lock_irqsave(&b->raw_lock, flags);
176 spin_lock_irqsave(&b->lock, flags);
182 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
183 struct bucket *b, u32 hash,
186 hash = hash & HASHTAB_MAP_LOCK_MASK;
187 if (htab_use_raw_lock(htab))
188 raw_spin_unlock_irqrestore(&b->raw_lock, flags);
190 spin_unlock_irqrestore(&b->lock, flags);
191 __this_cpu_dec(*(htab->map_locked[hash]));
195 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
197 static bool htab_is_lru(const struct bpf_htab *htab)
199 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
200 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
203 static bool htab_is_percpu(const struct bpf_htab *htab)
205 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
206 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
209 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
212 *(void __percpu **)(l->key + key_size) = pptr;
215 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
217 return *(void __percpu **)(l->key + key_size);
220 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
222 return *(void **)(l->key + roundup(map->key_size, 8));
225 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
227 return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
230 static void htab_free_elems(struct bpf_htab *htab)
234 if (!htab_is_percpu(htab))
237 for (i = 0; i < htab->map.max_entries; i++) {
240 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
246 bpf_map_area_free(htab->elems);
249 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
250 * (bucket_lock). If both locks need to be acquired together, the lock
251 * order is always lru_lock -> bucket_lock and this only happens in
252 * bpf_lru_list.c logic. For example, certain code path of
253 * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
254 * will acquire lru_lock first followed by acquiring bucket_lock.
256 * In hashtab.c, to avoid deadlock, lock acquisition of
257 * bucket_lock followed by lru_lock is not allowed. In such cases,
258 * bucket_lock needs to be released first before acquiring lru_lock.
260 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
263 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
267 l = container_of(node, struct htab_elem, lru_node);
268 memcpy(l->key, key, htab->map.key_size);
275 static int prealloc_init(struct bpf_htab *htab)
277 u32 num_entries = htab->map.max_entries;
278 int err = -ENOMEM, i;
280 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
281 num_entries += num_possible_cpus();
283 htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
284 htab->map.numa_node);
288 if (!htab_is_percpu(htab))
289 goto skip_percpu_elems;
291 for (i = 0; i < num_entries; i++) {
292 u32 size = round_up(htab->map.value_size, 8);
295 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
296 GFP_USER | __GFP_NOWARN);
299 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
305 if (htab_is_lru(htab))
306 err = bpf_lru_init(&htab->lru,
307 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
308 offsetof(struct htab_elem, hash) -
309 offsetof(struct htab_elem, lru_node),
310 htab_lru_map_delete_node,
313 err = pcpu_freelist_init(&htab->freelist);
318 if (htab_is_lru(htab))
319 bpf_lru_populate(&htab->lru, htab->elems,
320 offsetof(struct htab_elem, lru_node),
321 htab->elem_size, num_entries);
323 pcpu_freelist_populate(&htab->freelist,
324 htab->elems + offsetof(struct htab_elem, fnode),
325 htab->elem_size, num_entries);
330 htab_free_elems(htab);
334 static void prealloc_destroy(struct bpf_htab *htab)
336 htab_free_elems(htab);
338 if (htab_is_lru(htab))
339 bpf_lru_destroy(&htab->lru);
341 pcpu_freelist_destroy(&htab->freelist);
344 static int alloc_extra_elems(struct bpf_htab *htab)
346 struct htab_elem *__percpu *pptr, *l_new;
347 struct pcpu_freelist_node *l;
350 pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
351 GFP_USER | __GFP_NOWARN);
355 for_each_possible_cpu(cpu) {
356 l = pcpu_freelist_pop(&htab->freelist);
357 /* pop will succeed, since prealloc_init()
358 * preallocated extra num_possible_cpus elements
360 l_new = container_of(l, struct htab_elem, fnode);
361 *per_cpu_ptr(pptr, cpu) = l_new;
363 htab->extra_elems = pptr;
367 /* Called from syscall */
368 static int htab_map_alloc_check(union bpf_attr *attr)
370 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
371 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
372 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
373 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
374 /* percpu_lru means each cpu has its own LRU list.
375 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
376 * the map's value itself is percpu. percpu_lru has
377 * nothing to do with the map's value.
379 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
380 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
381 bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
382 int numa_node = bpf_map_attr_numa_node(attr);
384 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
385 offsetof(struct htab_elem, hash_node.pprev));
386 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
387 offsetof(struct htab_elem, hash_node.pprev));
389 if (lru && !bpf_capable())
390 /* LRU implementation is much complicated than other
391 * maps. Hence, limit to CAP_BPF.
395 if (zero_seed && !capable(CAP_SYS_ADMIN))
396 /* Guard against local DoS, and discourage production use. */
399 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
400 !bpf_map_flags_access_ok(attr->map_flags))
403 if (!lru && percpu_lru)
406 if (lru && !prealloc)
409 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
412 /* check sanity of attributes.
413 * value_size == 0 may be allowed in the future to use map as a set
415 if (attr->max_entries == 0 || attr->key_size == 0 ||
416 attr->value_size == 0)
419 if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
420 sizeof(struct htab_elem))
421 /* if key_size + value_size is bigger, the user space won't be
422 * able to access the elements via bpf syscall. This check
423 * also makes sure that the elem_size doesn't overflow and it's
424 * kmalloc-able later in htab_map_update_elem()
431 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
433 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
434 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
435 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
436 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
437 /* percpu_lru means each cpu has its own LRU list.
438 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
439 * the map's value itself is percpu. percpu_lru has
440 * nothing to do with the map's value.
442 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
443 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
444 struct bpf_htab *htab;
447 htab = kzalloc(sizeof(*htab), GFP_USER | __GFP_ACCOUNT);
449 return ERR_PTR(-ENOMEM);
451 lockdep_register_key(&htab->lockdep_key);
453 bpf_map_init_from_attr(&htab->map, attr);
456 /* ensure each CPU's lru list has >=1 elements.
457 * since we are at it, make each lru list has the same
458 * number of elements.
460 htab->map.max_entries = roundup(attr->max_entries,
461 num_possible_cpus());
462 if (htab->map.max_entries < attr->max_entries)
463 htab->map.max_entries = rounddown(attr->max_entries,
464 num_possible_cpus());
467 /* hash table size must be power of 2 */
468 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
470 htab->elem_size = sizeof(struct htab_elem) +
471 round_up(htab->map.key_size, 8);
473 htab->elem_size += sizeof(void *);
475 htab->elem_size += round_up(htab->map.value_size, 8);
478 /* prevent zero size kmalloc and check for u32 overflow */
479 if (htab->n_buckets == 0 ||
480 htab->n_buckets > U32_MAX / sizeof(struct bucket))
484 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
485 sizeof(struct bucket),
486 htab->map.numa_node);
490 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
491 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
495 if (!htab->map_locked[i])
496 goto free_map_locked;
499 if (htab->map.map_flags & BPF_F_ZERO_SEED)
502 htab->hashrnd = get_random_int();
504 htab_init_buckets(htab);
507 err = prealloc_init(htab);
509 goto free_map_locked;
511 if (!percpu && !lru) {
512 /* lru itself can remove the least used element, so
513 * there is no need for an extra elem during map_update.
515 err = alloc_extra_elems(htab);
524 prealloc_destroy(htab);
526 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
527 free_percpu(htab->map_locked[i]);
528 bpf_map_area_free(htab->buckets);
530 lockdep_unregister_key(&htab->lockdep_key);
535 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
537 return jhash(key, key_len, hashrnd);
540 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
542 return &htab->buckets[hash & (htab->n_buckets - 1)];
545 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
547 return &__select_bucket(htab, hash)->head;
550 /* this lookup function can only be called with bucket lock taken */
551 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
552 void *key, u32 key_size)
554 struct hlist_nulls_node *n;
557 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
558 if (l->hash == hash && !memcmp(&l->key, key, key_size))
564 /* can be called without bucket lock. it will repeat the loop in
565 * the unlikely event when elements moved from one bucket into another
566 * while link list is being walked
568 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
570 u32 key_size, u32 n_buckets)
572 struct hlist_nulls_node *n;
576 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
577 if (l->hash == hash && !memcmp(&l->key, key, key_size))
580 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
586 /* Called from syscall or from eBPF program directly, so
587 * arguments have to match bpf_map_lookup_elem() exactly.
588 * The return value is adjusted by BPF instructions
589 * in htab_map_gen_lookup().
591 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
593 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
594 struct hlist_nulls_head *head;
598 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
600 key_size = map->key_size;
602 hash = htab_map_hash(key, key_size, htab->hashrnd);
604 head = select_bucket(htab, hash);
606 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
611 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
613 struct htab_elem *l = __htab_map_lookup_elem(map, key);
616 return l->key + round_up(map->key_size, 8);
621 /* inline bpf_map_lookup_elem() call.
624 * bpf_map_lookup_elem
625 * map->ops->map_lookup_elem
626 * htab_map_lookup_elem
627 * __htab_map_lookup_elem
630 * __htab_map_lookup_elem
632 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
634 struct bpf_insn *insn = insn_buf;
635 const int ret = BPF_REG_0;
637 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
638 (void *(*)(struct bpf_map *map, void *key))NULL));
639 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
640 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
641 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
642 offsetof(struct htab_elem, key) +
643 round_up(map->key_size, 8));
644 return insn - insn_buf;
647 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
648 void *key, const bool mark)
650 struct htab_elem *l = __htab_map_lookup_elem(map, key);
654 bpf_lru_node_set_ref(&l->lru_node);
655 return l->key + round_up(map->key_size, 8);
661 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
663 return __htab_lru_map_lookup_elem(map, key, true);
666 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
668 return __htab_lru_map_lookup_elem(map, key, false);
671 static int htab_lru_map_gen_lookup(struct bpf_map *map,
672 struct bpf_insn *insn_buf)
674 struct bpf_insn *insn = insn_buf;
675 const int ret = BPF_REG_0;
676 const int ref_reg = BPF_REG_1;
678 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
679 (void *(*)(struct bpf_map *map, void *key))NULL));
680 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
681 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
682 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
683 offsetof(struct htab_elem, lru_node) +
684 offsetof(struct bpf_lru_node, ref));
685 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
686 *insn++ = BPF_ST_MEM(BPF_B, ret,
687 offsetof(struct htab_elem, lru_node) +
688 offsetof(struct bpf_lru_node, ref),
690 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
691 offsetof(struct htab_elem, key) +
692 round_up(map->key_size, 8));
693 return insn - insn_buf;
696 /* It is called from the bpf_lru_list when the LRU needs to delete
697 * older elements from the htab.
699 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
701 struct bpf_htab *htab = (struct bpf_htab *)arg;
702 struct htab_elem *l = NULL, *tgt_l;
703 struct hlist_nulls_head *head;
704 struct hlist_nulls_node *n;
709 tgt_l = container_of(node, struct htab_elem, lru_node);
710 b = __select_bucket(htab, tgt_l->hash);
713 ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
717 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
719 hlist_nulls_del_rcu(&l->hash_node);
723 htab_unlock_bucket(htab, b, tgt_l->hash, flags);
728 /* Called from syscall */
729 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
731 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
732 struct hlist_nulls_head *head;
733 struct htab_elem *l, *next_l;
737 WARN_ON_ONCE(!rcu_read_lock_held());
739 key_size = map->key_size;
742 goto find_first_elem;
744 hash = htab_map_hash(key, key_size, htab->hashrnd);
746 head = select_bucket(htab, hash);
749 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
752 goto find_first_elem;
754 /* key was found, get next key in the same bucket */
755 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
756 struct htab_elem, hash_node);
759 /* if next elem in this hash list is non-zero, just return it */
760 memcpy(next_key, next_l->key, key_size);
764 /* no more elements in this hash list, go to the next bucket */
765 i = hash & (htab->n_buckets - 1);
769 /* iterate over buckets */
770 for (; i < htab->n_buckets; i++) {
771 head = select_bucket(htab, i);
773 /* pick first element in the bucket */
774 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
775 struct htab_elem, hash_node);
777 /* if it's not empty, just return it */
778 memcpy(next_key, next_l->key, key_size);
783 /* iterated over all buckets and all elements */
787 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
789 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
790 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
794 static void htab_elem_free_rcu(struct rcu_head *head)
796 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
797 struct bpf_htab *htab = l->htab;
799 htab_elem_free(htab, l);
802 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
804 struct bpf_map *map = &htab->map;
807 if (map->ops->map_fd_put_ptr) {
808 ptr = fd_htab_map_get_ptr(map, l);
809 map->ops->map_fd_put_ptr(ptr);
813 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
815 htab_put_fd_value(htab, l);
817 if (htab_is_prealloc(htab)) {
818 __pcpu_freelist_push(&htab->freelist, &l->fnode);
820 atomic_dec(&htab->count);
822 call_rcu(&l->rcu, htab_elem_free_rcu);
826 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
827 void *value, bool onallcpus)
830 /* copy true value_size bytes */
831 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
833 u32 size = round_up(htab->map.value_size, 8);
836 for_each_possible_cpu(cpu) {
837 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
844 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
845 void *value, bool onallcpus)
847 /* When using prealloc and not setting the initial value on all cpus,
848 * zero-fill element values for other cpus (just as what happens when
849 * not using prealloc). Otherwise, bpf program has no way to ensure
850 * known initial values for cpus other than current one
851 * (onallcpus=false always when coming from bpf prog).
853 if (htab_is_prealloc(htab) && !onallcpus) {
854 u32 size = round_up(htab->map.value_size, 8);
855 int current_cpu = raw_smp_processor_id();
858 for_each_possible_cpu(cpu) {
859 if (cpu == current_cpu)
860 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value,
863 memset(per_cpu_ptr(pptr, cpu), 0, size);
866 pcpu_copy_value(htab, pptr, value, onallcpus);
870 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
872 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
876 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
877 void *value, u32 key_size, u32 hash,
878 bool percpu, bool onallcpus,
879 struct htab_elem *old_elem)
881 u32 size = htab->map.value_size;
882 bool prealloc = htab_is_prealloc(htab);
883 struct htab_elem *l_new, **pl_new;
888 /* if we're updating the existing element,
889 * use per-cpu extra elems to avoid freelist_pop/push
891 pl_new = this_cpu_ptr(htab->extra_elems);
893 htab_put_fd_value(htab, old_elem);
896 struct pcpu_freelist_node *l;
898 l = __pcpu_freelist_pop(&htab->freelist);
900 return ERR_PTR(-E2BIG);
901 l_new = container_of(l, struct htab_elem, fnode);
904 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
906 /* when map is full and update() is replacing
907 * old element, it's ok to allocate, since
908 * old element will be freed immediately.
909 * Otherwise return an error
911 l_new = ERR_PTR(-E2BIG);
914 l_new = bpf_map_kmalloc_node(&htab->map, htab->elem_size,
915 GFP_ATOMIC | __GFP_NOWARN,
916 htab->map.numa_node);
918 l_new = ERR_PTR(-ENOMEM);
921 check_and_init_map_lock(&htab->map,
922 l_new->key + round_up(key_size, 8));
925 memcpy(l_new->key, key, key_size);
927 size = round_up(size, 8);
929 pptr = htab_elem_get_ptr(l_new, key_size);
931 /* alloc_percpu zero-fills */
932 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
933 GFP_ATOMIC | __GFP_NOWARN);
936 l_new = ERR_PTR(-ENOMEM);
941 pcpu_init_value(htab, pptr, value, onallcpus);
944 htab_elem_set_ptr(l_new, key_size, pptr);
945 } else if (fd_htab_map_needs_adjust(htab)) {
946 size = round_up(size, 8);
947 memcpy(l_new->key + round_up(key_size, 8), value, size);
949 copy_map_value(&htab->map,
950 l_new->key + round_up(key_size, 8),
957 atomic_dec(&htab->count);
961 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
964 if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
965 /* elem already exists */
968 if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
969 /* elem doesn't exist, cannot update it */
975 /* Called from syscall or from eBPF program */
976 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
979 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
980 struct htab_elem *l_new = NULL, *l_old;
981 struct hlist_nulls_head *head;
987 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
991 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
993 key_size = map->key_size;
995 hash = htab_map_hash(key, key_size, htab->hashrnd);
997 b = __select_bucket(htab, hash);
1000 if (unlikely(map_flags & BPF_F_LOCK)) {
1001 if (unlikely(!map_value_has_spin_lock(map)))
1003 /* find an element without taking the bucket lock */
1004 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1006 ret = check_flags(htab, l_old, map_flags);
1010 /* grab the element lock and update value in place */
1011 copy_map_value_locked(map,
1012 l_old->key + round_up(key_size, 8),
1016 /* fall through, grab the bucket lock and lookup again.
1017 * 99.9% chance that the element won't be found,
1018 * but second lookup under lock has to be done.
1022 ret = htab_lock_bucket(htab, b, hash, &flags);
1026 l_old = lookup_elem_raw(head, hash, key, key_size);
1028 ret = check_flags(htab, l_old, map_flags);
1032 if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1033 /* first lookup without the bucket lock didn't find the element,
1034 * but second lookup with the bucket lock found it.
1035 * This case is highly unlikely, but has to be dealt with:
1036 * grab the element lock in addition to the bucket lock
1037 * and update element in place
1039 copy_map_value_locked(map,
1040 l_old->key + round_up(key_size, 8),
1046 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1048 if (IS_ERR(l_new)) {
1049 /* all pre-allocated elements are in use or memory exhausted */
1050 ret = PTR_ERR(l_new);
1054 /* add new element to the head of the list, so that
1055 * concurrent search will find it before old elem
1057 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1059 hlist_nulls_del_rcu(&l_old->hash_node);
1060 if (!htab_is_prealloc(htab))
1061 free_htab_elem(htab, l_old);
1065 htab_unlock_bucket(htab, b, hash, flags);
1069 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1072 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1073 struct htab_elem *l_new, *l_old = NULL;
1074 struct hlist_nulls_head *head;
1075 unsigned long flags;
1080 if (unlikely(map_flags > BPF_EXIST))
1084 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1086 key_size = map->key_size;
1088 hash = htab_map_hash(key, key_size, htab->hashrnd);
1090 b = __select_bucket(htab, hash);
1093 /* For LRU, we need to alloc before taking bucket's
1094 * spinlock because getting free nodes from LRU may need
1095 * to remove older elements from htab and this removal
1096 * operation will need a bucket lock.
1098 l_new = prealloc_lru_pop(htab, key, hash);
1101 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
1103 ret = htab_lock_bucket(htab, b, hash, &flags);
1107 l_old = lookup_elem_raw(head, hash, key, key_size);
1109 ret = check_flags(htab, l_old, map_flags);
1113 /* add new element to the head of the list, so that
1114 * concurrent search will find it before old elem
1116 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1118 bpf_lru_node_set_ref(&l_new->lru_node);
1119 hlist_nulls_del_rcu(&l_old->hash_node);
1124 htab_unlock_bucket(htab, b, hash, flags);
1127 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1129 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
1134 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1135 void *value, u64 map_flags,
1138 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1139 struct htab_elem *l_new = NULL, *l_old;
1140 struct hlist_nulls_head *head;
1141 unsigned long flags;
1146 if (unlikely(map_flags > BPF_EXIST))
1150 WARN_ON_ONCE(!rcu_read_lock_held());
1152 key_size = map->key_size;
1154 hash = htab_map_hash(key, key_size, htab->hashrnd);
1156 b = __select_bucket(htab, hash);
1159 ret = htab_lock_bucket(htab, b, hash, &flags);
1163 l_old = lookup_elem_raw(head, hash, key, key_size);
1165 ret = check_flags(htab, l_old, map_flags);
1170 /* per-cpu hash map can update value in-place */
1171 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1174 l_new = alloc_htab_elem(htab, key, value, key_size,
1175 hash, true, onallcpus, NULL);
1176 if (IS_ERR(l_new)) {
1177 ret = PTR_ERR(l_new);
1180 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1184 htab_unlock_bucket(htab, b, hash, flags);
1188 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1189 void *value, u64 map_flags,
1192 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1193 struct htab_elem *l_new = NULL, *l_old;
1194 struct hlist_nulls_head *head;
1195 unsigned long flags;
1200 if (unlikely(map_flags > BPF_EXIST))
1204 WARN_ON_ONCE(!rcu_read_lock_held());
1206 key_size = map->key_size;
1208 hash = htab_map_hash(key, key_size, htab->hashrnd);
1210 b = __select_bucket(htab, hash);
1213 /* For LRU, we need to alloc before taking bucket's
1214 * spinlock because LRU's elem alloc may need
1215 * to remove older elem from htab and this removal
1216 * operation will need a bucket lock.
1218 if (map_flags != BPF_EXIST) {
1219 l_new = prealloc_lru_pop(htab, key, hash);
1224 ret = htab_lock_bucket(htab, b, hash, &flags);
1228 l_old = lookup_elem_raw(head, hash, key, key_size);
1230 ret = check_flags(htab, l_old, map_flags);
1235 bpf_lru_node_set_ref(&l_old->lru_node);
1237 /* per-cpu hash map can update value in-place */
1238 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1241 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1243 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1248 htab_unlock_bucket(htab, b, hash, flags);
1250 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1254 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1255 void *value, u64 map_flags)
1257 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1260 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1261 void *value, u64 map_flags)
1263 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1267 /* Called from syscall or from eBPF program */
1268 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1270 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1271 struct hlist_nulls_head *head;
1273 struct htab_elem *l;
1274 unsigned long flags;
1278 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1280 key_size = map->key_size;
1282 hash = htab_map_hash(key, key_size, htab->hashrnd);
1283 b = __select_bucket(htab, hash);
1286 ret = htab_lock_bucket(htab, b, hash, &flags);
1290 l = lookup_elem_raw(head, hash, key, key_size);
1293 hlist_nulls_del_rcu(&l->hash_node);
1294 free_htab_elem(htab, l);
1299 htab_unlock_bucket(htab, b, hash, flags);
1303 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1305 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1306 struct hlist_nulls_head *head;
1308 struct htab_elem *l;
1309 unsigned long flags;
1313 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1315 key_size = map->key_size;
1317 hash = htab_map_hash(key, key_size, htab->hashrnd);
1318 b = __select_bucket(htab, hash);
1321 ret = htab_lock_bucket(htab, b, hash, &flags);
1325 l = lookup_elem_raw(head, hash, key, key_size);
1328 hlist_nulls_del_rcu(&l->hash_node);
1332 htab_unlock_bucket(htab, b, hash, flags);
1334 bpf_lru_push_free(&htab->lru, &l->lru_node);
1338 static void delete_all_elements(struct bpf_htab *htab)
1342 for (i = 0; i < htab->n_buckets; i++) {
1343 struct hlist_nulls_head *head = select_bucket(htab, i);
1344 struct hlist_nulls_node *n;
1345 struct htab_elem *l;
1347 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1348 hlist_nulls_del_rcu(&l->hash_node);
1349 htab_elem_free(htab, l);
1354 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1355 static void htab_map_free(struct bpf_map *map)
1357 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1360 /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1361 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1362 * There is no need to synchronize_rcu() here to protect map elements.
1365 /* some of free_htab_elem() callbacks for elements of this map may
1366 * not have executed. Wait for them.
1369 if (!htab_is_prealloc(htab))
1370 delete_all_elements(htab);
1372 prealloc_destroy(htab);
1374 free_percpu(htab->extra_elems);
1375 bpf_map_area_free(htab->buckets);
1376 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1377 free_percpu(htab->map_locked[i]);
1378 lockdep_unregister_key(&htab->lockdep_key);
1382 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1389 value = htab_map_lookup_elem(map, key);
1395 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1397 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1404 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1405 const union bpf_attr *attr,
1406 union bpf_attr __user *uattr,
1407 bool do_delete, bool is_lru_map,
1410 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1411 u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1412 void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1413 void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1414 void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1415 void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1416 u32 batch, max_count, size, bucket_size;
1417 struct htab_elem *node_to_free = NULL;
1418 u64 elem_map_flags, map_flags;
1419 struct hlist_nulls_head *head;
1420 struct hlist_nulls_node *n;
1421 unsigned long flags = 0;
1422 bool locked = false;
1423 struct htab_elem *l;
1427 elem_map_flags = attr->batch.elem_flags;
1428 if ((elem_map_flags & ~BPF_F_LOCK) ||
1429 ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1432 map_flags = attr->batch.flags;
1436 max_count = attr->batch.count;
1440 if (put_user(0, &uattr->batch.count))
1444 if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1447 if (batch >= htab->n_buckets)
1450 key_size = htab->map.key_size;
1451 roundup_key_size = round_up(htab->map.key_size, 8);
1452 value_size = htab->map.value_size;
1453 size = round_up(value_size, 8);
1455 value_size = size * num_possible_cpus();
1457 /* while experimenting with hash tables with sizes ranging from 10 to
1458 * 1000, it was observed that a bucket can have upto 5 entries.
1463 /* We cannot do copy_from_user or copy_to_user inside
1464 * the rcu_read_lock. Allocate enough space here.
1466 keys = kvmalloc(key_size * bucket_size, GFP_USER | __GFP_NOWARN);
1467 values = kvmalloc(value_size * bucket_size, GFP_USER | __GFP_NOWARN);
1468 if (!keys || !values) {
1474 bpf_disable_instrumentation();
1479 b = &htab->buckets[batch];
1481 /* do not grab the lock unless need it (bucket_cnt > 0). */
1483 ret = htab_lock_bucket(htab, b, batch, &flags);
1489 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1492 if (bucket_cnt && !locked) {
1497 if (bucket_cnt > (max_count - total)) {
1500 /* Note that since bucket_cnt > 0 here, it is implicit
1501 * that the locked was grabbed, so release it.
1503 htab_unlock_bucket(htab, b, batch, flags);
1505 bpf_enable_instrumentation();
1509 if (bucket_cnt > bucket_size) {
1510 bucket_size = bucket_cnt;
1511 /* Note that since bucket_cnt > 0 here, it is implicit
1512 * that the locked was grabbed, so release it.
1514 htab_unlock_bucket(htab, b, batch, flags);
1516 bpf_enable_instrumentation();
1522 /* Next block is only safe to run if you have grabbed the lock */
1526 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1527 memcpy(dst_key, l->key, key_size);
1531 void __percpu *pptr;
1533 pptr = htab_elem_get_ptr(l, map->key_size);
1534 for_each_possible_cpu(cpu) {
1535 bpf_long_memcpy(dst_val + off,
1536 per_cpu_ptr(pptr, cpu), size);
1540 value = l->key + roundup_key_size;
1541 if (elem_map_flags & BPF_F_LOCK)
1542 copy_map_value_locked(map, dst_val, value,
1545 copy_map_value(map, dst_val, value);
1546 check_and_init_map_lock(map, dst_val);
1549 hlist_nulls_del_rcu(&l->hash_node);
1551 /* bpf_lru_push_free() will acquire lru_lock, which
1552 * may cause deadlock. See comments in function
1553 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1554 * after releasing the bucket lock.
1557 l->batch_flink = node_to_free;
1560 free_htab_elem(htab, l);
1563 dst_key += key_size;
1564 dst_val += value_size;
1567 htab_unlock_bucket(htab, b, batch, flags);
1570 while (node_to_free) {
1572 node_to_free = node_to_free->batch_flink;
1573 bpf_lru_push_free(&htab->lru, &l->lru_node);
1577 /* If we are not copying data, we can go to next bucket and avoid
1578 * unlocking the rcu.
1580 if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1586 bpf_enable_instrumentation();
1587 if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1588 key_size * bucket_cnt) ||
1589 copy_to_user(uvalues + total * value_size, values,
1590 value_size * bucket_cnt))) {
1595 total += bucket_cnt;
1597 if (batch >= htab->n_buckets) {
1607 /* copy # of entries and next batch */
1608 ubatch = u64_to_user_ptr(attr->batch.out_batch);
1609 if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1610 put_user(total, &uattr->batch.count))
1620 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1621 union bpf_attr __user *uattr)
1623 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1628 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1629 const union bpf_attr *attr,
1630 union bpf_attr __user *uattr)
1632 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1637 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1638 union bpf_attr __user *uattr)
1640 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1645 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1646 const union bpf_attr *attr,
1647 union bpf_attr __user *uattr)
1649 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1654 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1655 const union bpf_attr *attr,
1656 union bpf_attr __user *uattr)
1658 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1663 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1664 const union bpf_attr *attr,
1665 union bpf_attr __user *uattr)
1667 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1672 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1673 union bpf_attr __user *uattr)
1675 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1680 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1681 const union bpf_attr *attr,
1682 union bpf_attr __user *uattr)
1684 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1688 struct bpf_iter_seq_hash_map_info {
1689 struct bpf_map *map;
1690 struct bpf_htab *htab;
1691 void *percpu_value_buf; // non-zero means percpu hash
1696 static struct htab_elem *
1697 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1698 struct htab_elem *prev_elem)
1700 const struct bpf_htab *htab = info->htab;
1701 u32 skip_elems = info->skip_elems;
1702 u32 bucket_id = info->bucket_id;
1703 struct hlist_nulls_head *head;
1704 struct hlist_nulls_node *n;
1705 struct htab_elem *elem;
1709 if (bucket_id >= htab->n_buckets)
1712 /* try to find next elem in the same bucket */
1714 /* no update/deletion on this bucket, prev_elem should be still valid
1715 * and we won't skip elements.
1717 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1718 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1722 /* not found, unlock and go to the next bucket */
1723 b = &htab->buckets[bucket_id++];
1728 for (i = bucket_id; i < htab->n_buckets; i++) {
1729 b = &htab->buckets[i];
1734 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1735 if (count >= skip_elems) {
1736 info->bucket_id = i;
1737 info->skip_elems = count;
1747 info->bucket_id = i;
1748 info->skip_elems = 0;
1752 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
1754 struct bpf_iter_seq_hash_map_info *info = seq->private;
1755 struct htab_elem *elem;
1757 elem = bpf_hash_map_seq_find_next(info, NULL);
1766 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1768 struct bpf_iter_seq_hash_map_info *info = seq->private;
1772 return bpf_hash_map_seq_find_next(info, v);
1775 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
1777 struct bpf_iter_seq_hash_map_info *info = seq->private;
1778 u32 roundup_key_size, roundup_value_size;
1779 struct bpf_iter__bpf_map_elem ctx = {};
1780 struct bpf_map *map = info->map;
1781 struct bpf_iter_meta meta;
1782 int ret = 0, off = 0, cpu;
1783 struct bpf_prog *prog;
1784 void __percpu *pptr;
1787 prog = bpf_iter_get_info(&meta, elem == NULL);
1790 ctx.map = info->map;
1792 roundup_key_size = round_up(map->key_size, 8);
1793 ctx.key = elem->key;
1794 if (!info->percpu_value_buf) {
1795 ctx.value = elem->key + roundup_key_size;
1797 roundup_value_size = round_up(map->value_size, 8);
1798 pptr = htab_elem_get_ptr(elem, map->key_size);
1799 for_each_possible_cpu(cpu) {
1800 bpf_long_memcpy(info->percpu_value_buf + off,
1801 per_cpu_ptr(pptr, cpu),
1802 roundup_value_size);
1803 off += roundup_value_size;
1805 ctx.value = info->percpu_value_buf;
1808 ret = bpf_iter_run_prog(prog, &ctx);
1814 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
1816 return __bpf_hash_map_seq_show(seq, v);
1819 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
1822 (void)__bpf_hash_map_seq_show(seq, NULL);
1827 static int bpf_iter_init_hash_map(void *priv_data,
1828 struct bpf_iter_aux_info *aux)
1830 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1831 struct bpf_map *map = aux->map;
1835 if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
1836 map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
1837 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
1838 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
1842 seq_info->percpu_value_buf = value_buf;
1845 seq_info->map = map;
1846 seq_info->htab = container_of(map, struct bpf_htab, map);
1850 static void bpf_iter_fini_hash_map(void *priv_data)
1852 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1854 kfree(seq_info->percpu_value_buf);
1857 static const struct seq_operations bpf_hash_map_seq_ops = {
1858 .start = bpf_hash_map_seq_start,
1859 .next = bpf_hash_map_seq_next,
1860 .stop = bpf_hash_map_seq_stop,
1861 .show = bpf_hash_map_seq_show,
1864 static const struct bpf_iter_seq_info iter_seq_info = {
1865 .seq_ops = &bpf_hash_map_seq_ops,
1866 .init_seq_private = bpf_iter_init_hash_map,
1867 .fini_seq_private = bpf_iter_fini_hash_map,
1868 .seq_priv_size = sizeof(struct bpf_iter_seq_hash_map_info),
1871 static int htab_map_btf_id;
1872 const struct bpf_map_ops htab_map_ops = {
1873 .map_meta_equal = bpf_map_meta_equal,
1874 .map_alloc_check = htab_map_alloc_check,
1875 .map_alloc = htab_map_alloc,
1876 .map_free = htab_map_free,
1877 .map_get_next_key = htab_map_get_next_key,
1878 .map_lookup_elem = htab_map_lookup_elem,
1879 .map_update_elem = htab_map_update_elem,
1880 .map_delete_elem = htab_map_delete_elem,
1881 .map_gen_lookup = htab_map_gen_lookup,
1882 .map_seq_show_elem = htab_map_seq_show_elem,
1884 .map_btf_name = "bpf_htab",
1885 .map_btf_id = &htab_map_btf_id,
1886 .iter_seq_info = &iter_seq_info,
1889 static int htab_lru_map_btf_id;
1890 const struct bpf_map_ops htab_lru_map_ops = {
1891 .map_meta_equal = bpf_map_meta_equal,
1892 .map_alloc_check = htab_map_alloc_check,
1893 .map_alloc = htab_map_alloc,
1894 .map_free = htab_map_free,
1895 .map_get_next_key = htab_map_get_next_key,
1896 .map_lookup_elem = htab_lru_map_lookup_elem,
1897 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1898 .map_update_elem = htab_lru_map_update_elem,
1899 .map_delete_elem = htab_lru_map_delete_elem,
1900 .map_gen_lookup = htab_lru_map_gen_lookup,
1901 .map_seq_show_elem = htab_map_seq_show_elem,
1902 BATCH_OPS(htab_lru),
1903 .map_btf_name = "bpf_htab",
1904 .map_btf_id = &htab_lru_map_btf_id,
1905 .iter_seq_info = &iter_seq_info,
1908 /* Called from eBPF program */
1909 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1911 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1914 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1919 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1921 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1924 bpf_lru_node_set_ref(&l->lru_node);
1925 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1931 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1933 struct htab_elem *l;
1934 void __percpu *pptr;
1939 /* per_cpu areas are zero-filled and bpf programs can only
1940 * access 'value_size' of them, so copying rounded areas
1941 * will not leak any kernel data
1943 size = round_up(map->value_size, 8);
1945 l = __htab_map_lookup_elem(map, key);
1948 /* We do not mark LRU map element here in order to not mess up
1949 * eviction heuristics when user space does a map walk.
1951 pptr = htab_elem_get_ptr(l, map->key_size);
1952 for_each_possible_cpu(cpu) {
1953 bpf_long_memcpy(value + off,
1954 per_cpu_ptr(pptr, cpu), size);
1963 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1966 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1970 if (htab_is_lru(htab))
1971 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1974 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1981 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1984 struct htab_elem *l;
1985 void __percpu *pptr;
1990 l = __htab_map_lookup_elem(map, key);
1996 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1997 seq_puts(m, ": {\n");
1998 pptr = htab_elem_get_ptr(l, map->key_size);
1999 for_each_possible_cpu(cpu) {
2000 seq_printf(m, "\tcpu%d: ", cpu);
2001 btf_type_seq_show(map->btf, map->btf_value_type_id,
2002 per_cpu_ptr(pptr, cpu), m);
2010 static int htab_percpu_map_btf_id;
2011 const struct bpf_map_ops htab_percpu_map_ops = {
2012 .map_meta_equal = bpf_map_meta_equal,
2013 .map_alloc_check = htab_map_alloc_check,
2014 .map_alloc = htab_map_alloc,
2015 .map_free = htab_map_free,
2016 .map_get_next_key = htab_map_get_next_key,
2017 .map_lookup_elem = htab_percpu_map_lookup_elem,
2018 .map_update_elem = htab_percpu_map_update_elem,
2019 .map_delete_elem = htab_map_delete_elem,
2020 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2021 BATCH_OPS(htab_percpu),
2022 .map_btf_name = "bpf_htab",
2023 .map_btf_id = &htab_percpu_map_btf_id,
2024 .iter_seq_info = &iter_seq_info,
2027 static int htab_lru_percpu_map_btf_id;
2028 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2029 .map_meta_equal = bpf_map_meta_equal,
2030 .map_alloc_check = htab_map_alloc_check,
2031 .map_alloc = htab_map_alloc,
2032 .map_free = htab_map_free,
2033 .map_get_next_key = htab_map_get_next_key,
2034 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2035 .map_update_elem = htab_lru_percpu_map_update_elem,
2036 .map_delete_elem = htab_lru_map_delete_elem,
2037 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2038 BATCH_OPS(htab_lru_percpu),
2039 .map_btf_name = "bpf_htab",
2040 .map_btf_id = &htab_lru_percpu_map_btf_id,
2041 .iter_seq_info = &iter_seq_info,
2044 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2046 if (attr->value_size != sizeof(u32))
2048 return htab_map_alloc_check(attr);
2051 static void fd_htab_map_free(struct bpf_map *map)
2053 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2054 struct hlist_nulls_node *n;
2055 struct hlist_nulls_head *head;
2056 struct htab_elem *l;
2059 for (i = 0; i < htab->n_buckets; i++) {
2060 head = select_bucket(htab, i);
2062 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2063 void *ptr = fd_htab_map_get_ptr(map, l);
2065 map->ops->map_fd_put_ptr(ptr);
2072 /* only called from syscall */
2073 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2078 if (!map->ops->map_fd_sys_lookup_elem)
2082 ptr = htab_map_lookup_elem(map, key);
2084 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2092 /* only called from syscall */
2093 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2094 void *key, void *value, u64 map_flags)
2098 u32 ufd = *(u32 *)value;
2100 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2102 return PTR_ERR(ptr);
2104 ret = htab_map_update_elem(map, key, &ptr, map_flags);
2106 map->ops->map_fd_put_ptr(ptr);
2111 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2113 struct bpf_map *map, *inner_map_meta;
2115 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2116 if (IS_ERR(inner_map_meta))
2117 return inner_map_meta;
2119 map = htab_map_alloc(attr);
2121 bpf_map_meta_free(inner_map_meta);
2125 map->inner_map_meta = inner_map_meta;
2130 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2132 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
2137 return READ_ONCE(*inner_map);
2140 static int htab_of_map_gen_lookup(struct bpf_map *map,
2141 struct bpf_insn *insn_buf)
2143 struct bpf_insn *insn = insn_buf;
2144 const int ret = BPF_REG_0;
2146 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2147 (void *(*)(struct bpf_map *map, void *key))NULL));
2148 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
2149 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2150 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2151 offsetof(struct htab_elem, key) +
2152 round_up(map->key_size, 8));
2153 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2155 return insn - insn_buf;
2158 static void htab_of_map_free(struct bpf_map *map)
2160 bpf_map_meta_free(map->inner_map_meta);
2161 fd_htab_map_free(map);
2164 static int htab_of_maps_map_btf_id;
2165 const struct bpf_map_ops htab_of_maps_map_ops = {
2166 .map_alloc_check = fd_htab_map_alloc_check,
2167 .map_alloc = htab_of_map_alloc,
2168 .map_free = htab_of_map_free,
2169 .map_get_next_key = htab_map_get_next_key,
2170 .map_lookup_elem = htab_of_map_lookup_elem,
2171 .map_delete_elem = htab_map_delete_elem,
2172 .map_fd_get_ptr = bpf_map_fd_get_ptr,
2173 .map_fd_put_ptr = bpf_map_fd_put_ptr,
2174 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2175 .map_gen_lookup = htab_of_map_gen_lookup,
2176 .map_check_btf = map_check_no_btf,
2177 .map_btf_name = "bpf_htab",
2178 .map_btf_id = &htab_of_maps_map_btf_id,