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 * 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(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 = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
298 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
304 if (htab_is_lru(htab))
305 err = bpf_lru_init(&htab->lru,
306 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
307 offsetof(struct htab_elem, hash) -
308 offsetof(struct htab_elem, lru_node),
309 htab_lru_map_delete_node,
312 err = pcpu_freelist_init(&htab->freelist);
317 if (htab_is_lru(htab))
318 bpf_lru_populate(&htab->lru, htab->elems,
319 offsetof(struct htab_elem, lru_node),
320 htab->elem_size, num_entries);
322 pcpu_freelist_populate(&htab->freelist,
323 htab->elems + offsetof(struct htab_elem, fnode),
324 htab->elem_size, num_entries);
329 htab_free_elems(htab);
333 static void prealloc_destroy(struct bpf_htab *htab)
335 htab_free_elems(htab);
337 if (htab_is_lru(htab))
338 bpf_lru_destroy(&htab->lru);
340 pcpu_freelist_destroy(&htab->freelist);
343 static int alloc_extra_elems(struct bpf_htab *htab)
345 struct htab_elem *__percpu *pptr, *l_new;
346 struct pcpu_freelist_node *l;
349 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
350 GFP_USER | __GFP_NOWARN);
354 for_each_possible_cpu(cpu) {
355 l = pcpu_freelist_pop(&htab->freelist);
356 /* pop will succeed, since prealloc_init()
357 * preallocated extra num_possible_cpus elements
359 l_new = container_of(l, struct htab_elem, fnode);
360 *per_cpu_ptr(pptr, cpu) = l_new;
362 htab->extra_elems = pptr;
366 /* Called from syscall */
367 static int htab_map_alloc_check(union bpf_attr *attr)
369 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
370 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
371 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
372 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
373 /* percpu_lru means each cpu has its own LRU list.
374 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
375 * the map's value itself is percpu. percpu_lru has
376 * nothing to do with the map's value.
378 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
379 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
380 bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
381 int numa_node = bpf_map_attr_numa_node(attr);
383 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
384 offsetof(struct htab_elem, hash_node.pprev));
385 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
386 offsetof(struct htab_elem, hash_node.pprev));
388 if (lru && !bpf_capable())
389 /* LRU implementation is much complicated than other
390 * maps. Hence, limit to CAP_BPF.
394 if (zero_seed && !capable(CAP_SYS_ADMIN))
395 /* Guard against local DoS, and discourage production use. */
398 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
399 !bpf_map_flags_access_ok(attr->map_flags))
402 if (!lru && percpu_lru)
405 if (lru && !prealloc)
408 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
411 /* check sanity of attributes.
412 * value_size == 0 may be allowed in the future to use map as a set
414 if (attr->max_entries == 0 || attr->key_size == 0 ||
415 attr->value_size == 0)
418 if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
419 sizeof(struct htab_elem))
420 /* if key_size + value_size is bigger, the user space won't be
421 * able to access the elements via bpf syscall. This check
422 * also makes sure that the elem_size doesn't overflow and it's
423 * kmalloc-able later in htab_map_update_elem()
430 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
432 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
433 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
434 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
435 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
436 /* percpu_lru means each cpu has its own LRU list.
437 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
438 * the map's value itself is percpu. percpu_lru has
439 * nothing to do with the map's value.
441 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
442 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
443 struct bpf_htab *htab;
447 htab = kzalloc(sizeof(*htab), GFP_USER);
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))
483 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
484 (u64) htab->elem_size * htab->map.max_entries;
487 cost += (u64) round_up(htab->map.value_size, 8) *
488 num_possible_cpus() * htab->map.max_entries;
490 cost += (u64) htab->elem_size * num_possible_cpus();
492 /* if map size is larger than memlock limit, reject it */
493 err = bpf_map_charge_init(&htab->map.memory, cost);
498 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
499 sizeof(struct bucket),
500 htab->map.numa_node);
504 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
505 htab->map_locked[i] = __alloc_percpu_gfp(sizeof(int),
506 sizeof(int), GFP_USER);
507 if (!htab->map_locked[i])
508 goto free_map_locked;
511 if (htab->map.map_flags & BPF_F_ZERO_SEED)
514 htab->hashrnd = get_random_int();
516 htab_init_buckets(htab);
519 err = prealloc_init(htab);
521 goto free_map_locked;
523 if (!percpu && !lru) {
524 /* lru itself can remove the least used element, so
525 * there is no need for an extra elem during map_update.
527 err = alloc_extra_elems(htab);
536 prealloc_destroy(htab);
538 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
539 free_percpu(htab->map_locked[i]);
540 bpf_map_area_free(htab->buckets);
542 bpf_map_charge_finish(&htab->map.memory);
544 lockdep_unregister_key(&htab->lockdep_key);
549 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
551 return jhash(key, key_len, hashrnd);
554 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
556 return &htab->buckets[hash & (htab->n_buckets - 1)];
559 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
561 return &__select_bucket(htab, hash)->head;
564 /* this lookup function can only be called with bucket lock taken */
565 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
566 void *key, u32 key_size)
568 struct hlist_nulls_node *n;
571 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
572 if (l->hash == hash && !memcmp(&l->key, key, key_size))
578 /* can be called without bucket lock. it will repeat the loop in
579 * the unlikely event when elements moved from one bucket into another
580 * while link list is being walked
582 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
584 u32 key_size, u32 n_buckets)
586 struct hlist_nulls_node *n;
590 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
591 if (l->hash == hash && !memcmp(&l->key, key, key_size))
594 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
600 /* Called from syscall or from eBPF program directly, so
601 * arguments have to match bpf_map_lookup_elem() exactly.
602 * The return value is adjusted by BPF instructions
603 * in htab_map_gen_lookup().
605 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
607 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
608 struct hlist_nulls_head *head;
612 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
614 key_size = map->key_size;
616 hash = htab_map_hash(key, key_size, htab->hashrnd);
618 head = select_bucket(htab, hash);
620 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
625 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
627 struct htab_elem *l = __htab_map_lookup_elem(map, key);
630 return l->key + round_up(map->key_size, 8);
635 /* inline bpf_map_lookup_elem() call.
638 * bpf_map_lookup_elem
639 * map->ops->map_lookup_elem
640 * htab_map_lookup_elem
641 * __htab_map_lookup_elem
644 * __htab_map_lookup_elem
646 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
648 struct bpf_insn *insn = insn_buf;
649 const int ret = BPF_REG_0;
651 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
652 (void *(*)(struct bpf_map *map, void *key))NULL));
653 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
654 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
655 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
656 offsetof(struct htab_elem, key) +
657 round_up(map->key_size, 8));
658 return insn - insn_buf;
661 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
662 void *key, const bool mark)
664 struct htab_elem *l = __htab_map_lookup_elem(map, key);
668 bpf_lru_node_set_ref(&l->lru_node);
669 return l->key + round_up(map->key_size, 8);
675 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
677 return __htab_lru_map_lookup_elem(map, key, true);
680 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
682 return __htab_lru_map_lookup_elem(map, key, false);
685 static int htab_lru_map_gen_lookup(struct bpf_map *map,
686 struct bpf_insn *insn_buf)
688 struct bpf_insn *insn = insn_buf;
689 const int ret = BPF_REG_0;
690 const int ref_reg = BPF_REG_1;
692 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
693 (void *(*)(struct bpf_map *map, void *key))NULL));
694 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
695 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
696 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
697 offsetof(struct htab_elem, lru_node) +
698 offsetof(struct bpf_lru_node, ref));
699 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
700 *insn++ = BPF_ST_MEM(BPF_B, ret,
701 offsetof(struct htab_elem, lru_node) +
702 offsetof(struct bpf_lru_node, ref),
704 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
705 offsetof(struct htab_elem, key) +
706 round_up(map->key_size, 8));
707 return insn - insn_buf;
710 /* It is called from the bpf_lru_list when the LRU needs to delete
711 * older elements from the htab.
713 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
715 struct bpf_htab *htab = (struct bpf_htab *)arg;
716 struct htab_elem *l = NULL, *tgt_l;
717 struct hlist_nulls_head *head;
718 struct hlist_nulls_node *n;
723 tgt_l = container_of(node, struct htab_elem, lru_node);
724 b = __select_bucket(htab, tgt_l->hash);
727 ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
731 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
733 hlist_nulls_del_rcu(&l->hash_node);
737 htab_unlock_bucket(htab, b, tgt_l->hash, flags);
742 /* Called from syscall */
743 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
745 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
746 struct hlist_nulls_head *head;
747 struct htab_elem *l, *next_l;
751 WARN_ON_ONCE(!rcu_read_lock_held());
753 key_size = map->key_size;
756 goto find_first_elem;
758 hash = htab_map_hash(key, key_size, htab->hashrnd);
760 head = select_bucket(htab, hash);
763 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
766 goto find_first_elem;
768 /* key was found, get next key in the same bucket */
769 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
770 struct htab_elem, hash_node);
773 /* if next elem in this hash list is non-zero, just return it */
774 memcpy(next_key, next_l->key, key_size);
778 /* no more elements in this hash list, go to the next bucket */
779 i = hash & (htab->n_buckets - 1);
783 /* iterate over buckets */
784 for (; i < htab->n_buckets; i++) {
785 head = select_bucket(htab, i);
787 /* pick first element in the bucket */
788 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
789 struct htab_elem, hash_node);
791 /* if it's not empty, just return it */
792 memcpy(next_key, next_l->key, key_size);
797 /* iterated over all buckets and all elements */
801 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
803 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
804 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
808 static void htab_elem_free_rcu(struct rcu_head *head)
810 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
811 struct bpf_htab *htab = l->htab;
813 htab_elem_free(htab, l);
816 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
818 struct bpf_map *map = &htab->map;
821 if (map->ops->map_fd_put_ptr) {
822 ptr = fd_htab_map_get_ptr(map, l);
823 map->ops->map_fd_put_ptr(ptr);
827 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
829 htab_put_fd_value(htab, l);
831 if (htab_is_prealloc(htab)) {
832 __pcpu_freelist_push(&htab->freelist, &l->fnode);
834 atomic_dec(&htab->count);
836 call_rcu(&l->rcu, htab_elem_free_rcu);
840 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
841 void *value, bool onallcpus)
844 /* copy true value_size bytes */
845 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
847 u32 size = round_up(htab->map.value_size, 8);
850 for_each_possible_cpu(cpu) {
851 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
858 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
859 void *value, bool onallcpus)
861 /* When using prealloc and not setting the initial value on all cpus,
862 * zero-fill element values for other cpus (just as what happens when
863 * not using prealloc). Otherwise, bpf program has no way to ensure
864 * known initial values for cpus other than current one
865 * (onallcpus=false always when coming from bpf prog).
867 if (htab_is_prealloc(htab) && !onallcpus) {
868 u32 size = round_up(htab->map.value_size, 8);
869 int current_cpu = raw_smp_processor_id();
872 for_each_possible_cpu(cpu) {
873 if (cpu == current_cpu)
874 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value,
877 memset(per_cpu_ptr(pptr, cpu), 0, size);
880 pcpu_copy_value(htab, pptr, value, onallcpus);
884 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
886 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
890 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
891 void *value, u32 key_size, u32 hash,
892 bool percpu, bool onallcpus,
893 struct htab_elem *old_elem)
895 u32 size = htab->map.value_size;
896 bool prealloc = htab_is_prealloc(htab);
897 struct htab_elem *l_new, **pl_new;
902 /* if we're updating the existing element,
903 * use per-cpu extra elems to avoid freelist_pop/push
905 pl_new = this_cpu_ptr(htab->extra_elems);
907 htab_put_fd_value(htab, old_elem);
910 struct pcpu_freelist_node *l;
912 l = __pcpu_freelist_pop(&htab->freelist);
914 return ERR_PTR(-E2BIG);
915 l_new = container_of(l, struct htab_elem, fnode);
918 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
920 /* when map is full and update() is replacing
921 * old element, it's ok to allocate, since
922 * old element will be freed immediately.
923 * Otherwise return an error
925 l_new = ERR_PTR(-E2BIG);
928 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
929 htab->map.numa_node);
931 l_new = ERR_PTR(-ENOMEM);
934 check_and_init_map_lock(&htab->map,
935 l_new->key + round_up(key_size, 8));
938 memcpy(l_new->key, key, key_size);
940 size = round_up(size, 8);
942 pptr = htab_elem_get_ptr(l_new, key_size);
944 /* alloc_percpu zero-fills */
945 pptr = __alloc_percpu_gfp(size, 8,
946 GFP_ATOMIC | __GFP_NOWARN);
949 l_new = ERR_PTR(-ENOMEM);
954 pcpu_init_value(htab, pptr, value, onallcpus);
957 htab_elem_set_ptr(l_new, key_size, pptr);
958 } else if (fd_htab_map_needs_adjust(htab)) {
959 size = round_up(size, 8);
960 memcpy(l_new->key + round_up(key_size, 8), value, size);
962 copy_map_value(&htab->map,
963 l_new->key + round_up(key_size, 8),
970 atomic_dec(&htab->count);
974 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
977 if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
978 /* elem already exists */
981 if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
982 /* elem doesn't exist, cannot update it */
988 /* Called from syscall or from eBPF program */
989 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
992 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
993 struct htab_elem *l_new = NULL, *l_old;
994 struct hlist_nulls_head *head;
1000 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1004 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1006 key_size = map->key_size;
1008 hash = htab_map_hash(key, key_size, htab->hashrnd);
1010 b = __select_bucket(htab, hash);
1013 if (unlikely(map_flags & BPF_F_LOCK)) {
1014 if (unlikely(!map_value_has_spin_lock(map)))
1016 /* find an element without taking the bucket lock */
1017 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1019 ret = check_flags(htab, l_old, map_flags);
1023 /* grab the element lock and update value in place */
1024 copy_map_value_locked(map,
1025 l_old->key + round_up(key_size, 8),
1029 /* fall through, grab the bucket lock and lookup again.
1030 * 99.9% chance that the element won't be found,
1031 * but second lookup under lock has to be done.
1035 ret = htab_lock_bucket(htab, b, hash, &flags);
1039 l_old = lookup_elem_raw(head, hash, key, key_size);
1041 ret = check_flags(htab, l_old, map_flags);
1045 if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1046 /* first lookup without the bucket lock didn't find the element,
1047 * but second lookup with the bucket lock found it.
1048 * This case is highly unlikely, but has to be dealt with:
1049 * grab the element lock in addition to the bucket lock
1050 * and update element in place
1052 copy_map_value_locked(map,
1053 l_old->key + round_up(key_size, 8),
1059 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1061 if (IS_ERR(l_new)) {
1062 /* all pre-allocated elements are in use or memory exhausted */
1063 ret = PTR_ERR(l_new);
1067 /* add new element to the head of the list, so that
1068 * concurrent search will find it before old elem
1070 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1072 hlist_nulls_del_rcu(&l_old->hash_node);
1073 if (!htab_is_prealloc(htab))
1074 free_htab_elem(htab, l_old);
1078 htab_unlock_bucket(htab, b, hash, flags);
1082 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1085 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1086 struct htab_elem *l_new, *l_old = NULL;
1087 struct hlist_nulls_head *head;
1088 unsigned long flags;
1093 if (unlikely(map_flags > BPF_EXIST))
1097 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1099 key_size = map->key_size;
1101 hash = htab_map_hash(key, key_size, htab->hashrnd);
1103 b = __select_bucket(htab, hash);
1106 /* For LRU, we need to alloc before taking bucket's
1107 * spinlock because getting free nodes from LRU may need
1108 * to remove older elements from htab and this removal
1109 * operation will need a bucket lock.
1111 l_new = prealloc_lru_pop(htab, key, hash);
1114 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
1116 ret = htab_lock_bucket(htab, b, hash, &flags);
1120 l_old = lookup_elem_raw(head, hash, key, key_size);
1122 ret = check_flags(htab, l_old, map_flags);
1126 /* add new element to the head of the list, so that
1127 * concurrent search will find it before old elem
1129 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1131 bpf_lru_node_set_ref(&l_new->lru_node);
1132 hlist_nulls_del_rcu(&l_old->hash_node);
1137 htab_unlock_bucket(htab, b, hash, flags);
1140 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1142 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
1147 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1148 void *value, u64 map_flags,
1151 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1152 struct htab_elem *l_new = NULL, *l_old;
1153 struct hlist_nulls_head *head;
1154 unsigned long flags;
1159 if (unlikely(map_flags > BPF_EXIST))
1163 WARN_ON_ONCE(!rcu_read_lock_held());
1165 key_size = map->key_size;
1167 hash = htab_map_hash(key, key_size, htab->hashrnd);
1169 b = __select_bucket(htab, hash);
1172 ret = htab_lock_bucket(htab, b, hash, &flags);
1176 l_old = lookup_elem_raw(head, hash, key, key_size);
1178 ret = check_flags(htab, l_old, map_flags);
1183 /* per-cpu hash map can update value in-place */
1184 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1187 l_new = alloc_htab_elem(htab, key, value, key_size,
1188 hash, true, onallcpus, NULL);
1189 if (IS_ERR(l_new)) {
1190 ret = PTR_ERR(l_new);
1193 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1197 htab_unlock_bucket(htab, b, hash, flags);
1201 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1202 void *value, u64 map_flags,
1205 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1206 struct htab_elem *l_new = NULL, *l_old;
1207 struct hlist_nulls_head *head;
1208 unsigned long flags;
1213 if (unlikely(map_flags > BPF_EXIST))
1217 WARN_ON_ONCE(!rcu_read_lock_held());
1219 key_size = map->key_size;
1221 hash = htab_map_hash(key, key_size, htab->hashrnd);
1223 b = __select_bucket(htab, hash);
1226 /* For LRU, we need to alloc before taking bucket's
1227 * spinlock because LRU's elem alloc may need
1228 * to remove older elem from htab and this removal
1229 * operation will need a bucket lock.
1231 if (map_flags != BPF_EXIST) {
1232 l_new = prealloc_lru_pop(htab, key, hash);
1237 ret = htab_lock_bucket(htab, b, hash, &flags);
1241 l_old = lookup_elem_raw(head, hash, key, key_size);
1243 ret = check_flags(htab, l_old, map_flags);
1248 bpf_lru_node_set_ref(&l_old->lru_node);
1250 /* per-cpu hash map can update value in-place */
1251 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1254 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1256 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1261 htab_unlock_bucket(htab, b, hash, flags);
1263 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1267 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1268 void *value, u64 map_flags)
1270 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1273 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1274 void *value, u64 map_flags)
1276 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1280 /* Called from syscall or from eBPF program */
1281 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1283 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1284 struct hlist_nulls_head *head;
1286 struct htab_elem *l;
1287 unsigned long flags;
1291 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1293 key_size = map->key_size;
1295 hash = htab_map_hash(key, key_size, htab->hashrnd);
1296 b = __select_bucket(htab, hash);
1299 ret = htab_lock_bucket(htab, b, hash, &flags);
1303 l = lookup_elem_raw(head, hash, key, key_size);
1306 hlist_nulls_del_rcu(&l->hash_node);
1307 free_htab_elem(htab, l);
1312 htab_unlock_bucket(htab, b, hash, flags);
1316 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1318 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1319 struct hlist_nulls_head *head;
1321 struct htab_elem *l;
1322 unsigned long flags;
1326 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1328 key_size = map->key_size;
1330 hash = htab_map_hash(key, key_size, htab->hashrnd);
1331 b = __select_bucket(htab, hash);
1334 ret = htab_lock_bucket(htab, b, hash, &flags);
1338 l = lookup_elem_raw(head, hash, key, key_size);
1341 hlist_nulls_del_rcu(&l->hash_node);
1345 htab_unlock_bucket(htab, b, hash, flags);
1347 bpf_lru_push_free(&htab->lru, &l->lru_node);
1351 static void delete_all_elements(struct bpf_htab *htab)
1355 for (i = 0; i < htab->n_buckets; i++) {
1356 struct hlist_nulls_head *head = select_bucket(htab, i);
1357 struct hlist_nulls_node *n;
1358 struct htab_elem *l;
1360 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1361 hlist_nulls_del_rcu(&l->hash_node);
1362 htab_elem_free(htab, l);
1367 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1368 static void htab_map_free(struct bpf_map *map)
1370 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1373 /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1374 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1375 * There is no need to synchronize_rcu() here to protect map elements.
1378 /* some of free_htab_elem() callbacks for elements of this map may
1379 * not have executed. Wait for them.
1382 if (!htab_is_prealloc(htab))
1383 delete_all_elements(htab);
1385 prealloc_destroy(htab);
1387 free_percpu(htab->extra_elems);
1388 bpf_map_area_free(htab->buckets);
1389 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1390 free_percpu(htab->map_locked[i]);
1391 lockdep_unregister_key(&htab->lockdep_key);
1395 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1402 value = htab_map_lookup_elem(map, key);
1408 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1410 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1417 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1418 const union bpf_attr *attr,
1419 union bpf_attr __user *uattr,
1420 bool do_delete, bool is_lru_map,
1423 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1424 u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1425 void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1426 void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1427 void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1428 void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1429 u32 batch, max_count, size, bucket_size;
1430 struct htab_elem *node_to_free = NULL;
1431 u64 elem_map_flags, map_flags;
1432 struct hlist_nulls_head *head;
1433 struct hlist_nulls_node *n;
1434 unsigned long flags = 0;
1435 bool locked = false;
1436 struct htab_elem *l;
1440 elem_map_flags = attr->batch.elem_flags;
1441 if ((elem_map_flags & ~BPF_F_LOCK) ||
1442 ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1445 map_flags = attr->batch.flags;
1449 max_count = attr->batch.count;
1453 if (put_user(0, &uattr->batch.count))
1457 if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1460 if (batch >= htab->n_buckets)
1463 key_size = htab->map.key_size;
1464 roundup_key_size = round_up(htab->map.key_size, 8);
1465 value_size = htab->map.value_size;
1466 size = round_up(value_size, 8);
1468 value_size = size * num_possible_cpus();
1470 /* while experimenting with hash tables with sizes ranging from 10 to
1471 * 1000, it was observed that a bucket can have upto 5 entries.
1476 /* We cannot do copy_from_user or copy_to_user inside
1477 * the rcu_read_lock. Allocate enough space here.
1479 keys = kvmalloc(key_size * bucket_size, GFP_USER | __GFP_NOWARN);
1480 values = kvmalloc(value_size * bucket_size, GFP_USER | __GFP_NOWARN);
1481 if (!keys || !values) {
1487 bpf_disable_instrumentation();
1492 b = &htab->buckets[batch];
1494 /* do not grab the lock unless need it (bucket_cnt > 0). */
1496 ret = htab_lock_bucket(htab, b, batch, &flags);
1502 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1505 if (bucket_cnt && !locked) {
1510 if (bucket_cnt > (max_count - total)) {
1513 /* Note that since bucket_cnt > 0 here, it is implicit
1514 * that the locked was grabbed, so release it.
1516 htab_unlock_bucket(htab, b, batch, flags);
1518 bpf_enable_instrumentation();
1522 if (bucket_cnt > bucket_size) {
1523 bucket_size = bucket_cnt;
1524 /* Note that since bucket_cnt > 0 here, it is implicit
1525 * that the locked was grabbed, so release it.
1527 htab_unlock_bucket(htab, b, batch, flags);
1529 bpf_enable_instrumentation();
1535 /* Next block is only safe to run if you have grabbed the lock */
1539 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1540 memcpy(dst_key, l->key, key_size);
1544 void __percpu *pptr;
1546 pptr = htab_elem_get_ptr(l, map->key_size);
1547 for_each_possible_cpu(cpu) {
1548 bpf_long_memcpy(dst_val + off,
1549 per_cpu_ptr(pptr, cpu), size);
1553 value = l->key + roundup_key_size;
1554 if (elem_map_flags & BPF_F_LOCK)
1555 copy_map_value_locked(map, dst_val, value,
1558 copy_map_value(map, dst_val, value);
1559 check_and_init_map_lock(map, dst_val);
1562 hlist_nulls_del_rcu(&l->hash_node);
1564 /* bpf_lru_push_free() will acquire lru_lock, which
1565 * may cause deadlock. See comments in function
1566 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1567 * after releasing the bucket lock.
1570 l->batch_flink = node_to_free;
1573 free_htab_elem(htab, l);
1576 dst_key += key_size;
1577 dst_val += value_size;
1580 htab_unlock_bucket(htab, b, batch, flags);
1583 while (node_to_free) {
1585 node_to_free = node_to_free->batch_flink;
1586 bpf_lru_push_free(&htab->lru, &l->lru_node);
1590 /* If we are not copying data, we can go to next bucket and avoid
1591 * unlocking the rcu.
1593 if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1599 bpf_enable_instrumentation();
1600 if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1601 key_size * bucket_cnt) ||
1602 copy_to_user(uvalues + total * value_size, values,
1603 value_size * bucket_cnt))) {
1608 total += bucket_cnt;
1610 if (batch >= htab->n_buckets) {
1620 /* copy # of entries and next batch */
1621 ubatch = u64_to_user_ptr(attr->batch.out_batch);
1622 if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1623 put_user(total, &uattr->batch.count))
1633 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1634 union bpf_attr __user *uattr)
1636 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1641 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1642 const union bpf_attr *attr,
1643 union bpf_attr __user *uattr)
1645 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1650 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1651 union bpf_attr __user *uattr)
1653 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1658 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1659 const union bpf_attr *attr,
1660 union bpf_attr __user *uattr)
1662 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1667 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1668 const union bpf_attr *attr,
1669 union bpf_attr __user *uattr)
1671 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1676 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1677 const union bpf_attr *attr,
1678 union bpf_attr __user *uattr)
1680 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1685 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1686 union bpf_attr __user *uattr)
1688 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1693 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1694 const union bpf_attr *attr,
1695 union bpf_attr __user *uattr)
1697 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1701 struct bpf_iter_seq_hash_map_info {
1702 struct bpf_map *map;
1703 struct bpf_htab *htab;
1704 void *percpu_value_buf; // non-zero means percpu hash
1709 static struct htab_elem *
1710 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1711 struct htab_elem *prev_elem)
1713 const struct bpf_htab *htab = info->htab;
1714 u32 skip_elems = info->skip_elems;
1715 u32 bucket_id = info->bucket_id;
1716 struct hlist_nulls_head *head;
1717 struct hlist_nulls_node *n;
1718 struct htab_elem *elem;
1722 if (bucket_id >= htab->n_buckets)
1725 /* try to find next elem in the same bucket */
1727 /* no update/deletion on this bucket, prev_elem should be still valid
1728 * and we won't skip elements.
1730 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1731 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1735 /* not found, unlock and go to the next bucket */
1736 b = &htab->buckets[bucket_id++];
1741 for (i = bucket_id; i < htab->n_buckets; i++) {
1742 b = &htab->buckets[i];
1747 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1748 if (count >= skip_elems) {
1749 info->bucket_id = i;
1750 info->skip_elems = count;
1760 info->bucket_id = i;
1761 info->skip_elems = 0;
1765 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
1767 struct bpf_iter_seq_hash_map_info *info = seq->private;
1768 struct htab_elem *elem;
1770 elem = bpf_hash_map_seq_find_next(info, NULL);
1779 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1781 struct bpf_iter_seq_hash_map_info *info = seq->private;
1785 return bpf_hash_map_seq_find_next(info, v);
1788 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
1790 struct bpf_iter_seq_hash_map_info *info = seq->private;
1791 u32 roundup_key_size, roundup_value_size;
1792 struct bpf_iter__bpf_map_elem ctx = {};
1793 struct bpf_map *map = info->map;
1794 struct bpf_iter_meta meta;
1795 int ret = 0, off = 0, cpu;
1796 struct bpf_prog *prog;
1797 void __percpu *pptr;
1800 prog = bpf_iter_get_info(&meta, elem == NULL);
1803 ctx.map = info->map;
1805 roundup_key_size = round_up(map->key_size, 8);
1806 ctx.key = elem->key;
1807 if (!info->percpu_value_buf) {
1808 ctx.value = elem->key + roundup_key_size;
1810 roundup_value_size = round_up(map->value_size, 8);
1811 pptr = htab_elem_get_ptr(elem, map->key_size);
1812 for_each_possible_cpu(cpu) {
1813 bpf_long_memcpy(info->percpu_value_buf + off,
1814 per_cpu_ptr(pptr, cpu),
1815 roundup_value_size);
1816 off += roundup_value_size;
1818 ctx.value = info->percpu_value_buf;
1821 ret = bpf_iter_run_prog(prog, &ctx);
1827 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
1829 return __bpf_hash_map_seq_show(seq, v);
1832 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
1835 (void)__bpf_hash_map_seq_show(seq, NULL);
1840 static int bpf_iter_init_hash_map(void *priv_data,
1841 struct bpf_iter_aux_info *aux)
1843 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1844 struct bpf_map *map = aux->map;
1848 if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
1849 map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
1850 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
1851 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
1855 seq_info->percpu_value_buf = value_buf;
1858 seq_info->map = map;
1859 seq_info->htab = container_of(map, struct bpf_htab, map);
1863 static void bpf_iter_fini_hash_map(void *priv_data)
1865 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1867 kfree(seq_info->percpu_value_buf);
1870 static const struct seq_operations bpf_hash_map_seq_ops = {
1871 .start = bpf_hash_map_seq_start,
1872 .next = bpf_hash_map_seq_next,
1873 .stop = bpf_hash_map_seq_stop,
1874 .show = bpf_hash_map_seq_show,
1877 static const struct bpf_iter_seq_info iter_seq_info = {
1878 .seq_ops = &bpf_hash_map_seq_ops,
1879 .init_seq_private = bpf_iter_init_hash_map,
1880 .fini_seq_private = bpf_iter_fini_hash_map,
1881 .seq_priv_size = sizeof(struct bpf_iter_seq_hash_map_info),
1884 static int htab_map_btf_id;
1885 const struct bpf_map_ops htab_map_ops = {
1886 .map_meta_equal = bpf_map_meta_equal,
1887 .map_alloc_check = htab_map_alloc_check,
1888 .map_alloc = htab_map_alloc,
1889 .map_free = htab_map_free,
1890 .map_get_next_key = htab_map_get_next_key,
1891 .map_lookup_elem = htab_map_lookup_elem,
1892 .map_update_elem = htab_map_update_elem,
1893 .map_delete_elem = htab_map_delete_elem,
1894 .map_gen_lookup = htab_map_gen_lookup,
1895 .map_seq_show_elem = htab_map_seq_show_elem,
1897 .map_btf_name = "bpf_htab",
1898 .map_btf_id = &htab_map_btf_id,
1899 .iter_seq_info = &iter_seq_info,
1902 static int htab_lru_map_btf_id;
1903 const struct bpf_map_ops htab_lru_map_ops = {
1904 .map_meta_equal = bpf_map_meta_equal,
1905 .map_alloc_check = htab_map_alloc_check,
1906 .map_alloc = htab_map_alloc,
1907 .map_free = htab_map_free,
1908 .map_get_next_key = htab_map_get_next_key,
1909 .map_lookup_elem = htab_lru_map_lookup_elem,
1910 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1911 .map_update_elem = htab_lru_map_update_elem,
1912 .map_delete_elem = htab_lru_map_delete_elem,
1913 .map_gen_lookup = htab_lru_map_gen_lookup,
1914 .map_seq_show_elem = htab_map_seq_show_elem,
1915 BATCH_OPS(htab_lru),
1916 .map_btf_name = "bpf_htab",
1917 .map_btf_id = &htab_lru_map_btf_id,
1918 .iter_seq_info = &iter_seq_info,
1921 /* Called from eBPF program */
1922 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1924 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1927 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1932 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1934 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1937 bpf_lru_node_set_ref(&l->lru_node);
1938 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1944 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1946 struct htab_elem *l;
1947 void __percpu *pptr;
1952 /* per_cpu areas are zero-filled and bpf programs can only
1953 * access 'value_size' of them, so copying rounded areas
1954 * will not leak any kernel data
1956 size = round_up(map->value_size, 8);
1958 l = __htab_map_lookup_elem(map, key);
1961 /* We do not mark LRU map element here in order to not mess up
1962 * eviction heuristics when user space does a map walk.
1964 pptr = htab_elem_get_ptr(l, map->key_size);
1965 for_each_possible_cpu(cpu) {
1966 bpf_long_memcpy(value + off,
1967 per_cpu_ptr(pptr, cpu), size);
1976 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1979 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1983 if (htab_is_lru(htab))
1984 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1987 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1994 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1997 struct htab_elem *l;
1998 void __percpu *pptr;
2003 l = __htab_map_lookup_elem(map, key);
2009 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2010 seq_puts(m, ": {\n");
2011 pptr = htab_elem_get_ptr(l, map->key_size);
2012 for_each_possible_cpu(cpu) {
2013 seq_printf(m, "\tcpu%d: ", cpu);
2014 btf_type_seq_show(map->btf, map->btf_value_type_id,
2015 per_cpu_ptr(pptr, cpu), m);
2023 static int htab_percpu_map_btf_id;
2024 const struct bpf_map_ops htab_percpu_map_ops = {
2025 .map_meta_equal = bpf_map_meta_equal,
2026 .map_alloc_check = htab_map_alloc_check,
2027 .map_alloc = htab_map_alloc,
2028 .map_free = htab_map_free,
2029 .map_get_next_key = htab_map_get_next_key,
2030 .map_lookup_elem = htab_percpu_map_lookup_elem,
2031 .map_update_elem = htab_percpu_map_update_elem,
2032 .map_delete_elem = htab_map_delete_elem,
2033 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2034 BATCH_OPS(htab_percpu),
2035 .map_btf_name = "bpf_htab",
2036 .map_btf_id = &htab_percpu_map_btf_id,
2037 .iter_seq_info = &iter_seq_info,
2040 static int htab_lru_percpu_map_btf_id;
2041 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2042 .map_meta_equal = bpf_map_meta_equal,
2043 .map_alloc_check = htab_map_alloc_check,
2044 .map_alloc = htab_map_alloc,
2045 .map_free = htab_map_free,
2046 .map_get_next_key = htab_map_get_next_key,
2047 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2048 .map_update_elem = htab_lru_percpu_map_update_elem,
2049 .map_delete_elem = htab_lru_map_delete_elem,
2050 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2051 BATCH_OPS(htab_lru_percpu),
2052 .map_btf_name = "bpf_htab",
2053 .map_btf_id = &htab_lru_percpu_map_btf_id,
2054 .iter_seq_info = &iter_seq_info,
2057 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2059 if (attr->value_size != sizeof(u32))
2061 return htab_map_alloc_check(attr);
2064 static void fd_htab_map_free(struct bpf_map *map)
2066 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2067 struct hlist_nulls_node *n;
2068 struct hlist_nulls_head *head;
2069 struct htab_elem *l;
2072 for (i = 0; i < htab->n_buckets; i++) {
2073 head = select_bucket(htab, i);
2075 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2076 void *ptr = fd_htab_map_get_ptr(map, l);
2078 map->ops->map_fd_put_ptr(ptr);
2085 /* only called from syscall */
2086 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2091 if (!map->ops->map_fd_sys_lookup_elem)
2095 ptr = htab_map_lookup_elem(map, key);
2097 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2105 /* only called from syscall */
2106 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2107 void *key, void *value, u64 map_flags)
2111 u32 ufd = *(u32 *)value;
2113 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2115 return PTR_ERR(ptr);
2117 ret = htab_map_update_elem(map, key, &ptr, map_flags);
2119 map->ops->map_fd_put_ptr(ptr);
2124 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2126 struct bpf_map *map, *inner_map_meta;
2128 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2129 if (IS_ERR(inner_map_meta))
2130 return inner_map_meta;
2132 map = htab_map_alloc(attr);
2134 bpf_map_meta_free(inner_map_meta);
2138 map->inner_map_meta = inner_map_meta;
2143 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2145 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
2150 return READ_ONCE(*inner_map);
2153 static int htab_of_map_gen_lookup(struct bpf_map *map,
2154 struct bpf_insn *insn_buf)
2156 struct bpf_insn *insn = insn_buf;
2157 const int ret = BPF_REG_0;
2159 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2160 (void *(*)(struct bpf_map *map, void *key))NULL));
2161 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
2162 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2163 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2164 offsetof(struct htab_elem, key) +
2165 round_up(map->key_size, 8));
2166 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2168 return insn - insn_buf;
2171 static void htab_of_map_free(struct bpf_map *map)
2173 bpf_map_meta_free(map->inner_map_meta);
2174 fd_htab_map_free(map);
2177 static int htab_of_maps_map_btf_id;
2178 const struct bpf_map_ops htab_of_maps_map_ops = {
2179 .map_alloc_check = fd_htab_map_alloc_check,
2180 .map_alloc = htab_of_map_alloc,
2181 .map_free = htab_of_map_free,
2182 .map_get_next_key = htab_map_get_next_key,
2183 .map_lookup_elem = htab_of_map_lookup_elem,
2184 .map_delete_elem = htab_map_delete_elem,
2185 .map_fd_get_ptr = bpf_map_fd_get_ptr,
2186 .map_fd_put_ptr = bpf_map_fd_put_ptr,
2187 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2188 .map_gen_lookup = htab_of_map_gen_lookup,
2189 .map_check_btf = map_check_no_btf,
2190 .map_btf_name = "bpf_htab",
2191 .map_btf_id = &htab_of_maps_map_btf_id,