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 different
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 across 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 paths 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,
159 static inline int htab_lock_bucket(const struct bpf_htab *htab,
160 struct bucket *b, u32 hash,
161 unsigned long *pflags)
165 hash = hash & HASHTAB_MAP_LOCK_MASK;
168 if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
169 __this_cpu_dec(*(htab->map_locked[hash]));
174 if (htab_use_raw_lock(htab))
175 raw_spin_lock_irqsave(&b->raw_lock, flags);
177 spin_lock_irqsave(&b->lock, flags);
183 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
184 struct bucket *b, u32 hash,
187 hash = hash & HASHTAB_MAP_LOCK_MASK;
188 if (htab_use_raw_lock(htab))
189 raw_spin_unlock_irqrestore(&b->raw_lock, flags);
191 spin_unlock_irqrestore(&b->lock, flags);
192 __this_cpu_dec(*(htab->map_locked[hash]));
196 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
198 static bool htab_is_lru(const struct bpf_htab *htab)
200 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
201 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
204 static bool htab_is_percpu(const struct bpf_htab *htab)
206 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
207 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
210 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
213 *(void __percpu **)(l->key + key_size) = pptr;
216 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
218 return *(void __percpu **)(l->key + key_size);
221 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
223 return *(void **)(l->key + roundup(map->key_size, 8));
226 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
228 return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
231 static bool htab_has_extra_elems(struct bpf_htab *htab)
233 return !htab_is_percpu(htab) && !htab_is_lru(htab);
236 static void htab_free_prealloced_timers(struct bpf_htab *htab)
238 u32 num_entries = htab->map.max_entries;
241 if (likely(!map_value_has_timer(&htab->map)))
243 if (htab_has_extra_elems(htab))
244 num_entries += num_possible_cpus();
246 for (i = 0; i < num_entries; i++) {
247 struct htab_elem *elem;
249 elem = get_htab_elem(htab, i);
250 bpf_timer_cancel_and_free(elem->key +
251 round_up(htab->map.key_size, 8) +
252 htab->map.timer_off);
257 static void htab_free_elems(struct bpf_htab *htab)
261 if (!htab_is_percpu(htab))
264 for (i = 0; i < htab->map.max_entries; i++) {
267 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
273 bpf_map_area_free(htab->elems);
276 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
277 * (bucket_lock). If both locks need to be acquired together, the lock
278 * order is always lru_lock -> bucket_lock and this only happens in
279 * bpf_lru_list.c logic. For example, certain code path of
280 * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
281 * will acquire lru_lock first followed by acquiring bucket_lock.
283 * In hashtab.c, to avoid deadlock, lock acquisition of
284 * bucket_lock followed by lru_lock is not allowed. In such cases,
285 * bucket_lock needs to be released first before acquiring lru_lock.
287 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
290 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
294 u32 key_size = htab->map.key_size;
296 l = container_of(node, struct htab_elem, lru_node);
297 memcpy(l->key, key, key_size);
298 check_and_init_map_value(&htab->map,
299 l->key + round_up(key_size, 8));
306 static int prealloc_init(struct bpf_htab *htab)
308 u32 num_entries = htab->map.max_entries;
309 int err = -ENOMEM, i;
311 if (htab_has_extra_elems(htab))
312 num_entries += num_possible_cpus();
314 htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
315 htab->map.numa_node);
319 if (!htab_is_percpu(htab))
320 goto skip_percpu_elems;
322 for (i = 0; i < num_entries; i++) {
323 u32 size = round_up(htab->map.value_size, 8);
326 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
327 GFP_USER | __GFP_NOWARN);
330 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
336 if (htab_is_lru(htab))
337 err = bpf_lru_init(&htab->lru,
338 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
339 offsetof(struct htab_elem, hash) -
340 offsetof(struct htab_elem, lru_node),
341 htab_lru_map_delete_node,
344 err = pcpu_freelist_init(&htab->freelist);
349 if (htab_is_lru(htab))
350 bpf_lru_populate(&htab->lru, htab->elems,
351 offsetof(struct htab_elem, lru_node),
352 htab->elem_size, num_entries);
354 pcpu_freelist_populate(&htab->freelist,
355 htab->elems + offsetof(struct htab_elem, fnode),
356 htab->elem_size, num_entries);
361 htab_free_elems(htab);
365 static void prealloc_destroy(struct bpf_htab *htab)
367 htab_free_elems(htab);
369 if (htab_is_lru(htab))
370 bpf_lru_destroy(&htab->lru);
372 pcpu_freelist_destroy(&htab->freelist);
375 static int alloc_extra_elems(struct bpf_htab *htab)
377 struct htab_elem *__percpu *pptr, *l_new;
378 struct pcpu_freelist_node *l;
381 pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
382 GFP_USER | __GFP_NOWARN);
386 for_each_possible_cpu(cpu) {
387 l = pcpu_freelist_pop(&htab->freelist);
388 /* pop will succeed, since prealloc_init()
389 * preallocated extra num_possible_cpus elements
391 l_new = container_of(l, struct htab_elem, fnode);
392 *per_cpu_ptr(pptr, cpu) = l_new;
394 htab->extra_elems = pptr;
398 /* Called from syscall */
399 static int htab_map_alloc_check(union bpf_attr *attr)
401 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
402 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
403 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
404 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
405 /* percpu_lru means each cpu has its own LRU list.
406 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
407 * the map's value itself is percpu. percpu_lru has
408 * nothing to do with the map's value.
410 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
411 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
412 bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
413 int numa_node = bpf_map_attr_numa_node(attr);
415 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
416 offsetof(struct htab_elem, hash_node.pprev));
417 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
418 offsetof(struct htab_elem, hash_node.pprev));
420 if (lru && !bpf_capable())
421 /* LRU implementation is much complicated than other
422 * maps. Hence, limit to CAP_BPF.
426 if (zero_seed && !capable(CAP_SYS_ADMIN))
427 /* Guard against local DoS, and discourage production use. */
430 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
431 !bpf_map_flags_access_ok(attr->map_flags))
434 if (!lru && percpu_lru)
437 if (lru && !prealloc)
440 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
443 /* check sanity of attributes.
444 * value_size == 0 may be allowed in the future to use map as a set
446 if (attr->max_entries == 0 || attr->key_size == 0 ||
447 attr->value_size == 0)
450 if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
451 sizeof(struct htab_elem))
452 /* if key_size + value_size is bigger, the user space won't be
453 * able to access the elements via bpf syscall. This check
454 * also makes sure that the elem_size doesn't overflow and it's
455 * kmalloc-able later in htab_map_update_elem()
462 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
464 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
465 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
466 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
467 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
468 /* percpu_lru means each cpu has its own LRU list.
469 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
470 * the map's value itself is percpu. percpu_lru has
471 * nothing to do with the map's value.
473 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
474 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
475 struct bpf_htab *htab;
478 htab = kzalloc(sizeof(*htab), GFP_USER | __GFP_ACCOUNT);
480 return ERR_PTR(-ENOMEM);
482 lockdep_register_key(&htab->lockdep_key);
484 bpf_map_init_from_attr(&htab->map, attr);
487 /* ensure each CPU's lru list has >=1 elements.
488 * since we are at it, make each lru list has the same
489 * number of elements.
491 htab->map.max_entries = roundup(attr->max_entries,
492 num_possible_cpus());
493 if (htab->map.max_entries < attr->max_entries)
494 htab->map.max_entries = rounddown(attr->max_entries,
495 num_possible_cpus());
498 /* hash table size must be power of 2 */
499 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
501 htab->elem_size = sizeof(struct htab_elem) +
502 round_up(htab->map.key_size, 8);
504 htab->elem_size += sizeof(void *);
506 htab->elem_size += round_up(htab->map.value_size, 8);
509 /* prevent zero size kmalloc and check for u32 overflow */
510 if (htab->n_buckets == 0 ||
511 htab->n_buckets > U32_MAX / sizeof(struct bucket))
515 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
516 sizeof(struct bucket),
517 htab->map.numa_node);
521 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
522 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
526 if (!htab->map_locked[i])
527 goto free_map_locked;
530 if (htab->map.map_flags & BPF_F_ZERO_SEED)
533 htab->hashrnd = get_random_int();
535 htab_init_buckets(htab);
538 err = prealloc_init(htab);
540 goto free_map_locked;
542 if (!percpu && !lru) {
543 /* lru itself can remove the least used element, so
544 * there is no need for an extra elem during map_update.
546 err = alloc_extra_elems(htab);
555 prealloc_destroy(htab);
557 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
558 free_percpu(htab->map_locked[i]);
559 bpf_map_area_free(htab->buckets);
561 lockdep_unregister_key(&htab->lockdep_key);
566 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
568 return jhash(key, key_len, hashrnd);
571 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
573 return &htab->buckets[hash & (htab->n_buckets - 1)];
576 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
578 return &__select_bucket(htab, hash)->head;
581 /* this lookup function can only be called with bucket lock taken */
582 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
583 void *key, u32 key_size)
585 struct hlist_nulls_node *n;
588 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
589 if (l->hash == hash && !memcmp(&l->key, key, key_size))
595 /* can be called without bucket lock. it will repeat the loop in
596 * the unlikely event when elements moved from one bucket into another
597 * while link list is being walked
599 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
601 u32 key_size, u32 n_buckets)
603 struct hlist_nulls_node *n;
607 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
608 if (l->hash == hash && !memcmp(&l->key, key, key_size))
611 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
617 /* Called from syscall or from eBPF program directly, so
618 * arguments have to match bpf_map_lookup_elem() exactly.
619 * The return value is adjusted by BPF instructions
620 * in htab_map_gen_lookup().
622 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
624 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
625 struct hlist_nulls_head *head;
629 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
630 !rcu_read_lock_bh_held());
632 key_size = map->key_size;
634 hash = htab_map_hash(key, key_size, htab->hashrnd);
636 head = select_bucket(htab, hash);
638 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
643 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
645 struct htab_elem *l = __htab_map_lookup_elem(map, key);
648 return l->key + round_up(map->key_size, 8);
653 /* inline bpf_map_lookup_elem() call.
656 * bpf_map_lookup_elem
657 * map->ops->map_lookup_elem
658 * htab_map_lookup_elem
659 * __htab_map_lookup_elem
662 * __htab_map_lookup_elem
664 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
666 struct bpf_insn *insn = insn_buf;
667 const int ret = BPF_REG_0;
669 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
670 (void *(*)(struct bpf_map *map, void *key))NULL));
671 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
672 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
673 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
674 offsetof(struct htab_elem, key) +
675 round_up(map->key_size, 8));
676 return insn - insn_buf;
679 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
680 void *key, const bool mark)
682 struct htab_elem *l = __htab_map_lookup_elem(map, key);
686 bpf_lru_node_set_ref(&l->lru_node);
687 return l->key + round_up(map->key_size, 8);
693 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
695 return __htab_lru_map_lookup_elem(map, key, true);
698 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
700 return __htab_lru_map_lookup_elem(map, key, false);
703 static int htab_lru_map_gen_lookup(struct bpf_map *map,
704 struct bpf_insn *insn_buf)
706 struct bpf_insn *insn = insn_buf;
707 const int ret = BPF_REG_0;
708 const int ref_reg = BPF_REG_1;
710 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
711 (void *(*)(struct bpf_map *map, void *key))NULL));
712 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
713 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
714 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
715 offsetof(struct htab_elem, lru_node) +
716 offsetof(struct bpf_lru_node, ref));
717 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
718 *insn++ = BPF_ST_MEM(BPF_B, ret,
719 offsetof(struct htab_elem, lru_node) +
720 offsetof(struct bpf_lru_node, ref),
722 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
723 offsetof(struct htab_elem, key) +
724 round_up(map->key_size, 8));
725 return insn - insn_buf;
728 static void check_and_free_timer(struct bpf_htab *htab, struct htab_elem *elem)
730 if (unlikely(map_value_has_timer(&htab->map)))
731 bpf_timer_cancel_and_free(elem->key +
732 round_up(htab->map.key_size, 8) +
733 htab->map.timer_off);
736 /* It is called from the bpf_lru_list when the LRU needs to delete
737 * older elements from the htab.
739 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
741 struct bpf_htab *htab = (struct bpf_htab *)arg;
742 struct htab_elem *l = NULL, *tgt_l;
743 struct hlist_nulls_head *head;
744 struct hlist_nulls_node *n;
749 tgt_l = container_of(node, struct htab_elem, lru_node);
750 b = __select_bucket(htab, tgt_l->hash);
753 ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
757 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
759 hlist_nulls_del_rcu(&l->hash_node);
760 check_and_free_timer(htab, l);
764 htab_unlock_bucket(htab, b, tgt_l->hash, flags);
769 /* Called from syscall */
770 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
772 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
773 struct hlist_nulls_head *head;
774 struct htab_elem *l, *next_l;
778 WARN_ON_ONCE(!rcu_read_lock_held());
780 key_size = map->key_size;
783 goto find_first_elem;
785 hash = htab_map_hash(key, key_size, htab->hashrnd);
787 head = select_bucket(htab, hash);
790 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
793 goto find_first_elem;
795 /* key was found, get next key in the same bucket */
796 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
797 struct htab_elem, hash_node);
800 /* if next elem in this hash list is non-zero, just return it */
801 memcpy(next_key, next_l->key, key_size);
805 /* no more elements in this hash list, go to the next bucket */
806 i = hash & (htab->n_buckets - 1);
810 /* iterate over buckets */
811 for (; i < htab->n_buckets; i++) {
812 head = select_bucket(htab, i);
814 /* pick first element in the bucket */
815 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
816 struct htab_elem, hash_node);
818 /* if it's not empty, just return it */
819 memcpy(next_key, next_l->key, key_size);
824 /* iterated over all buckets and all elements */
828 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
830 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
831 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
832 check_and_free_timer(htab, l);
836 static void htab_elem_free_rcu(struct rcu_head *head)
838 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
839 struct bpf_htab *htab = l->htab;
841 htab_elem_free(htab, l);
844 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
846 struct bpf_map *map = &htab->map;
849 if (map->ops->map_fd_put_ptr) {
850 ptr = fd_htab_map_get_ptr(map, l);
851 map->ops->map_fd_put_ptr(ptr);
855 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
857 htab_put_fd_value(htab, l);
859 if (htab_is_prealloc(htab)) {
860 check_and_free_timer(htab, l);
861 __pcpu_freelist_push(&htab->freelist, &l->fnode);
863 atomic_dec(&htab->count);
865 call_rcu(&l->rcu, htab_elem_free_rcu);
869 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
870 void *value, bool onallcpus)
873 /* copy true value_size bytes */
874 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
876 u32 size = round_up(htab->map.value_size, 8);
879 for_each_possible_cpu(cpu) {
880 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
887 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
888 void *value, bool onallcpus)
890 /* When using prealloc and not setting the initial value on all cpus,
891 * zero-fill element values for other cpus (just as what happens when
892 * not using prealloc). Otherwise, bpf program has no way to ensure
893 * known initial values for cpus other than current one
894 * (onallcpus=false always when coming from bpf prog).
896 if (htab_is_prealloc(htab) && !onallcpus) {
897 u32 size = round_up(htab->map.value_size, 8);
898 int current_cpu = raw_smp_processor_id();
901 for_each_possible_cpu(cpu) {
902 if (cpu == current_cpu)
903 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value,
906 memset(per_cpu_ptr(pptr, cpu), 0, size);
909 pcpu_copy_value(htab, pptr, value, onallcpus);
913 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
915 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
919 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
920 void *value, u32 key_size, u32 hash,
921 bool percpu, bool onallcpus,
922 struct htab_elem *old_elem)
924 u32 size = htab->map.value_size;
925 bool prealloc = htab_is_prealloc(htab);
926 struct htab_elem *l_new, **pl_new;
931 /* if we're updating the existing element,
932 * use per-cpu extra elems to avoid freelist_pop/push
934 pl_new = this_cpu_ptr(htab->extra_elems);
936 htab_put_fd_value(htab, old_elem);
939 struct pcpu_freelist_node *l;
941 l = __pcpu_freelist_pop(&htab->freelist);
943 return ERR_PTR(-E2BIG);
944 l_new = container_of(l, struct htab_elem, fnode);
947 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
949 /* when map is full and update() is replacing
950 * old element, it's ok to allocate, since
951 * old element will be freed immediately.
952 * Otherwise return an error
954 l_new = ERR_PTR(-E2BIG);
957 l_new = bpf_map_kmalloc_node(&htab->map, htab->elem_size,
958 GFP_ATOMIC | __GFP_NOWARN,
959 htab->map.numa_node);
961 l_new = ERR_PTR(-ENOMEM);
964 check_and_init_map_value(&htab->map,
965 l_new->key + round_up(key_size, 8));
968 memcpy(l_new->key, key, key_size);
970 size = round_up(size, 8);
972 pptr = htab_elem_get_ptr(l_new, key_size);
974 /* alloc_percpu zero-fills */
975 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
976 GFP_ATOMIC | __GFP_NOWARN);
979 l_new = ERR_PTR(-ENOMEM);
984 pcpu_init_value(htab, pptr, value, onallcpus);
987 htab_elem_set_ptr(l_new, key_size, pptr);
988 } else if (fd_htab_map_needs_adjust(htab)) {
989 size = round_up(size, 8);
990 memcpy(l_new->key + round_up(key_size, 8), value, size);
992 copy_map_value(&htab->map,
993 l_new->key + round_up(key_size, 8),
1000 atomic_dec(&htab->count);
1004 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1007 if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1008 /* elem already exists */
1011 if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1012 /* elem doesn't exist, cannot update it */
1018 /* Called from syscall or from eBPF program */
1019 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1022 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1023 struct htab_elem *l_new = NULL, *l_old;
1024 struct hlist_nulls_head *head;
1025 unsigned long flags;
1030 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1034 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1035 !rcu_read_lock_bh_held());
1037 key_size = map->key_size;
1039 hash = htab_map_hash(key, key_size, htab->hashrnd);
1041 b = __select_bucket(htab, hash);
1044 if (unlikely(map_flags & BPF_F_LOCK)) {
1045 if (unlikely(!map_value_has_spin_lock(map)))
1047 /* find an element without taking the bucket lock */
1048 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1050 ret = check_flags(htab, l_old, map_flags);
1054 /* grab the element lock and update value in place */
1055 copy_map_value_locked(map,
1056 l_old->key + round_up(key_size, 8),
1060 /* fall through, grab the bucket lock and lookup again.
1061 * 99.9% chance that the element won't be found,
1062 * but second lookup under lock has to be done.
1066 ret = htab_lock_bucket(htab, b, hash, &flags);
1070 l_old = lookup_elem_raw(head, hash, key, key_size);
1072 ret = check_flags(htab, l_old, map_flags);
1076 if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1077 /* first lookup without the bucket lock didn't find the element,
1078 * but second lookup with the bucket lock found it.
1079 * This case is highly unlikely, but has to be dealt with:
1080 * grab the element lock in addition to the bucket lock
1081 * and update element in place
1083 copy_map_value_locked(map,
1084 l_old->key + round_up(key_size, 8),
1090 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1092 if (IS_ERR(l_new)) {
1093 /* all pre-allocated elements are in use or memory exhausted */
1094 ret = PTR_ERR(l_new);
1098 /* add new element to the head of the list, so that
1099 * concurrent search will find it before old elem
1101 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1103 hlist_nulls_del_rcu(&l_old->hash_node);
1104 if (!htab_is_prealloc(htab))
1105 free_htab_elem(htab, l_old);
1107 check_and_free_timer(htab, l_old);
1111 htab_unlock_bucket(htab, b, hash, flags);
1115 static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1117 check_and_free_timer(htab, elem);
1118 bpf_lru_push_free(&htab->lru, &elem->lru_node);
1121 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1124 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1125 struct htab_elem *l_new, *l_old = NULL;
1126 struct hlist_nulls_head *head;
1127 unsigned long flags;
1132 if (unlikely(map_flags > BPF_EXIST))
1136 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1137 !rcu_read_lock_bh_held());
1139 key_size = map->key_size;
1141 hash = htab_map_hash(key, key_size, htab->hashrnd);
1143 b = __select_bucket(htab, hash);
1146 /* For LRU, we need to alloc before taking bucket's
1147 * spinlock because getting free nodes from LRU may need
1148 * to remove older elements from htab and this removal
1149 * operation will need a bucket lock.
1151 l_new = prealloc_lru_pop(htab, key, hash);
1154 copy_map_value(&htab->map,
1155 l_new->key + round_up(map->key_size, 8), value);
1157 ret = htab_lock_bucket(htab, b, hash, &flags);
1161 l_old = lookup_elem_raw(head, hash, key, key_size);
1163 ret = check_flags(htab, l_old, map_flags);
1167 /* add new element to the head of the list, so that
1168 * concurrent search will find it before old elem
1170 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1172 bpf_lru_node_set_ref(&l_new->lru_node);
1173 hlist_nulls_del_rcu(&l_old->hash_node);
1178 htab_unlock_bucket(htab, b, hash, flags);
1181 htab_lru_push_free(htab, l_new);
1183 htab_lru_push_free(htab, l_old);
1188 static int __htab_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() && !rcu_read_lock_trace_held() &&
1205 !rcu_read_lock_bh_held());
1207 key_size = map->key_size;
1209 hash = htab_map_hash(key, key_size, htab->hashrnd);
1211 b = __select_bucket(htab, hash);
1214 ret = htab_lock_bucket(htab, b, hash, &flags);
1218 l_old = lookup_elem_raw(head, hash, key, key_size);
1220 ret = check_flags(htab, l_old, map_flags);
1225 /* per-cpu hash map can update value in-place */
1226 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1229 l_new = alloc_htab_elem(htab, key, value, key_size,
1230 hash, true, onallcpus, NULL);
1231 if (IS_ERR(l_new)) {
1232 ret = PTR_ERR(l_new);
1235 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1239 htab_unlock_bucket(htab, b, hash, flags);
1243 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1244 void *value, u64 map_flags,
1247 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1248 struct htab_elem *l_new = NULL, *l_old;
1249 struct hlist_nulls_head *head;
1250 unsigned long flags;
1255 if (unlikely(map_flags > BPF_EXIST))
1259 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1260 !rcu_read_lock_bh_held());
1262 key_size = map->key_size;
1264 hash = htab_map_hash(key, key_size, htab->hashrnd);
1266 b = __select_bucket(htab, hash);
1269 /* For LRU, we need to alloc before taking bucket's
1270 * spinlock because LRU's elem alloc may need
1271 * to remove older elem from htab and this removal
1272 * operation will need a bucket lock.
1274 if (map_flags != BPF_EXIST) {
1275 l_new = prealloc_lru_pop(htab, key, hash);
1280 ret = htab_lock_bucket(htab, b, hash, &flags);
1284 l_old = lookup_elem_raw(head, hash, key, key_size);
1286 ret = check_flags(htab, l_old, map_flags);
1291 bpf_lru_node_set_ref(&l_old->lru_node);
1293 /* per-cpu hash map can update value in-place */
1294 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1297 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1299 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1304 htab_unlock_bucket(htab, b, hash, flags);
1306 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1310 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1311 void *value, u64 map_flags)
1313 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1316 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1317 void *value, u64 map_flags)
1319 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1323 /* Called from syscall or from eBPF program */
1324 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1326 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1327 struct hlist_nulls_head *head;
1329 struct htab_elem *l;
1330 unsigned long flags;
1334 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1335 !rcu_read_lock_bh_held());
1337 key_size = map->key_size;
1339 hash = htab_map_hash(key, key_size, htab->hashrnd);
1340 b = __select_bucket(htab, hash);
1343 ret = htab_lock_bucket(htab, b, hash, &flags);
1347 l = lookup_elem_raw(head, hash, key, key_size);
1350 hlist_nulls_del_rcu(&l->hash_node);
1351 free_htab_elem(htab, l);
1356 htab_unlock_bucket(htab, b, hash, flags);
1360 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1362 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1363 struct hlist_nulls_head *head;
1365 struct htab_elem *l;
1366 unsigned long flags;
1370 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1371 !rcu_read_lock_bh_held());
1373 key_size = map->key_size;
1375 hash = htab_map_hash(key, key_size, htab->hashrnd);
1376 b = __select_bucket(htab, hash);
1379 ret = htab_lock_bucket(htab, b, hash, &flags);
1383 l = lookup_elem_raw(head, hash, key, key_size);
1386 hlist_nulls_del_rcu(&l->hash_node);
1390 htab_unlock_bucket(htab, b, hash, flags);
1392 htab_lru_push_free(htab, l);
1396 static void delete_all_elements(struct bpf_htab *htab)
1400 for (i = 0; i < htab->n_buckets; i++) {
1401 struct hlist_nulls_head *head = select_bucket(htab, i);
1402 struct hlist_nulls_node *n;
1403 struct htab_elem *l;
1405 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1406 hlist_nulls_del_rcu(&l->hash_node);
1407 htab_elem_free(htab, l);
1412 static void htab_free_malloced_timers(struct bpf_htab *htab)
1417 for (i = 0; i < htab->n_buckets; i++) {
1418 struct hlist_nulls_head *head = select_bucket(htab, i);
1419 struct hlist_nulls_node *n;
1420 struct htab_elem *l;
1422 hlist_nulls_for_each_entry(l, n, head, hash_node)
1423 check_and_free_timer(htab, l);
1429 static void htab_map_free_timers(struct bpf_map *map)
1431 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1433 if (likely(!map_value_has_timer(&htab->map)))
1435 if (!htab_is_prealloc(htab))
1436 htab_free_malloced_timers(htab);
1438 htab_free_prealloced_timers(htab);
1441 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1442 static void htab_map_free(struct bpf_map *map)
1444 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1447 /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1448 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1449 * There is no need to synchronize_rcu() here to protect map elements.
1452 /* some of free_htab_elem() callbacks for elements of this map may
1453 * not have executed. Wait for them.
1456 if (!htab_is_prealloc(htab))
1457 delete_all_elements(htab);
1459 prealloc_destroy(htab);
1461 free_percpu(htab->extra_elems);
1462 bpf_map_area_free(htab->buckets);
1463 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1464 free_percpu(htab->map_locked[i]);
1465 lockdep_unregister_key(&htab->lockdep_key);
1469 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1476 value = htab_map_lookup_elem(map, key);
1482 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1484 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1490 static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1491 void *value, bool is_lru_map,
1492 bool is_percpu, u64 flags)
1494 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1495 struct hlist_nulls_head *head;
1496 unsigned long bflags;
1497 struct htab_elem *l;
1502 key_size = map->key_size;
1504 hash = htab_map_hash(key, key_size, htab->hashrnd);
1505 b = __select_bucket(htab, hash);
1508 ret = htab_lock_bucket(htab, b, hash, &bflags);
1512 l = lookup_elem_raw(head, hash, key, key_size);
1517 u32 roundup_value_size = round_up(map->value_size, 8);
1518 void __percpu *pptr;
1521 pptr = htab_elem_get_ptr(l, key_size);
1522 for_each_possible_cpu(cpu) {
1523 bpf_long_memcpy(value + off,
1524 per_cpu_ptr(pptr, cpu),
1525 roundup_value_size);
1526 off += roundup_value_size;
1529 u32 roundup_key_size = round_up(map->key_size, 8);
1531 if (flags & BPF_F_LOCK)
1532 copy_map_value_locked(map, value, l->key +
1536 copy_map_value(map, value, l->key +
1538 check_and_init_map_value(map, value);
1541 hlist_nulls_del_rcu(&l->hash_node);
1543 free_htab_elem(htab, l);
1546 htab_unlock_bucket(htab, b, hash, bflags);
1548 if (is_lru_map && l)
1549 htab_lru_push_free(htab, l);
1554 static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1555 void *value, u64 flags)
1557 return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1561 static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1562 void *key, void *value,
1565 return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1569 static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1570 void *value, u64 flags)
1572 return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1576 static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1577 void *key, void *value,
1580 return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1585 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1586 const union bpf_attr *attr,
1587 union bpf_attr __user *uattr,
1588 bool do_delete, bool is_lru_map,
1591 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1592 u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1593 void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1594 void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1595 void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1596 void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1597 u32 batch, max_count, size, bucket_size;
1598 struct htab_elem *node_to_free = NULL;
1599 u64 elem_map_flags, map_flags;
1600 struct hlist_nulls_head *head;
1601 struct hlist_nulls_node *n;
1602 unsigned long flags = 0;
1603 bool locked = false;
1604 struct htab_elem *l;
1608 elem_map_flags = attr->batch.elem_flags;
1609 if ((elem_map_flags & ~BPF_F_LOCK) ||
1610 ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1613 map_flags = attr->batch.flags;
1617 max_count = attr->batch.count;
1621 if (put_user(0, &uattr->batch.count))
1625 if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1628 if (batch >= htab->n_buckets)
1631 key_size = htab->map.key_size;
1632 roundup_key_size = round_up(htab->map.key_size, 8);
1633 value_size = htab->map.value_size;
1634 size = round_up(value_size, 8);
1636 value_size = size * num_possible_cpus();
1638 /* while experimenting with hash tables with sizes ranging from 10 to
1639 * 1000, it was observed that a bucket can have upto 5 entries.
1644 /* We cannot do copy_from_user or copy_to_user inside
1645 * the rcu_read_lock. Allocate enough space here.
1647 keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1648 values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1649 if (!keys || !values) {
1655 bpf_disable_instrumentation();
1660 b = &htab->buckets[batch];
1662 /* do not grab the lock unless need it (bucket_cnt > 0). */
1664 ret = htab_lock_bucket(htab, b, batch, &flags);
1670 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1673 if (bucket_cnt && !locked) {
1678 if (bucket_cnt > (max_count - total)) {
1681 /* Note that since bucket_cnt > 0 here, it is implicit
1682 * that the locked was grabbed, so release it.
1684 htab_unlock_bucket(htab, b, batch, flags);
1686 bpf_enable_instrumentation();
1690 if (bucket_cnt > bucket_size) {
1691 bucket_size = bucket_cnt;
1692 /* Note that since bucket_cnt > 0 here, it is implicit
1693 * that the locked was grabbed, so release it.
1695 htab_unlock_bucket(htab, b, batch, flags);
1697 bpf_enable_instrumentation();
1703 /* Next block is only safe to run if you have grabbed the lock */
1707 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1708 memcpy(dst_key, l->key, key_size);
1712 void __percpu *pptr;
1714 pptr = htab_elem_get_ptr(l, map->key_size);
1715 for_each_possible_cpu(cpu) {
1716 bpf_long_memcpy(dst_val + off,
1717 per_cpu_ptr(pptr, cpu), size);
1721 value = l->key + roundup_key_size;
1722 if (elem_map_flags & BPF_F_LOCK)
1723 copy_map_value_locked(map, dst_val, value,
1726 copy_map_value(map, dst_val, value);
1727 check_and_init_map_value(map, dst_val);
1730 hlist_nulls_del_rcu(&l->hash_node);
1732 /* bpf_lru_push_free() will acquire lru_lock, which
1733 * may cause deadlock. See comments in function
1734 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1735 * after releasing the bucket lock.
1738 l->batch_flink = node_to_free;
1741 free_htab_elem(htab, l);
1744 dst_key += key_size;
1745 dst_val += value_size;
1748 htab_unlock_bucket(htab, b, batch, flags);
1751 while (node_to_free) {
1753 node_to_free = node_to_free->batch_flink;
1754 htab_lru_push_free(htab, l);
1758 /* If we are not copying data, we can go to next bucket and avoid
1759 * unlocking the rcu.
1761 if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1767 bpf_enable_instrumentation();
1768 if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1769 key_size * bucket_cnt) ||
1770 copy_to_user(uvalues + total * value_size, values,
1771 value_size * bucket_cnt))) {
1776 total += bucket_cnt;
1778 if (batch >= htab->n_buckets) {
1788 /* copy # of entries and next batch */
1789 ubatch = u64_to_user_ptr(attr->batch.out_batch);
1790 if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1791 put_user(total, &uattr->batch.count))
1801 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1802 union bpf_attr __user *uattr)
1804 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1809 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1810 const union bpf_attr *attr,
1811 union bpf_attr __user *uattr)
1813 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1818 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1819 union bpf_attr __user *uattr)
1821 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1826 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1827 const union bpf_attr *attr,
1828 union bpf_attr __user *uattr)
1830 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1835 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1836 const union bpf_attr *attr,
1837 union bpf_attr __user *uattr)
1839 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1844 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1845 const union bpf_attr *attr,
1846 union bpf_attr __user *uattr)
1848 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1853 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1854 union bpf_attr __user *uattr)
1856 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1861 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1862 const union bpf_attr *attr,
1863 union bpf_attr __user *uattr)
1865 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1869 struct bpf_iter_seq_hash_map_info {
1870 struct bpf_map *map;
1871 struct bpf_htab *htab;
1872 void *percpu_value_buf; // non-zero means percpu hash
1877 static struct htab_elem *
1878 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1879 struct htab_elem *prev_elem)
1881 const struct bpf_htab *htab = info->htab;
1882 u32 skip_elems = info->skip_elems;
1883 u32 bucket_id = info->bucket_id;
1884 struct hlist_nulls_head *head;
1885 struct hlist_nulls_node *n;
1886 struct htab_elem *elem;
1890 if (bucket_id >= htab->n_buckets)
1893 /* try to find next elem in the same bucket */
1895 /* no update/deletion on this bucket, prev_elem should be still valid
1896 * and we won't skip elements.
1898 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1899 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1903 /* not found, unlock and go to the next bucket */
1904 b = &htab->buckets[bucket_id++];
1909 for (i = bucket_id; i < htab->n_buckets; i++) {
1910 b = &htab->buckets[i];
1915 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1916 if (count >= skip_elems) {
1917 info->bucket_id = i;
1918 info->skip_elems = count;
1928 info->bucket_id = i;
1929 info->skip_elems = 0;
1933 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
1935 struct bpf_iter_seq_hash_map_info *info = seq->private;
1936 struct htab_elem *elem;
1938 elem = bpf_hash_map_seq_find_next(info, NULL);
1947 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1949 struct bpf_iter_seq_hash_map_info *info = seq->private;
1953 return bpf_hash_map_seq_find_next(info, v);
1956 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
1958 struct bpf_iter_seq_hash_map_info *info = seq->private;
1959 u32 roundup_key_size, roundup_value_size;
1960 struct bpf_iter__bpf_map_elem ctx = {};
1961 struct bpf_map *map = info->map;
1962 struct bpf_iter_meta meta;
1963 int ret = 0, off = 0, cpu;
1964 struct bpf_prog *prog;
1965 void __percpu *pptr;
1968 prog = bpf_iter_get_info(&meta, elem == NULL);
1971 ctx.map = info->map;
1973 roundup_key_size = round_up(map->key_size, 8);
1974 ctx.key = elem->key;
1975 if (!info->percpu_value_buf) {
1976 ctx.value = elem->key + roundup_key_size;
1978 roundup_value_size = round_up(map->value_size, 8);
1979 pptr = htab_elem_get_ptr(elem, map->key_size);
1980 for_each_possible_cpu(cpu) {
1981 bpf_long_memcpy(info->percpu_value_buf + off,
1982 per_cpu_ptr(pptr, cpu),
1983 roundup_value_size);
1984 off += roundup_value_size;
1986 ctx.value = info->percpu_value_buf;
1989 ret = bpf_iter_run_prog(prog, &ctx);
1995 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
1997 return __bpf_hash_map_seq_show(seq, v);
2000 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2003 (void)__bpf_hash_map_seq_show(seq, NULL);
2008 static int bpf_iter_init_hash_map(void *priv_data,
2009 struct bpf_iter_aux_info *aux)
2011 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2012 struct bpf_map *map = aux->map;
2016 if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2017 map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2018 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2019 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2023 seq_info->percpu_value_buf = value_buf;
2026 seq_info->map = map;
2027 seq_info->htab = container_of(map, struct bpf_htab, map);
2031 static void bpf_iter_fini_hash_map(void *priv_data)
2033 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2035 kfree(seq_info->percpu_value_buf);
2038 static const struct seq_operations bpf_hash_map_seq_ops = {
2039 .start = bpf_hash_map_seq_start,
2040 .next = bpf_hash_map_seq_next,
2041 .stop = bpf_hash_map_seq_stop,
2042 .show = bpf_hash_map_seq_show,
2045 static const struct bpf_iter_seq_info iter_seq_info = {
2046 .seq_ops = &bpf_hash_map_seq_ops,
2047 .init_seq_private = bpf_iter_init_hash_map,
2048 .fini_seq_private = bpf_iter_fini_hash_map,
2049 .seq_priv_size = sizeof(struct bpf_iter_seq_hash_map_info),
2052 static int bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2053 void *callback_ctx, u64 flags)
2055 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2056 struct hlist_nulls_head *head;
2057 struct hlist_nulls_node *n;
2058 struct htab_elem *elem;
2059 u32 roundup_key_size;
2060 int i, num_elems = 0;
2061 void __percpu *pptr;
2070 is_percpu = htab_is_percpu(htab);
2072 roundup_key_size = round_up(map->key_size, 8);
2073 /* disable migration so percpu value prepared here will be the
2074 * same as the one seen by the bpf program with bpf_map_lookup_elem().
2078 for (i = 0; i < htab->n_buckets; i++) {
2079 b = &htab->buckets[i];
2082 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2085 /* current cpu value for percpu map */
2086 pptr = htab_elem_get_ptr(elem, map->key_size);
2087 val = this_cpu_ptr(pptr);
2089 val = elem->key + roundup_key_size;
2092 ret = callback_fn((u64)(long)map, (u64)(long)key,
2093 (u64)(long)val, (u64)(long)callback_ctx, 0);
2094 /* return value: 0 - continue, 1 - stop and return */
2108 static int htab_map_btf_id;
2109 const struct bpf_map_ops htab_map_ops = {
2110 .map_meta_equal = bpf_map_meta_equal,
2111 .map_alloc_check = htab_map_alloc_check,
2112 .map_alloc = htab_map_alloc,
2113 .map_free = htab_map_free,
2114 .map_get_next_key = htab_map_get_next_key,
2115 .map_release_uref = htab_map_free_timers,
2116 .map_lookup_elem = htab_map_lookup_elem,
2117 .map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2118 .map_update_elem = htab_map_update_elem,
2119 .map_delete_elem = htab_map_delete_elem,
2120 .map_gen_lookup = htab_map_gen_lookup,
2121 .map_seq_show_elem = htab_map_seq_show_elem,
2122 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2123 .map_for_each_callback = bpf_for_each_hash_elem,
2125 .map_btf_name = "bpf_htab",
2126 .map_btf_id = &htab_map_btf_id,
2127 .iter_seq_info = &iter_seq_info,
2130 static int htab_lru_map_btf_id;
2131 const struct bpf_map_ops htab_lru_map_ops = {
2132 .map_meta_equal = bpf_map_meta_equal,
2133 .map_alloc_check = htab_map_alloc_check,
2134 .map_alloc = htab_map_alloc,
2135 .map_free = htab_map_free,
2136 .map_get_next_key = htab_map_get_next_key,
2137 .map_release_uref = htab_map_free_timers,
2138 .map_lookup_elem = htab_lru_map_lookup_elem,
2139 .map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2140 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2141 .map_update_elem = htab_lru_map_update_elem,
2142 .map_delete_elem = htab_lru_map_delete_elem,
2143 .map_gen_lookup = htab_lru_map_gen_lookup,
2144 .map_seq_show_elem = htab_map_seq_show_elem,
2145 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2146 .map_for_each_callback = bpf_for_each_hash_elem,
2147 BATCH_OPS(htab_lru),
2148 .map_btf_name = "bpf_htab",
2149 .map_btf_id = &htab_lru_map_btf_id,
2150 .iter_seq_info = &iter_seq_info,
2153 /* Called from eBPF program */
2154 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2156 struct htab_elem *l = __htab_map_lookup_elem(map, key);
2159 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2164 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2166 struct htab_elem *l = __htab_map_lookup_elem(map, key);
2169 bpf_lru_node_set_ref(&l->lru_node);
2170 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2176 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2178 struct htab_elem *l;
2179 void __percpu *pptr;
2184 /* per_cpu areas are zero-filled and bpf programs can only
2185 * access 'value_size' of them, so copying rounded areas
2186 * will not leak any kernel data
2188 size = round_up(map->value_size, 8);
2190 l = __htab_map_lookup_elem(map, key);
2193 /* We do not mark LRU map element here in order to not mess up
2194 * eviction heuristics when user space does a map walk.
2196 pptr = htab_elem_get_ptr(l, map->key_size);
2197 for_each_possible_cpu(cpu) {
2198 bpf_long_memcpy(value + off,
2199 per_cpu_ptr(pptr, cpu), size);
2208 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2211 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2215 if (htab_is_lru(htab))
2216 ret = __htab_lru_percpu_map_update_elem(map, key, value,
2219 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2226 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2229 struct htab_elem *l;
2230 void __percpu *pptr;
2235 l = __htab_map_lookup_elem(map, key);
2241 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2242 seq_puts(m, ": {\n");
2243 pptr = htab_elem_get_ptr(l, map->key_size);
2244 for_each_possible_cpu(cpu) {
2245 seq_printf(m, "\tcpu%d: ", cpu);
2246 btf_type_seq_show(map->btf, map->btf_value_type_id,
2247 per_cpu_ptr(pptr, cpu), m);
2255 static int htab_percpu_map_btf_id;
2256 const struct bpf_map_ops htab_percpu_map_ops = {
2257 .map_meta_equal = bpf_map_meta_equal,
2258 .map_alloc_check = htab_map_alloc_check,
2259 .map_alloc = htab_map_alloc,
2260 .map_free = htab_map_free,
2261 .map_get_next_key = htab_map_get_next_key,
2262 .map_lookup_elem = htab_percpu_map_lookup_elem,
2263 .map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2264 .map_update_elem = htab_percpu_map_update_elem,
2265 .map_delete_elem = htab_map_delete_elem,
2266 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2267 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2268 .map_for_each_callback = bpf_for_each_hash_elem,
2269 BATCH_OPS(htab_percpu),
2270 .map_btf_name = "bpf_htab",
2271 .map_btf_id = &htab_percpu_map_btf_id,
2272 .iter_seq_info = &iter_seq_info,
2275 static int htab_lru_percpu_map_btf_id;
2276 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2277 .map_meta_equal = bpf_map_meta_equal,
2278 .map_alloc_check = htab_map_alloc_check,
2279 .map_alloc = htab_map_alloc,
2280 .map_free = htab_map_free,
2281 .map_get_next_key = htab_map_get_next_key,
2282 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2283 .map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2284 .map_update_elem = htab_lru_percpu_map_update_elem,
2285 .map_delete_elem = htab_lru_map_delete_elem,
2286 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2287 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2288 .map_for_each_callback = bpf_for_each_hash_elem,
2289 BATCH_OPS(htab_lru_percpu),
2290 .map_btf_name = "bpf_htab",
2291 .map_btf_id = &htab_lru_percpu_map_btf_id,
2292 .iter_seq_info = &iter_seq_info,
2295 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2297 if (attr->value_size != sizeof(u32))
2299 return htab_map_alloc_check(attr);
2302 static void fd_htab_map_free(struct bpf_map *map)
2304 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2305 struct hlist_nulls_node *n;
2306 struct hlist_nulls_head *head;
2307 struct htab_elem *l;
2310 for (i = 0; i < htab->n_buckets; i++) {
2311 head = select_bucket(htab, i);
2313 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2314 void *ptr = fd_htab_map_get_ptr(map, l);
2316 map->ops->map_fd_put_ptr(ptr);
2323 /* only called from syscall */
2324 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2329 if (!map->ops->map_fd_sys_lookup_elem)
2333 ptr = htab_map_lookup_elem(map, key);
2335 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2343 /* only called from syscall */
2344 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2345 void *key, void *value, u64 map_flags)
2349 u32 ufd = *(u32 *)value;
2351 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2353 return PTR_ERR(ptr);
2355 ret = htab_map_update_elem(map, key, &ptr, map_flags);
2357 map->ops->map_fd_put_ptr(ptr);
2362 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2364 struct bpf_map *map, *inner_map_meta;
2366 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2367 if (IS_ERR(inner_map_meta))
2368 return inner_map_meta;
2370 map = htab_map_alloc(attr);
2372 bpf_map_meta_free(inner_map_meta);
2376 map->inner_map_meta = inner_map_meta;
2381 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2383 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
2388 return READ_ONCE(*inner_map);
2391 static int htab_of_map_gen_lookup(struct bpf_map *map,
2392 struct bpf_insn *insn_buf)
2394 struct bpf_insn *insn = insn_buf;
2395 const int ret = BPF_REG_0;
2397 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2398 (void *(*)(struct bpf_map *map, void *key))NULL));
2399 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2400 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2401 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2402 offsetof(struct htab_elem, key) +
2403 round_up(map->key_size, 8));
2404 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2406 return insn - insn_buf;
2409 static void htab_of_map_free(struct bpf_map *map)
2411 bpf_map_meta_free(map->inner_map_meta);
2412 fd_htab_map_free(map);
2415 static int htab_of_maps_map_btf_id;
2416 const struct bpf_map_ops htab_of_maps_map_ops = {
2417 .map_alloc_check = fd_htab_map_alloc_check,
2418 .map_alloc = htab_of_map_alloc,
2419 .map_free = htab_of_map_free,
2420 .map_get_next_key = htab_map_get_next_key,
2421 .map_lookup_elem = htab_of_map_lookup_elem,
2422 .map_delete_elem = htab_map_delete_elem,
2423 .map_fd_get_ptr = bpf_map_fd_get_ptr,
2424 .map_fd_put_ptr = bpf_map_fd_put_ptr,
2425 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2426 .map_gen_lookup = htab_of_map_gen_lookup,
2427 .map_check_btf = map_check_no_btf,
2428 .map_btf_name = "bpf_htab",
2429 .map_btf_id = &htab_of_maps_map_btf_id,