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 <linux/btf_ids.h>
14 #include "percpu_freelist.h"
15 #include "bpf_lru_list.h"
16 #include "map_in_map.h"
18 #define HTAB_CREATE_FLAG_MASK \
19 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
20 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
22 #define BATCH_OPS(_name) \
24 _name##_map_lookup_batch, \
25 .map_lookup_and_delete_batch = \
26 _name##_map_lookup_and_delete_batch, \
28 generic_map_update_batch, \
30 generic_map_delete_batch
33 * The bucket lock has two protection scopes:
35 * 1) Serializing concurrent operations from BPF programs on different
38 * 2) Serializing concurrent operations from BPF programs and sys_bpf()
40 * BPF programs can execute in any context including perf, kprobes and
41 * tracing. As there are almost no limits where perf, kprobes and tracing
42 * can be invoked from the lock operations need to be protected against
43 * deadlocks. Deadlocks can be caused by recursion and by an invocation in
44 * the lock held section when functions which acquire this lock are invoked
45 * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
46 * variable bpf_prog_active, which prevents BPF programs attached to perf
47 * events, kprobes and tracing to be invoked before the prior invocation
48 * from one of these contexts completed. sys_bpf() uses the same mechanism
49 * by pinning the task to the current CPU and incrementing the recursion
50 * protection across the map operation.
52 * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
53 * operations like memory allocations (even with GFP_ATOMIC) from atomic
54 * contexts. This is required because even with GFP_ATOMIC the memory
55 * allocator calls into code paths which acquire locks with long held lock
56 * sections. To ensure the deterministic behaviour these locks are regular
57 * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
58 * true atomic contexts on an RT kernel are the low level hardware
59 * handling, scheduling, low level interrupt handling, NMIs etc. None of
60 * these contexts should ever do memory allocations.
62 * As regular device interrupt handlers and soft interrupts are forced into
63 * thread context, the existing code which does
64 * spin_lock*(); alloc(GFP_ATOMIC); spin_unlock*();
67 * In theory the BPF locks could be converted to regular spinlocks as well,
68 * but the bucket locks and percpu_freelist locks can be taken from
69 * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
70 * atomic contexts even on RT. These mechanisms require preallocated maps,
71 * so there is no need to invoke memory allocations within the lock held
74 * BPF maps which need dynamic allocation are only used from (forced)
75 * thread context on RT and can therefore use regular spinlocks which in
76 * turn allows to invoke memory allocations from the lock held section.
78 * On a non RT kernel this distinction is neither possible nor required.
79 * spinlock maps to raw_spinlock and the extra code is optimized out by the
83 struct hlist_nulls_head head;
85 raw_spinlock_t raw_lock;
90 #define HASHTAB_MAP_LOCK_COUNT 8
91 #define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
95 struct bucket *buckets;
98 struct pcpu_freelist freelist;
101 struct htab_elem *__percpu *extra_elems;
102 atomic_t count; /* number of elements in this hashtable */
103 u32 n_buckets; /* number of hash buckets */
104 u32 elem_size; /* size of each element in bytes */
106 struct lock_class_key lockdep_key;
107 int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
110 /* each htab element is struct htab_elem + key + value */
113 struct hlist_nulls_node hash_node;
117 struct bpf_htab *htab;
118 struct pcpu_freelist_node fnode;
119 struct htab_elem *batch_flink;
125 struct bpf_lru_node lru_node;
128 char key[] __aligned(8);
131 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
133 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
136 static inline bool htab_use_raw_lock(const struct bpf_htab *htab)
138 return (!IS_ENABLED(CONFIG_PREEMPT_RT) || htab_is_prealloc(htab));
141 static void htab_init_buckets(struct bpf_htab *htab)
145 for (i = 0; i < htab->n_buckets; i++) {
146 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
147 if (htab_use_raw_lock(htab)) {
148 raw_spin_lock_init(&htab->buckets[i].raw_lock);
149 lockdep_set_class(&htab->buckets[i].raw_lock,
152 spin_lock_init(&htab->buckets[i].lock);
153 lockdep_set_class(&htab->buckets[i].lock,
160 static inline int htab_lock_bucket(const struct bpf_htab *htab,
161 struct bucket *b, u32 hash,
162 unsigned long *pflags)
166 hash = hash & HASHTAB_MAP_LOCK_MASK;
169 if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
170 __this_cpu_dec(*(htab->map_locked[hash]));
175 if (htab_use_raw_lock(htab))
176 raw_spin_lock_irqsave(&b->raw_lock, flags);
178 spin_lock_irqsave(&b->lock, flags);
184 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
185 struct bucket *b, u32 hash,
188 hash = hash & HASHTAB_MAP_LOCK_MASK;
189 if (htab_use_raw_lock(htab))
190 raw_spin_unlock_irqrestore(&b->raw_lock, flags);
192 spin_unlock_irqrestore(&b->lock, flags);
193 __this_cpu_dec(*(htab->map_locked[hash]));
197 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
199 static bool htab_is_lru(const struct bpf_htab *htab)
201 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
202 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
205 static bool htab_is_percpu(const struct bpf_htab *htab)
207 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
208 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
211 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
214 *(void __percpu **)(l->key + key_size) = pptr;
217 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
219 return *(void __percpu **)(l->key + key_size);
222 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
224 return *(void **)(l->key + roundup(map->key_size, 8));
227 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
229 return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
232 static bool htab_has_extra_elems(struct bpf_htab *htab)
234 return !htab_is_percpu(htab) && !htab_is_lru(htab);
237 static void htab_free_prealloced_timers(struct bpf_htab *htab)
239 u32 num_entries = htab->map.max_entries;
242 if (!map_value_has_timer(&htab->map))
244 if (htab_has_extra_elems(htab))
245 num_entries += num_possible_cpus();
247 for (i = 0; i < num_entries; i++) {
248 struct htab_elem *elem;
250 elem = get_htab_elem(htab, i);
251 bpf_timer_cancel_and_free(elem->key +
252 round_up(htab->map.key_size, 8) +
253 htab->map.timer_off);
258 static void htab_free_prealloced_kptrs(struct bpf_htab *htab)
260 u32 num_entries = htab->map.max_entries;
263 if (!map_value_has_kptrs(&htab->map))
265 if (htab_has_extra_elems(htab))
266 num_entries += num_possible_cpus();
268 for (i = 0; i < num_entries; i++) {
269 struct htab_elem *elem;
271 elem = get_htab_elem(htab, i);
272 bpf_map_free_kptrs(&htab->map, elem->key + round_up(htab->map.key_size, 8));
277 static void htab_free_elems(struct bpf_htab *htab)
281 if (!htab_is_percpu(htab))
284 for (i = 0; i < htab->map.max_entries; i++) {
287 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
293 bpf_map_area_free(htab->elems);
296 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
297 * (bucket_lock). If both locks need to be acquired together, the lock
298 * order is always lru_lock -> bucket_lock and this only happens in
299 * bpf_lru_list.c logic. For example, certain code path of
300 * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
301 * will acquire lru_lock first followed by acquiring bucket_lock.
303 * In hashtab.c, to avoid deadlock, lock acquisition of
304 * bucket_lock followed by lru_lock is not allowed. In such cases,
305 * bucket_lock needs to be released first before acquiring lru_lock.
307 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
310 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
314 l = container_of(node, struct htab_elem, lru_node);
315 memcpy(l->key, key, htab->map.key_size);
322 static int prealloc_init(struct bpf_htab *htab)
324 u32 num_entries = htab->map.max_entries;
325 int err = -ENOMEM, i;
327 if (htab_has_extra_elems(htab))
328 num_entries += num_possible_cpus();
330 htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
331 htab->map.numa_node);
335 if (!htab_is_percpu(htab))
336 goto skip_percpu_elems;
338 for (i = 0; i < num_entries; i++) {
339 u32 size = round_up(htab->map.value_size, 8);
342 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
343 GFP_USER | __GFP_NOWARN);
346 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
352 if (htab_is_lru(htab))
353 err = bpf_lru_init(&htab->lru,
354 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
355 offsetof(struct htab_elem, hash) -
356 offsetof(struct htab_elem, lru_node),
357 htab_lru_map_delete_node,
360 err = pcpu_freelist_init(&htab->freelist);
365 if (htab_is_lru(htab))
366 bpf_lru_populate(&htab->lru, htab->elems,
367 offsetof(struct htab_elem, lru_node),
368 htab->elem_size, num_entries);
370 pcpu_freelist_populate(&htab->freelist,
371 htab->elems + offsetof(struct htab_elem, fnode),
372 htab->elem_size, num_entries);
377 htab_free_elems(htab);
381 static void prealloc_destroy(struct bpf_htab *htab)
383 htab_free_elems(htab);
385 if (htab_is_lru(htab))
386 bpf_lru_destroy(&htab->lru);
388 pcpu_freelist_destroy(&htab->freelist);
391 static int alloc_extra_elems(struct bpf_htab *htab)
393 struct htab_elem *__percpu *pptr, *l_new;
394 struct pcpu_freelist_node *l;
397 pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
398 GFP_USER | __GFP_NOWARN);
402 for_each_possible_cpu(cpu) {
403 l = pcpu_freelist_pop(&htab->freelist);
404 /* pop will succeed, since prealloc_init()
405 * preallocated extra num_possible_cpus elements
407 l_new = container_of(l, struct htab_elem, fnode);
408 *per_cpu_ptr(pptr, cpu) = l_new;
410 htab->extra_elems = pptr;
414 /* Called from syscall */
415 static int htab_map_alloc_check(union bpf_attr *attr)
417 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
418 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
419 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
420 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
421 /* percpu_lru means each cpu has its own LRU list.
422 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
423 * the map's value itself is percpu. percpu_lru has
424 * nothing to do with the map's value.
426 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
427 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
428 bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
429 int numa_node = bpf_map_attr_numa_node(attr);
431 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
432 offsetof(struct htab_elem, hash_node.pprev));
433 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
434 offsetof(struct htab_elem, hash_node.pprev));
436 if (lru && !bpf_capable())
437 /* LRU implementation is much complicated than other
438 * maps. Hence, limit to CAP_BPF.
442 if (zero_seed && !capable(CAP_SYS_ADMIN))
443 /* Guard against local DoS, and discourage production use. */
446 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
447 !bpf_map_flags_access_ok(attr->map_flags))
450 if (!lru && percpu_lru)
453 if (lru && !prealloc)
456 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
459 /* check sanity of attributes.
460 * value_size == 0 may be allowed in the future to use map as a set
462 if (attr->max_entries == 0 || attr->key_size == 0 ||
463 attr->value_size == 0)
466 if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
467 sizeof(struct htab_elem))
468 /* if key_size + value_size is bigger, the user space won't be
469 * able to access the elements via bpf syscall. This check
470 * also makes sure that the elem_size doesn't overflow and it's
471 * kmalloc-able later in htab_map_update_elem()
478 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
480 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
481 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
482 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
483 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
484 /* percpu_lru means each cpu has its own LRU list.
485 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
486 * the map's value itself is percpu. percpu_lru has
487 * nothing to do with the map's value.
489 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
490 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
491 struct bpf_htab *htab;
494 htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
496 return ERR_PTR(-ENOMEM);
498 lockdep_register_key(&htab->lockdep_key);
500 bpf_map_init_from_attr(&htab->map, attr);
503 /* ensure each CPU's lru list has >=1 elements.
504 * since we are at it, make each lru list has the same
505 * number of elements.
507 htab->map.max_entries = roundup(attr->max_entries,
508 num_possible_cpus());
509 if (htab->map.max_entries < attr->max_entries)
510 htab->map.max_entries = rounddown(attr->max_entries,
511 num_possible_cpus());
514 /* hash table size must be power of 2 */
515 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
517 htab->elem_size = sizeof(struct htab_elem) +
518 round_up(htab->map.key_size, 8);
520 htab->elem_size += sizeof(void *);
522 htab->elem_size += round_up(htab->map.value_size, 8);
525 /* prevent zero size kmalloc and check for u32 overflow */
526 if (htab->n_buckets == 0 ||
527 htab->n_buckets > U32_MAX / sizeof(struct bucket))
531 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
532 sizeof(struct bucket),
533 htab->map.numa_node);
537 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
538 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
542 if (!htab->map_locked[i])
543 goto free_map_locked;
546 if (htab->map.map_flags & BPF_F_ZERO_SEED)
549 htab->hashrnd = get_random_int();
551 htab_init_buckets(htab);
554 err = prealloc_init(htab);
556 goto free_map_locked;
558 if (!percpu && !lru) {
559 /* lru itself can remove the least used element, so
560 * there is no need for an extra elem during map_update.
562 err = alloc_extra_elems(htab);
571 prealloc_destroy(htab);
573 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
574 free_percpu(htab->map_locked[i]);
575 bpf_map_area_free(htab->buckets);
577 lockdep_unregister_key(&htab->lockdep_key);
578 bpf_map_area_free(htab);
582 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
584 return jhash(key, key_len, hashrnd);
587 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
589 return &htab->buckets[hash & (htab->n_buckets - 1)];
592 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
594 return &__select_bucket(htab, hash)->head;
597 /* this lookup function can only be called with bucket lock taken */
598 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
599 void *key, u32 key_size)
601 struct hlist_nulls_node *n;
604 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
605 if (l->hash == hash && !memcmp(&l->key, key, key_size))
611 /* can be called without bucket lock. it will repeat the loop in
612 * the unlikely event when elements moved from one bucket into another
613 * while link list is being walked
615 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
617 u32 key_size, u32 n_buckets)
619 struct hlist_nulls_node *n;
623 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
624 if (l->hash == hash && !memcmp(&l->key, key, key_size))
627 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
633 /* Called from syscall or from eBPF program directly, so
634 * arguments have to match bpf_map_lookup_elem() exactly.
635 * The return value is adjusted by BPF instructions
636 * in htab_map_gen_lookup().
638 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
640 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
641 struct hlist_nulls_head *head;
645 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
646 !rcu_read_lock_bh_held());
648 key_size = map->key_size;
650 hash = htab_map_hash(key, key_size, htab->hashrnd);
652 head = select_bucket(htab, hash);
654 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
659 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
661 struct htab_elem *l = __htab_map_lookup_elem(map, key);
664 return l->key + round_up(map->key_size, 8);
669 /* inline bpf_map_lookup_elem() call.
672 * bpf_map_lookup_elem
673 * map->ops->map_lookup_elem
674 * htab_map_lookup_elem
675 * __htab_map_lookup_elem
678 * __htab_map_lookup_elem
680 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
682 struct bpf_insn *insn = insn_buf;
683 const int ret = BPF_REG_0;
685 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
686 (void *(*)(struct bpf_map *map, void *key))NULL));
687 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
688 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
689 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
690 offsetof(struct htab_elem, key) +
691 round_up(map->key_size, 8));
692 return insn - insn_buf;
695 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
696 void *key, const bool mark)
698 struct htab_elem *l = __htab_map_lookup_elem(map, key);
702 bpf_lru_node_set_ref(&l->lru_node);
703 return l->key + round_up(map->key_size, 8);
709 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
711 return __htab_lru_map_lookup_elem(map, key, true);
714 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
716 return __htab_lru_map_lookup_elem(map, key, false);
719 static int htab_lru_map_gen_lookup(struct bpf_map *map,
720 struct bpf_insn *insn_buf)
722 struct bpf_insn *insn = insn_buf;
723 const int ret = BPF_REG_0;
724 const int ref_reg = BPF_REG_1;
726 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
727 (void *(*)(struct bpf_map *map, void *key))NULL));
728 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
729 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
730 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
731 offsetof(struct htab_elem, lru_node) +
732 offsetof(struct bpf_lru_node, ref));
733 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
734 *insn++ = BPF_ST_MEM(BPF_B, ret,
735 offsetof(struct htab_elem, lru_node) +
736 offsetof(struct bpf_lru_node, ref),
738 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
739 offsetof(struct htab_elem, key) +
740 round_up(map->key_size, 8));
741 return insn - insn_buf;
744 static void check_and_free_fields(struct bpf_htab *htab,
745 struct htab_elem *elem)
747 void *map_value = elem->key + round_up(htab->map.key_size, 8);
749 if (map_value_has_timer(&htab->map))
750 bpf_timer_cancel_and_free(map_value + htab->map.timer_off);
751 if (map_value_has_kptrs(&htab->map))
752 bpf_map_free_kptrs(&htab->map, map_value);
755 /* It is called from the bpf_lru_list when the LRU needs to delete
756 * older elements from the htab.
758 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
760 struct bpf_htab *htab = arg;
761 struct htab_elem *l = NULL, *tgt_l;
762 struct hlist_nulls_head *head;
763 struct hlist_nulls_node *n;
768 tgt_l = container_of(node, struct htab_elem, lru_node);
769 b = __select_bucket(htab, tgt_l->hash);
772 ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
776 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
778 hlist_nulls_del_rcu(&l->hash_node);
779 check_and_free_fields(htab, l);
783 htab_unlock_bucket(htab, b, tgt_l->hash, flags);
788 /* Called from syscall */
789 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
791 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
792 struct hlist_nulls_head *head;
793 struct htab_elem *l, *next_l;
797 WARN_ON_ONCE(!rcu_read_lock_held());
799 key_size = map->key_size;
802 goto find_first_elem;
804 hash = htab_map_hash(key, key_size, htab->hashrnd);
806 head = select_bucket(htab, hash);
809 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
812 goto find_first_elem;
814 /* key was found, get next key in the same bucket */
815 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
816 struct htab_elem, hash_node);
819 /* if next elem in this hash list is non-zero, just return it */
820 memcpy(next_key, next_l->key, key_size);
824 /* no more elements in this hash list, go to the next bucket */
825 i = hash & (htab->n_buckets - 1);
829 /* iterate over buckets */
830 for (; i < htab->n_buckets; i++) {
831 head = select_bucket(htab, i);
833 /* pick first element in the bucket */
834 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
835 struct htab_elem, hash_node);
837 /* if it's not empty, just return it */
838 memcpy(next_key, next_l->key, key_size);
843 /* iterated over all buckets and all elements */
847 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
849 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
850 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
851 check_and_free_fields(htab, l);
855 static void htab_elem_free_rcu(struct rcu_head *head)
857 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
858 struct bpf_htab *htab = l->htab;
860 htab_elem_free(htab, l);
863 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
865 struct bpf_map *map = &htab->map;
868 if (map->ops->map_fd_put_ptr) {
869 ptr = fd_htab_map_get_ptr(map, l);
870 map->ops->map_fd_put_ptr(ptr);
874 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
876 htab_put_fd_value(htab, l);
878 if (htab_is_prealloc(htab)) {
879 check_and_free_fields(htab, l);
880 __pcpu_freelist_push(&htab->freelist, &l->fnode);
882 atomic_dec(&htab->count);
884 call_rcu(&l->rcu, htab_elem_free_rcu);
888 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
889 void *value, bool onallcpus)
892 /* copy true value_size bytes */
893 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
895 u32 size = round_up(htab->map.value_size, 8);
898 for_each_possible_cpu(cpu) {
899 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
906 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
907 void *value, bool onallcpus)
909 /* When using prealloc and not setting the initial value on all cpus,
910 * zero-fill element values for other cpus (just as what happens when
911 * not using prealloc). Otherwise, bpf program has no way to ensure
912 * known initial values for cpus other than current one
913 * (onallcpus=false always when coming from bpf prog).
915 if (htab_is_prealloc(htab) && !onallcpus) {
916 u32 size = round_up(htab->map.value_size, 8);
917 int current_cpu = raw_smp_processor_id();
920 for_each_possible_cpu(cpu) {
921 if (cpu == current_cpu)
922 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value,
925 memset(per_cpu_ptr(pptr, cpu), 0, size);
928 pcpu_copy_value(htab, pptr, value, onallcpus);
932 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
934 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
938 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
939 void *value, u32 key_size, u32 hash,
940 bool percpu, bool onallcpus,
941 struct htab_elem *old_elem)
943 u32 size = htab->map.value_size;
944 bool prealloc = htab_is_prealloc(htab);
945 struct htab_elem *l_new, **pl_new;
950 /* if we're updating the existing element,
951 * use per-cpu extra elems to avoid freelist_pop/push
953 pl_new = this_cpu_ptr(htab->extra_elems);
955 htab_put_fd_value(htab, old_elem);
958 struct pcpu_freelist_node *l;
960 l = __pcpu_freelist_pop(&htab->freelist);
962 return ERR_PTR(-E2BIG);
963 l_new = container_of(l, struct htab_elem, fnode);
966 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
968 /* when map is full and update() is replacing
969 * old element, it's ok to allocate, since
970 * old element will be freed immediately.
971 * Otherwise return an error
973 l_new = ERR_PTR(-E2BIG);
976 l_new = bpf_map_kmalloc_node(&htab->map, htab->elem_size,
977 GFP_NOWAIT | __GFP_NOWARN,
978 htab->map.numa_node);
980 l_new = ERR_PTR(-ENOMEM);
983 check_and_init_map_value(&htab->map,
984 l_new->key + round_up(key_size, 8));
987 memcpy(l_new->key, key, key_size);
989 size = round_up(size, 8);
991 pptr = htab_elem_get_ptr(l_new, key_size);
993 /* alloc_percpu zero-fills */
994 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
995 GFP_NOWAIT | __GFP_NOWARN);
998 l_new = ERR_PTR(-ENOMEM);
1003 pcpu_init_value(htab, pptr, value, onallcpus);
1006 htab_elem_set_ptr(l_new, key_size, pptr);
1007 } else if (fd_htab_map_needs_adjust(htab)) {
1008 size = round_up(size, 8);
1009 memcpy(l_new->key + round_up(key_size, 8), value, size);
1011 copy_map_value(&htab->map,
1012 l_new->key + round_up(key_size, 8),
1019 atomic_dec(&htab->count);
1023 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1026 if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1027 /* elem already exists */
1030 if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1031 /* elem doesn't exist, cannot update it */
1037 /* Called from syscall or from eBPF program */
1038 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1041 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1042 struct htab_elem *l_new = NULL, *l_old;
1043 struct hlist_nulls_head *head;
1044 unsigned long flags;
1049 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1053 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1054 !rcu_read_lock_bh_held());
1056 key_size = map->key_size;
1058 hash = htab_map_hash(key, key_size, htab->hashrnd);
1060 b = __select_bucket(htab, hash);
1063 if (unlikely(map_flags & BPF_F_LOCK)) {
1064 if (unlikely(!map_value_has_spin_lock(map)))
1066 /* find an element without taking the bucket lock */
1067 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1069 ret = check_flags(htab, l_old, map_flags);
1073 /* grab the element lock and update value in place */
1074 copy_map_value_locked(map,
1075 l_old->key + round_up(key_size, 8),
1079 /* fall through, grab the bucket lock and lookup again.
1080 * 99.9% chance that the element won't be found,
1081 * but second lookup under lock has to be done.
1085 ret = htab_lock_bucket(htab, b, hash, &flags);
1089 l_old = lookup_elem_raw(head, hash, key, key_size);
1091 ret = check_flags(htab, l_old, map_flags);
1095 if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1096 /* first lookup without the bucket lock didn't find the element,
1097 * but second lookup with the bucket lock found it.
1098 * This case is highly unlikely, but has to be dealt with:
1099 * grab the element lock in addition to the bucket lock
1100 * and update element in place
1102 copy_map_value_locked(map,
1103 l_old->key + round_up(key_size, 8),
1109 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1111 if (IS_ERR(l_new)) {
1112 /* all pre-allocated elements are in use or memory exhausted */
1113 ret = PTR_ERR(l_new);
1117 /* add new element to the head of the list, so that
1118 * concurrent search will find it before old elem
1120 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1122 hlist_nulls_del_rcu(&l_old->hash_node);
1123 if (!htab_is_prealloc(htab))
1124 free_htab_elem(htab, l_old);
1126 check_and_free_fields(htab, l_old);
1130 htab_unlock_bucket(htab, b, hash, flags);
1134 static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1136 check_and_free_fields(htab, elem);
1137 bpf_lru_push_free(&htab->lru, &elem->lru_node);
1140 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1143 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1144 struct htab_elem *l_new, *l_old = NULL;
1145 struct hlist_nulls_head *head;
1146 unsigned long flags;
1151 if (unlikely(map_flags > BPF_EXIST))
1155 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1156 !rcu_read_lock_bh_held());
1158 key_size = map->key_size;
1160 hash = htab_map_hash(key, key_size, htab->hashrnd);
1162 b = __select_bucket(htab, hash);
1165 /* For LRU, we need to alloc before taking bucket's
1166 * spinlock because getting free nodes from LRU may need
1167 * to remove older elements from htab and this removal
1168 * operation will need a bucket lock.
1170 l_new = prealloc_lru_pop(htab, key, hash);
1173 copy_map_value(&htab->map,
1174 l_new->key + round_up(map->key_size, 8), value);
1176 ret = htab_lock_bucket(htab, b, hash, &flags);
1180 l_old = lookup_elem_raw(head, hash, key, key_size);
1182 ret = check_flags(htab, l_old, map_flags);
1186 /* add new element to the head of the list, so that
1187 * concurrent search will find it before old elem
1189 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1191 bpf_lru_node_set_ref(&l_new->lru_node);
1192 hlist_nulls_del_rcu(&l_old->hash_node);
1197 htab_unlock_bucket(htab, b, hash, flags);
1200 htab_lru_push_free(htab, l_new);
1202 htab_lru_push_free(htab, l_old);
1207 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1208 void *value, u64 map_flags,
1211 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1212 struct htab_elem *l_new = NULL, *l_old;
1213 struct hlist_nulls_head *head;
1214 unsigned long flags;
1219 if (unlikely(map_flags > BPF_EXIST))
1223 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1224 !rcu_read_lock_bh_held());
1226 key_size = map->key_size;
1228 hash = htab_map_hash(key, key_size, htab->hashrnd);
1230 b = __select_bucket(htab, hash);
1233 ret = htab_lock_bucket(htab, b, hash, &flags);
1237 l_old = lookup_elem_raw(head, hash, key, key_size);
1239 ret = check_flags(htab, l_old, map_flags);
1244 /* per-cpu hash map can update value in-place */
1245 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1248 l_new = alloc_htab_elem(htab, key, value, key_size,
1249 hash, true, onallcpus, NULL);
1250 if (IS_ERR(l_new)) {
1251 ret = PTR_ERR(l_new);
1254 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1258 htab_unlock_bucket(htab, b, hash, flags);
1262 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1263 void *value, u64 map_flags,
1266 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1267 struct htab_elem *l_new = NULL, *l_old;
1268 struct hlist_nulls_head *head;
1269 unsigned long flags;
1274 if (unlikely(map_flags > BPF_EXIST))
1278 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1279 !rcu_read_lock_bh_held());
1281 key_size = map->key_size;
1283 hash = htab_map_hash(key, key_size, htab->hashrnd);
1285 b = __select_bucket(htab, hash);
1288 /* For LRU, we need to alloc before taking bucket's
1289 * spinlock because LRU's elem alloc may need
1290 * to remove older elem from htab and this removal
1291 * operation will need a bucket lock.
1293 if (map_flags != BPF_EXIST) {
1294 l_new = prealloc_lru_pop(htab, key, hash);
1299 ret = htab_lock_bucket(htab, b, hash, &flags);
1303 l_old = lookup_elem_raw(head, hash, key, key_size);
1305 ret = check_flags(htab, l_old, map_flags);
1310 bpf_lru_node_set_ref(&l_old->lru_node);
1312 /* per-cpu hash map can update value in-place */
1313 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1316 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1318 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1323 htab_unlock_bucket(htab, b, hash, flags);
1325 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1329 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1330 void *value, u64 map_flags)
1332 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1335 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1336 void *value, u64 map_flags)
1338 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1342 /* Called from syscall or from eBPF program */
1343 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1345 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1346 struct hlist_nulls_head *head;
1348 struct htab_elem *l;
1349 unsigned long flags;
1353 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1354 !rcu_read_lock_bh_held());
1356 key_size = map->key_size;
1358 hash = htab_map_hash(key, key_size, htab->hashrnd);
1359 b = __select_bucket(htab, hash);
1362 ret = htab_lock_bucket(htab, b, hash, &flags);
1366 l = lookup_elem_raw(head, hash, key, key_size);
1369 hlist_nulls_del_rcu(&l->hash_node);
1370 free_htab_elem(htab, l);
1375 htab_unlock_bucket(htab, b, hash, flags);
1379 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1381 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1382 struct hlist_nulls_head *head;
1384 struct htab_elem *l;
1385 unsigned long flags;
1389 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1390 !rcu_read_lock_bh_held());
1392 key_size = map->key_size;
1394 hash = htab_map_hash(key, key_size, htab->hashrnd);
1395 b = __select_bucket(htab, hash);
1398 ret = htab_lock_bucket(htab, b, hash, &flags);
1402 l = lookup_elem_raw(head, hash, key, key_size);
1405 hlist_nulls_del_rcu(&l->hash_node);
1409 htab_unlock_bucket(htab, b, hash, flags);
1411 htab_lru_push_free(htab, l);
1415 static void delete_all_elements(struct bpf_htab *htab)
1419 for (i = 0; i < htab->n_buckets; i++) {
1420 struct hlist_nulls_head *head = select_bucket(htab, i);
1421 struct hlist_nulls_node *n;
1422 struct htab_elem *l;
1424 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1425 hlist_nulls_del_rcu(&l->hash_node);
1426 htab_elem_free(htab, l);
1431 static void htab_free_malloced_timers(struct bpf_htab *htab)
1436 for (i = 0; i < htab->n_buckets; i++) {
1437 struct hlist_nulls_head *head = select_bucket(htab, i);
1438 struct hlist_nulls_node *n;
1439 struct htab_elem *l;
1441 hlist_nulls_for_each_entry(l, n, head, hash_node) {
1442 /* We don't reset or free kptr on uref dropping to zero,
1443 * hence just free timer.
1445 bpf_timer_cancel_and_free(l->key +
1446 round_up(htab->map.key_size, 8) +
1447 htab->map.timer_off);
1454 static void htab_map_free_timers(struct bpf_map *map)
1456 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1458 /* We don't reset or free kptr on uref dropping to zero. */
1459 if (!map_value_has_timer(&htab->map))
1461 if (!htab_is_prealloc(htab))
1462 htab_free_malloced_timers(htab);
1464 htab_free_prealloced_timers(htab);
1467 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1468 static void htab_map_free(struct bpf_map *map)
1470 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1473 /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1474 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1475 * There is no need to synchronize_rcu() here to protect map elements.
1478 /* some of free_htab_elem() callbacks for elements of this map may
1479 * not have executed. Wait for them.
1482 if (!htab_is_prealloc(htab)) {
1483 delete_all_elements(htab);
1485 htab_free_prealloced_kptrs(htab);
1486 prealloc_destroy(htab);
1489 bpf_map_free_kptr_off_tab(map);
1490 free_percpu(htab->extra_elems);
1491 bpf_map_area_free(htab->buckets);
1492 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1493 free_percpu(htab->map_locked[i]);
1494 lockdep_unregister_key(&htab->lockdep_key);
1495 bpf_map_area_free(htab);
1498 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1505 value = htab_map_lookup_elem(map, key);
1511 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1513 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1519 static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1520 void *value, bool is_lru_map,
1521 bool is_percpu, u64 flags)
1523 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1524 struct hlist_nulls_head *head;
1525 unsigned long bflags;
1526 struct htab_elem *l;
1531 key_size = map->key_size;
1533 hash = htab_map_hash(key, key_size, htab->hashrnd);
1534 b = __select_bucket(htab, hash);
1537 ret = htab_lock_bucket(htab, b, hash, &bflags);
1541 l = lookup_elem_raw(head, hash, key, key_size);
1546 u32 roundup_value_size = round_up(map->value_size, 8);
1547 void __percpu *pptr;
1550 pptr = htab_elem_get_ptr(l, key_size);
1551 for_each_possible_cpu(cpu) {
1552 bpf_long_memcpy(value + off,
1553 per_cpu_ptr(pptr, cpu),
1554 roundup_value_size);
1555 off += roundup_value_size;
1558 u32 roundup_key_size = round_up(map->key_size, 8);
1560 if (flags & BPF_F_LOCK)
1561 copy_map_value_locked(map, value, l->key +
1565 copy_map_value(map, value, l->key +
1567 check_and_init_map_value(map, value);
1570 hlist_nulls_del_rcu(&l->hash_node);
1572 free_htab_elem(htab, l);
1575 htab_unlock_bucket(htab, b, hash, bflags);
1577 if (is_lru_map && l)
1578 htab_lru_push_free(htab, l);
1583 static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1584 void *value, u64 flags)
1586 return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1590 static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1591 void *key, void *value,
1594 return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1598 static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1599 void *value, u64 flags)
1601 return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1605 static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1606 void *key, void *value,
1609 return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1614 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1615 const union bpf_attr *attr,
1616 union bpf_attr __user *uattr,
1617 bool do_delete, bool is_lru_map,
1620 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1621 u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1622 void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1623 void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1624 void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1625 void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1626 u32 batch, max_count, size, bucket_size, map_id;
1627 struct htab_elem *node_to_free = NULL;
1628 u64 elem_map_flags, map_flags;
1629 struct hlist_nulls_head *head;
1630 struct hlist_nulls_node *n;
1631 unsigned long flags = 0;
1632 bool locked = false;
1633 struct htab_elem *l;
1637 elem_map_flags = attr->batch.elem_flags;
1638 if ((elem_map_flags & ~BPF_F_LOCK) ||
1639 ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1642 map_flags = attr->batch.flags;
1646 max_count = attr->batch.count;
1650 if (put_user(0, &uattr->batch.count))
1654 if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1657 if (batch >= htab->n_buckets)
1660 key_size = htab->map.key_size;
1661 roundup_key_size = round_up(htab->map.key_size, 8);
1662 value_size = htab->map.value_size;
1663 size = round_up(value_size, 8);
1665 value_size = size * num_possible_cpus();
1667 /* while experimenting with hash tables with sizes ranging from 10 to
1668 * 1000, it was observed that a bucket can have up to 5 entries.
1673 /* We cannot do copy_from_user or copy_to_user inside
1674 * the rcu_read_lock. Allocate enough space here.
1676 keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1677 values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1678 if (!keys || !values) {
1684 bpf_disable_instrumentation();
1689 b = &htab->buckets[batch];
1691 /* do not grab the lock unless need it (bucket_cnt > 0). */
1693 ret = htab_lock_bucket(htab, b, batch, &flags);
1699 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1702 if (bucket_cnt && !locked) {
1707 if (bucket_cnt > (max_count - total)) {
1710 /* Note that since bucket_cnt > 0 here, it is implicit
1711 * that the locked was grabbed, so release it.
1713 htab_unlock_bucket(htab, b, batch, flags);
1715 bpf_enable_instrumentation();
1719 if (bucket_cnt > bucket_size) {
1720 bucket_size = bucket_cnt;
1721 /* Note that since bucket_cnt > 0 here, it is implicit
1722 * that the locked was grabbed, so release it.
1724 htab_unlock_bucket(htab, b, batch, flags);
1726 bpf_enable_instrumentation();
1732 /* Next block is only safe to run if you have grabbed the lock */
1736 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1737 memcpy(dst_key, l->key, key_size);
1741 void __percpu *pptr;
1743 pptr = htab_elem_get_ptr(l, map->key_size);
1744 for_each_possible_cpu(cpu) {
1745 bpf_long_memcpy(dst_val + off,
1746 per_cpu_ptr(pptr, cpu), size);
1750 value = l->key + roundup_key_size;
1751 if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
1752 struct bpf_map **inner_map = value;
1754 /* Actual value is the id of the inner map */
1755 map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
1759 if (elem_map_flags & BPF_F_LOCK)
1760 copy_map_value_locked(map, dst_val, value,
1763 copy_map_value(map, dst_val, value);
1764 check_and_init_map_value(map, dst_val);
1767 hlist_nulls_del_rcu(&l->hash_node);
1769 /* bpf_lru_push_free() will acquire lru_lock, which
1770 * may cause deadlock. See comments in function
1771 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1772 * after releasing the bucket lock.
1775 l->batch_flink = node_to_free;
1778 free_htab_elem(htab, l);
1781 dst_key += key_size;
1782 dst_val += value_size;
1785 htab_unlock_bucket(htab, b, batch, flags);
1788 while (node_to_free) {
1790 node_to_free = node_to_free->batch_flink;
1791 htab_lru_push_free(htab, l);
1795 /* If we are not copying data, we can go to next bucket and avoid
1796 * unlocking the rcu.
1798 if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1804 bpf_enable_instrumentation();
1805 if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1806 key_size * bucket_cnt) ||
1807 copy_to_user(uvalues + total * value_size, values,
1808 value_size * bucket_cnt))) {
1813 total += bucket_cnt;
1815 if (batch >= htab->n_buckets) {
1825 /* copy # of entries and next batch */
1826 ubatch = u64_to_user_ptr(attr->batch.out_batch);
1827 if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1828 put_user(total, &uattr->batch.count))
1838 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1839 union bpf_attr __user *uattr)
1841 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1846 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1847 const union bpf_attr *attr,
1848 union bpf_attr __user *uattr)
1850 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1855 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1856 union bpf_attr __user *uattr)
1858 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1863 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1864 const union bpf_attr *attr,
1865 union bpf_attr __user *uattr)
1867 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1872 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1873 const union bpf_attr *attr,
1874 union bpf_attr __user *uattr)
1876 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1881 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1882 const union bpf_attr *attr,
1883 union bpf_attr __user *uattr)
1885 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1890 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1891 union bpf_attr __user *uattr)
1893 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1898 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1899 const union bpf_attr *attr,
1900 union bpf_attr __user *uattr)
1902 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1906 struct bpf_iter_seq_hash_map_info {
1907 struct bpf_map *map;
1908 struct bpf_htab *htab;
1909 void *percpu_value_buf; // non-zero means percpu hash
1914 static struct htab_elem *
1915 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1916 struct htab_elem *prev_elem)
1918 const struct bpf_htab *htab = info->htab;
1919 u32 skip_elems = info->skip_elems;
1920 u32 bucket_id = info->bucket_id;
1921 struct hlist_nulls_head *head;
1922 struct hlist_nulls_node *n;
1923 struct htab_elem *elem;
1927 if (bucket_id >= htab->n_buckets)
1930 /* try to find next elem in the same bucket */
1932 /* no update/deletion on this bucket, prev_elem should be still valid
1933 * and we won't skip elements.
1935 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1936 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1940 /* not found, unlock and go to the next bucket */
1941 b = &htab->buckets[bucket_id++];
1946 for (i = bucket_id; i < htab->n_buckets; i++) {
1947 b = &htab->buckets[i];
1952 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1953 if (count >= skip_elems) {
1954 info->bucket_id = i;
1955 info->skip_elems = count;
1965 info->bucket_id = i;
1966 info->skip_elems = 0;
1970 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
1972 struct bpf_iter_seq_hash_map_info *info = seq->private;
1973 struct htab_elem *elem;
1975 elem = bpf_hash_map_seq_find_next(info, NULL);
1984 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1986 struct bpf_iter_seq_hash_map_info *info = seq->private;
1990 return bpf_hash_map_seq_find_next(info, v);
1993 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
1995 struct bpf_iter_seq_hash_map_info *info = seq->private;
1996 u32 roundup_key_size, roundup_value_size;
1997 struct bpf_iter__bpf_map_elem ctx = {};
1998 struct bpf_map *map = info->map;
1999 struct bpf_iter_meta meta;
2000 int ret = 0, off = 0, cpu;
2001 struct bpf_prog *prog;
2002 void __percpu *pptr;
2005 prog = bpf_iter_get_info(&meta, elem == NULL);
2008 ctx.map = info->map;
2010 roundup_key_size = round_up(map->key_size, 8);
2011 ctx.key = elem->key;
2012 if (!info->percpu_value_buf) {
2013 ctx.value = elem->key + roundup_key_size;
2015 roundup_value_size = round_up(map->value_size, 8);
2016 pptr = htab_elem_get_ptr(elem, map->key_size);
2017 for_each_possible_cpu(cpu) {
2018 bpf_long_memcpy(info->percpu_value_buf + off,
2019 per_cpu_ptr(pptr, cpu),
2020 roundup_value_size);
2021 off += roundup_value_size;
2023 ctx.value = info->percpu_value_buf;
2026 ret = bpf_iter_run_prog(prog, &ctx);
2032 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
2034 return __bpf_hash_map_seq_show(seq, v);
2037 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2040 (void)__bpf_hash_map_seq_show(seq, NULL);
2045 static int bpf_iter_init_hash_map(void *priv_data,
2046 struct bpf_iter_aux_info *aux)
2048 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2049 struct bpf_map *map = aux->map;
2053 if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2054 map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2055 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2056 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2060 seq_info->percpu_value_buf = value_buf;
2063 bpf_map_inc_with_uref(map);
2064 seq_info->map = map;
2065 seq_info->htab = container_of(map, struct bpf_htab, map);
2069 static void bpf_iter_fini_hash_map(void *priv_data)
2071 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2073 bpf_map_put_with_uref(seq_info->map);
2074 kfree(seq_info->percpu_value_buf);
2077 static const struct seq_operations bpf_hash_map_seq_ops = {
2078 .start = bpf_hash_map_seq_start,
2079 .next = bpf_hash_map_seq_next,
2080 .stop = bpf_hash_map_seq_stop,
2081 .show = bpf_hash_map_seq_show,
2084 static const struct bpf_iter_seq_info iter_seq_info = {
2085 .seq_ops = &bpf_hash_map_seq_ops,
2086 .init_seq_private = bpf_iter_init_hash_map,
2087 .fini_seq_private = bpf_iter_fini_hash_map,
2088 .seq_priv_size = sizeof(struct bpf_iter_seq_hash_map_info),
2091 static int bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2092 void *callback_ctx, u64 flags)
2094 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2095 struct hlist_nulls_head *head;
2096 struct hlist_nulls_node *n;
2097 struct htab_elem *elem;
2098 u32 roundup_key_size;
2099 int i, num_elems = 0;
2100 void __percpu *pptr;
2109 is_percpu = htab_is_percpu(htab);
2111 roundup_key_size = round_up(map->key_size, 8);
2112 /* disable migration so percpu value prepared here will be the
2113 * same as the one seen by the bpf program with bpf_map_lookup_elem().
2117 for (i = 0; i < htab->n_buckets; i++) {
2118 b = &htab->buckets[i];
2121 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2124 /* current cpu value for percpu map */
2125 pptr = htab_elem_get_ptr(elem, map->key_size);
2126 val = this_cpu_ptr(pptr);
2128 val = elem->key + roundup_key_size;
2131 ret = callback_fn((u64)(long)map, (u64)(long)key,
2132 (u64)(long)val, (u64)(long)callback_ctx, 0);
2133 /* return value: 0 - continue, 1 - stop and return */
2147 BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
2148 const struct bpf_map_ops htab_map_ops = {
2149 .map_meta_equal = bpf_map_meta_equal,
2150 .map_alloc_check = htab_map_alloc_check,
2151 .map_alloc = htab_map_alloc,
2152 .map_free = htab_map_free,
2153 .map_get_next_key = htab_map_get_next_key,
2154 .map_release_uref = htab_map_free_timers,
2155 .map_lookup_elem = htab_map_lookup_elem,
2156 .map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2157 .map_update_elem = htab_map_update_elem,
2158 .map_delete_elem = htab_map_delete_elem,
2159 .map_gen_lookup = htab_map_gen_lookup,
2160 .map_seq_show_elem = htab_map_seq_show_elem,
2161 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2162 .map_for_each_callback = bpf_for_each_hash_elem,
2164 .map_btf_id = &htab_map_btf_ids[0],
2165 .iter_seq_info = &iter_seq_info,
2168 const struct bpf_map_ops htab_lru_map_ops = {
2169 .map_meta_equal = bpf_map_meta_equal,
2170 .map_alloc_check = htab_map_alloc_check,
2171 .map_alloc = htab_map_alloc,
2172 .map_free = htab_map_free,
2173 .map_get_next_key = htab_map_get_next_key,
2174 .map_release_uref = htab_map_free_timers,
2175 .map_lookup_elem = htab_lru_map_lookup_elem,
2176 .map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2177 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2178 .map_update_elem = htab_lru_map_update_elem,
2179 .map_delete_elem = htab_lru_map_delete_elem,
2180 .map_gen_lookup = htab_lru_map_gen_lookup,
2181 .map_seq_show_elem = htab_map_seq_show_elem,
2182 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2183 .map_for_each_callback = bpf_for_each_hash_elem,
2184 BATCH_OPS(htab_lru),
2185 .map_btf_id = &htab_map_btf_ids[0],
2186 .iter_seq_info = &iter_seq_info,
2189 /* Called from eBPF program */
2190 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2192 struct htab_elem *l = __htab_map_lookup_elem(map, key);
2195 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2200 static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2202 struct htab_elem *l;
2204 if (cpu >= nr_cpu_ids)
2207 l = __htab_map_lookup_elem(map, key);
2209 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2214 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2216 struct htab_elem *l = __htab_map_lookup_elem(map, key);
2219 bpf_lru_node_set_ref(&l->lru_node);
2220 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2226 static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2228 struct htab_elem *l;
2230 if (cpu >= nr_cpu_ids)
2233 l = __htab_map_lookup_elem(map, key);
2235 bpf_lru_node_set_ref(&l->lru_node);
2236 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2242 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2244 struct htab_elem *l;
2245 void __percpu *pptr;
2250 /* per_cpu areas are zero-filled and bpf programs can only
2251 * access 'value_size' of them, so copying rounded areas
2252 * will not leak any kernel data
2254 size = round_up(map->value_size, 8);
2256 l = __htab_map_lookup_elem(map, key);
2259 /* We do not mark LRU map element here in order to not mess up
2260 * eviction heuristics when user space does a map walk.
2262 pptr = htab_elem_get_ptr(l, map->key_size);
2263 for_each_possible_cpu(cpu) {
2264 bpf_long_memcpy(value + off,
2265 per_cpu_ptr(pptr, cpu), size);
2274 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2277 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2281 if (htab_is_lru(htab))
2282 ret = __htab_lru_percpu_map_update_elem(map, key, value,
2285 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2292 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2295 struct htab_elem *l;
2296 void __percpu *pptr;
2301 l = __htab_map_lookup_elem(map, key);
2307 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2308 seq_puts(m, ": {\n");
2309 pptr = htab_elem_get_ptr(l, map->key_size);
2310 for_each_possible_cpu(cpu) {
2311 seq_printf(m, "\tcpu%d: ", cpu);
2312 btf_type_seq_show(map->btf, map->btf_value_type_id,
2313 per_cpu_ptr(pptr, cpu), m);
2321 const struct bpf_map_ops htab_percpu_map_ops = {
2322 .map_meta_equal = bpf_map_meta_equal,
2323 .map_alloc_check = htab_map_alloc_check,
2324 .map_alloc = htab_map_alloc,
2325 .map_free = htab_map_free,
2326 .map_get_next_key = htab_map_get_next_key,
2327 .map_lookup_elem = htab_percpu_map_lookup_elem,
2328 .map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2329 .map_update_elem = htab_percpu_map_update_elem,
2330 .map_delete_elem = htab_map_delete_elem,
2331 .map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
2332 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2333 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2334 .map_for_each_callback = bpf_for_each_hash_elem,
2335 BATCH_OPS(htab_percpu),
2336 .map_btf_id = &htab_map_btf_ids[0],
2337 .iter_seq_info = &iter_seq_info,
2340 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2341 .map_meta_equal = bpf_map_meta_equal,
2342 .map_alloc_check = htab_map_alloc_check,
2343 .map_alloc = htab_map_alloc,
2344 .map_free = htab_map_free,
2345 .map_get_next_key = htab_map_get_next_key,
2346 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2347 .map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2348 .map_update_elem = htab_lru_percpu_map_update_elem,
2349 .map_delete_elem = htab_lru_map_delete_elem,
2350 .map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
2351 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2352 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2353 .map_for_each_callback = bpf_for_each_hash_elem,
2354 BATCH_OPS(htab_lru_percpu),
2355 .map_btf_id = &htab_map_btf_ids[0],
2356 .iter_seq_info = &iter_seq_info,
2359 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2361 if (attr->value_size != sizeof(u32))
2363 return htab_map_alloc_check(attr);
2366 static void fd_htab_map_free(struct bpf_map *map)
2368 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2369 struct hlist_nulls_node *n;
2370 struct hlist_nulls_head *head;
2371 struct htab_elem *l;
2374 for (i = 0; i < htab->n_buckets; i++) {
2375 head = select_bucket(htab, i);
2377 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2378 void *ptr = fd_htab_map_get_ptr(map, l);
2380 map->ops->map_fd_put_ptr(ptr);
2387 /* only called from syscall */
2388 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2393 if (!map->ops->map_fd_sys_lookup_elem)
2397 ptr = htab_map_lookup_elem(map, key);
2399 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2407 /* only called from syscall */
2408 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2409 void *key, void *value, u64 map_flags)
2413 u32 ufd = *(u32 *)value;
2415 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2417 return PTR_ERR(ptr);
2419 ret = htab_map_update_elem(map, key, &ptr, map_flags);
2421 map->ops->map_fd_put_ptr(ptr);
2426 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2428 struct bpf_map *map, *inner_map_meta;
2430 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2431 if (IS_ERR(inner_map_meta))
2432 return inner_map_meta;
2434 map = htab_map_alloc(attr);
2436 bpf_map_meta_free(inner_map_meta);
2440 map->inner_map_meta = inner_map_meta;
2445 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2447 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
2452 return READ_ONCE(*inner_map);
2455 static int htab_of_map_gen_lookup(struct bpf_map *map,
2456 struct bpf_insn *insn_buf)
2458 struct bpf_insn *insn = insn_buf;
2459 const int ret = BPF_REG_0;
2461 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2462 (void *(*)(struct bpf_map *map, void *key))NULL));
2463 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2464 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2465 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2466 offsetof(struct htab_elem, key) +
2467 round_up(map->key_size, 8));
2468 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2470 return insn - insn_buf;
2473 static void htab_of_map_free(struct bpf_map *map)
2475 bpf_map_meta_free(map->inner_map_meta);
2476 fd_htab_map_free(map);
2479 const struct bpf_map_ops htab_of_maps_map_ops = {
2480 .map_alloc_check = fd_htab_map_alloc_check,
2481 .map_alloc = htab_of_map_alloc,
2482 .map_free = htab_of_map_free,
2483 .map_get_next_key = htab_map_get_next_key,
2484 .map_lookup_elem = htab_of_map_lookup_elem,
2485 .map_delete_elem = htab_map_delete_elem,
2486 .map_fd_get_ptr = bpf_map_fd_get_ptr,
2487 .map_fd_put_ptr = bpf_map_fd_put_ptr,
2488 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2489 .map_gen_lookup = htab_of_map_gen_lookup,
2490 .map_check_btf = map_check_no_btf,
2492 .map_btf_id = &htab_map_btf_ids[0],