1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
7 #include <linux/jhash.h>
8 #include <linux/filter.h>
9 #include <linux/rculist_nulls.h>
10 #include <linux/random.h>
11 #include <uapi/linux/btf.h>
12 #include <linux/rcupdate_trace.h>
13 #include "percpu_freelist.h"
14 #include "bpf_lru_list.h"
15 #include "map_in_map.h"
17 #define HTAB_CREATE_FLAG_MASK \
18 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
19 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
21 #define BATCH_OPS(_name) \
23 _name##_map_lookup_batch, \
24 .map_lookup_and_delete_batch = \
25 _name##_map_lookup_and_delete_batch, \
27 generic_map_update_batch, \
29 generic_map_delete_batch
32 * The bucket lock has two protection scopes:
34 * 1) Serializing concurrent operations from BPF programs on differrent
37 * 2) Serializing concurrent operations from BPF programs and sys_bpf()
39 * BPF programs can execute in any context including perf, kprobes and
40 * tracing. As there are almost no limits where perf, kprobes and tracing
41 * can be invoked from the lock operations need to be protected against
42 * deadlocks. Deadlocks can be caused by recursion and by an invocation in
43 * the lock held section when functions which acquire this lock are invoked
44 * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
45 * variable bpf_prog_active, which prevents BPF programs attached to perf
46 * events, kprobes and tracing to be invoked before the prior invocation
47 * from one of these contexts completed. sys_bpf() uses the same mechanism
48 * by pinning the task to the current CPU and incrementing the recursion
49 * protection accross the map operation.
51 * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
52 * operations like memory allocations (even with GFP_ATOMIC) from atomic
53 * contexts. This is required because even with GFP_ATOMIC the memory
54 * allocator calls into code pathes which acquire locks with long held lock
55 * sections. To ensure the deterministic behaviour these locks are regular
56 * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
57 * true atomic contexts on an RT kernel are the low level hardware
58 * handling, scheduling, low level interrupt handling, NMIs etc. None of
59 * these contexts should ever do memory allocations.
61 * As regular device interrupt handlers and soft interrupts are forced into
62 * thread context, the existing code which does
63 * spin_lock*(); alloc(GPF_ATOMIC); spin_unlock*();
66 * In theory the BPF locks could be converted to regular spinlocks as well,
67 * but the bucket locks and percpu_freelist locks can be taken from
68 * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
69 * atomic contexts even on RT. These mechanisms require preallocated maps,
70 * so there is no need to invoke memory allocations within the lock held
73 * BPF maps which need dynamic allocation are only used from (forced)
74 * thread context on RT and can therefore use regular spinlocks which in
75 * turn allows to invoke memory allocations from the lock held section.
77 * On a non RT kernel this distinction is neither possible nor required.
78 * spinlock maps to raw_spinlock and the extra code is optimized out by the
82 struct hlist_nulls_head head;
84 raw_spinlock_t raw_lock;
91 struct bucket *buckets;
94 struct pcpu_freelist freelist;
97 struct htab_elem *__percpu *extra_elems;
98 atomic_t count; /* number of elements in this hashtable */
99 u32 n_buckets; /* number of hash buckets */
100 u32 elem_size; /* size of each element in bytes */
104 /* each htab element is struct htab_elem + key + value */
107 struct hlist_nulls_node hash_node;
111 struct bpf_htab *htab;
112 struct pcpu_freelist_node fnode;
113 struct htab_elem *batch_flink;
119 struct bpf_lru_node lru_node;
122 char key[] __aligned(8);
125 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
127 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
130 static inline bool htab_use_raw_lock(const struct bpf_htab *htab)
132 return (!IS_ENABLED(CONFIG_PREEMPT_RT) || htab_is_prealloc(htab));
135 static void htab_init_buckets(struct bpf_htab *htab)
139 for (i = 0; i < htab->n_buckets; i++) {
140 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
141 if (htab_use_raw_lock(htab))
142 raw_spin_lock_init(&htab->buckets[i].raw_lock);
144 spin_lock_init(&htab->buckets[i].lock);
148 static inline unsigned long htab_lock_bucket(const struct bpf_htab *htab,
153 if (htab_use_raw_lock(htab))
154 raw_spin_lock_irqsave(&b->raw_lock, flags);
156 spin_lock_irqsave(&b->lock, flags);
160 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
164 if (htab_use_raw_lock(htab))
165 raw_spin_unlock_irqrestore(&b->raw_lock, flags);
167 spin_unlock_irqrestore(&b->lock, flags);
170 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
172 static bool htab_is_lru(const struct bpf_htab *htab)
174 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
175 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
178 static bool htab_is_percpu(const struct bpf_htab *htab)
180 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
181 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
184 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
187 *(void __percpu **)(l->key + key_size) = pptr;
190 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
192 return *(void __percpu **)(l->key + key_size);
195 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
197 return *(void **)(l->key + roundup(map->key_size, 8));
200 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
202 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
205 static void htab_free_elems(struct bpf_htab *htab)
209 if (!htab_is_percpu(htab))
212 for (i = 0; i < htab->map.max_entries; i++) {
215 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
221 bpf_map_area_free(htab->elems);
224 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
225 * (bucket_lock). If both locks need to be acquired together, the lock
226 * order is always lru_lock -> bucket_lock and this only happens in
227 * bpf_lru_list.c logic. For example, certain code path of
228 * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
229 * will acquire lru_lock first followed by acquiring bucket_lock.
231 * In hashtab.c, to avoid deadlock, lock acquisition of
232 * bucket_lock followed by lru_lock is not allowed. In such cases,
233 * bucket_lock needs to be released first before acquiring lru_lock.
235 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
238 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
242 l = container_of(node, struct htab_elem, lru_node);
243 memcpy(l->key, key, htab->map.key_size);
250 static int prealloc_init(struct bpf_htab *htab)
252 u32 num_entries = htab->map.max_entries;
253 int err = -ENOMEM, i;
255 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
256 num_entries += num_possible_cpus();
258 htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
259 htab->map.numa_node);
263 if (!htab_is_percpu(htab))
264 goto skip_percpu_elems;
266 for (i = 0; i < num_entries; i++) {
267 u32 size = round_up(htab->map.value_size, 8);
270 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
273 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
279 if (htab_is_lru(htab))
280 err = bpf_lru_init(&htab->lru,
281 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
282 offsetof(struct htab_elem, hash) -
283 offsetof(struct htab_elem, lru_node),
284 htab_lru_map_delete_node,
287 err = pcpu_freelist_init(&htab->freelist);
292 if (htab_is_lru(htab))
293 bpf_lru_populate(&htab->lru, htab->elems,
294 offsetof(struct htab_elem, lru_node),
295 htab->elem_size, num_entries);
297 pcpu_freelist_populate(&htab->freelist,
298 htab->elems + offsetof(struct htab_elem, fnode),
299 htab->elem_size, num_entries);
304 htab_free_elems(htab);
308 static void prealloc_destroy(struct bpf_htab *htab)
310 htab_free_elems(htab);
312 if (htab_is_lru(htab))
313 bpf_lru_destroy(&htab->lru);
315 pcpu_freelist_destroy(&htab->freelist);
318 static int alloc_extra_elems(struct bpf_htab *htab)
320 struct htab_elem *__percpu *pptr, *l_new;
321 struct pcpu_freelist_node *l;
324 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
325 GFP_USER | __GFP_NOWARN);
329 for_each_possible_cpu(cpu) {
330 l = pcpu_freelist_pop(&htab->freelist);
331 /* pop will succeed, since prealloc_init()
332 * preallocated extra num_possible_cpus elements
334 l_new = container_of(l, struct htab_elem, fnode);
335 *per_cpu_ptr(pptr, cpu) = l_new;
337 htab->extra_elems = pptr;
341 /* Called from syscall */
342 static int htab_map_alloc_check(union bpf_attr *attr)
344 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
345 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
346 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
347 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
348 /* percpu_lru means each cpu has its own LRU list.
349 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
350 * the map's value itself is percpu. percpu_lru has
351 * nothing to do with the map's value.
353 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
354 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
355 bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
356 int numa_node = bpf_map_attr_numa_node(attr);
358 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
359 offsetof(struct htab_elem, hash_node.pprev));
360 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
361 offsetof(struct htab_elem, hash_node.pprev));
363 if (lru && !bpf_capable())
364 /* LRU implementation is much complicated than other
365 * maps. Hence, limit to CAP_BPF.
369 if (zero_seed && !capable(CAP_SYS_ADMIN))
370 /* Guard against local DoS, and discourage production use. */
373 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
374 !bpf_map_flags_access_ok(attr->map_flags))
377 if (!lru && percpu_lru)
380 if (lru && !prealloc)
383 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
386 /* check sanity of attributes.
387 * value_size == 0 may be allowed in the future to use map as a set
389 if (attr->max_entries == 0 || attr->key_size == 0 ||
390 attr->value_size == 0)
393 if (attr->key_size > MAX_BPF_STACK)
394 /* eBPF programs initialize keys on stack, so they cannot be
395 * larger than max stack size
399 if (attr->value_size >= KMALLOC_MAX_SIZE -
400 MAX_BPF_STACK - sizeof(struct htab_elem))
401 /* if value_size is bigger, the user space won't be able to
402 * access the elements via bpf syscall. This check also makes
403 * sure that the elem_size doesn't overflow and it's
404 * kmalloc-able later in htab_map_update_elem()
411 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
413 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
414 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
415 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
416 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
417 /* percpu_lru means each cpu has its own LRU list.
418 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
419 * the map's value itself is percpu. percpu_lru has
420 * nothing to do with the map's value.
422 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
423 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
424 struct bpf_htab *htab;
428 htab = kzalloc(sizeof(*htab), GFP_USER);
430 return ERR_PTR(-ENOMEM);
432 bpf_map_init_from_attr(&htab->map, attr);
435 /* ensure each CPU's lru list has >=1 elements.
436 * since we are at it, make each lru list has the same
437 * number of elements.
439 htab->map.max_entries = roundup(attr->max_entries,
440 num_possible_cpus());
441 if (htab->map.max_entries < attr->max_entries)
442 htab->map.max_entries = rounddown(attr->max_entries,
443 num_possible_cpus());
446 /* hash table size must be power of 2 */
447 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
449 htab->elem_size = sizeof(struct htab_elem) +
450 round_up(htab->map.key_size, 8);
452 htab->elem_size += sizeof(void *);
454 htab->elem_size += round_up(htab->map.value_size, 8);
457 /* prevent zero size kmalloc and check for u32 overflow */
458 if (htab->n_buckets == 0 ||
459 htab->n_buckets > U32_MAX / sizeof(struct bucket))
462 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
463 (u64) htab->elem_size * htab->map.max_entries;
466 cost += (u64) round_up(htab->map.value_size, 8) *
467 num_possible_cpus() * htab->map.max_entries;
469 cost += (u64) htab->elem_size * num_possible_cpus();
471 /* if map size is larger than memlock limit, reject it */
472 err = bpf_map_charge_init(&htab->map.memory, cost);
477 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
478 sizeof(struct bucket),
479 htab->map.numa_node);
483 if (htab->map.map_flags & BPF_F_ZERO_SEED)
486 htab->hashrnd = get_random_int();
488 htab_init_buckets(htab);
491 err = prealloc_init(htab);
495 if (!percpu && !lru) {
496 /* lru itself can remove the least used element, so
497 * there is no need for an extra elem during map_update.
499 err = alloc_extra_elems(htab);
508 prealloc_destroy(htab);
510 bpf_map_area_free(htab->buckets);
512 bpf_map_charge_finish(&htab->map.memory);
518 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
520 return jhash(key, key_len, hashrnd);
523 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
525 return &htab->buckets[hash & (htab->n_buckets - 1)];
528 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
530 return &__select_bucket(htab, hash)->head;
533 /* this lookup function can only be called with bucket lock taken */
534 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
535 void *key, u32 key_size)
537 struct hlist_nulls_node *n;
540 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
541 if (l->hash == hash && !memcmp(&l->key, key, key_size))
547 /* can be called without bucket lock. it will repeat the loop in
548 * the unlikely event when elements moved from one bucket into another
549 * while link list is being walked
551 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
553 u32 key_size, u32 n_buckets)
555 struct hlist_nulls_node *n;
559 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
560 if (l->hash == hash && !memcmp(&l->key, key, key_size))
563 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
569 /* Called from syscall or from eBPF program directly, so
570 * arguments have to match bpf_map_lookup_elem() exactly.
571 * The return value is adjusted by BPF instructions
572 * in htab_map_gen_lookup().
574 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
576 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
577 struct hlist_nulls_head *head;
581 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
583 key_size = map->key_size;
585 hash = htab_map_hash(key, key_size, htab->hashrnd);
587 head = select_bucket(htab, hash);
589 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
594 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
596 struct htab_elem *l = __htab_map_lookup_elem(map, key);
599 return l->key + round_up(map->key_size, 8);
604 /* inline bpf_map_lookup_elem() call.
607 * bpf_map_lookup_elem
608 * map->ops->map_lookup_elem
609 * htab_map_lookup_elem
610 * __htab_map_lookup_elem
613 * __htab_map_lookup_elem
615 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
617 struct bpf_insn *insn = insn_buf;
618 const int ret = BPF_REG_0;
620 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
621 (void *(*)(struct bpf_map *map, void *key))NULL));
622 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
623 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
624 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
625 offsetof(struct htab_elem, key) +
626 round_up(map->key_size, 8));
627 return insn - insn_buf;
630 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
631 void *key, const bool mark)
633 struct htab_elem *l = __htab_map_lookup_elem(map, key);
637 bpf_lru_node_set_ref(&l->lru_node);
638 return l->key + round_up(map->key_size, 8);
644 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
646 return __htab_lru_map_lookup_elem(map, key, true);
649 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
651 return __htab_lru_map_lookup_elem(map, key, false);
654 static int htab_lru_map_gen_lookup(struct bpf_map *map,
655 struct bpf_insn *insn_buf)
657 struct bpf_insn *insn = insn_buf;
658 const int ret = BPF_REG_0;
659 const int ref_reg = BPF_REG_1;
661 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
662 (void *(*)(struct bpf_map *map, void *key))NULL));
663 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
664 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
665 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
666 offsetof(struct htab_elem, lru_node) +
667 offsetof(struct bpf_lru_node, ref));
668 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
669 *insn++ = BPF_ST_MEM(BPF_B, ret,
670 offsetof(struct htab_elem, lru_node) +
671 offsetof(struct bpf_lru_node, ref),
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 /* It is called from the bpf_lru_list when the LRU needs to delete
680 * older elements from the htab.
682 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
684 struct bpf_htab *htab = (struct bpf_htab *)arg;
685 struct htab_elem *l = NULL, *tgt_l;
686 struct hlist_nulls_head *head;
687 struct hlist_nulls_node *n;
691 tgt_l = container_of(node, struct htab_elem, lru_node);
692 b = __select_bucket(htab, tgt_l->hash);
695 flags = htab_lock_bucket(htab, b);
697 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
699 hlist_nulls_del_rcu(&l->hash_node);
703 htab_unlock_bucket(htab, b, flags);
708 /* Called from syscall */
709 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
711 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
712 struct hlist_nulls_head *head;
713 struct htab_elem *l, *next_l;
717 WARN_ON_ONCE(!rcu_read_lock_held());
719 key_size = map->key_size;
722 goto find_first_elem;
724 hash = htab_map_hash(key, key_size, htab->hashrnd);
726 head = select_bucket(htab, hash);
729 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
732 goto find_first_elem;
734 /* key was found, get next key in the same bucket */
735 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
736 struct htab_elem, hash_node);
739 /* if next elem in this hash list is non-zero, just return it */
740 memcpy(next_key, next_l->key, key_size);
744 /* no more elements in this hash list, go to the next bucket */
745 i = hash & (htab->n_buckets - 1);
749 /* iterate over buckets */
750 for (; i < htab->n_buckets; i++) {
751 head = select_bucket(htab, i);
753 /* pick first element in the bucket */
754 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
755 struct htab_elem, hash_node);
757 /* if it's not empty, just return it */
758 memcpy(next_key, next_l->key, key_size);
763 /* iterated over all buckets and all elements */
767 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
769 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
770 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
774 static void htab_elem_free_rcu(struct rcu_head *head)
776 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
777 struct bpf_htab *htab = l->htab;
779 htab_elem_free(htab, l);
782 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
784 struct bpf_map *map = &htab->map;
787 if (map->ops->map_fd_put_ptr) {
788 ptr = fd_htab_map_get_ptr(map, l);
789 map->ops->map_fd_put_ptr(ptr);
793 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
795 htab_put_fd_value(htab, l);
797 if (htab_is_prealloc(htab)) {
798 __pcpu_freelist_push(&htab->freelist, &l->fnode);
800 atomic_dec(&htab->count);
802 call_rcu(&l->rcu, htab_elem_free_rcu);
806 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
807 void *value, bool onallcpus)
810 /* copy true value_size bytes */
811 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
813 u32 size = round_up(htab->map.value_size, 8);
816 for_each_possible_cpu(cpu) {
817 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
824 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
825 void *value, bool onallcpus)
827 /* When using prealloc and not setting the initial value on all cpus,
828 * zero-fill element values for other cpus (just as what happens when
829 * not using prealloc). Otherwise, bpf program has no way to ensure
830 * known initial values for cpus other than current one
831 * (onallcpus=false always when coming from bpf prog).
833 if (htab_is_prealloc(htab) && !onallcpus) {
834 u32 size = round_up(htab->map.value_size, 8);
835 int current_cpu = raw_smp_processor_id();
838 for_each_possible_cpu(cpu) {
839 if (cpu == current_cpu)
840 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value,
843 memset(per_cpu_ptr(pptr, cpu), 0, size);
846 pcpu_copy_value(htab, pptr, value, onallcpus);
850 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
852 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
856 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
857 void *value, u32 key_size, u32 hash,
858 bool percpu, bool onallcpus,
859 struct htab_elem *old_elem)
861 u32 size = htab->map.value_size;
862 bool prealloc = htab_is_prealloc(htab);
863 struct htab_elem *l_new, **pl_new;
868 /* if we're updating the existing element,
869 * use per-cpu extra elems to avoid freelist_pop/push
871 pl_new = this_cpu_ptr(htab->extra_elems);
873 htab_put_fd_value(htab, old_elem);
876 struct pcpu_freelist_node *l;
878 l = __pcpu_freelist_pop(&htab->freelist);
880 return ERR_PTR(-E2BIG);
881 l_new = container_of(l, struct htab_elem, fnode);
884 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
886 /* when map is full and update() is replacing
887 * old element, it's ok to allocate, since
888 * old element will be freed immediately.
889 * Otherwise return an error
891 l_new = ERR_PTR(-E2BIG);
894 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
895 htab->map.numa_node);
897 l_new = ERR_PTR(-ENOMEM);
900 check_and_init_map_lock(&htab->map,
901 l_new->key + round_up(key_size, 8));
904 memcpy(l_new->key, key, key_size);
906 size = round_up(size, 8);
908 pptr = htab_elem_get_ptr(l_new, key_size);
910 /* alloc_percpu zero-fills */
911 pptr = __alloc_percpu_gfp(size, 8,
912 GFP_ATOMIC | __GFP_NOWARN);
915 l_new = ERR_PTR(-ENOMEM);
920 pcpu_init_value(htab, pptr, value, onallcpus);
923 htab_elem_set_ptr(l_new, key_size, pptr);
924 } else if (fd_htab_map_needs_adjust(htab)) {
925 size = round_up(size, 8);
926 memcpy(l_new->key + round_up(key_size, 8), value, size);
928 copy_map_value(&htab->map,
929 l_new->key + round_up(key_size, 8),
936 atomic_dec(&htab->count);
940 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
943 if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
944 /* elem already exists */
947 if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
948 /* elem doesn't exist, cannot update it */
954 /* Called from syscall or from eBPF program */
955 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
958 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
959 struct htab_elem *l_new = NULL, *l_old;
960 struct hlist_nulls_head *head;
966 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
970 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
972 key_size = map->key_size;
974 hash = htab_map_hash(key, key_size, htab->hashrnd);
976 b = __select_bucket(htab, hash);
979 if (unlikely(map_flags & BPF_F_LOCK)) {
980 if (unlikely(!map_value_has_spin_lock(map)))
982 /* find an element without taking the bucket lock */
983 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
985 ret = check_flags(htab, l_old, map_flags);
989 /* grab the element lock and update value in place */
990 copy_map_value_locked(map,
991 l_old->key + round_up(key_size, 8),
995 /* fall through, grab the bucket lock and lookup again.
996 * 99.9% chance that the element won't be found,
997 * but second lookup under lock has to be done.
1001 flags = htab_lock_bucket(htab, b);
1003 l_old = lookup_elem_raw(head, hash, key, key_size);
1005 ret = check_flags(htab, l_old, map_flags);
1009 if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1010 /* first lookup without the bucket lock didn't find the element,
1011 * but second lookup with the bucket lock found it.
1012 * This case is highly unlikely, but has to be dealt with:
1013 * grab the element lock in addition to the bucket lock
1014 * and update element in place
1016 copy_map_value_locked(map,
1017 l_old->key + round_up(key_size, 8),
1023 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1025 if (IS_ERR(l_new)) {
1026 /* all pre-allocated elements are in use or memory exhausted */
1027 ret = PTR_ERR(l_new);
1031 /* add new element to the head of the list, so that
1032 * concurrent search will find it before old elem
1034 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1036 hlist_nulls_del_rcu(&l_old->hash_node);
1037 if (!htab_is_prealloc(htab))
1038 free_htab_elem(htab, l_old);
1042 htab_unlock_bucket(htab, b, flags);
1046 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1049 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1050 struct htab_elem *l_new, *l_old = NULL;
1051 struct hlist_nulls_head *head;
1052 unsigned long flags;
1057 if (unlikely(map_flags > BPF_EXIST))
1061 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1063 key_size = map->key_size;
1065 hash = htab_map_hash(key, key_size, htab->hashrnd);
1067 b = __select_bucket(htab, hash);
1070 /* For LRU, we need to alloc before taking bucket's
1071 * spinlock because getting free nodes from LRU may need
1072 * to remove older elements from htab and this removal
1073 * operation will need a bucket lock.
1075 l_new = prealloc_lru_pop(htab, key, hash);
1078 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
1080 flags = htab_lock_bucket(htab, b);
1082 l_old = lookup_elem_raw(head, hash, key, key_size);
1084 ret = check_flags(htab, l_old, map_flags);
1088 /* add new element to the head of the list, so that
1089 * concurrent search will find it before old elem
1091 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1093 bpf_lru_node_set_ref(&l_new->lru_node);
1094 hlist_nulls_del_rcu(&l_old->hash_node);
1099 htab_unlock_bucket(htab, b, flags);
1102 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1104 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
1109 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1110 void *value, u64 map_flags,
1113 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1114 struct htab_elem *l_new = NULL, *l_old;
1115 struct hlist_nulls_head *head;
1116 unsigned long flags;
1121 if (unlikely(map_flags > BPF_EXIST))
1125 WARN_ON_ONCE(!rcu_read_lock_held());
1127 key_size = map->key_size;
1129 hash = htab_map_hash(key, key_size, htab->hashrnd);
1131 b = __select_bucket(htab, hash);
1134 flags = htab_lock_bucket(htab, b);
1136 l_old = lookup_elem_raw(head, hash, key, key_size);
1138 ret = check_flags(htab, l_old, map_flags);
1143 /* per-cpu hash map can update value in-place */
1144 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1147 l_new = alloc_htab_elem(htab, key, value, key_size,
1148 hash, true, onallcpus, NULL);
1149 if (IS_ERR(l_new)) {
1150 ret = PTR_ERR(l_new);
1153 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1157 htab_unlock_bucket(htab, b, flags);
1161 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1162 void *value, u64 map_flags,
1165 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1166 struct htab_elem *l_new = NULL, *l_old;
1167 struct hlist_nulls_head *head;
1168 unsigned long flags;
1173 if (unlikely(map_flags > BPF_EXIST))
1177 WARN_ON_ONCE(!rcu_read_lock_held());
1179 key_size = map->key_size;
1181 hash = htab_map_hash(key, key_size, htab->hashrnd);
1183 b = __select_bucket(htab, hash);
1186 /* For LRU, we need to alloc before taking bucket's
1187 * spinlock because LRU's elem alloc may need
1188 * to remove older elem from htab and this removal
1189 * operation will need a bucket lock.
1191 if (map_flags != BPF_EXIST) {
1192 l_new = prealloc_lru_pop(htab, key, hash);
1197 flags = htab_lock_bucket(htab, b);
1199 l_old = lookup_elem_raw(head, hash, key, key_size);
1201 ret = check_flags(htab, l_old, map_flags);
1206 bpf_lru_node_set_ref(&l_old->lru_node);
1208 /* per-cpu hash map can update value in-place */
1209 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1212 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1214 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1219 htab_unlock_bucket(htab, b, flags);
1221 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1225 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1226 void *value, u64 map_flags)
1228 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1231 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1232 void *value, u64 map_flags)
1234 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1238 /* Called from syscall or from eBPF program */
1239 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1241 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1242 struct hlist_nulls_head *head;
1244 struct htab_elem *l;
1245 unsigned long flags;
1249 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1251 key_size = map->key_size;
1253 hash = htab_map_hash(key, key_size, htab->hashrnd);
1254 b = __select_bucket(htab, hash);
1257 flags = htab_lock_bucket(htab, b);
1259 l = lookup_elem_raw(head, hash, key, key_size);
1262 hlist_nulls_del_rcu(&l->hash_node);
1263 free_htab_elem(htab, l);
1267 htab_unlock_bucket(htab, b, flags);
1271 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1273 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1274 struct hlist_nulls_head *head;
1276 struct htab_elem *l;
1277 unsigned long flags;
1281 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1283 key_size = map->key_size;
1285 hash = htab_map_hash(key, key_size, htab->hashrnd);
1286 b = __select_bucket(htab, hash);
1289 flags = htab_lock_bucket(htab, b);
1291 l = lookup_elem_raw(head, hash, key, key_size);
1294 hlist_nulls_del_rcu(&l->hash_node);
1298 htab_unlock_bucket(htab, b, flags);
1300 bpf_lru_push_free(&htab->lru, &l->lru_node);
1304 static void delete_all_elements(struct bpf_htab *htab)
1308 for (i = 0; i < htab->n_buckets; i++) {
1309 struct hlist_nulls_head *head = select_bucket(htab, i);
1310 struct hlist_nulls_node *n;
1311 struct htab_elem *l;
1313 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1314 hlist_nulls_del_rcu(&l->hash_node);
1315 htab_elem_free(htab, l);
1320 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1321 static void htab_map_free(struct bpf_map *map)
1323 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1325 /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1326 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1327 * There is no need to synchronize_rcu() here to protect map elements.
1330 /* some of free_htab_elem() callbacks for elements of this map may
1331 * not have executed. Wait for them.
1334 if (!htab_is_prealloc(htab))
1335 delete_all_elements(htab);
1337 prealloc_destroy(htab);
1339 free_percpu(htab->extra_elems);
1340 bpf_map_area_free(htab->buckets);
1344 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1351 value = htab_map_lookup_elem(map, key);
1357 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1359 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1366 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1367 const union bpf_attr *attr,
1368 union bpf_attr __user *uattr,
1369 bool do_delete, bool is_lru_map,
1372 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1373 u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1374 void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1375 void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1376 void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1377 void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1378 u32 batch, max_count, size, bucket_size;
1379 struct htab_elem *node_to_free = NULL;
1380 u64 elem_map_flags, map_flags;
1381 struct hlist_nulls_head *head;
1382 struct hlist_nulls_node *n;
1383 unsigned long flags = 0;
1384 bool locked = false;
1385 struct htab_elem *l;
1389 elem_map_flags = attr->batch.elem_flags;
1390 if ((elem_map_flags & ~BPF_F_LOCK) ||
1391 ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1394 map_flags = attr->batch.flags;
1398 max_count = attr->batch.count;
1402 if (put_user(0, &uattr->batch.count))
1406 if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1409 if (batch >= htab->n_buckets)
1412 key_size = htab->map.key_size;
1413 roundup_key_size = round_up(htab->map.key_size, 8);
1414 value_size = htab->map.value_size;
1415 size = round_up(value_size, 8);
1417 value_size = size * num_possible_cpus();
1419 /* while experimenting with hash tables with sizes ranging from 10 to
1420 * 1000, it was observed that a bucket can have upto 5 entries.
1425 /* We cannot do copy_from_user or copy_to_user inside
1426 * the rcu_read_lock. Allocate enough space here.
1428 keys = kvmalloc(key_size * bucket_size, GFP_USER | __GFP_NOWARN);
1429 values = kvmalloc(value_size * bucket_size, GFP_USER | __GFP_NOWARN);
1430 if (!keys || !values) {
1436 bpf_disable_instrumentation();
1441 b = &htab->buckets[batch];
1443 /* do not grab the lock unless need it (bucket_cnt > 0). */
1445 flags = htab_lock_bucket(htab, b);
1448 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1451 if (bucket_cnt && !locked) {
1456 if (bucket_cnt > (max_count - total)) {
1459 /* Note that since bucket_cnt > 0 here, it is implicit
1460 * that the locked was grabbed, so release it.
1462 htab_unlock_bucket(htab, b, flags);
1464 bpf_enable_instrumentation();
1468 if (bucket_cnt > bucket_size) {
1469 bucket_size = bucket_cnt;
1470 /* Note that since bucket_cnt > 0 here, it is implicit
1471 * that the locked was grabbed, so release it.
1473 htab_unlock_bucket(htab, b, flags);
1475 bpf_enable_instrumentation();
1481 /* Next block is only safe to run if you have grabbed the lock */
1485 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1486 memcpy(dst_key, l->key, key_size);
1490 void __percpu *pptr;
1492 pptr = htab_elem_get_ptr(l, map->key_size);
1493 for_each_possible_cpu(cpu) {
1494 bpf_long_memcpy(dst_val + off,
1495 per_cpu_ptr(pptr, cpu), size);
1499 value = l->key + roundup_key_size;
1500 if (elem_map_flags & BPF_F_LOCK)
1501 copy_map_value_locked(map, dst_val, value,
1504 copy_map_value(map, dst_val, value);
1505 check_and_init_map_lock(map, dst_val);
1508 hlist_nulls_del_rcu(&l->hash_node);
1510 /* bpf_lru_push_free() will acquire lru_lock, which
1511 * may cause deadlock. See comments in function
1512 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1513 * after releasing the bucket lock.
1516 l->batch_flink = node_to_free;
1519 free_htab_elem(htab, l);
1522 dst_key += key_size;
1523 dst_val += value_size;
1526 htab_unlock_bucket(htab, b, flags);
1529 while (node_to_free) {
1531 node_to_free = node_to_free->batch_flink;
1532 bpf_lru_push_free(&htab->lru, &l->lru_node);
1536 /* If we are not copying data, we can go to next bucket and avoid
1537 * unlocking the rcu.
1539 if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1545 bpf_enable_instrumentation();
1546 if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1547 key_size * bucket_cnt) ||
1548 copy_to_user(uvalues + total * value_size, values,
1549 value_size * bucket_cnt))) {
1554 total += bucket_cnt;
1556 if (batch >= htab->n_buckets) {
1566 /* copy # of entries and next batch */
1567 ubatch = u64_to_user_ptr(attr->batch.out_batch);
1568 if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1569 put_user(total, &uattr->batch.count))
1579 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1580 union bpf_attr __user *uattr)
1582 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1587 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1588 const union bpf_attr *attr,
1589 union bpf_attr __user *uattr)
1591 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1596 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1597 union bpf_attr __user *uattr)
1599 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1604 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1605 const union bpf_attr *attr,
1606 union bpf_attr __user *uattr)
1608 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1613 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1614 const union bpf_attr *attr,
1615 union bpf_attr __user *uattr)
1617 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1622 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1623 const union bpf_attr *attr,
1624 union bpf_attr __user *uattr)
1626 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1631 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1632 union bpf_attr __user *uattr)
1634 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1639 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1640 const union bpf_attr *attr,
1641 union bpf_attr __user *uattr)
1643 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1647 struct bpf_iter_seq_hash_map_info {
1648 struct bpf_map *map;
1649 struct bpf_htab *htab;
1650 void *percpu_value_buf; // non-zero means percpu hash
1655 static struct htab_elem *
1656 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1657 struct htab_elem *prev_elem)
1659 const struct bpf_htab *htab = info->htab;
1660 u32 skip_elems = info->skip_elems;
1661 u32 bucket_id = info->bucket_id;
1662 struct hlist_nulls_head *head;
1663 struct hlist_nulls_node *n;
1664 struct htab_elem *elem;
1668 if (bucket_id >= htab->n_buckets)
1671 /* try to find next elem in the same bucket */
1673 /* no update/deletion on this bucket, prev_elem should be still valid
1674 * and we won't skip elements.
1676 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1677 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1681 /* not found, unlock and go to the next bucket */
1682 b = &htab->buckets[bucket_id++];
1687 for (i = bucket_id; i < htab->n_buckets; i++) {
1688 b = &htab->buckets[i];
1693 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1694 if (count >= skip_elems) {
1695 info->bucket_id = i;
1696 info->skip_elems = count;
1706 info->bucket_id = i;
1707 info->skip_elems = 0;
1711 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
1713 struct bpf_iter_seq_hash_map_info *info = seq->private;
1714 struct htab_elem *elem;
1716 elem = bpf_hash_map_seq_find_next(info, NULL);
1725 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1727 struct bpf_iter_seq_hash_map_info *info = seq->private;
1731 return bpf_hash_map_seq_find_next(info, v);
1734 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
1736 struct bpf_iter_seq_hash_map_info *info = seq->private;
1737 u32 roundup_key_size, roundup_value_size;
1738 struct bpf_iter__bpf_map_elem ctx = {};
1739 struct bpf_map *map = info->map;
1740 struct bpf_iter_meta meta;
1741 int ret = 0, off = 0, cpu;
1742 struct bpf_prog *prog;
1743 void __percpu *pptr;
1746 prog = bpf_iter_get_info(&meta, elem == NULL);
1749 ctx.map = info->map;
1751 roundup_key_size = round_up(map->key_size, 8);
1752 ctx.key = elem->key;
1753 if (!info->percpu_value_buf) {
1754 ctx.value = elem->key + roundup_key_size;
1756 roundup_value_size = round_up(map->value_size, 8);
1757 pptr = htab_elem_get_ptr(elem, map->key_size);
1758 for_each_possible_cpu(cpu) {
1759 bpf_long_memcpy(info->percpu_value_buf + off,
1760 per_cpu_ptr(pptr, cpu),
1761 roundup_value_size);
1762 off += roundup_value_size;
1764 ctx.value = info->percpu_value_buf;
1767 ret = bpf_iter_run_prog(prog, &ctx);
1773 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
1775 return __bpf_hash_map_seq_show(seq, v);
1778 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
1781 (void)__bpf_hash_map_seq_show(seq, NULL);
1786 static int bpf_iter_init_hash_map(void *priv_data,
1787 struct bpf_iter_aux_info *aux)
1789 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1790 struct bpf_map *map = aux->map;
1794 if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
1795 map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
1796 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
1797 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
1801 seq_info->percpu_value_buf = value_buf;
1804 seq_info->map = map;
1805 seq_info->htab = container_of(map, struct bpf_htab, map);
1809 static void bpf_iter_fini_hash_map(void *priv_data)
1811 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1813 kfree(seq_info->percpu_value_buf);
1816 static const struct seq_operations bpf_hash_map_seq_ops = {
1817 .start = bpf_hash_map_seq_start,
1818 .next = bpf_hash_map_seq_next,
1819 .stop = bpf_hash_map_seq_stop,
1820 .show = bpf_hash_map_seq_show,
1823 static const struct bpf_iter_seq_info iter_seq_info = {
1824 .seq_ops = &bpf_hash_map_seq_ops,
1825 .init_seq_private = bpf_iter_init_hash_map,
1826 .fini_seq_private = bpf_iter_fini_hash_map,
1827 .seq_priv_size = sizeof(struct bpf_iter_seq_hash_map_info),
1830 static int htab_map_btf_id;
1831 const struct bpf_map_ops htab_map_ops = {
1832 .map_meta_equal = bpf_map_meta_equal,
1833 .map_alloc_check = htab_map_alloc_check,
1834 .map_alloc = htab_map_alloc,
1835 .map_free = htab_map_free,
1836 .map_get_next_key = htab_map_get_next_key,
1837 .map_lookup_elem = htab_map_lookup_elem,
1838 .map_update_elem = htab_map_update_elem,
1839 .map_delete_elem = htab_map_delete_elem,
1840 .map_gen_lookup = htab_map_gen_lookup,
1841 .map_seq_show_elem = htab_map_seq_show_elem,
1843 .map_btf_name = "bpf_htab",
1844 .map_btf_id = &htab_map_btf_id,
1845 .iter_seq_info = &iter_seq_info,
1848 static int htab_lru_map_btf_id;
1849 const struct bpf_map_ops htab_lru_map_ops = {
1850 .map_meta_equal = bpf_map_meta_equal,
1851 .map_alloc_check = htab_map_alloc_check,
1852 .map_alloc = htab_map_alloc,
1853 .map_free = htab_map_free,
1854 .map_get_next_key = htab_map_get_next_key,
1855 .map_lookup_elem = htab_lru_map_lookup_elem,
1856 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1857 .map_update_elem = htab_lru_map_update_elem,
1858 .map_delete_elem = htab_lru_map_delete_elem,
1859 .map_gen_lookup = htab_lru_map_gen_lookup,
1860 .map_seq_show_elem = htab_map_seq_show_elem,
1861 BATCH_OPS(htab_lru),
1862 .map_btf_name = "bpf_htab",
1863 .map_btf_id = &htab_lru_map_btf_id,
1864 .iter_seq_info = &iter_seq_info,
1867 /* Called from eBPF program */
1868 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1870 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1873 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1878 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1880 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1883 bpf_lru_node_set_ref(&l->lru_node);
1884 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1890 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1892 struct htab_elem *l;
1893 void __percpu *pptr;
1898 /* per_cpu areas are zero-filled and bpf programs can only
1899 * access 'value_size' of them, so copying rounded areas
1900 * will not leak any kernel data
1902 size = round_up(map->value_size, 8);
1904 l = __htab_map_lookup_elem(map, key);
1907 /* We do not mark LRU map element here in order to not mess up
1908 * eviction heuristics when user space does a map walk.
1910 pptr = htab_elem_get_ptr(l, map->key_size);
1911 for_each_possible_cpu(cpu) {
1912 bpf_long_memcpy(value + off,
1913 per_cpu_ptr(pptr, cpu), size);
1922 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1925 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1929 if (htab_is_lru(htab))
1930 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1933 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1940 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1943 struct htab_elem *l;
1944 void __percpu *pptr;
1949 l = __htab_map_lookup_elem(map, key);
1955 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1956 seq_puts(m, ": {\n");
1957 pptr = htab_elem_get_ptr(l, map->key_size);
1958 for_each_possible_cpu(cpu) {
1959 seq_printf(m, "\tcpu%d: ", cpu);
1960 btf_type_seq_show(map->btf, map->btf_value_type_id,
1961 per_cpu_ptr(pptr, cpu), m);
1969 static int htab_percpu_map_btf_id;
1970 const struct bpf_map_ops htab_percpu_map_ops = {
1971 .map_meta_equal = bpf_map_meta_equal,
1972 .map_alloc_check = htab_map_alloc_check,
1973 .map_alloc = htab_map_alloc,
1974 .map_free = htab_map_free,
1975 .map_get_next_key = htab_map_get_next_key,
1976 .map_lookup_elem = htab_percpu_map_lookup_elem,
1977 .map_update_elem = htab_percpu_map_update_elem,
1978 .map_delete_elem = htab_map_delete_elem,
1979 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1980 BATCH_OPS(htab_percpu),
1981 .map_btf_name = "bpf_htab",
1982 .map_btf_id = &htab_percpu_map_btf_id,
1983 .iter_seq_info = &iter_seq_info,
1986 static int htab_lru_percpu_map_btf_id;
1987 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1988 .map_meta_equal = bpf_map_meta_equal,
1989 .map_alloc_check = htab_map_alloc_check,
1990 .map_alloc = htab_map_alloc,
1991 .map_free = htab_map_free,
1992 .map_get_next_key = htab_map_get_next_key,
1993 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1994 .map_update_elem = htab_lru_percpu_map_update_elem,
1995 .map_delete_elem = htab_lru_map_delete_elem,
1996 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1997 BATCH_OPS(htab_lru_percpu),
1998 .map_btf_name = "bpf_htab",
1999 .map_btf_id = &htab_lru_percpu_map_btf_id,
2000 .iter_seq_info = &iter_seq_info,
2003 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2005 if (attr->value_size != sizeof(u32))
2007 return htab_map_alloc_check(attr);
2010 static void fd_htab_map_free(struct bpf_map *map)
2012 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2013 struct hlist_nulls_node *n;
2014 struct hlist_nulls_head *head;
2015 struct htab_elem *l;
2018 for (i = 0; i < htab->n_buckets; i++) {
2019 head = select_bucket(htab, i);
2021 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2022 void *ptr = fd_htab_map_get_ptr(map, l);
2024 map->ops->map_fd_put_ptr(ptr);
2031 /* only called from syscall */
2032 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2037 if (!map->ops->map_fd_sys_lookup_elem)
2041 ptr = htab_map_lookup_elem(map, key);
2043 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2051 /* only called from syscall */
2052 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2053 void *key, void *value, u64 map_flags)
2057 u32 ufd = *(u32 *)value;
2059 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2061 return PTR_ERR(ptr);
2063 ret = htab_map_update_elem(map, key, &ptr, map_flags);
2065 map->ops->map_fd_put_ptr(ptr);
2070 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2072 struct bpf_map *map, *inner_map_meta;
2074 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2075 if (IS_ERR(inner_map_meta))
2076 return inner_map_meta;
2078 map = htab_map_alloc(attr);
2080 bpf_map_meta_free(inner_map_meta);
2084 map->inner_map_meta = inner_map_meta;
2089 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2091 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
2096 return READ_ONCE(*inner_map);
2099 static int htab_of_map_gen_lookup(struct bpf_map *map,
2100 struct bpf_insn *insn_buf)
2102 struct bpf_insn *insn = insn_buf;
2103 const int ret = BPF_REG_0;
2105 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2106 (void *(*)(struct bpf_map *map, void *key))NULL));
2107 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
2108 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2109 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2110 offsetof(struct htab_elem, key) +
2111 round_up(map->key_size, 8));
2112 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2114 return insn - insn_buf;
2117 static void htab_of_map_free(struct bpf_map *map)
2119 bpf_map_meta_free(map->inner_map_meta);
2120 fd_htab_map_free(map);
2123 static int htab_of_maps_map_btf_id;
2124 const struct bpf_map_ops htab_of_maps_map_ops = {
2125 .map_alloc_check = fd_htab_map_alloc_check,
2126 .map_alloc = htab_of_map_alloc,
2127 .map_free = htab_of_map_free,
2128 .map_get_next_key = htab_map_get_next_key,
2129 .map_lookup_elem = htab_of_map_lookup_elem,
2130 .map_delete_elem = htab_map_delete_elem,
2131 .map_fd_get_ptr = bpf_map_fd_get_ptr,
2132 .map_fd_put_ptr = bpf_map_fd_put_ptr,
2133 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2134 .map_gen_lookup = htab_of_map_gen_lookup,
2135 .map_check_btf = map_check_no_btf,
2136 .map_btf_name = "bpf_htab",
2137 .map_btf_id = &htab_of_maps_map_btf_id,