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; roundup_pow_of_two() can overflow
447 * into UB on 32-bit arches, so check that first
450 if (htab->map.max_entries > 1UL << 31)
453 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
455 htab->elem_size = sizeof(struct htab_elem) +
456 round_up(htab->map.key_size, 8);
458 htab->elem_size += sizeof(void *);
460 htab->elem_size += round_up(htab->map.value_size, 8);
462 /* check for u32 overflow */
463 if (htab->n_buckets > U32_MAX / sizeof(struct bucket))
466 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
467 (u64) htab->elem_size * htab->map.max_entries;
470 cost += (u64) round_up(htab->map.value_size, 8) *
471 num_possible_cpus() * htab->map.max_entries;
473 cost += (u64) htab->elem_size * num_possible_cpus();
475 /* if map size is larger than memlock limit, reject it */
476 err = bpf_map_charge_init(&htab->map.memory, cost);
481 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
482 sizeof(struct bucket),
483 htab->map.numa_node);
487 if (htab->map.map_flags & BPF_F_ZERO_SEED)
490 htab->hashrnd = get_random_int();
492 htab_init_buckets(htab);
495 err = prealloc_init(htab);
499 if (!percpu && !lru) {
500 /* lru itself can remove the least used element, so
501 * there is no need for an extra elem during map_update.
503 err = alloc_extra_elems(htab);
512 prealloc_destroy(htab);
514 bpf_map_area_free(htab->buckets);
516 bpf_map_charge_finish(&htab->map.memory);
522 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
524 return jhash(key, key_len, hashrnd);
527 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
529 return &htab->buckets[hash & (htab->n_buckets - 1)];
532 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
534 return &__select_bucket(htab, hash)->head;
537 /* this lookup function can only be called with bucket lock taken */
538 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
539 void *key, u32 key_size)
541 struct hlist_nulls_node *n;
544 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
545 if (l->hash == hash && !memcmp(&l->key, key, key_size))
551 /* can be called without bucket lock. it will repeat the loop in
552 * the unlikely event when elements moved from one bucket into another
553 * while link list is being walked
555 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
557 u32 key_size, u32 n_buckets)
559 struct hlist_nulls_node *n;
563 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
564 if (l->hash == hash && !memcmp(&l->key, key, key_size))
567 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
573 /* Called from syscall or from eBPF program directly, so
574 * arguments have to match bpf_map_lookup_elem() exactly.
575 * The return value is adjusted by BPF instructions
576 * in htab_map_gen_lookup().
578 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
580 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
581 struct hlist_nulls_head *head;
585 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
587 key_size = map->key_size;
589 hash = htab_map_hash(key, key_size, htab->hashrnd);
591 head = select_bucket(htab, hash);
593 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
598 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
600 struct htab_elem *l = __htab_map_lookup_elem(map, key);
603 return l->key + round_up(map->key_size, 8);
608 /* inline bpf_map_lookup_elem() call.
611 * bpf_map_lookup_elem
612 * map->ops->map_lookup_elem
613 * htab_map_lookup_elem
614 * __htab_map_lookup_elem
617 * __htab_map_lookup_elem
619 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
621 struct bpf_insn *insn = insn_buf;
622 const int ret = BPF_REG_0;
624 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
625 (void *(*)(struct bpf_map *map, void *key))NULL));
626 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
627 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
628 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
629 offsetof(struct htab_elem, key) +
630 round_up(map->key_size, 8));
631 return insn - insn_buf;
634 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
635 void *key, const bool mark)
637 struct htab_elem *l = __htab_map_lookup_elem(map, key);
641 bpf_lru_node_set_ref(&l->lru_node);
642 return l->key + round_up(map->key_size, 8);
648 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
650 return __htab_lru_map_lookup_elem(map, key, true);
653 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
655 return __htab_lru_map_lookup_elem(map, key, false);
658 static int htab_lru_map_gen_lookup(struct bpf_map *map,
659 struct bpf_insn *insn_buf)
661 struct bpf_insn *insn = insn_buf;
662 const int ret = BPF_REG_0;
663 const int ref_reg = BPF_REG_1;
665 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
666 (void *(*)(struct bpf_map *map, void *key))NULL));
667 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
668 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
669 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
670 offsetof(struct htab_elem, lru_node) +
671 offsetof(struct bpf_lru_node, ref));
672 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
673 *insn++ = BPF_ST_MEM(BPF_B, ret,
674 offsetof(struct htab_elem, lru_node) +
675 offsetof(struct bpf_lru_node, ref),
677 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
678 offsetof(struct htab_elem, key) +
679 round_up(map->key_size, 8));
680 return insn - insn_buf;
683 /* It is called from the bpf_lru_list when the LRU needs to delete
684 * older elements from the htab.
686 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
688 struct bpf_htab *htab = (struct bpf_htab *)arg;
689 struct htab_elem *l = NULL, *tgt_l;
690 struct hlist_nulls_head *head;
691 struct hlist_nulls_node *n;
695 tgt_l = container_of(node, struct htab_elem, lru_node);
696 b = __select_bucket(htab, tgt_l->hash);
699 flags = htab_lock_bucket(htab, b);
701 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
703 hlist_nulls_del_rcu(&l->hash_node);
707 htab_unlock_bucket(htab, b, flags);
712 /* Called from syscall */
713 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
715 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
716 struct hlist_nulls_head *head;
717 struct htab_elem *l, *next_l;
721 WARN_ON_ONCE(!rcu_read_lock_held());
723 key_size = map->key_size;
726 goto find_first_elem;
728 hash = htab_map_hash(key, key_size, htab->hashrnd);
730 head = select_bucket(htab, hash);
733 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
736 goto find_first_elem;
738 /* key was found, get next key in the same bucket */
739 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
740 struct htab_elem, hash_node);
743 /* if next elem in this hash list is non-zero, just return it */
744 memcpy(next_key, next_l->key, key_size);
748 /* no more elements in this hash list, go to the next bucket */
749 i = hash & (htab->n_buckets - 1);
753 /* iterate over buckets */
754 for (; i < htab->n_buckets; i++) {
755 head = select_bucket(htab, i);
757 /* pick first element in the bucket */
758 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
759 struct htab_elem, hash_node);
761 /* if it's not empty, just return it */
762 memcpy(next_key, next_l->key, key_size);
767 /* iterated over all buckets and all elements */
771 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
773 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
774 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
778 static void htab_elem_free_rcu(struct rcu_head *head)
780 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
781 struct bpf_htab *htab = l->htab;
783 htab_elem_free(htab, l);
786 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
788 struct bpf_map *map = &htab->map;
791 if (map->ops->map_fd_put_ptr) {
792 ptr = fd_htab_map_get_ptr(map, l);
793 map->ops->map_fd_put_ptr(map, ptr, true);
797 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
799 htab_put_fd_value(htab, l);
801 if (htab_is_prealloc(htab)) {
802 __pcpu_freelist_push(&htab->freelist, &l->fnode);
804 atomic_dec(&htab->count);
806 call_rcu(&l->rcu, htab_elem_free_rcu);
810 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
811 void *value, bool onallcpus)
814 /* copy true value_size bytes */
815 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
817 u32 size = round_up(htab->map.value_size, 8);
820 for_each_possible_cpu(cpu) {
821 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
828 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
829 void *value, bool onallcpus)
831 /* When using prealloc and not setting the initial value on all cpus,
832 * zero-fill element values for other cpus (just as what happens when
833 * not using prealloc). Otherwise, bpf program has no way to ensure
834 * known initial values for cpus other than current one
835 * (onallcpus=false always when coming from bpf prog).
837 if (htab_is_prealloc(htab) && !onallcpus) {
838 u32 size = round_up(htab->map.value_size, 8);
839 int current_cpu = raw_smp_processor_id();
842 for_each_possible_cpu(cpu) {
843 if (cpu == current_cpu)
844 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value,
847 memset(per_cpu_ptr(pptr, cpu), 0, size);
850 pcpu_copy_value(htab, pptr, value, onallcpus);
854 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
856 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
860 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
861 void *value, u32 key_size, u32 hash,
862 bool percpu, bool onallcpus,
863 struct htab_elem *old_elem)
865 u32 size = htab->map.value_size;
866 bool prealloc = htab_is_prealloc(htab);
867 struct htab_elem *l_new, **pl_new;
872 /* if we're updating the existing element,
873 * use per-cpu extra elems to avoid freelist_pop/push
875 pl_new = this_cpu_ptr(htab->extra_elems);
877 htab_put_fd_value(htab, old_elem);
880 struct pcpu_freelist_node *l;
882 l = __pcpu_freelist_pop(&htab->freelist);
884 return ERR_PTR(-E2BIG);
885 l_new = container_of(l, struct htab_elem, fnode);
888 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
890 /* when map is full and update() is replacing
891 * old element, it's ok to allocate, since
892 * old element will be freed immediately.
893 * Otherwise return an error
895 l_new = ERR_PTR(-E2BIG);
898 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
899 htab->map.numa_node);
901 l_new = ERR_PTR(-ENOMEM);
904 check_and_init_map_lock(&htab->map,
905 l_new->key + round_up(key_size, 8));
908 memcpy(l_new->key, key, key_size);
910 size = round_up(size, 8);
912 pptr = htab_elem_get_ptr(l_new, key_size);
914 /* alloc_percpu zero-fills */
915 pptr = __alloc_percpu_gfp(size, 8,
916 GFP_ATOMIC | __GFP_NOWARN);
919 l_new = ERR_PTR(-ENOMEM);
924 pcpu_init_value(htab, pptr, value, onallcpus);
927 htab_elem_set_ptr(l_new, key_size, pptr);
928 } else if (fd_htab_map_needs_adjust(htab)) {
929 size = round_up(size, 8);
930 memcpy(l_new->key + round_up(key_size, 8), value, size);
932 copy_map_value(&htab->map,
933 l_new->key + round_up(key_size, 8),
940 atomic_dec(&htab->count);
944 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
947 if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
948 /* elem already exists */
951 if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
952 /* elem doesn't exist, cannot update it */
958 /* Called from syscall or from eBPF program */
959 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
962 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
963 struct htab_elem *l_new = NULL, *l_old;
964 struct hlist_nulls_head *head;
970 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
974 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
976 key_size = map->key_size;
978 hash = htab_map_hash(key, key_size, htab->hashrnd);
980 b = __select_bucket(htab, hash);
983 if (unlikely(map_flags & BPF_F_LOCK)) {
984 if (unlikely(!map_value_has_spin_lock(map)))
986 /* find an element without taking the bucket lock */
987 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
989 ret = check_flags(htab, l_old, map_flags);
993 /* grab the element lock and update value in place */
994 copy_map_value_locked(map,
995 l_old->key + round_up(key_size, 8),
999 /* fall through, grab the bucket lock and lookup again.
1000 * 99.9% chance that the element won't be found,
1001 * but second lookup under lock has to be done.
1005 flags = htab_lock_bucket(htab, b);
1007 l_old = lookup_elem_raw(head, hash, key, key_size);
1009 ret = check_flags(htab, l_old, map_flags);
1013 if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1014 /* first lookup without the bucket lock didn't find the element,
1015 * but second lookup with the bucket lock found it.
1016 * This case is highly unlikely, but has to be dealt with:
1017 * grab the element lock in addition to the bucket lock
1018 * and update element in place
1020 copy_map_value_locked(map,
1021 l_old->key + round_up(key_size, 8),
1027 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1029 if (IS_ERR(l_new)) {
1030 /* all pre-allocated elements are in use or memory exhausted */
1031 ret = PTR_ERR(l_new);
1035 /* add new element to the head of the list, so that
1036 * concurrent search will find it before old elem
1038 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1040 hlist_nulls_del_rcu(&l_old->hash_node);
1041 if (!htab_is_prealloc(htab))
1042 free_htab_elem(htab, l_old);
1046 htab_unlock_bucket(htab, b, flags);
1050 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1053 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1054 struct htab_elem *l_new, *l_old = NULL;
1055 struct hlist_nulls_head *head;
1056 unsigned long flags;
1061 if (unlikely(map_flags > BPF_EXIST))
1065 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1067 key_size = map->key_size;
1069 hash = htab_map_hash(key, key_size, htab->hashrnd);
1071 b = __select_bucket(htab, hash);
1074 /* For LRU, we need to alloc before taking bucket's
1075 * spinlock because getting free nodes from LRU may need
1076 * to remove older elements from htab and this removal
1077 * operation will need a bucket lock.
1079 l_new = prealloc_lru_pop(htab, key, hash);
1082 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
1084 flags = htab_lock_bucket(htab, b);
1086 l_old = lookup_elem_raw(head, hash, key, key_size);
1088 ret = check_flags(htab, l_old, map_flags);
1092 /* add new element to the head of the list, so that
1093 * concurrent search will find it before old elem
1095 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1097 bpf_lru_node_set_ref(&l_new->lru_node);
1098 hlist_nulls_del_rcu(&l_old->hash_node);
1103 htab_unlock_bucket(htab, b, flags);
1106 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1108 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
1113 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1114 void *value, u64 map_flags,
1117 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1118 struct htab_elem *l_new = NULL, *l_old;
1119 struct hlist_nulls_head *head;
1120 unsigned long flags;
1125 if (unlikely(map_flags > BPF_EXIST))
1129 WARN_ON_ONCE(!rcu_read_lock_held());
1131 key_size = map->key_size;
1133 hash = htab_map_hash(key, key_size, htab->hashrnd);
1135 b = __select_bucket(htab, hash);
1138 flags = htab_lock_bucket(htab, b);
1140 l_old = lookup_elem_raw(head, hash, key, key_size);
1142 ret = check_flags(htab, l_old, map_flags);
1147 /* per-cpu hash map can update value in-place */
1148 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1151 l_new = alloc_htab_elem(htab, key, value, key_size,
1152 hash, true, onallcpus, NULL);
1153 if (IS_ERR(l_new)) {
1154 ret = PTR_ERR(l_new);
1157 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1161 htab_unlock_bucket(htab, b, flags);
1165 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1166 void *value, u64 map_flags,
1169 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1170 struct htab_elem *l_new = NULL, *l_old;
1171 struct hlist_nulls_head *head;
1172 unsigned long flags;
1177 if (unlikely(map_flags > BPF_EXIST))
1181 WARN_ON_ONCE(!rcu_read_lock_held());
1183 key_size = map->key_size;
1185 hash = htab_map_hash(key, key_size, htab->hashrnd);
1187 b = __select_bucket(htab, hash);
1190 /* For LRU, we need to alloc before taking bucket's
1191 * spinlock because LRU's elem alloc may need
1192 * to remove older elem from htab and this removal
1193 * operation will need a bucket lock.
1195 if (map_flags != BPF_EXIST) {
1196 l_new = prealloc_lru_pop(htab, key, hash);
1201 flags = htab_lock_bucket(htab, b);
1203 l_old = lookup_elem_raw(head, hash, key, key_size);
1205 ret = check_flags(htab, l_old, map_flags);
1210 bpf_lru_node_set_ref(&l_old->lru_node);
1212 /* per-cpu hash map can update value in-place */
1213 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1216 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1218 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1223 htab_unlock_bucket(htab, b, flags);
1225 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1229 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1230 void *value, u64 map_flags)
1232 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1235 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1236 void *value, u64 map_flags)
1238 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1242 /* Called from syscall or from eBPF program */
1243 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1245 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1246 struct hlist_nulls_head *head;
1248 struct htab_elem *l;
1249 unsigned long flags;
1253 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1255 key_size = map->key_size;
1257 hash = htab_map_hash(key, key_size, htab->hashrnd);
1258 b = __select_bucket(htab, hash);
1261 flags = htab_lock_bucket(htab, b);
1263 l = lookup_elem_raw(head, hash, key, key_size);
1266 hlist_nulls_del_rcu(&l->hash_node);
1267 free_htab_elem(htab, l);
1271 htab_unlock_bucket(htab, b, flags);
1275 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1277 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1278 struct hlist_nulls_head *head;
1280 struct htab_elem *l;
1281 unsigned long flags;
1285 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1287 key_size = map->key_size;
1289 hash = htab_map_hash(key, key_size, htab->hashrnd);
1290 b = __select_bucket(htab, hash);
1293 flags = htab_lock_bucket(htab, b);
1295 l = lookup_elem_raw(head, hash, key, key_size);
1298 hlist_nulls_del_rcu(&l->hash_node);
1302 htab_unlock_bucket(htab, b, flags);
1304 bpf_lru_push_free(&htab->lru, &l->lru_node);
1308 static void delete_all_elements(struct bpf_htab *htab)
1312 for (i = 0; i < htab->n_buckets; i++) {
1313 struct hlist_nulls_head *head = select_bucket(htab, i);
1314 struct hlist_nulls_node *n;
1315 struct htab_elem *l;
1317 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1318 hlist_nulls_del_rcu(&l->hash_node);
1319 htab_elem_free(htab, l);
1324 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1325 static void htab_map_free(struct bpf_map *map)
1327 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1329 /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1330 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1331 * There is no need to synchronize_rcu() here to protect map elements.
1334 /* some of free_htab_elem() callbacks for elements of this map may
1335 * not have executed. Wait for them.
1338 if (!htab_is_prealloc(htab))
1339 delete_all_elements(htab);
1341 prealloc_destroy(htab);
1343 free_percpu(htab->extra_elems);
1344 bpf_map_area_free(htab->buckets);
1348 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1355 value = htab_map_lookup_elem(map, key);
1361 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1363 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1370 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1371 const union bpf_attr *attr,
1372 union bpf_attr __user *uattr,
1373 bool do_delete, bool is_lru_map,
1376 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1377 u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1378 void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1379 void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1380 void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1381 void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1382 u32 batch, max_count, size, bucket_size;
1383 struct htab_elem *node_to_free = NULL;
1384 u64 elem_map_flags, map_flags;
1385 struct hlist_nulls_head *head;
1386 struct hlist_nulls_node *n;
1387 unsigned long flags = 0;
1388 bool locked = false;
1389 struct htab_elem *l;
1393 elem_map_flags = attr->batch.elem_flags;
1394 if ((elem_map_flags & ~BPF_F_LOCK) ||
1395 ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1398 map_flags = attr->batch.flags;
1402 max_count = attr->batch.count;
1406 if (put_user(0, &uattr->batch.count))
1410 if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1413 if (batch >= htab->n_buckets)
1416 key_size = htab->map.key_size;
1417 roundup_key_size = round_up(htab->map.key_size, 8);
1418 value_size = htab->map.value_size;
1419 size = round_up(value_size, 8);
1421 value_size = size * num_possible_cpus();
1423 /* while experimenting with hash tables with sizes ranging from 10 to
1424 * 1000, it was observed that a bucket can have upto 5 entries.
1429 /* We cannot do copy_from_user or copy_to_user inside
1430 * the rcu_read_lock. Allocate enough space here.
1432 keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1433 values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1434 if (!keys || !values) {
1440 bpf_disable_instrumentation();
1445 b = &htab->buckets[batch];
1447 /* do not grab the lock unless need it (bucket_cnt > 0). */
1449 flags = htab_lock_bucket(htab, b);
1452 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1455 if (bucket_cnt && !locked) {
1460 if (bucket_cnt > (max_count - total)) {
1463 /* Note that since bucket_cnt > 0 here, it is implicit
1464 * that the locked was grabbed, so release it.
1466 htab_unlock_bucket(htab, b, flags);
1468 bpf_enable_instrumentation();
1472 if (bucket_cnt > bucket_size) {
1473 bucket_size = bucket_cnt;
1474 /* Note that since bucket_cnt > 0 here, it is implicit
1475 * that the locked was grabbed, so release it.
1477 htab_unlock_bucket(htab, b, flags);
1479 bpf_enable_instrumentation();
1485 /* Next block is only safe to run if you have grabbed the lock */
1489 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1490 memcpy(dst_key, l->key, key_size);
1494 void __percpu *pptr;
1496 pptr = htab_elem_get_ptr(l, map->key_size);
1497 for_each_possible_cpu(cpu) {
1498 bpf_long_memcpy(dst_val + off,
1499 per_cpu_ptr(pptr, cpu), size);
1503 value = l->key + roundup_key_size;
1504 if (elem_map_flags & BPF_F_LOCK)
1505 copy_map_value_locked(map, dst_val, value,
1508 copy_map_value(map, dst_val, value);
1509 check_and_init_map_lock(map, dst_val);
1512 hlist_nulls_del_rcu(&l->hash_node);
1514 /* bpf_lru_push_free() will acquire lru_lock, which
1515 * may cause deadlock. See comments in function
1516 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1517 * after releasing the bucket lock.
1520 l->batch_flink = node_to_free;
1523 free_htab_elem(htab, l);
1526 dst_key += key_size;
1527 dst_val += value_size;
1530 htab_unlock_bucket(htab, b, flags);
1533 while (node_to_free) {
1535 node_to_free = node_to_free->batch_flink;
1536 bpf_lru_push_free(&htab->lru, &l->lru_node);
1540 /* If we are not copying data, we can go to next bucket and avoid
1541 * unlocking the rcu.
1543 if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1549 bpf_enable_instrumentation();
1550 if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1551 key_size * bucket_cnt) ||
1552 copy_to_user(uvalues + total * value_size, values,
1553 value_size * bucket_cnt))) {
1558 total += bucket_cnt;
1560 if (batch >= htab->n_buckets) {
1570 /* copy # of entries and next batch */
1571 ubatch = u64_to_user_ptr(attr->batch.out_batch);
1572 if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1573 put_user(total, &uattr->batch.count))
1583 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1584 union bpf_attr __user *uattr)
1586 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1591 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1592 const union bpf_attr *attr,
1593 union bpf_attr __user *uattr)
1595 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1600 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1601 union bpf_attr __user *uattr)
1603 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1608 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1609 const union bpf_attr *attr,
1610 union bpf_attr __user *uattr)
1612 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1617 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1618 const union bpf_attr *attr,
1619 union bpf_attr __user *uattr)
1621 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1626 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1627 const union bpf_attr *attr,
1628 union bpf_attr __user *uattr)
1630 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1635 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1636 union bpf_attr __user *uattr)
1638 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1643 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1644 const union bpf_attr *attr,
1645 union bpf_attr __user *uattr)
1647 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1651 struct bpf_iter_seq_hash_map_info {
1652 struct bpf_map *map;
1653 struct bpf_htab *htab;
1654 void *percpu_value_buf; // non-zero means percpu hash
1659 static struct htab_elem *
1660 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1661 struct htab_elem *prev_elem)
1663 const struct bpf_htab *htab = info->htab;
1664 u32 skip_elems = info->skip_elems;
1665 u32 bucket_id = info->bucket_id;
1666 struct hlist_nulls_head *head;
1667 struct hlist_nulls_node *n;
1668 struct htab_elem *elem;
1672 if (bucket_id >= htab->n_buckets)
1675 /* try to find next elem in the same bucket */
1677 /* no update/deletion on this bucket, prev_elem should be still valid
1678 * and we won't skip elements.
1680 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1681 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1685 /* not found, unlock and go to the next bucket */
1686 b = &htab->buckets[bucket_id++];
1691 for (i = bucket_id; i < htab->n_buckets; i++) {
1692 b = &htab->buckets[i];
1697 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1698 if (count >= skip_elems) {
1699 info->bucket_id = i;
1700 info->skip_elems = count;
1710 info->bucket_id = i;
1711 info->skip_elems = 0;
1715 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
1717 struct bpf_iter_seq_hash_map_info *info = seq->private;
1718 struct htab_elem *elem;
1720 elem = bpf_hash_map_seq_find_next(info, NULL);
1729 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1731 struct bpf_iter_seq_hash_map_info *info = seq->private;
1735 return bpf_hash_map_seq_find_next(info, v);
1738 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
1740 struct bpf_iter_seq_hash_map_info *info = seq->private;
1741 u32 roundup_key_size, roundup_value_size;
1742 struct bpf_iter__bpf_map_elem ctx = {};
1743 struct bpf_map *map = info->map;
1744 struct bpf_iter_meta meta;
1745 int ret = 0, off = 0, cpu;
1746 struct bpf_prog *prog;
1747 void __percpu *pptr;
1750 prog = bpf_iter_get_info(&meta, elem == NULL);
1753 ctx.map = info->map;
1755 roundup_key_size = round_up(map->key_size, 8);
1756 ctx.key = elem->key;
1757 if (!info->percpu_value_buf) {
1758 ctx.value = elem->key + roundup_key_size;
1760 roundup_value_size = round_up(map->value_size, 8);
1761 pptr = htab_elem_get_ptr(elem, map->key_size);
1762 for_each_possible_cpu(cpu) {
1763 bpf_long_memcpy(info->percpu_value_buf + off,
1764 per_cpu_ptr(pptr, cpu),
1765 roundup_value_size);
1766 off += roundup_value_size;
1768 ctx.value = info->percpu_value_buf;
1771 ret = bpf_iter_run_prog(prog, &ctx);
1777 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
1779 return __bpf_hash_map_seq_show(seq, v);
1782 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
1785 (void)__bpf_hash_map_seq_show(seq, NULL);
1790 static int bpf_iter_init_hash_map(void *priv_data,
1791 struct bpf_iter_aux_info *aux)
1793 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1794 struct bpf_map *map = aux->map;
1798 if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
1799 map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
1800 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
1801 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
1805 seq_info->percpu_value_buf = value_buf;
1808 bpf_map_inc_with_uref(map);
1809 seq_info->map = map;
1810 seq_info->htab = container_of(map, struct bpf_htab, map);
1814 static void bpf_iter_fini_hash_map(void *priv_data)
1816 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1818 bpf_map_put_with_uref(seq_info->map);
1819 kfree(seq_info->percpu_value_buf);
1822 static const struct seq_operations bpf_hash_map_seq_ops = {
1823 .start = bpf_hash_map_seq_start,
1824 .next = bpf_hash_map_seq_next,
1825 .stop = bpf_hash_map_seq_stop,
1826 .show = bpf_hash_map_seq_show,
1829 static const struct bpf_iter_seq_info iter_seq_info = {
1830 .seq_ops = &bpf_hash_map_seq_ops,
1831 .init_seq_private = bpf_iter_init_hash_map,
1832 .fini_seq_private = bpf_iter_fini_hash_map,
1833 .seq_priv_size = sizeof(struct bpf_iter_seq_hash_map_info),
1836 static int htab_map_btf_id;
1837 const struct bpf_map_ops htab_map_ops = {
1838 .map_meta_equal = bpf_map_meta_equal,
1839 .map_alloc_check = htab_map_alloc_check,
1840 .map_alloc = htab_map_alloc,
1841 .map_free = htab_map_free,
1842 .map_get_next_key = htab_map_get_next_key,
1843 .map_lookup_elem = htab_map_lookup_elem,
1844 .map_update_elem = htab_map_update_elem,
1845 .map_delete_elem = htab_map_delete_elem,
1846 .map_gen_lookup = htab_map_gen_lookup,
1847 .map_seq_show_elem = htab_map_seq_show_elem,
1849 .map_btf_name = "bpf_htab",
1850 .map_btf_id = &htab_map_btf_id,
1851 .iter_seq_info = &iter_seq_info,
1854 static int htab_lru_map_btf_id;
1855 const struct bpf_map_ops htab_lru_map_ops = {
1856 .map_meta_equal = bpf_map_meta_equal,
1857 .map_alloc_check = htab_map_alloc_check,
1858 .map_alloc = htab_map_alloc,
1859 .map_free = htab_map_free,
1860 .map_get_next_key = htab_map_get_next_key,
1861 .map_lookup_elem = htab_lru_map_lookup_elem,
1862 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1863 .map_update_elem = htab_lru_map_update_elem,
1864 .map_delete_elem = htab_lru_map_delete_elem,
1865 .map_gen_lookup = htab_lru_map_gen_lookup,
1866 .map_seq_show_elem = htab_map_seq_show_elem,
1867 BATCH_OPS(htab_lru),
1868 .map_btf_name = "bpf_htab",
1869 .map_btf_id = &htab_lru_map_btf_id,
1870 .iter_seq_info = &iter_seq_info,
1873 /* Called from eBPF program */
1874 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1876 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1879 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1884 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1886 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1889 bpf_lru_node_set_ref(&l->lru_node);
1890 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1896 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1898 struct htab_elem *l;
1899 void __percpu *pptr;
1904 /* per_cpu areas are zero-filled and bpf programs can only
1905 * access 'value_size' of them, so copying rounded areas
1906 * will not leak any kernel data
1908 size = round_up(map->value_size, 8);
1910 l = __htab_map_lookup_elem(map, key);
1913 /* We do not mark LRU map element here in order to not mess up
1914 * eviction heuristics when user space does a map walk.
1916 pptr = htab_elem_get_ptr(l, map->key_size);
1917 for_each_possible_cpu(cpu) {
1918 bpf_long_memcpy(value + off,
1919 per_cpu_ptr(pptr, cpu), size);
1928 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1931 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1935 if (htab_is_lru(htab))
1936 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1939 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1946 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1949 struct htab_elem *l;
1950 void __percpu *pptr;
1955 l = __htab_map_lookup_elem(map, key);
1961 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1962 seq_puts(m, ": {\n");
1963 pptr = htab_elem_get_ptr(l, map->key_size);
1964 for_each_possible_cpu(cpu) {
1965 seq_printf(m, "\tcpu%d: ", cpu);
1966 btf_type_seq_show(map->btf, map->btf_value_type_id,
1967 per_cpu_ptr(pptr, cpu), m);
1975 static int htab_percpu_map_btf_id;
1976 const struct bpf_map_ops htab_percpu_map_ops = {
1977 .map_meta_equal = bpf_map_meta_equal,
1978 .map_alloc_check = htab_map_alloc_check,
1979 .map_alloc = htab_map_alloc,
1980 .map_free = htab_map_free,
1981 .map_get_next_key = htab_map_get_next_key,
1982 .map_lookup_elem = htab_percpu_map_lookup_elem,
1983 .map_update_elem = htab_percpu_map_update_elem,
1984 .map_delete_elem = htab_map_delete_elem,
1985 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1986 BATCH_OPS(htab_percpu),
1987 .map_btf_name = "bpf_htab",
1988 .map_btf_id = &htab_percpu_map_btf_id,
1989 .iter_seq_info = &iter_seq_info,
1992 static int htab_lru_percpu_map_btf_id;
1993 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1994 .map_meta_equal = bpf_map_meta_equal,
1995 .map_alloc_check = htab_map_alloc_check,
1996 .map_alloc = htab_map_alloc,
1997 .map_free = htab_map_free,
1998 .map_get_next_key = htab_map_get_next_key,
1999 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2000 .map_update_elem = htab_lru_percpu_map_update_elem,
2001 .map_delete_elem = htab_lru_map_delete_elem,
2002 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2003 BATCH_OPS(htab_lru_percpu),
2004 .map_btf_name = "bpf_htab",
2005 .map_btf_id = &htab_lru_percpu_map_btf_id,
2006 .iter_seq_info = &iter_seq_info,
2009 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2011 if (attr->value_size != sizeof(u32))
2013 return htab_map_alloc_check(attr);
2016 static void fd_htab_map_free(struct bpf_map *map)
2018 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2019 struct hlist_nulls_node *n;
2020 struct hlist_nulls_head *head;
2021 struct htab_elem *l;
2024 for (i = 0; i < htab->n_buckets; i++) {
2025 head = select_bucket(htab, i);
2027 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2028 void *ptr = fd_htab_map_get_ptr(map, l);
2030 map->ops->map_fd_put_ptr(map, ptr, false);
2037 /* only called from syscall */
2038 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2043 if (!map->ops->map_fd_sys_lookup_elem)
2047 ptr = htab_map_lookup_elem(map, key);
2049 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2057 /* only called from syscall */
2058 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2059 void *key, void *value, u64 map_flags)
2063 u32 ufd = *(u32 *)value;
2065 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2067 return PTR_ERR(ptr);
2069 ret = htab_map_update_elem(map, key, &ptr, map_flags);
2071 map->ops->map_fd_put_ptr(map, ptr, false);
2076 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2078 struct bpf_map *map, *inner_map_meta;
2080 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2081 if (IS_ERR(inner_map_meta))
2082 return inner_map_meta;
2084 map = htab_map_alloc(attr);
2086 bpf_map_meta_free(inner_map_meta);
2090 map->inner_map_meta = inner_map_meta;
2095 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2097 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
2102 return READ_ONCE(*inner_map);
2105 static int htab_of_map_gen_lookup(struct bpf_map *map,
2106 struct bpf_insn *insn_buf)
2108 struct bpf_insn *insn = insn_buf;
2109 const int ret = BPF_REG_0;
2111 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2112 (void *(*)(struct bpf_map *map, void *key))NULL));
2113 *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
2114 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2115 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2116 offsetof(struct htab_elem, key) +
2117 round_up(map->key_size, 8));
2118 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2120 return insn - insn_buf;
2123 static void htab_of_map_free(struct bpf_map *map)
2125 bpf_map_meta_free(map->inner_map_meta);
2126 fd_htab_map_free(map);
2129 static int htab_of_maps_map_btf_id;
2130 const struct bpf_map_ops htab_of_maps_map_ops = {
2131 .map_alloc_check = fd_htab_map_alloc_check,
2132 .map_alloc = htab_of_map_alloc,
2133 .map_free = htab_of_map_free,
2134 .map_get_next_key = htab_map_get_next_key,
2135 .map_lookup_elem = htab_of_map_lookup_elem,
2136 .map_delete_elem = htab_map_delete_elem,
2137 .map_fd_get_ptr = bpf_map_fd_get_ptr,
2138 .map_fd_put_ptr = bpf_map_fd_put_ptr,
2139 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2140 .map_gen_lookup = htab_of_map_gen_lookup,
2141 .map_check_btf = map_check_no_btf,
2142 .map_btf_name = "bpf_htab",
2143 .map_btf_id = &htab_of_maps_map_btf_id,