1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
7 #include <linux/jhash.h>
8 #include <linux/filter.h>
9 #include <linux/rculist_nulls.h>
10 #include <linux/random.h>
11 #include <uapi/linux/btf.h>
12 #include <linux/rcupdate_trace.h>
13 #include <linux/btf_ids.h>
14 #include "percpu_freelist.h"
15 #include "bpf_lru_list.h"
16 #include "map_in_map.h"
17 #include <linux/bpf_mem_alloc.h>
19 #define HTAB_CREATE_FLAG_MASK \
20 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
21 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
23 #define BATCH_OPS(_name) \
25 _name##_map_lookup_batch, \
26 .map_lookup_and_delete_batch = \
27 _name##_map_lookup_and_delete_batch, \
29 generic_map_update_batch, \
31 generic_map_delete_batch
34 * The bucket lock has two protection scopes:
36 * 1) Serializing concurrent operations from BPF programs on different
39 * 2) Serializing concurrent operations from BPF programs and sys_bpf()
41 * BPF programs can execute in any context including perf, kprobes and
42 * tracing. As there are almost no limits where perf, kprobes and tracing
43 * can be invoked from the lock operations need to be protected against
44 * deadlocks. Deadlocks can be caused by recursion and by an invocation in
45 * the lock held section when functions which acquire this lock are invoked
46 * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
47 * variable bpf_prog_active, which prevents BPF programs attached to perf
48 * events, kprobes and tracing to be invoked before the prior invocation
49 * from one of these contexts completed. sys_bpf() uses the same mechanism
50 * by pinning the task to the current CPU and incrementing the recursion
51 * protection across the map operation.
53 * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
54 * operations like memory allocations (even with GFP_ATOMIC) from atomic
55 * contexts. This is required because even with GFP_ATOMIC the memory
56 * allocator calls into code paths which acquire locks with long held lock
57 * sections. To ensure the deterministic behaviour these locks are regular
58 * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
59 * true atomic contexts on an RT kernel are the low level hardware
60 * handling, scheduling, low level interrupt handling, NMIs etc. None of
61 * these contexts should ever do memory allocations.
63 * As regular device interrupt handlers and soft interrupts are forced into
64 * thread context, the existing code which does
65 * spin_lock*(); alloc(GFP_ATOMIC); spin_unlock*();
68 * In theory the BPF locks could be converted to regular spinlocks as well,
69 * but the bucket locks and percpu_freelist locks can be taken from
70 * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
71 * atomic contexts even on RT. Before the introduction of bpf_mem_alloc,
72 * it is only safe to use raw spinlock for preallocated hash map on a RT kernel,
73 * because there is no memory allocation within the lock held sections. However
74 * after hash map was fully converted to use bpf_mem_alloc, there will be
75 * non-synchronous memory allocation for non-preallocated hash map, so it is
76 * safe to always use raw spinlock for bucket lock.
79 struct hlist_nulls_head head;
80 raw_spinlock_t raw_lock;
83 #define HASHTAB_MAP_LOCK_COUNT 8
84 #define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
88 struct bpf_mem_alloc ma;
89 struct bpf_mem_alloc pcpu_ma;
90 struct bucket *buckets;
93 struct pcpu_freelist freelist;
96 struct htab_elem *__percpu *extra_elems;
97 /* number of elements in non-preallocated hashtable are kept
98 * in either pcount or count
100 struct percpu_counter pcount;
102 bool use_percpu_counter;
103 u32 n_buckets; /* number of hash buckets */
104 u32 elem_size; /* size of each element in bytes */
106 struct lock_class_key lockdep_key;
107 int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
110 /* each htab element is struct htab_elem + key + value */
113 struct hlist_nulls_node hash_node;
117 struct pcpu_freelist_node fnode;
118 struct htab_elem *batch_flink;
123 /* pointer to per-cpu pointer */
125 struct bpf_lru_node lru_node;
128 char key[] __aligned(8);
131 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
133 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
136 static void htab_init_buckets(struct bpf_htab *htab)
140 for (i = 0; i < htab->n_buckets; i++) {
141 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
142 raw_spin_lock_init(&htab->buckets[i].raw_lock);
143 lockdep_set_class(&htab->buckets[i].raw_lock,
149 static inline int htab_lock_bucket(const struct bpf_htab *htab,
150 struct bucket *b, u32 hash,
151 unsigned long *pflags)
155 hash = hash & HASHTAB_MAP_LOCK_MASK;
158 if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
159 __this_cpu_dec(*(htab->map_locked[hash]));
164 raw_spin_lock_irqsave(&b->raw_lock, flags);
170 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
171 struct bucket *b, u32 hash,
174 hash = hash & HASHTAB_MAP_LOCK_MASK;
175 raw_spin_unlock_irqrestore(&b->raw_lock, flags);
176 __this_cpu_dec(*(htab->map_locked[hash]));
180 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
182 static bool htab_is_lru(const struct bpf_htab *htab)
184 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
185 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
188 static bool htab_is_percpu(const struct bpf_htab *htab)
190 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
191 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
194 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
197 *(void __percpu **)(l->key + key_size) = pptr;
200 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
202 return *(void __percpu **)(l->key + key_size);
205 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
207 return *(void **)(l->key + roundup(map->key_size, 8));
210 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
212 return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
215 static bool htab_has_extra_elems(struct bpf_htab *htab)
217 return !htab_is_percpu(htab) && !htab_is_lru(htab);
220 static void htab_free_prealloced_timers(struct bpf_htab *htab)
222 u32 num_entries = htab->map.max_entries;
225 if (!map_value_has_timer(&htab->map))
227 if (htab_has_extra_elems(htab))
228 num_entries += num_possible_cpus();
230 for (i = 0; i < num_entries; i++) {
231 struct htab_elem *elem;
233 elem = get_htab_elem(htab, i);
234 bpf_timer_cancel_and_free(elem->key +
235 round_up(htab->map.key_size, 8) +
236 htab->map.timer_off);
241 static void htab_free_prealloced_kptrs(struct bpf_htab *htab)
243 u32 num_entries = htab->map.max_entries;
246 if (!map_value_has_kptrs(&htab->map))
248 if (htab_has_extra_elems(htab))
249 num_entries += num_possible_cpus();
251 for (i = 0; i < num_entries; i++) {
252 struct htab_elem *elem;
254 elem = get_htab_elem(htab, i);
255 bpf_map_free_kptrs(&htab->map, elem->key + round_up(htab->map.key_size, 8));
260 static void htab_free_elems(struct bpf_htab *htab)
264 if (!htab_is_percpu(htab))
267 for (i = 0; i < htab->map.max_entries; i++) {
270 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
276 bpf_map_area_free(htab->elems);
279 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
280 * (bucket_lock). If both locks need to be acquired together, the lock
281 * order is always lru_lock -> bucket_lock and this only happens in
282 * bpf_lru_list.c logic. For example, certain code path of
283 * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
284 * will acquire lru_lock first followed by acquiring bucket_lock.
286 * In hashtab.c, to avoid deadlock, lock acquisition of
287 * bucket_lock followed by lru_lock is not allowed. In such cases,
288 * bucket_lock needs to be released first before acquiring lru_lock.
290 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
293 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
297 l = container_of(node, struct htab_elem, lru_node);
298 memcpy(l->key, key, htab->map.key_size);
305 static int prealloc_init(struct bpf_htab *htab)
307 u32 num_entries = htab->map.max_entries;
308 int err = -ENOMEM, i;
310 if (htab_has_extra_elems(htab))
311 num_entries += num_possible_cpus();
313 htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
314 htab->map.numa_node);
318 if (!htab_is_percpu(htab))
319 goto skip_percpu_elems;
321 for (i = 0; i < num_entries; i++) {
322 u32 size = round_up(htab->map.value_size, 8);
325 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
326 GFP_USER | __GFP_NOWARN);
329 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
335 if (htab_is_lru(htab))
336 err = bpf_lru_init(&htab->lru,
337 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
338 offsetof(struct htab_elem, hash) -
339 offsetof(struct htab_elem, lru_node),
340 htab_lru_map_delete_node,
343 err = pcpu_freelist_init(&htab->freelist);
348 if (htab_is_lru(htab))
349 bpf_lru_populate(&htab->lru, htab->elems,
350 offsetof(struct htab_elem, lru_node),
351 htab->elem_size, num_entries);
353 pcpu_freelist_populate(&htab->freelist,
354 htab->elems + offsetof(struct htab_elem, fnode),
355 htab->elem_size, num_entries);
360 htab_free_elems(htab);
364 static void prealloc_destroy(struct bpf_htab *htab)
366 htab_free_elems(htab);
368 if (htab_is_lru(htab))
369 bpf_lru_destroy(&htab->lru);
371 pcpu_freelist_destroy(&htab->freelist);
374 static int alloc_extra_elems(struct bpf_htab *htab)
376 struct htab_elem *__percpu *pptr, *l_new;
377 struct pcpu_freelist_node *l;
380 pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
381 GFP_USER | __GFP_NOWARN);
385 for_each_possible_cpu(cpu) {
386 l = pcpu_freelist_pop(&htab->freelist);
387 /* pop will succeed, since prealloc_init()
388 * preallocated extra num_possible_cpus elements
390 l_new = container_of(l, struct htab_elem, fnode);
391 *per_cpu_ptr(pptr, cpu) = l_new;
393 htab->extra_elems = pptr;
397 /* Called from syscall */
398 static int htab_map_alloc_check(union bpf_attr *attr)
400 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
401 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
402 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
403 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
404 /* percpu_lru means each cpu has its own LRU list.
405 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
406 * the map's value itself is percpu. percpu_lru has
407 * nothing to do with the map's value.
409 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
410 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
411 bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
412 int numa_node = bpf_map_attr_numa_node(attr);
414 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
415 offsetof(struct htab_elem, hash_node.pprev));
417 if (lru && !bpf_capable())
418 /* LRU implementation is much complicated than other
419 * maps. Hence, limit to CAP_BPF.
423 if (zero_seed && !capable(CAP_SYS_ADMIN))
424 /* Guard against local DoS, and discourage production use. */
427 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
428 !bpf_map_flags_access_ok(attr->map_flags))
431 if (!lru && percpu_lru)
434 if (lru && !prealloc)
437 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
440 /* check sanity of attributes.
441 * value_size == 0 may be allowed in the future to use map as a set
443 if (attr->max_entries == 0 || attr->key_size == 0 ||
444 attr->value_size == 0)
447 if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
448 sizeof(struct htab_elem))
449 /* if key_size + value_size is bigger, the user space won't be
450 * able to access the elements via bpf syscall. This check
451 * also makes sure that the elem_size doesn't overflow and it's
452 * kmalloc-able later in htab_map_update_elem()
459 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
461 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
462 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
463 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
464 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
465 /* percpu_lru means each cpu has its own LRU list.
466 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
467 * the map's value itself is percpu. percpu_lru has
468 * nothing to do with the map's value.
470 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
471 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
472 struct bpf_htab *htab;
475 htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
477 return ERR_PTR(-ENOMEM);
479 lockdep_register_key(&htab->lockdep_key);
481 bpf_map_init_from_attr(&htab->map, attr);
484 /* ensure each CPU's lru list has >=1 elements.
485 * since we are at it, make each lru list has the same
486 * number of elements.
488 htab->map.max_entries = roundup(attr->max_entries,
489 num_possible_cpus());
490 if (htab->map.max_entries < attr->max_entries)
491 htab->map.max_entries = rounddown(attr->max_entries,
492 num_possible_cpus());
495 /* hash table size must be power of 2 */
496 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
498 htab->elem_size = sizeof(struct htab_elem) +
499 round_up(htab->map.key_size, 8);
501 htab->elem_size += sizeof(void *);
503 htab->elem_size += round_up(htab->map.value_size, 8);
506 /* prevent zero size kmalloc and check for u32 overflow */
507 if (htab->n_buckets == 0 ||
508 htab->n_buckets > U32_MAX / sizeof(struct bucket))
512 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
513 sizeof(struct bucket),
514 htab->map.numa_node);
518 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
519 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
523 if (!htab->map_locked[i])
524 goto free_map_locked;
527 if (htab->map.map_flags & BPF_F_ZERO_SEED)
530 htab->hashrnd = get_random_u32();
532 htab_init_buckets(htab);
534 /* compute_batch_value() computes batch value as num_online_cpus() * 2
535 * and __percpu_counter_compare() needs
536 * htab->max_entries - cur_number_of_elems to be more than batch * num_online_cpus()
537 * for percpu_counter to be faster than atomic_t. In practice the average bpf
538 * hash map size is 10k, which means that a system with 64 cpus will fill
539 * hashmap to 20% of 10k before percpu_counter becomes ineffective. Therefore
540 * define our own batch count as 32 then 10k hash map can be filled up to 80%:
541 * 10k - 8k > 32 _batch_ * 64 _cpus_
542 * and __percpu_counter_compare() will still be fast. At that point hash map
543 * collisions will dominate its performance anyway. Assume that hash map filled
544 * to 50+% isn't going to be O(1) and use the following formula to choose
545 * between percpu_counter and atomic_t.
547 #define PERCPU_COUNTER_BATCH 32
548 if (attr->max_entries / 2 > num_online_cpus() * PERCPU_COUNTER_BATCH)
549 htab->use_percpu_counter = true;
551 if (htab->use_percpu_counter) {
552 err = percpu_counter_init(&htab->pcount, 0, GFP_KERNEL);
554 goto free_map_locked;
558 err = prealloc_init(htab);
560 goto free_map_locked;
562 if (!percpu && !lru) {
563 /* lru itself can remove the least used element, so
564 * there is no need for an extra elem during map_update.
566 err = alloc_extra_elems(htab);
571 err = bpf_mem_alloc_init(&htab->ma, htab->elem_size, false);
573 goto free_map_locked;
575 err = bpf_mem_alloc_init(&htab->pcpu_ma,
576 round_up(htab->map.value_size, 8), true);
578 goto free_map_locked;
585 prealloc_destroy(htab);
587 if (htab->use_percpu_counter)
588 percpu_counter_destroy(&htab->pcount);
589 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
590 free_percpu(htab->map_locked[i]);
591 bpf_map_area_free(htab->buckets);
592 bpf_mem_alloc_destroy(&htab->pcpu_ma);
593 bpf_mem_alloc_destroy(&htab->ma);
595 lockdep_unregister_key(&htab->lockdep_key);
596 bpf_map_area_free(htab);
600 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
602 return jhash(key, key_len, hashrnd);
605 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
607 return &htab->buckets[hash & (htab->n_buckets - 1)];
610 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
612 return &__select_bucket(htab, hash)->head;
615 /* this lookup function can only be called with bucket lock taken */
616 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
617 void *key, u32 key_size)
619 struct hlist_nulls_node *n;
622 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
623 if (l->hash == hash && !memcmp(&l->key, key, key_size))
629 /* can be called without bucket lock. it will repeat the loop in
630 * the unlikely event when elements moved from one bucket into another
631 * while link list is being walked
633 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
635 u32 key_size, u32 n_buckets)
637 struct hlist_nulls_node *n;
641 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
642 if (l->hash == hash && !memcmp(&l->key, key, key_size))
645 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
651 /* Called from syscall or from eBPF program directly, so
652 * arguments have to match bpf_map_lookup_elem() exactly.
653 * The return value is adjusted by BPF instructions
654 * in htab_map_gen_lookup().
656 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
658 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
659 struct hlist_nulls_head *head;
663 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
664 !rcu_read_lock_bh_held());
666 key_size = map->key_size;
668 hash = htab_map_hash(key, key_size, htab->hashrnd);
670 head = select_bucket(htab, hash);
672 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
677 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
679 struct htab_elem *l = __htab_map_lookup_elem(map, key);
682 return l->key + round_up(map->key_size, 8);
687 /* inline bpf_map_lookup_elem() call.
690 * bpf_map_lookup_elem
691 * map->ops->map_lookup_elem
692 * htab_map_lookup_elem
693 * __htab_map_lookup_elem
696 * __htab_map_lookup_elem
698 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
700 struct bpf_insn *insn = insn_buf;
701 const int ret = BPF_REG_0;
703 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
704 (void *(*)(struct bpf_map *map, void *key))NULL));
705 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
706 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
707 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
708 offsetof(struct htab_elem, key) +
709 round_up(map->key_size, 8));
710 return insn - insn_buf;
713 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
714 void *key, const bool mark)
716 struct htab_elem *l = __htab_map_lookup_elem(map, key);
720 bpf_lru_node_set_ref(&l->lru_node);
721 return l->key + round_up(map->key_size, 8);
727 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
729 return __htab_lru_map_lookup_elem(map, key, true);
732 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
734 return __htab_lru_map_lookup_elem(map, key, false);
737 static int htab_lru_map_gen_lookup(struct bpf_map *map,
738 struct bpf_insn *insn_buf)
740 struct bpf_insn *insn = insn_buf;
741 const int ret = BPF_REG_0;
742 const int ref_reg = BPF_REG_1;
744 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
745 (void *(*)(struct bpf_map *map, void *key))NULL));
746 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
747 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
748 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
749 offsetof(struct htab_elem, lru_node) +
750 offsetof(struct bpf_lru_node, ref));
751 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
752 *insn++ = BPF_ST_MEM(BPF_B, ret,
753 offsetof(struct htab_elem, lru_node) +
754 offsetof(struct bpf_lru_node, ref),
756 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
757 offsetof(struct htab_elem, key) +
758 round_up(map->key_size, 8));
759 return insn - insn_buf;
762 static void check_and_free_fields(struct bpf_htab *htab,
763 struct htab_elem *elem)
765 void *map_value = elem->key + round_up(htab->map.key_size, 8);
767 if (map_value_has_timer(&htab->map))
768 bpf_timer_cancel_and_free(map_value + htab->map.timer_off);
769 if (map_value_has_kptrs(&htab->map))
770 bpf_map_free_kptrs(&htab->map, map_value);
773 /* It is called from the bpf_lru_list when the LRU needs to delete
774 * older elements from the htab.
776 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
778 struct bpf_htab *htab = arg;
779 struct htab_elem *l = NULL, *tgt_l;
780 struct hlist_nulls_head *head;
781 struct hlist_nulls_node *n;
786 tgt_l = container_of(node, struct htab_elem, lru_node);
787 b = __select_bucket(htab, tgt_l->hash);
790 ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
794 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
796 hlist_nulls_del_rcu(&l->hash_node);
797 check_and_free_fields(htab, l);
801 htab_unlock_bucket(htab, b, tgt_l->hash, flags);
806 /* Called from syscall */
807 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
809 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
810 struct hlist_nulls_head *head;
811 struct htab_elem *l, *next_l;
815 WARN_ON_ONCE(!rcu_read_lock_held());
817 key_size = map->key_size;
820 goto find_first_elem;
822 hash = htab_map_hash(key, key_size, htab->hashrnd);
824 head = select_bucket(htab, hash);
827 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
830 goto find_first_elem;
832 /* key was found, get next key in the same bucket */
833 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
834 struct htab_elem, hash_node);
837 /* if next elem in this hash list is non-zero, just return it */
838 memcpy(next_key, next_l->key, key_size);
842 /* no more elements in this hash list, go to the next bucket */
843 i = hash & (htab->n_buckets - 1);
847 /* iterate over buckets */
848 for (; i < htab->n_buckets; i++) {
849 head = select_bucket(htab, i);
851 /* pick first element in the bucket */
852 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
853 struct htab_elem, hash_node);
855 /* if it's not empty, just return it */
856 memcpy(next_key, next_l->key, key_size);
861 /* iterated over all buckets and all elements */
865 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
867 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
868 bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
869 check_and_free_fields(htab, l);
870 bpf_mem_cache_free(&htab->ma, l);
873 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
875 struct bpf_map *map = &htab->map;
878 if (map->ops->map_fd_put_ptr) {
879 ptr = fd_htab_map_get_ptr(map, l);
880 map->ops->map_fd_put_ptr(ptr);
884 static bool is_map_full(struct bpf_htab *htab)
886 if (htab->use_percpu_counter)
887 return __percpu_counter_compare(&htab->pcount, htab->map.max_entries,
888 PERCPU_COUNTER_BATCH) >= 0;
889 return atomic_read(&htab->count) >= htab->map.max_entries;
892 static void inc_elem_count(struct bpf_htab *htab)
894 if (htab->use_percpu_counter)
895 percpu_counter_add_batch(&htab->pcount, 1, PERCPU_COUNTER_BATCH);
897 atomic_inc(&htab->count);
900 static void dec_elem_count(struct bpf_htab *htab)
902 if (htab->use_percpu_counter)
903 percpu_counter_add_batch(&htab->pcount, -1, PERCPU_COUNTER_BATCH);
905 atomic_dec(&htab->count);
909 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
911 htab_put_fd_value(htab, l);
913 if (htab_is_prealloc(htab)) {
914 check_and_free_fields(htab, l);
915 __pcpu_freelist_push(&htab->freelist, &l->fnode);
917 dec_elem_count(htab);
918 htab_elem_free(htab, l);
922 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
923 void *value, bool onallcpus)
926 /* copy true value_size bytes */
927 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
929 u32 size = round_up(htab->map.value_size, 8);
932 for_each_possible_cpu(cpu) {
933 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
940 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
941 void *value, bool onallcpus)
943 /* When not setting the initial value on all cpus, zero-fill element
944 * values for other cpus. Otherwise, bpf program has no way to ensure
945 * known initial values for cpus other than current one
946 * (onallcpus=false always when coming from bpf prog).
949 u32 size = round_up(htab->map.value_size, 8);
950 int current_cpu = raw_smp_processor_id();
953 for_each_possible_cpu(cpu) {
954 if (cpu == current_cpu)
955 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value,
958 memset(per_cpu_ptr(pptr, cpu), 0, size);
961 pcpu_copy_value(htab, pptr, value, onallcpus);
965 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
967 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
971 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
972 void *value, u32 key_size, u32 hash,
973 bool percpu, bool onallcpus,
974 struct htab_elem *old_elem)
976 u32 size = htab->map.value_size;
977 bool prealloc = htab_is_prealloc(htab);
978 struct htab_elem *l_new, **pl_new;
983 /* if we're updating the existing element,
984 * use per-cpu extra elems to avoid freelist_pop/push
986 pl_new = this_cpu_ptr(htab->extra_elems);
988 htab_put_fd_value(htab, old_elem);
991 struct pcpu_freelist_node *l;
993 l = __pcpu_freelist_pop(&htab->freelist);
995 return ERR_PTR(-E2BIG);
996 l_new = container_of(l, struct htab_elem, fnode);
999 if (is_map_full(htab))
1001 /* when map is full and update() is replacing
1002 * old element, it's ok to allocate, since
1003 * old element will be freed immediately.
1004 * Otherwise return an error
1006 return ERR_PTR(-E2BIG);
1007 inc_elem_count(htab);
1008 l_new = bpf_mem_cache_alloc(&htab->ma);
1010 l_new = ERR_PTR(-ENOMEM);
1013 check_and_init_map_value(&htab->map,
1014 l_new->key + round_up(key_size, 8));
1017 memcpy(l_new->key, key, key_size);
1020 pptr = htab_elem_get_ptr(l_new, key_size);
1022 /* alloc_percpu zero-fills */
1023 pptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
1025 bpf_mem_cache_free(&htab->ma, l_new);
1026 l_new = ERR_PTR(-ENOMEM);
1029 l_new->ptr_to_pptr = pptr;
1030 pptr = *(void **)pptr;
1033 pcpu_init_value(htab, pptr, value, onallcpus);
1036 htab_elem_set_ptr(l_new, key_size, pptr);
1037 } else if (fd_htab_map_needs_adjust(htab)) {
1038 size = round_up(size, 8);
1039 memcpy(l_new->key + round_up(key_size, 8), value, size);
1041 copy_map_value(&htab->map,
1042 l_new->key + round_up(key_size, 8),
1049 dec_elem_count(htab);
1053 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1056 if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1057 /* elem already exists */
1060 if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1061 /* elem doesn't exist, cannot update it */
1067 /* Called from syscall or from eBPF program */
1068 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1071 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1072 struct htab_elem *l_new = NULL, *l_old;
1073 struct hlist_nulls_head *head;
1074 unsigned long flags;
1079 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1083 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1084 !rcu_read_lock_bh_held());
1086 key_size = map->key_size;
1088 hash = htab_map_hash(key, key_size, htab->hashrnd);
1090 b = __select_bucket(htab, hash);
1093 if (unlikely(map_flags & BPF_F_LOCK)) {
1094 if (unlikely(!map_value_has_spin_lock(map)))
1096 /* find an element without taking the bucket lock */
1097 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1099 ret = check_flags(htab, l_old, map_flags);
1103 /* grab the element lock and update value in place */
1104 copy_map_value_locked(map,
1105 l_old->key + round_up(key_size, 8),
1109 /* fall through, grab the bucket lock and lookup again.
1110 * 99.9% chance that the element won't be found,
1111 * but second lookup under lock has to be done.
1115 ret = htab_lock_bucket(htab, b, hash, &flags);
1119 l_old = lookup_elem_raw(head, hash, key, key_size);
1121 ret = check_flags(htab, l_old, map_flags);
1125 if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1126 /* first lookup without the bucket lock didn't find the element,
1127 * but second lookup with the bucket lock found it.
1128 * This case is highly unlikely, but has to be dealt with:
1129 * grab the element lock in addition to the bucket lock
1130 * and update element in place
1132 copy_map_value_locked(map,
1133 l_old->key + round_up(key_size, 8),
1139 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1141 if (IS_ERR(l_new)) {
1142 /* all pre-allocated elements are in use or memory exhausted */
1143 ret = PTR_ERR(l_new);
1147 /* add new element to the head of the list, so that
1148 * concurrent search will find it before old elem
1150 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1152 hlist_nulls_del_rcu(&l_old->hash_node);
1153 if (!htab_is_prealloc(htab))
1154 free_htab_elem(htab, l_old);
1156 check_and_free_fields(htab, l_old);
1160 htab_unlock_bucket(htab, b, hash, flags);
1164 static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1166 check_and_free_fields(htab, elem);
1167 bpf_lru_push_free(&htab->lru, &elem->lru_node);
1170 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1173 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1174 struct htab_elem *l_new, *l_old = NULL;
1175 struct hlist_nulls_head *head;
1176 unsigned long flags;
1181 if (unlikely(map_flags > BPF_EXIST))
1185 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1186 !rcu_read_lock_bh_held());
1188 key_size = map->key_size;
1190 hash = htab_map_hash(key, key_size, htab->hashrnd);
1192 b = __select_bucket(htab, hash);
1195 /* For LRU, we need to alloc before taking bucket's
1196 * spinlock because getting free nodes from LRU may need
1197 * to remove older elements from htab and this removal
1198 * operation will need a bucket lock.
1200 l_new = prealloc_lru_pop(htab, key, hash);
1203 copy_map_value(&htab->map,
1204 l_new->key + round_up(map->key_size, 8), value);
1206 ret = htab_lock_bucket(htab, b, hash, &flags);
1210 l_old = lookup_elem_raw(head, hash, key, key_size);
1212 ret = check_flags(htab, l_old, map_flags);
1216 /* add new element to the head of the list, so that
1217 * concurrent search will find it before old elem
1219 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1221 bpf_lru_node_set_ref(&l_new->lru_node);
1222 hlist_nulls_del_rcu(&l_old->hash_node);
1227 htab_unlock_bucket(htab, b, hash, flags);
1230 htab_lru_push_free(htab, l_new);
1232 htab_lru_push_free(htab, l_old);
1237 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1238 void *value, u64 map_flags,
1241 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1242 struct htab_elem *l_new = NULL, *l_old;
1243 struct hlist_nulls_head *head;
1244 unsigned long flags;
1249 if (unlikely(map_flags > BPF_EXIST))
1253 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1254 !rcu_read_lock_bh_held());
1256 key_size = map->key_size;
1258 hash = htab_map_hash(key, key_size, htab->hashrnd);
1260 b = __select_bucket(htab, hash);
1263 ret = htab_lock_bucket(htab, b, hash, &flags);
1267 l_old = lookup_elem_raw(head, hash, key, key_size);
1269 ret = check_flags(htab, l_old, map_flags);
1274 /* per-cpu hash map can update value in-place */
1275 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1278 l_new = alloc_htab_elem(htab, key, value, key_size,
1279 hash, true, onallcpus, NULL);
1280 if (IS_ERR(l_new)) {
1281 ret = PTR_ERR(l_new);
1284 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1288 htab_unlock_bucket(htab, b, hash, flags);
1292 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1293 void *value, u64 map_flags,
1296 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1297 struct htab_elem *l_new = NULL, *l_old;
1298 struct hlist_nulls_head *head;
1299 unsigned long flags;
1304 if (unlikely(map_flags > BPF_EXIST))
1308 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1309 !rcu_read_lock_bh_held());
1311 key_size = map->key_size;
1313 hash = htab_map_hash(key, key_size, htab->hashrnd);
1315 b = __select_bucket(htab, hash);
1318 /* For LRU, we need to alloc before taking bucket's
1319 * spinlock because LRU's elem alloc may need
1320 * to remove older elem from htab and this removal
1321 * operation will need a bucket lock.
1323 if (map_flags != BPF_EXIST) {
1324 l_new = prealloc_lru_pop(htab, key, hash);
1329 ret = htab_lock_bucket(htab, b, hash, &flags);
1333 l_old = lookup_elem_raw(head, hash, key, key_size);
1335 ret = check_flags(htab, l_old, map_flags);
1340 bpf_lru_node_set_ref(&l_old->lru_node);
1342 /* per-cpu hash map can update value in-place */
1343 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1346 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1348 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1353 htab_unlock_bucket(htab, b, hash, flags);
1355 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1359 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1360 void *value, u64 map_flags)
1362 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1365 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1366 void *value, u64 map_flags)
1368 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1372 /* Called from syscall or from eBPF program */
1373 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1375 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1376 struct hlist_nulls_head *head;
1378 struct htab_elem *l;
1379 unsigned long flags;
1383 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1384 !rcu_read_lock_bh_held());
1386 key_size = map->key_size;
1388 hash = htab_map_hash(key, key_size, htab->hashrnd);
1389 b = __select_bucket(htab, hash);
1392 ret = htab_lock_bucket(htab, b, hash, &flags);
1396 l = lookup_elem_raw(head, hash, key, key_size);
1399 hlist_nulls_del_rcu(&l->hash_node);
1400 free_htab_elem(htab, l);
1405 htab_unlock_bucket(htab, b, hash, flags);
1409 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1411 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1412 struct hlist_nulls_head *head;
1414 struct htab_elem *l;
1415 unsigned long flags;
1419 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1420 !rcu_read_lock_bh_held());
1422 key_size = map->key_size;
1424 hash = htab_map_hash(key, key_size, htab->hashrnd);
1425 b = __select_bucket(htab, hash);
1428 ret = htab_lock_bucket(htab, b, hash, &flags);
1432 l = lookup_elem_raw(head, hash, key, key_size);
1435 hlist_nulls_del_rcu(&l->hash_node);
1439 htab_unlock_bucket(htab, b, hash, flags);
1441 htab_lru_push_free(htab, l);
1445 static void delete_all_elements(struct bpf_htab *htab)
1449 /* It's called from a worker thread, so disable migration here,
1450 * since bpf_mem_cache_free() relies on that.
1453 for (i = 0; i < htab->n_buckets; i++) {
1454 struct hlist_nulls_head *head = select_bucket(htab, i);
1455 struct hlist_nulls_node *n;
1456 struct htab_elem *l;
1458 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1459 hlist_nulls_del_rcu(&l->hash_node);
1460 htab_elem_free(htab, l);
1466 static void htab_free_malloced_timers(struct bpf_htab *htab)
1471 for (i = 0; i < htab->n_buckets; i++) {
1472 struct hlist_nulls_head *head = select_bucket(htab, i);
1473 struct hlist_nulls_node *n;
1474 struct htab_elem *l;
1476 hlist_nulls_for_each_entry(l, n, head, hash_node) {
1477 /* We don't reset or free kptr on uref dropping to zero,
1478 * hence just free timer.
1480 bpf_timer_cancel_and_free(l->key +
1481 round_up(htab->map.key_size, 8) +
1482 htab->map.timer_off);
1489 static void htab_map_free_timers(struct bpf_map *map)
1491 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1493 /* We don't reset or free kptr on uref dropping to zero. */
1494 if (!map_value_has_timer(&htab->map))
1496 if (!htab_is_prealloc(htab))
1497 htab_free_malloced_timers(htab);
1499 htab_free_prealloced_timers(htab);
1502 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1503 static void htab_map_free(struct bpf_map *map)
1505 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1508 /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1509 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1510 * There is no need to synchronize_rcu() here to protect map elements.
1513 /* htab no longer uses call_rcu() directly. bpf_mem_alloc does it
1514 * underneath and is reponsible for waiting for callbacks to finish
1515 * during bpf_mem_alloc_destroy().
1517 if (!htab_is_prealloc(htab)) {
1518 delete_all_elements(htab);
1520 htab_free_prealloced_kptrs(htab);
1521 prealloc_destroy(htab);
1524 bpf_map_free_kptr_off_tab(map);
1525 free_percpu(htab->extra_elems);
1526 bpf_map_area_free(htab->buckets);
1527 bpf_mem_alloc_destroy(&htab->pcpu_ma);
1528 bpf_mem_alloc_destroy(&htab->ma);
1529 if (htab->use_percpu_counter)
1530 percpu_counter_destroy(&htab->pcount);
1531 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1532 free_percpu(htab->map_locked[i]);
1533 lockdep_unregister_key(&htab->lockdep_key);
1534 bpf_map_area_free(htab);
1537 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1544 value = htab_map_lookup_elem(map, key);
1550 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1552 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1558 static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1559 void *value, bool is_lru_map,
1560 bool is_percpu, u64 flags)
1562 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1563 struct hlist_nulls_head *head;
1564 unsigned long bflags;
1565 struct htab_elem *l;
1570 key_size = map->key_size;
1572 hash = htab_map_hash(key, key_size, htab->hashrnd);
1573 b = __select_bucket(htab, hash);
1576 ret = htab_lock_bucket(htab, b, hash, &bflags);
1580 l = lookup_elem_raw(head, hash, key, key_size);
1585 u32 roundup_value_size = round_up(map->value_size, 8);
1586 void __percpu *pptr;
1589 pptr = htab_elem_get_ptr(l, key_size);
1590 for_each_possible_cpu(cpu) {
1591 bpf_long_memcpy(value + off,
1592 per_cpu_ptr(pptr, cpu),
1593 roundup_value_size);
1594 off += roundup_value_size;
1597 u32 roundup_key_size = round_up(map->key_size, 8);
1599 if (flags & BPF_F_LOCK)
1600 copy_map_value_locked(map, value, l->key +
1604 copy_map_value(map, value, l->key +
1606 check_and_init_map_value(map, value);
1609 hlist_nulls_del_rcu(&l->hash_node);
1611 free_htab_elem(htab, l);
1614 htab_unlock_bucket(htab, b, hash, bflags);
1616 if (is_lru_map && l)
1617 htab_lru_push_free(htab, l);
1622 static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1623 void *value, u64 flags)
1625 return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1629 static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1630 void *key, void *value,
1633 return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1637 static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1638 void *value, u64 flags)
1640 return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1644 static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1645 void *key, void *value,
1648 return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1653 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1654 const union bpf_attr *attr,
1655 union bpf_attr __user *uattr,
1656 bool do_delete, bool is_lru_map,
1659 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1660 u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1661 void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1662 void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1663 void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1664 void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1665 u32 batch, max_count, size, bucket_size, map_id;
1666 struct htab_elem *node_to_free = NULL;
1667 u64 elem_map_flags, map_flags;
1668 struct hlist_nulls_head *head;
1669 struct hlist_nulls_node *n;
1670 unsigned long flags = 0;
1671 bool locked = false;
1672 struct htab_elem *l;
1676 elem_map_flags = attr->batch.elem_flags;
1677 if ((elem_map_flags & ~BPF_F_LOCK) ||
1678 ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1681 map_flags = attr->batch.flags;
1685 max_count = attr->batch.count;
1689 if (put_user(0, &uattr->batch.count))
1693 if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1696 if (batch >= htab->n_buckets)
1699 key_size = htab->map.key_size;
1700 roundup_key_size = round_up(htab->map.key_size, 8);
1701 value_size = htab->map.value_size;
1702 size = round_up(value_size, 8);
1704 value_size = size * num_possible_cpus();
1706 /* while experimenting with hash tables with sizes ranging from 10 to
1707 * 1000, it was observed that a bucket can have up to 5 entries.
1712 /* We cannot do copy_from_user or copy_to_user inside
1713 * the rcu_read_lock. Allocate enough space here.
1715 keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1716 values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1717 if (!keys || !values) {
1723 bpf_disable_instrumentation();
1728 b = &htab->buckets[batch];
1730 /* do not grab the lock unless need it (bucket_cnt > 0). */
1732 ret = htab_lock_bucket(htab, b, batch, &flags);
1735 bpf_enable_instrumentation();
1741 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1744 if (bucket_cnt && !locked) {
1749 if (bucket_cnt > (max_count - total)) {
1752 /* Note that since bucket_cnt > 0 here, it is implicit
1753 * that the locked was grabbed, so release it.
1755 htab_unlock_bucket(htab, b, batch, flags);
1757 bpf_enable_instrumentation();
1761 if (bucket_cnt > bucket_size) {
1762 bucket_size = bucket_cnt;
1763 /* Note that since bucket_cnt > 0 here, it is implicit
1764 * that the locked was grabbed, so release it.
1766 htab_unlock_bucket(htab, b, batch, flags);
1768 bpf_enable_instrumentation();
1774 /* Next block is only safe to run if you have grabbed the lock */
1778 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1779 memcpy(dst_key, l->key, key_size);
1783 void __percpu *pptr;
1785 pptr = htab_elem_get_ptr(l, map->key_size);
1786 for_each_possible_cpu(cpu) {
1787 bpf_long_memcpy(dst_val + off,
1788 per_cpu_ptr(pptr, cpu), size);
1792 value = l->key + roundup_key_size;
1793 if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
1794 struct bpf_map **inner_map = value;
1796 /* Actual value is the id of the inner map */
1797 map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
1801 if (elem_map_flags & BPF_F_LOCK)
1802 copy_map_value_locked(map, dst_val, value,
1805 copy_map_value(map, dst_val, value);
1806 check_and_init_map_value(map, dst_val);
1809 hlist_nulls_del_rcu(&l->hash_node);
1811 /* bpf_lru_push_free() will acquire lru_lock, which
1812 * may cause deadlock. See comments in function
1813 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1814 * after releasing the bucket lock.
1817 l->batch_flink = node_to_free;
1820 free_htab_elem(htab, l);
1823 dst_key += key_size;
1824 dst_val += value_size;
1827 htab_unlock_bucket(htab, b, batch, flags);
1830 while (node_to_free) {
1832 node_to_free = node_to_free->batch_flink;
1833 htab_lru_push_free(htab, l);
1837 /* If we are not copying data, we can go to next bucket and avoid
1838 * unlocking the rcu.
1840 if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1846 bpf_enable_instrumentation();
1847 if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1848 key_size * bucket_cnt) ||
1849 copy_to_user(uvalues + total * value_size, values,
1850 value_size * bucket_cnt))) {
1855 total += bucket_cnt;
1857 if (batch >= htab->n_buckets) {
1867 /* copy # of entries and next batch */
1868 ubatch = u64_to_user_ptr(attr->batch.out_batch);
1869 if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1870 put_user(total, &uattr->batch.count))
1880 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1881 union bpf_attr __user *uattr)
1883 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1888 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1889 const union bpf_attr *attr,
1890 union bpf_attr __user *uattr)
1892 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1897 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1898 union bpf_attr __user *uattr)
1900 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1905 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1906 const union bpf_attr *attr,
1907 union bpf_attr __user *uattr)
1909 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1914 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1915 const union bpf_attr *attr,
1916 union bpf_attr __user *uattr)
1918 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1923 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1924 const union bpf_attr *attr,
1925 union bpf_attr __user *uattr)
1927 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1932 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1933 union bpf_attr __user *uattr)
1935 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1940 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1941 const union bpf_attr *attr,
1942 union bpf_attr __user *uattr)
1944 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1948 struct bpf_iter_seq_hash_map_info {
1949 struct bpf_map *map;
1950 struct bpf_htab *htab;
1951 void *percpu_value_buf; // non-zero means percpu hash
1956 static struct htab_elem *
1957 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1958 struct htab_elem *prev_elem)
1960 const struct bpf_htab *htab = info->htab;
1961 u32 skip_elems = info->skip_elems;
1962 u32 bucket_id = info->bucket_id;
1963 struct hlist_nulls_head *head;
1964 struct hlist_nulls_node *n;
1965 struct htab_elem *elem;
1969 if (bucket_id >= htab->n_buckets)
1972 /* try to find next elem in the same bucket */
1974 /* no update/deletion on this bucket, prev_elem should be still valid
1975 * and we won't skip elements.
1977 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1978 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1982 /* not found, unlock and go to the next bucket */
1983 b = &htab->buckets[bucket_id++];
1988 for (i = bucket_id; i < htab->n_buckets; i++) {
1989 b = &htab->buckets[i];
1994 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1995 if (count >= skip_elems) {
1996 info->bucket_id = i;
1997 info->skip_elems = count;
2007 info->bucket_id = i;
2008 info->skip_elems = 0;
2012 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
2014 struct bpf_iter_seq_hash_map_info *info = seq->private;
2015 struct htab_elem *elem;
2017 elem = bpf_hash_map_seq_find_next(info, NULL);
2026 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2028 struct bpf_iter_seq_hash_map_info *info = seq->private;
2032 return bpf_hash_map_seq_find_next(info, v);
2035 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
2037 struct bpf_iter_seq_hash_map_info *info = seq->private;
2038 u32 roundup_key_size, roundup_value_size;
2039 struct bpf_iter__bpf_map_elem ctx = {};
2040 struct bpf_map *map = info->map;
2041 struct bpf_iter_meta meta;
2042 int ret = 0, off = 0, cpu;
2043 struct bpf_prog *prog;
2044 void __percpu *pptr;
2047 prog = bpf_iter_get_info(&meta, elem == NULL);
2050 ctx.map = info->map;
2052 roundup_key_size = round_up(map->key_size, 8);
2053 ctx.key = elem->key;
2054 if (!info->percpu_value_buf) {
2055 ctx.value = elem->key + roundup_key_size;
2057 roundup_value_size = round_up(map->value_size, 8);
2058 pptr = htab_elem_get_ptr(elem, map->key_size);
2059 for_each_possible_cpu(cpu) {
2060 bpf_long_memcpy(info->percpu_value_buf + off,
2061 per_cpu_ptr(pptr, cpu),
2062 roundup_value_size);
2063 off += roundup_value_size;
2065 ctx.value = info->percpu_value_buf;
2068 ret = bpf_iter_run_prog(prog, &ctx);
2074 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
2076 return __bpf_hash_map_seq_show(seq, v);
2079 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2082 (void)__bpf_hash_map_seq_show(seq, NULL);
2087 static int bpf_iter_init_hash_map(void *priv_data,
2088 struct bpf_iter_aux_info *aux)
2090 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2091 struct bpf_map *map = aux->map;
2095 if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2096 map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2097 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2098 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2102 seq_info->percpu_value_buf = value_buf;
2105 bpf_map_inc_with_uref(map);
2106 seq_info->map = map;
2107 seq_info->htab = container_of(map, struct bpf_htab, map);
2111 static void bpf_iter_fini_hash_map(void *priv_data)
2113 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2115 bpf_map_put_with_uref(seq_info->map);
2116 kfree(seq_info->percpu_value_buf);
2119 static const struct seq_operations bpf_hash_map_seq_ops = {
2120 .start = bpf_hash_map_seq_start,
2121 .next = bpf_hash_map_seq_next,
2122 .stop = bpf_hash_map_seq_stop,
2123 .show = bpf_hash_map_seq_show,
2126 static const struct bpf_iter_seq_info iter_seq_info = {
2127 .seq_ops = &bpf_hash_map_seq_ops,
2128 .init_seq_private = bpf_iter_init_hash_map,
2129 .fini_seq_private = bpf_iter_fini_hash_map,
2130 .seq_priv_size = sizeof(struct bpf_iter_seq_hash_map_info),
2133 static int bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2134 void *callback_ctx, u64 flags)
2136 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2137 struct hlist_nulls_head *head;
2138 struct hlist_nulls_node *n;
2139 struct htab_elem *elem;
2140 u32 roundup_key_size;
2141 int i, num_elems = 0;
2142 void __percpu *pptr;
2151 is_percpu = htab_is_percpu(htab);
2153 roundup_key_size = round_up(map->key_size, 8);
2154 /* disable migration so percpu value prepared here will be the
2155 * same as the one seen by the bpf program with bpf_map_lookup_elem().
2159 for (i = 0; i < htab->n_buckets; i++) {
2160 b = &htab->buckets[i];
2163 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2166 /* current cpu value for percpu map */
2167 pptr = htab_elem_get_ptr(elem, map->key_size);
2168 val = this_cpu_ptr(pptr);
2170 val = elem->key + roundup_key_size;
2173 ret = callback_fn((u64)(long)map, (u64)(long)key,
2174 (u64)(long)val, (u64)(long)callback_ctx, 0);
2175 /* return value: 0 - continue, 1 - stop and return */
2189 BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
2190 const struct bpf_map_ops htab_map_ops = {
2191 .map_meta_equal = bpf_map_meta_equal,
2192 .map_alloc_check = htab_map_alloc_check,
2193 .map_alloc = htab_map_alloc,
2194 .map_free = htab_map_free,
2195 .map_get_next_key = htab_map_get_next_key,
2196 .map_release_uref = htab_map_free_timers,
2197 .map_lookup_elem = htab_map_lookup_elem,
2198 .map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2199 .map_update_elem = htab_map_update_elem,
2200 .map_delete_elem = htab_map_delete_elem,
2201 .map_gen_lookup = htab_map_gen_lookup,
2202 .map_seq_show_elem = htab_map_seq_show_elem,
2203 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2204 .map_for_each_callback = bpf_for_each_hash_elem,
2206 .map_btf_id = &htab_map_btf_ids[0],
2207 .iter_seq_info = &iter_seq_info,
2210 const struct bpf_map_ops htab_lru_map_ops = {
2211 .map_meta_equal = bpf_map_meta_equal,
2212 .map_alloc_check = htab_map_alloc_check,
2213 .map_alloc = htab_map_alloc,
2214 .map_free = htab_map_free,
2215 .map_get_next_key = htab_map_get_next_key,
2216 .map_release_uref = htab_map_free_timers,
2217 .map_lookup_elem = htab_lru_map_lookup_elem,
2218 .map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2219 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2220 .map_update_elem = htab_lru_map_update_elem,
2221 .map_delete_elem = htab_lru_map_delete_elem,
2222 .map_gen_lookup = htab_lru_map_gen_lookup,
2223 .map_seq_show_elem = htab_map_seq_show_elem,
2224 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2225 .map_for_each_callback = bpf_for_each_hash_elem,
2226 BATCH_OPS(htab_lru),
2227 .map_btf_id = &htab_map_btf_ids[0],
2228 .iter_seq_info = &iter_seq_info,
2231 /* Called from eBPF program */
2232 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2234 struct htab_elem *l = __htab_map_lookup_elem(map, key);
2237 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2242 static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2244 struct htab_elem *l;
2246 if (cpu >= nr_cpu_ids)
2249 l = __htab_map_lookup_elem(map, key);
2251 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2256 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2258 struct htab_elem *l = __htab_map_lookup_elem(map, key);
2261 bpf_lru_node_set_ref(&l->lru_node);
2262 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2268 static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2270 struct htab_elem *l;
2272 if (cpu >= nr_cpu_ids)
2275 l = __htab_map_lookup_elem(map, key);
2277 bpf_lru_node_set_ref(&l->lru_node);
2278 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2284 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2286 struct htab_elem *l;
2287 void __percpu *pptr;
2292 /* per_cpu areas are zero-filled and bpf programs can only
2293 * access 'value_size' of them, so copying rounded areas
2294 * will not leak any kernel data
2296 size = round_up(map->value_size, 8);
2298 l = __htab_map_lookup_elem(map, key);
2301 /* We do not mark LRU map element here in order to not mess up
2302 * eviction heuristics when user space does a map walk.
2304 pptr = htab_elem_get_ptr(l, map->key_size);
2305 for_each_possible_cpu(cpu) {
2306 bpf_long_memcpy(value + off,
2307 per_cpu_ptr(pptr, cpu), size);
2316 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2319 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2323 if (htab_is_lru(htab))
2324 ret = __htab_lru_percpu_map_update_elem(map, key, value,
2327 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2334 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2337 struct htab_elem *l;
2338 void __percpu *pptr;
2343 l = __htab_map_lookup_elem(map, key);
2349 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2350 seq_puts(m, ": {\n");
2351 pptr = htab_elem_get_ptr(l, map->key_size);
2352 for_each_possible_cpu(cpu) {
2353 seq_printf(m, "\tcpu%d: ", cpu);
2354 btf_type_seq_show(map->btf, map->btf_value_type_id,
2355 per_cpu_ptr(pptr, cpu), m);
2363 const struct bpf_map_ops htab_percpu_map_ops = {
2364 .map_meta_equal = bpf_map_meta_equal,
2365 .map_alloc_check = htab_map_alloc_check,
2366 .map_alloc = htab_map_alloc,
2367 .map_free = htab_map_free,
2368 .map_get_next_key = htab_map_get_next_key,
2369 .map_lookup_elem = htab_percpu_map_lookup_elem,
2370 .map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2371 .map_update_elem = htab_percpu_map_update_elem,
2372 .map_delete_elem = htab_map_delete_elem,
2373 .map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
2374 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2375 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2376 .map_for_each_callback = bpf_for_each_hash_elem,
2377 BATCH_OPS(htab_percpu),
2378 .map_btf_id = &htab_map_btf_ids[0],
2379 .iter_seq_info = &iter_seq_info,
2382 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2383 .map_meta_equal = bpf_map_meta_equal,
2384 .map_alloc_check = htab_map_alloc_check,
2385 .map_alloc = htab_map_alloc,
2386 .map_free = htab_map_free,
2387 .map_get_next_key = htab_map_get_next_key,
2388 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2389 .map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2390 .map_update_elem = htab_lru_percpu_map_update_elem,
2391 .map_delete_elem = htab_lru_map_delete_elem,
2392 .map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
2393 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2394 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2395 .map_for_each_callback = bpf_for_each_hash_elem,
2396 BATCH_OPS(htab_lru_percpu),
2397 .map_btf_id = &htab_map_btf_ids[0],
2398 .iter_seq_info = &iter_seq_info,
2401 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2403 if (attr->value_size != sizeof(u32))
2405 return htab_map_alloc_check(attr);
2408 static void fd_htab_map_free(struct bpf_map *map)
2410 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2411 struct hlist_nulls_node *n;
2412 struct hlist_nulls_head *head;
2413 struct htab_elem *l;
2416 for (i = 0; i < htab->n_buckets; i++) {
2417 head = select_bucket(htab, i);
2419 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2420 void *ptr = fd_htab_map_get_ptr(map, l);
2422 map->ops->map_fd_put_ptr(ptr);
2429 /* only called from syscall */
2430 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2435 if (!map->ops->map_fd_sys_lookup_elem)
2439 ptr = htab_map_lookup_elem(map, key);
2441 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2449 /* only called from syscall */
2450 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2451 void *key, void *value, u64 map_flags)
2455 u32 ufd = *(u32 *)value;
2457 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2459 return PTR_ERR(ptr);
2461 ret = htab_map_update_elem(map, key, &ptr, map_flags);
2463 map->ops->map_fd_put_ptr(ptr);
2468 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2470 struct bpf_map *map, *inner_map_meta;
2472 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2473 if (IS_ERR(inner_map_meta))
2474 return inner_map_meta;
2476 map = htab_map_alloc(attr);
2478 bpf_map_meta_free(inner_map_meta);
2482 map->inner_map_meta = inner_map_meta;
2487 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2489 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
2494 return READ_ONCE(*inner_map);
2497 static int htab_of_map_gen_lookup(struct bpf_map *map,
2498 struct bpf_insn *insn_buf)
2500 struct bpf_insn *insn = insn_buf;
2501 const int ret = BPF_REG_0;
2503 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2504 (void *(*)(struct bpf_map *map, void *key))NULL));
2505 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2506 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2507 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2508 offsetof(struct htab_elem, key) +
2509 round_up(map->key_size, 8));
2510 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2512 return insn - insn_buf;
2515 static void htab_of_map_free(struct bpf_map *map)
2517 bpf_map_meta_free(map->inner_map_meta);
2518 fd_htab_map_free(map);
2521 const struct bpf_map_ops htab_of_maps_map_ops = {
2522 .map_alloc_check = fd_htab_map_alloc_check,
2523 .map_alloc = htab_of_map_alloc,
2524 .map_free = htab_of_map_free,
2525 .map_get_next_key = htab_map_get_next_key,
2526 .map_lookup_elem = htab_of_map_lookup_elem,
2527 .map_delete_elem = htab_map_delete_elem,
2528 .map_fd_get_ptr = bpf_map_fd_get_ptr,
2529 .map_fd_put_ptr = bpf_map_fd_put_ptr,
2530 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2531 .map_gen_lookup = htab_of_map_gen_lookup,
2532 .map_check_btf = map_check_no_btf,
2534 .map_btf_id = &htab_map_btf_ids[0],