1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2019 Facebook */
3 #include <linux/rculist.h>
4 #include <linux/list.h>
5 #include <linux/hash.h>
6 #include <linux/types.h>
7 #include <linux/spinlock.h>
9 #include <linux/btf_ids.h>
10 #include <linux/bpf_local_storage.h>
12 #include <uapi/linux/sock_diag.h>
13 #include <uapi/linux/btf.h>
14 #include <linux/rcupdate.h>
15 #include <linux/rcupdate_trace.h>
16 #include <linux/rcupdate_wait.h>
18 #define BPF_LOCAL_STORAGE_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_CLONE)
20 static struct bpf_local_storage_map_bucket *
21 select_bucket(struct bpf_local_storage_map *smap,
22 struct bpf_local_storage_elem *selem)
24 return &smap->buckets[hash_ptr(selem, smap->bucket_log)];
27 static int mem_charge(struct bpf_local_storage_map *smap, void *owner, u32 size)
29 struct bpf_map *map = &smap->map;
31 if (!map->ops->map_local_storage_charge)
34 return map->ops->map_local_storage_charge(smap, owner, size);
37 static void mem_uncharge(struct bpf_local_storage_map *smap, void *owner,
40 struct bpf_map *map = &smap->map;
42 if (map->ops->map_local_storage_uncharge)
43 map->ops->map_local_storage_uncharge(smap, owner, size);
46 static struct bpf_local_storage __rcu **
47 owner_storage(struct bpf_local_storage_map *smap, void *owner)
49 struct bpf_map *map = &smap->map;
51 return map->ops->map_owner_storage_ptr(owner);
54 static bool selem_linked_to_storage_lockless(const struct bpf_local_storage_elem *selem)
56 return !hlist_unhashed_lockless(&selem->snode);
59 static bool selem_linked_to_storage(const struct bpf_local_storage_elem *selem)
61 return !hlist_unhashed(&selem->snode);
64 static bool selem_linked_to_map_lockless(const struct bpf_local_storage_elem *selem)
66 return !hlist_unhashed_lockless(&selem->map_node);
69 static bool selem_linked_to_map(const struct bpf_local_storage_elem *selem)
71 return !hlist_unhashed(&selem->map_node);
74 struct bpf_local_storage_elem *
75 bpf_selem_alloc(struct bpf_local_storage_map *smap, void *owner,
76 void *value, bool charge_mem, gfp_t gfp_flags)
78 struct bpf_local_storage_elem *selem;
80 if (charge_mem && mem_charge(smap, owner, smap->elem_size))
85 selem = bpf_mem_cache_alloc_flags(&smap->selem_ma, gfp_flags);
88 /* Keep the original bpf_map_kzalloc behavior
89 * before started using the bpf_mem_cache_alloc.
91 * No need to use zero_map_value. The bpf_selem_free()
92 * only does bpf_mem_cache_free when there is
93 * no other bpf prog is using the selem.
95 memset(SDATA(selem)->data, 0, smap->map.value_size);
97 selem = bpf_map_kzalloc(&smap->map, smap->elem_size,
98 gfp_flags | __GFP_NOWARN);
103 copy_map_value(&smap->map, SDATA(selem)->data, value);
104 /* No need to call check_and_init_map_value as memory is zero init */
109 mem_uncharge(smap, owner, smap->elem_size);
114 /* rcu tasks trace callback for bpf_ma == false */
115 static void __bpf_local_storage_free_trace_rcu(struct rcu_head *rcu)
117 struct bpf_local_storage *local_storage;
119 /* If RCU Tasks Trace grace period implies RCU grace period, do
120 * kfree(), else do kfree_rcu().
122 local_storage = container_of(rcu, struct bpf_local_storage, rcu);
123 if (rcu_trace_implies_rcu_gp())
124 kfree(local_storage);
126 kfree_rcu(local_storage, rcu);
129 static void bpf_local_storage_free_rcu(struct rcu_head *rcu)
131 struct bpf_local_storage *local_storage;
133 local_storage = container_of(rcu, struct bpf_local_storage, rcu);
134 bpf_mem_cache_raw_free(local_storage);
137 static void bpf_local_storage_free_trace_rcu(struct rcu_head *rcu)
139 if (rcu_trace_implies_rcu_gp())
140 bpf_local_storage_free_rcu(rcu);
142 call_rcu(rcu, bpf_local_storage_free_rcu);
145 /* Handle bpf_ma == false */
146 static void __bpf_local_storage_free(struct bpf_local_storage *local_storage,
150 kfree_rcu(local_storage, rcu);
152 call_rcu_tasks_trace(&local_storage->rcu,
153 __bpf_local_storage_free_trace_rcu);
156 static void bpf_local_storage_free(struct bpf_local_storage *local_storage,
157 struct bpf_local_storage_map *smap,
158 bool bpf_ma, bool reuse_now)
161 __bpf_local_storage_free(local_storage, reuse_now);
166 call_rcu_tasks_trace(&local_storage->rcu,
167 bpf_local_storage_free_trace_rcu);
173 bpf_mem_cache_free(&smap->storage_ma, local_storage);
176 /* smap could be NULL if the selem that triggered
177 * this 'local_storage' creation had been long gone.
178 * In this case, directly do call_rcu().
180 call_rcu(&local_storage->rcu, bpf_local_storage_free_rcu);
184 /* rcu tasks trace callback for bpf_ma == false */
185 static void __bpf_selem_free_trace_rcu(struct rcu_head *rcu)
187 struct bpf_local_storage_elem *selem;
189 selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
190 if (rcu_trace_implies_rcu_gp())
193 kfree_rcu(selem, rcu);
196 /* Handle bpf_ma == false */
197 static void __bpf_selem_free(struct bpf_local_storage_elem *selem,
201 kfree_rcu(selem, rcu);
203 call_rcu_tasks_trace(&selem->rcu, __bpf_selem_free_trace_rcu);
206 static void bpf_selem_free_rcu(struct rcu_head *rcu)
208 struct bpf_local_storage_elem *selem;
210 selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
211 bpf_mem_cache_raw_free(selem);
214 static void bpf_selem_free_trace_rcu(struct rcu_head *rcu)
216 if (rcu_trace_implies_rcu_gp())
217 bpf_selem_free_rcu(rcu);
219 call_rcu(rcu, bpf_selem_free_rcu);
222 void bpf_selem_free(struct bpf_local_storage_elem *selem,
223 struct bpf_local_storage_map *smap,
226 bpf_obj_free_fields(smap->map.record, SDATA(selem)->data);
229 __bpf_selem_free(selem, reuse_now);
234 call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_trace_rcu);
236 /* Instead of using the vanilla call_rcu(),
237 * bpf_mem_cache_free will be able to reuse selem
241 bpf_mem_cache_free(&smap->selem_ma, selem);
246 /* local_storage->lock must be held and selem->local_storage == local_storage.
247 * The caller must ensure selem->smap is still valid to be
248 * dereferenced for its smap->elem_size and smap->cache_idx.
250 static bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
251 struct bpf_local_storage_elem *selem,
252 bool uncharge_mem, bool reuse_now)
254 struct bpf_local_storage_map *smap;
255 bool free_local_storage;
258 smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
259 owner = local_storage->owner;
261 /* All uncharging on the owner must be done first.
262 * The owner may be freed once the last selem is unlinked
263 * from local_storage.
266 mem_uncharge(smap, owner, smap->elem_size);
268 free_local_storage = hlist_is_singular_node(&selem->snode,
269 &local_storage->list);
270 if (free_local_storage) {
271 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
272 local_storage->owner = NULL;
274 /* After this RCU_INIT, owner may be freed and cannot be used */
275 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
277 /* local_storage is not freed now. local_storage->lock is
278 * still held and raw_spin_unlock_bh(&local_storage->lock)
279 * will be done by the caller.
281 * Although the unlock will be done under
282 * rcu_read_lock(), it is more intuitive to
283 * read if the freeing of the storage is done
284 * after the raw_spin_unlock_bh(&local_storage->lock).
286 * Hence, a "bool free_local_storage" is returned
287 * to the caller which then calls then frees the storage after
288 * all the RCU grace periods have expired.
291 hlist_del_init_rcu(&selem->snode);
292 if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
294 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
296 bpf_selem_free(selem, smap, reuse_now);
298 if (rcu_access_pointer(local_storage->smap) == smap)
299 RCU_INIT_POINTER(local_storage->smap, NULL);
301 return free_local_storage;
304 static bool check_storage_bpf_ma(struct bpf_local_storage *local_storage,
305 struct bpf_local_storage_map *storage_smap,
306 struct bpf_local_storage_elem *selem)
309 struct bpf_local_storage_map *selem_smap;
311 /* local_storage->smap may be NULL. If it is, get the bpf_ma
312 * from any selem in the local_storage->list. The bpf_ma of all
313 * local_storage and selem should have the same value
314 * for the same map type.
316 * If the local_storage->list is already empty, the caller will not
317 * care about the bpf_ma value also because the caller is not
318 * responsibile to free the local_storage.
322 return storage_smap->bpf_ma;
325 struct hlist_node *n;
327 n = rcu_dereference_check(hlist_first_rcu(&local_storage->list),
328 bpf_rcu_lock_held());
332 selem = hlist_entry(n, struct bpf_local_storage_elem, snode);
334 selem_smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
336 return selem_smap->bpf_ma;
339 static void bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem,
342 struct bpf_local_storage_map *storage_smap;
343 struct bpf_local_storage *local_storage;
344 bool bpf_ma, free_local_storage = false;
347 if (unlikely(!selem_linked_to_storage_lockless(selem)))
348 /* selem has already been unlinked from sk */
351 local_storage = rcu_dereference_check(selem->local_storage,
352 bpf_rcu_lock_held());
353 storage_smap = rcu_dereference_check(local_storage->smap,
354 bpf_rcu_lock_held());
355 bpf_ma = check_storage_bpf_ma(local_storage, storage_smap, selem);
357 raw_spin_lock_irqsave(&local_storage->lock, flags);
358 if (likely(selem_linked_to_storage(selem)))
359 free_local_storage = bpf_selem_unlink_storage_nolock(
360 local_storage, selem, true, reuse_now);
361 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
363 if (free_local_storage)
364 bpf_local_storage_free(local_storage, storage_smap, bpf_ma, reuse_now);
367 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
368 struct bpf_local_storage_elem *selem)
370 RCU_INIT_POINTER(selem->local_storage, local_storage);
371 hlist_add_head_rcu(&selem->snode, &local_storage->list);
374 static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
376 struct bpf_local_storage_map *smap;
377 struct bpf_local_storage_map_bucket *b;
380 if (unlikely(!selem_linked_to_map_lockless(selem)))
381 /* selem has already be unlinked from smap */
384 smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
385 b = select_bucket(smap, selem);
386 raw_spin_lock_irqsave(&b->lock, flags);
387 if (likely(selem_linked_to_map(selem)))
388 hlist_del_init_rcu(&selem->map_node);
389 raw_spin_unlock_irqrestore(&b->lock, flags);
392 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
393 struct bpf_local_storage_elem *selem)
395 struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
398 raw_spin_lock_irqsave(&b->lock, flags);
399 RCU_INIT_POINTER(SDATA(selem)->smap, smap);
400 hlist_add_head_rcu(&selem->map_node, &b->list);
401 raw_spin_unlock_irqrestore(&b->lock, flags);
404 void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool reuse_now)
406 /* Always unlink from map before unlinking from local_storage
407 * because selem will be freed after successfully unlinked from
410 bpf_selem_unlink_map(selem);
411 bpf_selem_unlink_storage(selem, reuse_now);
414 /* If cacheit_lockit is false, this lookup function is lockless */
415 struct bpf_local_storage_data *
416 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
417 struct bpf_local_storage_map *smap,
420 struct bpf_local_storage_data *sdata;
421 struct bpf_local_storage_elem *selem;
423 /* Fast path (cache hit) */
424 sdata = rcu_dereference_check(local_storage->cache[smap->cache_idx],
425 bpf_rcu_lock_held());
426 if (sdata && rcu_access_pointer(sdata->smap) == smap)
429 /* Slow path (cache miss) */
430 hlist_for_each_entry_rcu(selem, &local_storage->list, snode,
431 rcu_read_lock_trace_held())
432 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
438 sdata = SDATA(selem);
439 if (cacheit_lockit) {
442 /* spinlock is needed to avoid racing with the
443 * parallel delete. Otherwise, publishing an already
444 * deleted sdata to the cache will become a use-after-free
445 * problem in the next bpf_local_storage_lookup().
447 raw_spin_lock_irqsave(&local_storage->lock, flags);
448 if (selem_linked_to_storage(selem))
449 rcu_assign_pointer(local_storage->cache[smap->cache_idx],
451 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
457 static int check_flags(const struct bpf_local_storage_data *old_sdata,
460 if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
461 /* elem already exists */
464 if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
465 /* elem doesn't exist, cannot update it */
471 int bpf_local_storage_alloc(void *owner,
472 struct bpf_local_storage_map *smap,
473 struct bpf_local_storage_elem *first_selem,
476 struct bpf_local_storage *prev_storage, *storage;
477 struct bpf_local_storage **owner_storage_ptr;
480 err = mem_charge(smap, owner, sizeof(*storage));
486 storage = bpf_mem_cache_alloc_flags(&smap->storage_ma, gfp_flags);
489 storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
490 gfp_flags | __GFP_NOWARN);
498 RCU_INIT_POINTER(storage->smap, smap);
499 INIT_HLIST_HEAD(&storage->list);
500 raw_spin_lock_init(&storage->lock);
501 storage->owner = owner;
503 bpf_selem_link_storage_nolock(storage, first_selem);
504 bpf_selem_link_map(smap, first_selem);
507 (struct bpf_local_storage **)owner_storage(smap, owner);
508 /* Publish storage to the owner.
509 * Instead of using any lock of the kernel object (i.e. owner),
510 * cmpxchg will work with any kernel object regardless what
511 * the running context is, bh, irq...etc.
513 * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
514 * is protected by the storage->lock. Hence, when freeing
515 * the owner->storage, the storage->lock must be held before
516 * setting owner->storage ptr to NULL.
518 prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
519 if (unlikely(prev_storage)) {
520 bpf_selem_unlink_map(first_selem);
524 /* Note that even first_selem was linked to smap's
525 * bucket->list, first_selem can be freed immediately
526 * (instead of kfree_rcu) because
527 * bpf_local_storage_map_free() does a
528 * synchronize_rcu_mult (waiting for both sleepable and
529 * normal programs) before walking the bucket->list.
530 * Hence, no one is accessing selem from the
531 * bucket->list under rcu_read_lock().
538 bpf_local_storage_free(storage, smap, smap->bpf_ma, true);
539 mem_uncharge(smap, owner, sizeof(*storage));
543 /* sk cannot be going away because it is linking new elem
544 * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
545 * Otherwise, it will become a leak (and other memory issues
546 * during map destruction).
548 struct bpf_local_storage_data *
549 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
550 void *value, u64 map_flags, gfp_t gfp_flags)
552 struct bpf_local_storage_data *old_sdata = NULL;
553 struct bpf_local_storage_elem *selem = NULL;
554 struct bpf_local_storage *local_storage;
558 /* BPF_EXIST and BPF_NOEXIST cannot be both set */
559 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
560 /* BPF_F_LOCK can only be used in a value with spin_lock */
561 unlikely((map_flags & BPF_F_LOCK) &&
562 !btf_record_has_field(smap->map.record, BPF_SPIN_LOCK)))
563 return ERR_PTR(-EINVAL);
565 if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST)
566 return ERR_PTR(-EINVAL);
568 local_storage = rcu_dereference_check(*owner_storage(smap, owner),
569 bpf_rcu_lock_held());
570 if (!local_storage || hlist_empty(&local_storage->list)) {
571 /* Very first elem for the owner */
572 err = check_flags(NULL, map_flags);
576 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
578 return ERR_PTR(-ENOMEM);
580 err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags);
582 bpf_selem_free(selem, smap, true);
583 mem_uncharge(smap, owner, smap->elem_size);
590 if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
591 /* Hoping to find an old_sdata to do inline update
592 * such that it can avoid taking the local_storage->lock
593 * and changing the lists.
596 bpf_local_storage_lookup(local_storage, smap, false);
597 err = check_flags(old_sdata, map_flags);
600 if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
601 copy_map_value_locked(&smap->map, old_sdata->data,
607 if (gfp_flags == GFP_KERNEL) {
608 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
610 return ERR_PTR(-ENOMEM);
613 raw_spin_lock_irqsave(&local_storage->lock, flags);
615 /* Recheck local_storage->list under local_storage->lock */
616 if (unlikely(hlist_empty(&local_storage->list))) {
617 /* A parallel del is happening and local_storage is going
618 * away. It has just been checked before, so very
619 * unlikely. Return instead of retry to keep things
626 old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
627 err = check_flags(old_sdata, map_flags);
631 if (old_sdata && (map_flags & BPF_F_LOCK)) {
632 copy_map_value_locked(&smap->map, old_sdata->data, value,
634 selem = SELEM(old_sdata);
638 if (gfp_flags != GFP_KERNEL) {
639 /* local_storage->lock is held. Hence, we are sure
640 * we can unlink and uncharge the old_sdata successfully
641 * later. Hence, instead of charging the new selem now
642 * and then uncharge the old selem later (which may cause
643 * a potential but unnecessary charge failure), avoid taking
644 * a charge at all here (the "!old_sdata" check) and the
645 * old_sdata will not be uncharged later during
646 * bpf_selem_unlink_storage_nolock().
648 selem = bpf_selem_alloc(smap, owner, value, !old_sdata, gfp_flags);
655 /* First, link the new selem to the map */
656 bpf_selem_link_map(smap, selem);
658 /* Second, link (and publish) the new selem to local_storage */
659 bpf_selem_link_storage_nolock(local_storage, selem);
661 /* Third, remove old selem, SELEM(old_sdata) */
663 bpf_selem_unlink_map(SELEM(old_sdata));
664 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
669 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
673 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
675 mem_uncharge(smap, owner, smap->elem_size);
676 bpf_selem_free(selem, smap, true);
681 static u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
683 u64 min_usage = U64_MAX;
686 spin_lock(&cache->idx_lock);
688 for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
689 if (cache->idx_usage_counts[i] < min_usage) {
690 min_usage = cache->idx_usage_counts[i];
693 /* Found a free cache_idx */
698 cache->idx_usage_counts[res]++;
700 spin_unlock(&cache->idx_lock);
705 static void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
708 spin_lock(&cache->idx_lock);
709 cache->idx_usage_counts[idx]--;
710 spin_unlock(&cache->idx_lock);
713 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
715 if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
716 !(attr->map_flags & BPF_F_NO_PREALLOC) ||
718 attr->key_size != sizeof(int) || !attr->value_size ||
719 /* Enforce BTF for userspace sk dumping */
720 !attr->btf_key_type_id || !attr->btf_value_type_id)
726 if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
732 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
733 const struct btf *btf,
734 const struct btf_type *key_type,
735 const struct btf_type *value_type)
739 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
742 int_data = *(u32 *)(key_type + 1);
743 if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
749 void bpf_local_storage_destroy(struct bpf_local_storage *local_storage)
751 struct bpf_local_storage_map *storage_smap;
752 struct bpf_local_storage_elem *selem;
753 bool bpf_ma, free_storage = false;
754 struct hlist_node *n;
757 storage_smap = rcu_dereference_check(local_storage->smap, bpf_rcu_lock_held());
758 bpf_ma = check_storage_bpf_ma(local_storage, storage_smap, NULL);
760 /* Neither the bpf_prog nor the bpf_map's syscall
761 * could be modifying the local_storage->list now.
762 * Thus, no elem can be added to or deleted from the
763 * local_storage->list by the bpf_prog or by the bpf_map's syscall.
765 * It is racing with bpf_local_storage_map_free() alone
766 * when unlinking elem from the local_storage->list and
767 * the map's bucket->list.
769 raw_spin_lock_irqsave(&local_storage->lock, flags);
770 hlist_for_each_entry_safe(selem, n, &local_storage->list, snode) {
771 /* Always unlink from map before unlinking from
774 bpf_selem_unlink_map(selem);
775 /* If local_storage list has only one element, the
776 * bpf_selem_unlink_storage_nolock() will return true.
777 * Otherwise, it will return false. The current loop iteration
778 * intends to remove all local storage. So the last iteration
779 * of the loop will set the free_cgroup_storage to true.
781 free_storage = bpf_selem_unlink_storage_nolock(
782 local_storage, selem, false, true);
784 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
787 bpf_local_storage_free(local_storage, storage_smap, bpf_ma, true);
790 u64 bpf_local_storage_map_mem_usage(const struct bpf_map *map)
792 struct bpf_local_storage_map *smap = (struct bpf_local_storage_map *)map;
793 u64 usage = sizeof(*smap);
795 /* The dynamically callocated selems are not counted currently. */
796 usage += sizeof(*smap->buckets) * (1ULL << smap->bucket_log);
800 /* When bpf_ma == true, the bpf_mem_alloc is used to allocate and free memory.
801 * A deadlock free allocator is useful for storage that the bpf prog can easily
802 * get a hold of the owner PTR_TO_BTF_ID in any context. eg. bpf_get_current_task_btf.
803 * The task and cgroup storage fall into this case. The bpf_mem_alloc reuses
804 * memory immediately. To be reuse-immediate safe, the owner destruction
805 * code path needs to go through a rcu grace period before calling
806 * bpf_local_storage_destroy().
808 * When bpf_ma == false, the kmalloc and kfree are used.
811 bpf_local_storage_map_alloc(union bpf_attr *attr,
812 struct bpf_local_storage_cache *cache,
815 struct bpf_local_storage_map *smap;
820 smap = bpf_map_area_alloc(sizeof(*smap), NUMA_NO_NODE);
822 return ERR_PTR(-ENOMEM);
823 bpf_map_init_from_attr(&smap->map, attr);
825 nbuckets = roundup_pow_of_two(num_possible_cpus());
826 /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
827 nbuckets = max_t(u32, 2, nbuckets);
828 smap->bucket_log = ilog2(nbuckets);
830 smap->buckets = bpf_map_kvcalloc(&smap->map, sizeof(*smap->buckets),
831 nbuckets, GFP_USER | __GFP_NOWARN);
832 if (!smap->buckets) {
837 for (i = 0; i < nbuckets; i++) {
838 INIT_HLIST_HEAD(&smap->buckets[i].list);
839 raw_spin_lock_init(&smap->buckets[i].lock);
842 smap->elem_size = offsetof(struct bpf_local_storage_elem,
843 sdata.data[attr->value_size]);
845 smap->bpf_ma = bpf_ma;
847 err = bpf_mem_alloc_init(&smap->selem_ma, smap->elem_size, false);
851 err = bpf_mem_alloc_init(&smap->storage_ma, sizeof(struct bpf_local_storage), false);
853 bpf_mem_alloc_destroy(&smap->selem_ma);
858 smap->cache_idx = bpf_local_storage_cache_idx_get(cache);
862 kvfree(smap->buckets);
863 bpf_map_area_free(smap);
867 void bpf_local_storage_map_free(struct bpf_map *map,
868 struct bpf_local_storage_cache *cache,
869 int __percpu *busy_counter)
871 struct bpf_local_storage_map_bucket *b;
872 struct bpf_local_storage_elem *selem;
873 struct bpf_local_storage_map *smap;
876 smap = (struct bpf_local_storage_map *)map;
877 bpf_local_storage_cache_idx_free(cache, smap->cache_idx);
879 /* Note that this map might be concurrently cloned from
880 * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
881 * RCU read section to finish before proceeding. New RCU
882 * read sections should be prevented via bpf_map_inc_not_zero.
886 /* bpf prog and the userspace can no longer access this map
887 * now. No new selem (of this map) can be added
888 * to the owner->storage or to the map bucket's list.
890 * The elem of this map can be cleaned up here
891 * or when the storage is freed e.g.
892 * by bpf_sk_storage_free() during __sk_destruct().
894 for (i = 0; i < (1U << smap->bucket_log); i++) {
895 b = &smap->buckets[i];
898 /* No one is adding to b->list now */
899 while ((selem = hlist_entry_safe(
900 rcu_dereference_raw(hlist_first_rcu(&b->list)),
901 struct bpf_local_storage_elem, map_node))) {
904 this_cpu_inc(*busy_counter);
906 bpf_selem_unlink(selem, true);
908 this_cpu_dec(*busy_counter);
916 /* While freeing the storage we may still need to access the map.
918 * e.g. when bpf_sk_storage_free() has unlinked selem from the map
919 * which then made the above while((selem = ...)) loop
922 * However, while freeing the storage one still needs to access the
923 * smap->elem_size to do the uncharging in
924 * bpf_selem_unlink_storage_nolock().
926 * Hence, wait another rcu grace period for the storage to be freed.
931 bpf_mem_alloc_destroy(&smap->selem_ma);
932 bpf_mem_alloc_destroy(&smap->storage_ma);
934 kvfree(smap->buckets);
935 bpf_map_area_free(smap);