1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
5 #include <linux/btf_ids.h>
6 #include <linux/filter.h>
7 #include <linux/errno.h>
8 #include <linux/file.h>
10 #include <linux/workqueue.h>
11 #include <linux/skmsg.h>
12 #include <linux/list.h>
13 #include <linux/jhash.h>
14 #include <linux/sock_diag.h>
20 struct sk_psock_progs progs;
24 #define SOCK_CREATE_FLAG_MASK \
25 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
27 static int sock_map_prog_update(struct bpf_map *map, struct bpf_prog *prog,
28 struct bpf_prog *old, u32 which);
29 static struct sk_psock_progs *sock_map_progs(struct bpf_map *map);
31 static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
33 struct bpf_stab *stab;
35 if (attr->max_entries == 0 ||
36 attr->key_size != 4 ||
37 (attr->value_size != sizeof(u32) &&
38 attr->value_size != sizeof(u64)) ||
39 attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
40 return ERR_PTR(-EINVAL);
42 stab = bpf_map_area_alloc(sizeof(*stab), NUMA_NO_NODE);
44 return ERR_PTR(-ENOMEM);
46 bpf_map_init_from_attr(&stab->map, attr);
47 spin_lock_init(&stab->lock);
49 stab->sks = bpf_map_area_alloc((u64) stab->map.max_entries *
50 sizeof(struct sock *),
53 bpf_map_area_free(stab);
54 return ERR_PTR(-ENOMEM);
60 int sock_map_get_from_fd(const union bpf_attr *attr, struct bpf_prog *prog)
62 u32 ufd = attr->target_fd;
67 if (attr->attach_flags || attr->replace_bpf_fd)
71 map = __bpf_map_get(f);
74 ret = sock_map_prog_update(map, prog, NULL, attr->attach_type);
79 int sock_map_prog_detach(const union bpf_attr *attr, enum bpf_prog_type ptype)
81 u32 ufd = attr->target_fd;
82 struct bpf_prog *prog;
87 if (attr->attach_flags || attr->replace_bpf_fd)
91 map = __bpf_map_get(f);
95 prog = bpf_prog_get(attr->attach_bpf_fd);
101 if (prog->type != ptype) {
106 ret = sock_map_prog_update(map, NULL, prog, attr->attach_type);
114 static void sock_map_sk_acquire(struct sock *sk)
115 __acquires(&sk->sk_lock.slock)
121 static void sock_map_sk_release(struct sock *sk)
122 __releases(&sk->sk_lock.slock)
128 static void sock_map_add_link(struct sk_psock *psock,
129 struct sk_psock_link *link,
130 struct bpf_map *map, void *link_raw)
132 link->link_raw = link_raw;
134 spin_lock_bh(&psock->link_lock);
135 list_add_tail(&link->list, &psock->link);
136 spin_unlock_bh(&psock->link_lock);
139 static void sock_map_del_link(struct sock *sk,
140 struct sk_psock *psock, void *link_raw)
142 bool strp_stop = false, verdict_stop = false;
143 struct sk_psock_link *link, *tmp;
145 spin_lock_bh(&psock->link_lock);
146 list_for_each_entry_safe(link, tmp, &psock->link, list) {
147 if (link->link_raw == link_raw) {
148 struct bpf_map *map = link->map;
149 struct sk_psock_progs *progs = sock_map_progs(map);
151 if (psock->saved_data_ready && progs->stream_parser)
153 if (psock->saved_data_ready && progs->stream_verdict)
155 if (psock->saved_data_ready && progs->skb_verdict)
157 list_del(&link->list);
158 sk_psock_free_link(link);
161 spin_unlock_bh(&psock->link_lock);
162 if (strp_stop || verdict_stop) {
163 write_lock_bh(&sk->sk_callback_lock);
165 sk_psock_stop_strp(sk, psock);
167 sk_psock_stop_verdict(sk, psock);
169 if (psock->psock_update_sk_prot)
170 psock->psock_update_sk_prot(sk, psock, false);
171 write_unlock_bh(&sk->sk_callback_lock);
175 static void sock_map_unref(struct sock *sk, void *link_raw)
177 struct sk_psock *psock = sk_psock(sk);
180 sock_map_del_link(sk, psock, link_raw);
181 sk_psock_put(sk, psock);
185 static int sock_map_init_proto(struct sock *sk, struct sk_psock *psock)
187 if (!sk->sk_prot->psock_update_sk_prot)
189 psock->psock_update_sk_prot = sk->sk_prot->psock_update_sk_prot;
190 return sk->sk_prot->psock_update_sk_prot(sk, psock, false);
193 static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
195 struct sk_psock *psock;
198 psock = sk_psock(sk);
200 if (sk->sk_prot->close != sock_map_close) {
201 psock = ERR_PTR(-EBUSY);
205 if (!refcount_inc_not_zero(&psock->refcnt))
206 psock = ERR_PTR(-EBUSY);
213 static int sock_map_link(struct bpf_map *map, struct sock *sk)
215 struct sk_psock_progs *progs = sock_map_progs(map);
216 struct bpf_prog *stream_verdict = NULL;
217 struct bpf_prog *stream_parser = NULL;
218 struct bpf_prog *skb_verdict = NULL;
219 struct bpf_prog *msg_parser = NULL;
220 struct sk_psock *psock;
223 stream_verdict = READ_ONCE(progs->stream_verdict);
224 if (stream_verdict) {
225 stream_verdict = bpf_prog_inc_not_zero(stream_verdict);
226 if (IS_ERR(stream_verdict))
227 return PTR_ERR(stream_verdict);
230 stream_parser = READ_ONCE(progs->stream_parser);
232 stream_parser = bpf_prog_inc_not_zero(stream_parser);
233 if (IS_ERR(stream_parser)) {
234 ret = PTR_ERR(stream_parser);
235 goto out_put_stream_verdict;
239 msg_parser = READ_ONCE(progs->msg_parser);
241 msg_parser = bpf_prog_inc_not_zero(msg_parser);
242 if (IS_ERR(msg_parser)) {
243 ret = PTR_ERR(msg_parser);
244 goto out_put_stream_parser;
248 skb_verdict = READ_ONCE(progs->skb_verdict);
250 skb_verdict = bpf_prog_inc_not_zero(skb_verdict);
251 if (IS_ERR(skb_verdict)) {
252 ret = PTR_ERR(skb_verdict);
253 goto out_put_msg_parser;
257 psock = sock_map_psock_get_checked(sk);
259 ret = PTR_ERR(psock);
264 if ((msg_parser && READ_ONCE(psock->progs.msg_parser)) ||
265 (stream_parser && READ_ONCE(psock->progs.stream_parser)) ||
266 (skb_verdict && READ_ONCE(psock->progs.skb_verdict)) ||
267 (skb_verdict && READ_ONCE(psock->progs.stream_verdict)) ||
268 (stream_verdict && READ_ONCE(psock->progs.skb_verdict)) ||
269 (stream_verdict && READ_ONCE(psock->progs.stream_verdict))) {
270 sk_psock_put(sk, psock);
275 psock = sk_psock_init(sk, map->numa_node);
277 ret = PTR_ERR(psock);
283 psock_set_prog(&psock->progs.msg_parser, msg_parser);
285 psock_set_prog(&psock->progs.stream_parser, stream_parser);
287 psock_set_prog(&psock->progs.stream_verdict, stream_verdict);
289 psock_set_prog(&psock->progs.skb_verdict, skb_verdict);
291 /* msg_* and stream_* programs references tracked in psock after this
292 * point. Reference dec and cleanup will occur through psock destructor
294 ret = sock_map_init_proto(sk, psock);
296 sk_psock_put(sk, psock);
300 write_lock_bh(&sk->sk_callback_lock);
301 if (stream_parser && stream_verdict && !psock->saved_data_ready) {
302 ret = sk_psock_init_strp(sk, psock);
304 write_unlock_bh(&sk->sk_callback_lock);
305 sk_psock_put(sk, psock);
308 sk_psock_start_strp(sk, psock);
309 } else if (!stream_parser && stream_verdict && !psock->saved_data_ready) {
310 sk_psock_start_verdict(sk,psock);
311 } else if (!stream_verdict && skb_verdict && !psock->saved_data_ready) {
312 sk_psock_start_verdict(sk, psock);
314 write_unlock_bh(&sk->sk_callback_lock);
318 bpf_prog_put(skb_verdict);
321 bpf_prog_put(msg_parser);
322 out_put_stream_parser:
324 bpf_prog_put(stream_parser);
325 out_put_stream_verdict:
327 bpf_prog_put(stream_verdict);
332 static void sock_map_free(struct bpf_map *map)
334 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
337 /* After the sync no updates or deletes will be in-flight so it
338 * is safe to walk map and remove entries without risking a race
339 * in EEXIST update case.
342 for (i = 0; i < stab->map.max_entries; i++) {
343 struct sock **psk = &stab->sks[i];
346 sk = xchg(psk, NULL);
351 sock_map_unref(sk, psk);
358 /* wait for psock readers accessing its map link */
361 bpf_map_area_free(stab->sks);
362 bpf_map_area_free(stab);
365 static void sock_map_release_progs(struct bpf_map *map)
367 psock_progs_drop(&container_of(map, struct bpf_stab, map)->progs);
370 static struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key)
372 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
374 WARN_ON_ONCE(!rcu_read_lock_held());
376 if (unlikely(key >= map->max_entries))
378 return READ_ONCE(stab->sks[key]);
381 static void *sock_map_lookup(struct bpf_map *map, void *key)
385 sk = __sock_map_lookup_elem(map, *(u32 *)key);
388 if (sk_is_refcounted(sk) && !refcount_inc_not_zero(&sk->sk_refcnt))
393 static void *sock_map_lookup_sys(struct bpf_map *map, void *key)
397 if (map->value_size != sizeof(u64))
398 return ERR_PTR(-ENOSPC);
400 sk = __sock_map_lookup_elem(map, *(u32 *)key);
402 return ERR_PTR(-ENOENT);
404 __sock_gen_cookie(sk);
405 return &sk->sk_cookie;
408 static int __sock_map_delete(struct bpf_stab *stab, struct sock *sk_test,
415 return -EOPNOTSUPP; /* locks here are hardirq-unsafe */
417 spin_lock_bh(&stab->lock);
419 if (!sk_test || sk_test == sk)
420 sk = xchg(psk, NULL);
423 sock_map_unref(sk, psk);
427 spin_unlock_bh(&stab->lock);
431 static void sock_map_delete_from_link(struct bpf_map *map, struct sock *sk,
434 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
436 __sock_map_delete(stab, sk, link_raw);
439 static long sock_map_delete_elem(struct bpf_map *map, void *key)
441 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
445 if (unlikely(i >= map->max_entries))
449 return __sock_map_delete(stab, NULL, psk);
452 static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next)
454 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
455 u32 i = key ? *(u32 *)key : U32_MAX;
456 u32 *key_next = next;
458 if (i == stab->map.max_entries - 1)
460 if (i >= stab->map.max_entries)
467 static int sock_map_update_common(struct bpf_map *map, u32 idx,
468 struct sock *sk, u64 flags)
470 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
471 struct sk_psock_link *link;
472 struct sk_psock *psock;
476 WARN_ON_ONCE(!rcu_read_lock_held());
477 if (unlikely(flags > BPF_EXIST))
479 if (unlikely(idx >= map->max_entries))
482 link = sk_psock_init_link();
486 ret = sock_map_link(map, sk);
490 psock = sk_psock(sk);
491 WARN_ON_ONCE(!psock);
493 spin_lock_bh(&stab->lock);
494 osk = stab->sks[idx];
495 if (osk && flags == BPF_NOEXIST) {
498 } else if (!osk && flags == BPF_EXIST) {
503 sock_map_add_link(psock, link, map, &stab->sks[idx]);
506 sock_map_unref(osk, &stab->sks[idx]);
507 spin_unlock_bh(&stab->lock);
510 spin_unlock_bh(&stab->lock);
512 sk_psock_put(sk, psock);
514 sk_psock_free_link(link);
518 static bool sock_map_op_okay(const struct bpf_sock_ops_kern *ops)
520 return ops->op == BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB ||
521 ops->op == BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB ||
522 ops->op == BPF_SOCK_OPS_TCP_LISTEN_CB;
525 static bool sock_map_redirect_allowed(const struct sock *sk)
528 return sk->sk_state != TCP_LISTEN;
530 return sk->sk_state == TCP_ESTABLISHED;
533 static bool sock_map_sk_is_suitable(const struct sock *sk)
535 return !!sk->sk_prot->psock_update_sk_prot;
538 static bool sock_map_sk_state_allowed(const struct sock *sk)
541 return (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_LISTEN);
542 if (sk_is_stream_unix(sk))
543 return (1 << sk->sk_state) & TCPF_ESTABLISHED;
547 static int sock_hash_update_common(struct bpf_map *map, void *key,
548 struct sock *sk, u64 flags);
550 int sock_map_update_elem_sys(struct bpf_map *map, void *key, void *value,
558 if (map->value_size == sizeof(u64))
565 sock = sockfd_lookup(ufd, &ret);
573 if (!sock_map_sk_is_suitable(sk)) {
578 sock_map_sk_acquire(sk);
579 if (!sock_map_sk_state_allowed(sk))
581 else if (map->map_type == BPF_MAP_TYPE_SOCKMAP)
582 ret = sock_map_update_common(map, *(u32 *)key, sk, flags);
584 ret = sock_hash_update_common(map, key, sk, flags);
585 sock_map_sk_release(sk);
591 static long sock_map_update_elem(struct bpf_map *map, void *key,
592 void *value, u64 flags)
594 struct sock *sk = (struct sock *)value;
597 if (unlikely(!sk || !sk_fullsock(sk)))
600 if (!sock_map_sk_is_suitable(sk))
605 if (!sock_map_sk_state_allowed(sk))
607 else if (map->map_type == BPF_MAP_TYPE_SOCKMAP)
608 ret = sock_map_update_common(map, *(u32 *)key, sk, flags);
610 ret = sock_hash_update_common(map, key, sk, flags);
616 BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, sops,
617 struct bpf_map *, map, void *, key, u64, flags)
619 WARN_ON_ONCE(!rcu_read_lock_held());
621 if (likely(sock_map_sk_is_suitable(sops->sk) &&
622 sock_map_op_okay(sops)))
623 return sock_map_update_common(map, *(u32 *)key, sops->sk,
628 const struct bpf_func_proto bpf_sock_map_update_proto = {
629 .func = bpf_sock_map_update,
632 .ret_type = RET_INTEGER,
633 .arg1_type = ARG_PTR_TO_CTX,
634 .arg2_type = ARG_CONST_MAP_PTR,
635 .arg3_type = ARG_PTR_TO_MAP_KEY,
636 .arg4_type = ARG_ANYTHING,
639 BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
640 struct bpf_map *, map, u32, key, u64, flags)
644 if (unlikely(flags & ~(BPF_F_INGRESS)))
647 sk = __sock_map_lookup_elem(map, key);
648 if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
651 skb_bpf_set_redir(skb, sk, flags & BPF_F_INGRESS);
655 const struct bpf_func_proto bpf_sk_redirect_map_proto = {
656 .func = bpf_sk_redirect_map,
658 .ret_type = RET_INTEGER,
659 .arg1_type = ARG_PTR_TO_CTX,
660 .arg2_type = ARG_CONST_MAP_PTR,
661 .arg3_type = ARG_ANYTHING,
662 .arg4_type = ARG_ANYTHING,
665 BPF_CALL_4(bpf_msg_redirect_map, struct sk_msg *, msg,
666 struct bpf_map *, map, u32, key, u64, flags)
670 if (unlikely(flags & ~(BPF_F_INGRESS)))
673 sk = __sock_map_lookup_elem(map, key);
674 if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
676 if (!(flags & BPF_F_INGRESS) && !sk_is_tcp(sk))
684 const struct bpf_func_proto bpf_msg_redirect_map_proto = {
685 .func = bpf_msg_redirect_map,
687 .ret_type = RET_INTEGER,
688 .arg1_type = ARG_PTR_TO_CTX,
689 .arg2_type = ARG_CONST_MAP_PTR,
690 .arg3_type = ARG_ANYTHING,
691 .arg4_type = ARG_ANYTHING,
694 struct sock_map_seq_info {
700 struct bpf_iter__sockmap {
701 __bpf_md_ptr(struct bpf_iter_meta *, meta);
702 __bpf_md_ptr(struct bpf_map *, map);
703 __bpf_md_ptr(void *, key);
704 __bpf_md_ptr(struct sock *, sk);
707 DEFINE_BPF_ITER_FUNC(sockmap, struct bpf_iter_meta *meta,
708 struct bpf_map *map, void *key,
711 static void *sock_map_seq_lookup_elem(struct sock_map_seq_info *info)
713 if (unlikely(info->index >= info->map->max_entries))
716 info->sk = __sock_map_lookup_elem(info->map, info->index);
718 /* can't return sk directly, since that might be NULL */
722 static void *sock_map_seq_start(struct seq_file *seq, loff_t *pos)
725 struct sock_map_seq_info *info = seq->private;
730 /* pairs with sock_map_seq_stop */
732 return sock_map_seq_lookup_elem(info);
735 static void *sock_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
738 struct sock_map_seq_info *info = seq->private;
743 return sock_map_seq_lookup_elem(info);
746 static int sock_map_seq_show(struct seq_file *seq, void *v)
749 struct sock_map_seq_info *info = seq->private;
750 struct bpf_iter__sockmap ctx = {};
751 struct bpf_iter_meta meta;
752 struct bpf_prog *prog;
755 prog = bpf_iter_get_info(&meta, !v);
762 ctx.key = &info->index;
766 return bpf_iter_run_prog(prog, &ctx);
769 static void sock_map_seq_stop(struct seq_file *seq, void *v)
773 (void)sock_map_seq_show(seq, NULL);
775 /* pairs with sock_map_seq_start */
779 static const struct seq_operations sock_map_seq_ops = {
780 .start = sock_map_seq_start,
781 .next = sock_map_seq_next,
782 .stop = sock_map_seq_stop,
783 .show = sock_map_seq_show,
786 static int sock_map_init_seq_private(void *priv_data,
787 struct bpf_iter_aux_info *aux)
789 struct sock_map_seq_info *info = priv_data;
791 bpf_map_inc_with_uref(aux->map);
792 info->map = aux->map;
796 static void sock_map_fini_seq_private(void *priv_data)
798 struct sock_map_seq_info *info = priv_data;
800 bpf_map_put_with_uref(info->map);
803 static u64 sock_map_mem_usage(const struct bpf_map *map)
805 u64 usage = sizeof(struct bpf_stab);
807 usage += (u64)map->max_entries * sizeof(struct sock *);
811 static const struct bpf_iter_seq_info sock_map_iter_seq_info = {
812 .seq_ops = &sock_map_seq_ops,
813 .init_seq_private = sock_map_init_seq_private,
814 .fini_seq_private = sock_map_fini_seq_private,
815 .seq_priv_size = sizeof(struct sock_map_seq_info),
818 BTF_ID_LIST_SINGLE(sock_map_btf_ids, struct, bpf_stab)
819 const struct bpf_map_ops sock_map_ops = {
820 .map_meta_equal = bpf_map_meta_equal,
821 .map_alloc = sock_map_alloc,
822 .map_free = sock_map_free,
823 .map_get_next_key = sock_map_get_next_key,
824 .map_lookup_elem_sys_only = sock_map_lookup_sys,
825 .map_update_elem = sock_map_update_elem,
826 .map_delete_elem = sock_map_delete_elem,
827 .map_lookup_elem = sock_map_lookup,
828 .map_release_uref = sock_map_release_progs,
829 .map_check_btf = map_check_no_btf,
830 .map_mem_usage = sock_map_mem_usage,
831 .map_btf_id = &sock_map_btf_ids[0],
832 .iter_seq_info = &sock_map_iter_seq_info,
835 struct bpf_shtab_elem {
839 struct hlist_node node;
843 struct bpf_shtab_bucket {
844 struct hlist_head head;
850 struct bpf_shtab_bucket *buckets;
853 struct sk_psock_progs progs;
857 static inline u32 sock_hash_bucket_hash(const void *key, u32 len)
859 return jhash(key, len, 0);
862 static struct bpf_shtab_bucket *sock_hash_select_bucket(struct bpf_shtab *htab,
865 return &htab->buckets[hash & (htab->buckets_num - 1)];
868 static struct bpf_shtab_elem *
869 sock_hash_lookup_elem_raw(struct hlist_head *head, u32 hash, void *key,
872 struct bpf_shtab_elem *elem;
874 hlist_for_each_entry_rcu(elem, head, node) {
875 if (elem->hash == hash &&
876 !memcmp(&elem->key, key, key_size))
883 static struct sock *__sock_hash_lookup_elem(struct bpf_map *map, void *key)
885 struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
886 u32 key_size = map->key_size, hash;
887 struct bpf_shtab_bucket *bucket;
888 struct bpf_shtab_elem *elem;
890 WARN_ON_ONCE(!rcu_read_lock_held());
892 hash = sock_hash_bucket_hash(key, key_size);
893 bucket = sock_hash_select_bucket(htab, hash);
894 elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
896 return elem ? elem->sk : NULL;
899 static void sock_hash_free_elem(struct bpf_shtab *htab,
900 struct bpf_shtab_elem *elem)
902 atomic_dec(&htab->count);
903 kfree_rcu(elem, rcu);
906 static void sock_hash_delete_from_link(struct bpf_map *map, struct sock *sk,
909 struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
910 struct bpf_shtab_elem *elem_probe, *elem = link_raw;
911 struct bpf_shtab_bucket *bucket;
913 WARN_ON_ONCE(!rcu_read_lock_held());
914 bucket = sock_hash_select_bucket(htab, elem->hash);
916 /* elem may be deleted in parallel from the map, but access here
917 * is okay since it's going away only after RCU grace period.
918 * However, we need to check whether it's still present.
920 spin_lock_bh(&bucket->lock);
921 elem_probe = sock_hash_lookup_elem_raw(&bucket->head, elem->hash,
922 elem->key, map->key_size);
923 if (elem_probe && elem_probe == elem) {
924 hlist_del_rcu(&elem->node);
925 sock_map_unref(elem->sk, elem);
926 sock_hash_free_elem(htab, elem);
928 spin_unlock_bh(&bucket->lock);
931 static long sock_hash_delete_elem(struct bpf_map *map, void *key)
933 struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
934 u32 hash, key_size = map->key_size;
935 struct bpf_shtab_bucket *bucket;
936 struct bpf_shtab_elem *elem;
940 return -EOPNOTSUPP; /* locks here are hardirq-unsafe */
942 hash = sock_hash_bucket_hash(key, key_size);
943 bucket = sock_hash_select_bucket(htab, hash);
945 spin_lock_bh(&bucket->lock);
946 elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
948 hlist_del_rcu(&elem->node);
949 sock_map_unref(elem->sk, elem);
950 sock_hash_free_elem(htab, elem);
953 spin_unlock_bh(&bucket->lock);
957 static struct bpf_shtab_elem *sock_hash_alloc_elem(struct bpf_shtab *htab,
958 void *key, u32 key_size,
959 u32 hash, struct sock *sk,
960 struct bpf_shtab_elem *old)
962 struct bpf_shtab_elem *new;
964 if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
966 atomic_dec(&htab->count);
967 return ERR_PTR(-E2BIG);
971 new = bpf_map_kmalloc_node(&htab->map, htab->elem_size,
972 GFP_ATOMIC | __GFP_NOWARN,
973 htab->map.numa_node);
975 atomic_dec(&htab->count);
976 return ERR_PTR(-ENOMEM);
978 memcpy(new->key, key, key_size);
984 static int sock_hash_update_common(struct bpf_map *map, void *key,
985 struct sock *sk, u64 flags)
987 struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
988 u32 key_size = map->key_size, hash;
989 struct bpf_shtab_elem *elem, *elem_new;
990 struct bpf_shtab_bucket *bucket;
991 struct sk_psock_link *link;
992 struct sk_psock *psock;
995 WARN_ON_ONCE(!rcu_read_lock_held());
996 if (unlikely(flags > BPF_EXIST))
999 link = sk_psock_init_link();
1003 ret = sock_map_link(map, sk);
1007 psock = sk_psock(sk);
1008 WARN_ON_ONCE(!psock);
1010 hash = sock_hash_bucket_hash(key, key_size);
1011 bucket = sock_hash_select_bucket(htab, hash);
1013 spin_lock_bh(&bucket->lock);
1014 elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
1015 if (elem && flags == BPF_NOEXIST) {
1018 } else if (!elem && flags == BPF_EXIST) {
1023 elem_new = sock_hash_alloc_elem(htab, key, key_size, hash, sk, elem);
1024 if (IS_ERR(elem_new)) {
1025 ret = PTR_ERR(elem_new);
1029 sock_map_add_link(psock, link, map, elem_new);
1030 /* Add new element to the head of the list, so that
1031 * concurrent search will find it before old elem.
1033 hlist_add_head_rcu(&elem_new->node, &bucket->head);
1035 hlist_del_rcu(&elem->node);
1036 sock_map_unref(elem->sk, elem);
1037 sock_hash_free_elem(htab, elem);
1039 spin_unlock_bh(&bucket->lock);
1042 spin_unlock_bh(&bucket->lock);
1043 sk_psock_put(sk, psock);
1045 sk_psock_free_link(link);
1049 static int sock_hash_get_next_key(struct bpf_map *map, void *key,
1052 struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
1053 struct bpf_shtab_elem *elem, *elem_next;
1054 u32 hash, key_size = map->key_size;
1055 struct hlist_head *head;
1059 goto find_first_elem;
1060 hash = sock_hash_bucket_hash(key, key_size);
1061 head = &sock_hash_select_bucket(htab, hash)->head;
1062 elem = sock_hash_lookup_elem_raw(head, hash, key, key_size);
1064 goto find_first_elem;
1066 elem_next = hlist_entry_safe(rcu_dereference(hlist_next_rcu(&elem->node)),
1067 struct bpf_shtab_elem, node);
1069 memcpy(key_next, elem_next->key, key_size);
1073 i = hash & (htab->buckets_num - 1);
1076 for (; i < htab->buckets_num; i++) {
1077 head = &sock_hash_select_bucket(htab, i)->head;
1078 elem_next = hlist_entry_safe(rcu_dereference(hlist_first_rcu(head)),
1079 struct bpf_shtab_elem, node);
1081 memcpy(key_next, elem_next->key, key_size);
1089 static struct bpf_map *sock_hash_alloc(union bpf_attr *attr)
1091 struct bpf_shtab *htab;
1094 if (attr->max_entries == 0 ||
1095 attr->key_size == 0 ||
1096 (attr->value_size != sizeof(u32) &&
1097 attr->value_size != sizeof(u64)) ||
1098 attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
1099 return ERR_PTR(-EINVAL);
1100 if (attr->key_size > MAX_BPF_STACK)
1101 return ERR_PTR(-E2BIG);
1103 htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
1105 return ERR_PTR(-ENOMEM);
1107 bpf_map_init_from_attr(&htab->map, attr);
1109 htab->buckets_num = roundup_pow_of_two(htab->map.max_entries);
1110 htab->elem_size = sizeof(struct bpf_shtab_elem) +
1111 round_up(htab->map.key_size, 8);
1112 if (htab->buckets_num == 0 ||
1113 htab->buckets_num > U32_MAX / sizeof(struct bpf_shtab_bucket)) {
1118 htab->buckets = bpf_map_area_alloc(htab->buckets_num *
1119 sizeof(struct bpf_shtab_bucket),
1120 htab->map.numa_node);
1121 if (!htab->buckets) {
1126 for (i = 0; i < htab->buckets_num; i++) {
1127 INIT_HLIST_HEAD(&htab->buckets[i].head);
1128 spin_lock_init(&htab->buckets[i].lock);
1133 bpf_map_area_free(htab);
1134 return ERR_PTR(err);
1137 static void sock_hash_free(struct bpf_map *map)
1139 struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
1140 struct bpf_shtab_bucket *bucket;
1141 struct hlist_head unlink_list;
1142 struct bpf_shtab_elem *elem;
1143 struct hlist_node *node;
1146 /* After the sync no updates or deletes will be in-flight so it
1147 * is safe to walk map and remove entries without risking a race
1148 * in EEXIST update case.
1151 for (i = 0; i < htab->buckets_num; i++) {
1152 bucket = sock_hash_select_bucket(htab, i);
1154 /* We are racing with sock_hash_delete_from_link to
1155 * enter the spin-lock critical section. Every socket on
1156 * the list is still linked to sockhash. Since link
1157 * exists, psock exists and holds a ref to socket. That
1158 * lets us to grab a socket ref too.
1160 spin_lock_bh(&bucket->lock);
1161 hlist_for_each_entry(elem, &bucket->head, node)
1162 sock_hold(elem->sk);
1163 hlist_move_list(&bucket->head, &unlink_list);
1164 spin_unlock_bh(&bucket->lock);
1166 /* Process removed entries out of atomic context to
1167 * block for socket lock before deleting the psock's
1170 hlist_for_each_entry_safe(elem, node, &unlink_list, node) {
1171 hlist_del(&elem->node);
1172 lock_sock(elem->sk);
1174 sock_map_unref(elem->sk, elem);
1176 release_sock(elem->sk);
1178 sock_hash_free_elem(htab, elem);
1182 /* wait for psock readers accessing its map link */
1185 bpf_map_area_free(htab->buckets);
1186 bpf_map_area_free(htab);
1189 static void *sock_hash_lookup_sys(struct bpf_map *map, void *key)
1193 if (map->value_size != sizeof(u64))
1194 return ERR_PTR(-ENOSPC);
1196 sk = __sock_hash_lookup_elem(map, key);
1198 return ERR_PTR(-ENOENT);
1200 __sock_gen_cookie(sk);
1201 return &sk->sk_cookie;
1204 static void *sock_hash_lookup(struct bpf_map *map, void *key)
1208 sk = __sock_hash_lookup_elem(map, key);
1211 if (sk_is_refcounted(sk) && !refcount_inc_not_zero(&sk->sk_refcnt))
1216 static void sock_hash_release_progs(struct bpf_map *map)
1218 psock_progs_drop(&container_of(map, struct bpf_shtab, map)->progs);
1221 BPF_CALL_4(bpf_sock_hash_update, struct bpf_sock_ops_kern *, sops,
1222 struct bpf_map *, map, void *, key, u64, flags)
1224 WARN_ON_ONCE(!rcu_read_lock_held());
1226 if (likely(sock_map_sk_is_suitable(sops->sk) &&
1227 sock_map_op_okay(sops)))
1228 return sock_hash_update_common(map, key, sops->sk, flags);
1232 const struct bpf_func_proto bpf_sock_hash_update_proto = {
1233 .func = bpf_sock_hash_update,
1236 .ret_type = RET_INTEGER,
1237 .arg1_type = ARG_PTR_TO_CTX,
1238 .arg2_type = ARG_CONST_MAP_PTR,
1239 .arg3_type = ARG_PTR_TO_MAP_KEY,
1240 .arg4_type = ARG_ANYTHING,
1243 BPF_CALL_4(bpf_sk_redirect_hash, struct sk_buff *, skb,
1244 struct bpf_map *, map, void *, key, u64, flags)
1248 if (unlikely(flags & ~(BPF_F_INGRESS)))
1251 sk = __sock_hash_lookup_elem(map, key);
1252 if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1255 skb_bpf_set_redir(skb, sk, flags & BPF_F_INGRESS);
1259 const struct bpf_func_proto bpf_sk_redirect_hash_proto = {
1260 .func = bpf_sk_redirect_hash,
1262 .ret_type = RET_INTEGER,
1263 .arg1_type = ARG_PTR_TO_CTX,
1264 .arg2_type = ARG_CONST_MAP_PTR,
1265 .arg3_type = ARG_PTR_TO_MAP_KEY,
1266 .arg4_type = ARG_ANYTHING,
1269 BPF_CALL_4(bpf_msg_redirect_hash, struct sk_msg *, msg,
1270 struct bpf_map *, map, void *, key, u64, flags)
1274 if (unlikely(flags & ~(BPF_F_INGRESS)))
1277 sk = __sock_hash_lookup_elem(map, key);
1278 if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1280 if (!(flags & BPF_F_INGRESS) && !sk_is_tcp(sk))
1288 const struct bpf_func_proto bpf_msg_redirect_hash_proto = {
1289 .func = bpf_msg_redirect_hash,
1291 .ret_type = RET_INTEGER,
1292 .arg1_type = ARG_PTR_TO_CTX,
1293 .arg2_type = ARG_CONST_MAP_PTR,
1294 .arg3_type = ARG_PTR_TO_MAP_KEY,
1295 .arg4_type = ARG_ANYTHING,
1298 struct sock_hash_seq_info {
1299 struct bpf_map *map;
1300 struct bpf_shtab *htab;
1304 static void *sock_hash_seq_find_next(struct sock_hash_seq_info *info,
1305 struct bpf_shtab_elem *prev_elem)
1307 const struct bpf_shtab *htab = info->htab;
1308 struct bpf_shtab_bucket *bucket;
1309 struct bpf_shtab_elem *elem;
1310 struct hlist_node *node;
1312 /* try to find next elem in the same bucket */
1314 node = rcu_dereference(hlist_next_rcu(&prev_elem->node));
1315 elem = hlist_entry_safe(node, struct bpf_shtab_elem, node);
1319 /* no more elements, continue in the next bucket */
1323 for (; info->bucket_id < htab->buckets_num; info->bucket_id++) {
1324 bucket = &htab->buckets[info->bucket_id];
1325 node = rcu_dereference(hlist_first_rcu(&bucket->head));
1326 elem = hlist_entry_safe(node, struct bpf_shtab_elem, node);
1334 static void *sock_hash_seq_start(struct seq_file *seq, loff_t *pos)
1337 struct sock_hash_seq_info *info = seq->private;
1342 /* pairs with sock_hash_seq_stop */
1344 return sock_hash_seq_find_next(info, NULL);
1347 static void *sock_hash_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1350 struct sock_hash_seq_info *info = seq->private;
1353 return sock_hash_seq_find_next(info, v);
1356 static int sock_hash_seq_show(struct seq_file *seq, void *v)
1359 struct sock_hash_seq_info *info = seq->private;
1360 struct bpf_iter__sockmap ctx = {};
1361 struct bpf_shtab_elem *elem = v;
1362 struct bpf_iter_meta meta;
1363 struct bpf_prog *prog;
1366 prog = bpf_iter_get_info(&meta, !elem);
1371 ctx.map = info->map;
1373 ctx.key = elem->key;
1377 return bpf_iter_run_prog(prog, &ctx);
1380 static void sock_hash_seq_stop(struct seq_file *seq, void *v)
1384 (void)sock_hash_seq_show(seq, NULL);
1386 /* pairs with sock_hash_seq_start */
1390 static const struct seq_operations sock_hash_seq_ops = {
1391 .start = sock_hash_seq_start,
1392 .next = sock_hash_seq_next,
1393 .stop = sock_hash_seq_stop,
1394 .show = sock_hash_seq_show,
1397 static int sock_hash_init_seq_private(void *priv_data,
1398 struct bpf_iter_aux_info *aux)
1400 struct sock_hash_seq_info *info = priv_data;
1402 bpf_map_inc_with_uref(aux->map);
1403 info->map = aux->map;
1404 info->htab = container_of(aux->map, struct bpf_shtab, map);
1408 static void sock_hash_fini_seq_private(void *priv_data)
1410 struct sock_hash_seq_info *info = priv_data;
1412 bpf_map_put_with_uref(info->map);
1415 static u64 sock_hash_mem_usage(const struct bpf_map *map)
1417 struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
1418 u64 usage = sizeof(*htab);
1420 usage += htab->buckets_num * sizeof(struct bpf_shtab_bucket);
1421 usage += atomic_read(&htab->count) * (u64)htab->elem_size;
1425 static const struct bpf_iter_seq_info sock_hash_iter_seq_info = {
1426 .seq_ops = &sock_hash_seq_ops,
1427 .init_seq_private = sock_hash_init_seq_private,
1428 .fini_seq_private = sock_hash_fini_seq_private,
1429 .seq_priv_size = sizeof(struct sock_hash_seq_info),
1432 BTF_ID_LIST_SINGLE(sock_hash_map_btf_ids, struct, bpf_shtab)
1433 const struct bpf_map_ops sock_hash_ops = {
1434 .map_meta_equal = bpf_map_meta_equal,
1435 .map_alloc = sock_hash_alloc,
1436 .map_free = sock_hash_free,
1437 .map_get_next_key = sock_hash_get_next_key,
1438 .map_update_elem = sock_map_update_elem,
1439 .map_delete_elem = sock_hash_delete_elem,
1440 .map_lookup_elem = sock_hash_lookup,
1441 .map_lookup_elem_sys_only = sock_hash_lookup_sys,
1442 .map_release_uref = sock_hash_release_progs,
1443 .map_check_btf = map_check_no_btf,
1444 .map_mem_usage = sock_hash_mem_usage,
1445 .map_btf_id = &sock_hash_map_btf_ids[0],
1446 .iter_seq_info = &sock_hash_iter_seq_info,
1449 static struct sk_psock_progs *sock_map_progs(struct bpf_map *map)
1451 switch (map->map_type) {
1452 case BPF_MAP_TYPE_SOCKMAP:
1453 return &container_of(map, struct bpf_stab, map)->progs;
1454 case BPF_MAP_TYPE_SOCKHASH:
1455 return &container_of(map, struct bpf_shtab, map)->progs;
1463 static int sock_map_prog_lookup(struct bpf_map *map, struct bpf_prog ***pprog,
1466 struct sk_psock_progs *progs = sock_map_progs(map);
1472 case BPF_SK_MSG_VERDICT:
1473 *pprog = &progs->msg_parser;
1475 #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
1476 case BPF_SK_SKB_STREAM_PARSER:
1477 *pprog = &progs->stream_parser;
1480 case BPF_SK_SKB_STREAM_VERDICT:
1481 if (progs->skb_verdict)
1483 *pprog = &progs->stream_verdict;
1485 case BPF_SK_SKB_VERDICT:
1486 if (progs->stream_verdict)
1488 *pprog = &progs->skb_verdict;
1497 static int sock_map_prog_update(struct bpf_map *map, struct bpf_prog *prog,
1498 struct bpf_prog *old, u32 which)
1500 struct bpf_prog **pprog;
1503 ret = sock_map_prog_lookup(map, &pprog, which);
1508 return psock_replace_prog(pprog, prog, old);
1510 psock_set_prog(pprog, prog);
1514 int sock_map_bpf_prog_query(const union bpf_attr *attr,
1515 union bpf_attr __user *uattr)
1517 __u32 __user *prog_ids = u64_to_user_ptr(attr->query.prog_ids);
1518 u32 prog_cnt = 0, flags = 0, ufd = attr->target_fd;
1519 struct bpf_prog **pprog;
1520 struct bpf_prog *prog;
1521 struct bpf_map *map;
1526 if (attr->query.query_flags)
1530 map = __bpf_map_get(f);
1532 return PTR_ERR(map);
1536 ret = sock_map_prog_lookup(map, &pprog, attr->query.attach_type);
1541 prog_cnt = !prog ? 0 : 1;
1543 if (!attr->query.prog_cnt || !prog_ids || !prog_cnt)
1546 /* we do not hold the refcnt, the bpf prog may be released
1547 * asynchronously and the id would be set to 0.
1549 id = data_race(prog->aux->id);
1556 if (copy_to_user(&uattr->query.attach_flags, &flags, sizeof(flags)) ||
1557 (id != 0 && copy_to_user(prog_ids, &id, sizeof(u32))) ||
1558 copy_to_user(&uattr->query.prog_cnt, &prog_cnt, sizeof(prog_cnt)))
1565 static void sock_map_unlink(struct sock *sk, struct sk_psock_link *link)
1567 switch (link->map->map_type) {
1568 case BPF_MAP_TYPE_SOCKMAP:
1569 return sock_map_delete_from_link(link->map, sk,
1571 case BPF_MAP_TYPE_SOCKHASH:
1572 return sock_hash_delete_from_link(link->map, sk,
1579 static void sock_map_remove_links(struct sock *sk, struct sk_psock *psock)
1581 struct sk_psock_link *link;
1583 while ((link = sk_psock_link_pop(psock))) {
1584 sock_map_unlink(sk, link);
1585 sk_psock_free_link(link);
1589 void sock_map_unhash(struct sock *sk)
1591 void (*saved_unhash)(struct sock *sk);
1592 struct sk_psock *psock;
1595 psock = sk_psock(sk);
1596 if (unlikely(!psock)) {
1598 saved_unhash = READ_ONCE(sk->sk_prot)->unhash;
1600 saved_unhash = psock->saved_unhash;
1601 sock_map_remove_links(sk, psock);
1604 if (WARN_ON_ONCE(saved_unhash == sock_map_unhash))
1609 EXPORT_SYMBOL_GPL(sock_map_unhash);
1611 void sock_map_destroy(struct sock *sk)
1613 void (*saved_destroy)(struct sock *sk);
1614 struct sk_psock *psock;
1617 psock = sk_psock_get(sk);
1618 if (unlikely(!psock)) {
1620 saved_destroy = READ_ONCE(sk->sk_prot)->destroy;
1622 saved_destroy = psock->saved_destroy;
1623 sock_map_remove_links(sk, psock);
1625 sk_psock_stop(psock);
1626 sk_psock_put(sk, psock);
1628 if (WARN_ON_ONCE(saved_destroy == sock_map_destroy))
1633 EXPORT_SYMBOL_GPL(sock_map_destroy);
1635 void sock_map_close(struct sock *sk, long timeout)
1637 void (*saved_close)(struct sock *sk, long timeout);
1638 struct sk_psock *psock;
1642 psock = sk_psock_get(sk);
1643 if (unlikely(!psock)) {
1646 saved_close = READ_ONCE(sk->sk_prot)->close;
1648 saved_close = psock->saved_close;
1649 sock_map_remove_links(sk, psock);
1651 sk_psock_stop(psock);
1653 cancel_delayed_work_sync(&psock->work);
1654 sk_psock_put(sk, psock);
1657 /* Make sure we do not recurse. This is a bug.
1658 * Leak the socket instead of crashing on a stack overflow.
1660 if (WARN_ON_ONCE(saved_close == sock_map_close))
1662 saved_close(sk, timeout);
1664 EXPORT_SYMBOL_GPL(sock_map_close);
1666 static int sock_map_iter_attach_target(struct bpf_prog *prog,
1667 union bpf_iter_link_info *linfo,
1668 struct bpf_iter_aux_info *aux)
1670 struct bpf_map *map;
1673 if (!linfo->map.map_fd)
1676 map = bpf_map_get_with_uref(linfo->map.map_fd);
1678 return PTR_ERR(map);
1680 if (map->map_type != BPF_MAP_TYPE_SOCKMAP &&
1681 map->map_type != BPF_MAP_TYPE_SOCKHASH)
1684 if (prog->aux->max_rdonly_access > map->key_size) {
1693 bpf_map_put_with_uref(map);
1697 static void sock_map_iter_detach_target(struct bpf_iter_aux_info *aux)
1699 bpf_map_put_with_uref(aux->map);
1702 static struct bpf_iter_reg sock_map_iter_reg = {
1703 .target = "sockmap",
1704 .attach_target = sock_map_iter_attach_target,
1705 .detach_target = sock_map_iter_detach_target,
1706 .show_fdinfo = bpf_iter_map_show_fdinfo,
1707 .fill_link_info = bpf_iter_map_fill_link_info,
1708 .ctx_arg_info_size = 2,
1710 { offsetof(struct bpf_iter__sockmap, key),
1711 PTR_TO_BUF | PTR_MAYBE_NULL | MEM_RDONLY },
1712 { offsetof(struct bpf_iter__sockmap, sk),
1713 PTR_TO_BTF_ID_OR_NULL },
1717 static int __init bpf_sockmap_iter_init(void)
1719 sock_map_iter_reg.ctx_arg_info[1].btf_id =
1720 btf_sock_ids[BTF_SOCK_TYPE_SOCK];
1721 return bpf_iter_reg_target(&sock_map_iter_reg);
1723 late_initcall(bpf_sockmap_iter_init);