xref: /openbmc/linux/net/core/sock_map.c (revision 9c3d6497) 
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
3 
4 #include <linux/bpf.h>
5 #include <linux/filter.h>
6 #include <linux/errno.h>
7 #include <linux/file.h>
8 #include <linux/net.h>
9 #include <linux/workqueue.h>
10 #include <linux/skmsg.h>
11 #include <linux/list.h>
12 #include <linux/jhash.h>
13 #include <linux/sock_diag.h>
14 #include <net/udp.h>
15 
16 struct bpf_stab {
17 	struct bpf_map map;
18 	struct sock **sks;
19 	struct sk_psock_progs progs;
20 	raw_spinlock_t lock;
21 };
22 
23 #define SOCK_CREATE_FLAG_MASK				\
24 	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
25 
26 static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
27 {
28 	struct bpf_stab *stab;
29 	u64 cost;
30 	int err;
31 
32 	if (!capable(CAP_NET_ADMIN))
33 		return ERR_PTR(-EPERM);
34 	if (attr->max_entries == 0 ||
35 	    attr->key_size    != 4 ||
36 	    (attr->value_size != sizeof(u32) &&
37 	     attr->value_size != sizeof(u64)) ||
38 	    attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
39 		return ERR_PTR(-EINVAL);
40 
41 	stab = kzalloc(sizeof(*stab), GFP_USER);
42 	if (!stab)
43 		return ERR_PTR(-ENOMEM);
44 
45 	bpf_map_init_from_attr(&stab->map, attr);
46 	raw_spin_lock_init(&stab->lock);
47 
48 	/* Make sure page count doesn't overflow. */
49 	cost = (u64) stab->map.max_entries * sizeof(struct sock *);
50 	err = bpf_map_charge_init(&stab->map.memory, cost);
51 	if (err)
52 		goto free_stab;
53 
54 	stab->sks = bpf_map_area_alloc(stab->map.max_entries *
55 				       sizeof(struct sock *),
56 				       stab->map.numa_node);
57 	if (stab->sks)
58 		return &stab->map;
59 	err = -ENOMEM;
60 	bpf_map_charge_finish(&stab->map.memory);
61 free_stab:
62 	kfree(stab);
63 	return ERR_PTR(err);
64 }
65 
66 int sock_map_get_from_fd(const union bpf_attr *attr, struct bpf_prog *prog)
67 {
68 	u32 ufd = attr->target_fd;
69 	struct bpf_map *map;
70 	struct fd f;
71 	int ret;
72 
73 	f = fdget(ufd);
74 	map = __bpf_map_get(f);
75 	if (IS_ERR(map))
76 		return PTR_ERR(map);
77 	ret = sock_map_prog_update(map, prog, attr->attach_type);
78 	fdput(f);
79 	return ret;
80 }
81 
82 static void sock_map_sk_acquire(struct sock *sk)
83 	__acquires(&sk->sk_lock.slock)
84 {
85 	lock_sock(sk);
86 	preempt_disable();
87 	rcu_read_lock();
88 }
89 
90 static void sock_map_sk_release(struct sock *sk)
91 	__releases(&sk->sk_lock.slock)
92 {
93 	rcu_read_unlock();
94 	preempt_enable();
95 	release_sock(sk);
96 }
97 
98 static void sock_map_add_link(struct sk_psock *psock,
99 			      struct sk_psock_link *link,
100 			      struct bpf_map *map, void *link_raw)
101 {
102 	link->link_raw = link_raw;
103 	link->map = map;
104 	spin_lock_bh(&psock->link_lock);
105 	list_add_tail(&link->list, &psock->link);
106 	spin_unlock_bh(&psock->link_lock);
107 }
108 
109 static void sock_map_del_link(struct sock *sk,
110 			      struct sk_psock *psock, void *link_raw)
111 {
112 	struct sk_psock_link *link, *tmp;
113 	bool strp_stop = false;
114 
115 	spin_lock_bh(&psock->link_lock);
116 	list_for_each_entry_safe(link, tmp, &psock->link, list) {
117 		if (link->link_raw == link_raw) {
118 			struct bpf_map *map = link->map;
119 			struct bpf_stab *stab = container_of(map, struct bpf_stab,
120 							     map);
121 			if (psock->parser.enabled && stab->progs.skb_parser)
122 				strp_stop = true;
123 			list_del(&link->list);
124 			sk_psock_free_link(link);
125 		}
126 	}
127 	spin_unlock_bh(&psock->link_lock);
128 	if (strp_stop) {
129 		write_lock_bh(&sk->sk_callback_lock);
130 		sk_psock_stop_strp(sk, psock);
131 		write_unlock_bh(&sk->sk_callback_lock);
132 	}
133 }
134 
135 static void sock_map_unref(struct sock *sk, void *link_raw)
136 {
137 	struct sk_psock *psock = sk_psock(sk);
138 
139 	if (likely(psock)) {
140 		sock_map_del_link(sk, psock, link_raw);
141 		sk_psock_put(sk, psock);
142 	}
143 }
144 
145 static int sock_map_init_proto(struct sock *sk, struct sk_psock *psock)
146 {
147 	struct proto *prot;
148 
149 	sock_owned_by_me(sk);
150 
151 	switch (sk->sk_type) {
152 	case SOCK_STREAM:
153 		prot = tcp_bpf_get_proto(sk, psock);
154 		break;
155 
156 	case SOCK_DGRAM:
157 		prot = udp_bpf_get_proto(sk, psock);
158 		break;
159 
160 	default:
161 		return -EINVAL;
162 	}
163 
164 	if (IS_ERR(prot))
165 		return PTR_ERR(prot);
166 
167 	sk_psock_update_proto(sk, psock, prot);
168 	return 0;
169 }
170 
171 static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
172 {
173 	struct sk_psock *psock;
174 
175 	rcu_read_lock();
176 	psock = sk_psock(sk);
177 	if (psock) {
178 		if (sk->sk_prot->close != sock_map_close) {
179 			psock = ERR_PTR(-EBUSY);
180 			goto out;
181 		}
182 
183 		if (!refcount_inc_not_zero(&psock->refcnt))
184 			psock = ERR_PTR(-EBUSY);
185 	}
186 out:
187 	rcu_read_unlock();
188 	return psock;
189 }
190 
191 static int sock_map_link(struct bpf_map *map, struct sk_psock_progs *progs,
192 			 struct sock *sk)
193 {
194 	struct bpf_prog *msg_parser, *skb_parser, *skb_verdict;
195 	struct sk_psock *psock;
196 	bool skb_progs;
197 	int ret;
198 
199 	skb_verdict = READ_ONCE(progs->skb_verdict);
200 	skb_parser = READ_ONCE(progs->skb_parser);
201 	skb_progs = skb_parser && skb_verdict;
202 	if (skb_progs) {
203 		skb_verdict = bpf_prog_inc_not_zero(skb_verdict);
204 		if (IS_ERR(skb_verdict))
205 			return PTR_ERR(skb_verdict);
206 		skb_parser = bpf_prog_inc_not_zero(skb_parser);
207 		if (IS_ERR(skb_parser)) {
208 			bpf_prog_put(skb_verdict);
209 			return PTR_ERR(skb_parser);
210 		}
211 	}
212 
213 	msg_parser = READ_ONCE(progs->msg_parser);
214 	if (msg_parser) {
215 		msg_parser = bpf_prog_inc_not_zero(msg_parser);
216 		if (IS_ERR(msg_parser)) {
217 			ret = PTR_ERR(msg_parser);
218 			goto out;
219 		}
220 	}
221 
222 	psock = sock_map_psock_get_checked(sk);
223 	if (IS_ERR(psock)) {
224 		ret = PTR_ERR(psock);
225 		goto out_progs;
226 	}
227 
228 	if (psock) {
229 		if ((msg_parser && READ_ONCE(psock->progs.msg_parser)) ||
230 		    (skb_progs  && READ_ONCE(psock->progs.skb_parser))) {
231 			sk_psock_put(sk, psock);
232 			ret = -EBUSY;
233 			goto out_progs;
234 		}
235 	} else {
236 		psock = sk_psock_init(sk, map->numa_node);
237 		if (!psock) {
238 			ret = -ENOMEM;
239 			goto out_progs;
240 		}
241 	}
242 
243 	if (msg_parser)
244 		psock_set_prog(&psock->progs.msg_parser, msg_parser);
245 
246 	ret = sock_map_init_proto(sk, psock);
247 	if (ret < 0)
248 		goto out_drop;
249 
250 	write_lock_bh(&sk->sk_callback_lock);
251 	if (skb_progs && !psock->parser.enabled) {
252 		ret = sk_psock_init_strp(sk, psock);
253 		if (ret) {
254 			write_unlock_bh(&sk->sk_callback_lock);
255 			goto out_drop;
256 		}
257 		psock_set_prog(&psock->progs.skb_verdict, skb_verdict);
258 		psock_set_prog(&psock->progs.skb_parser, skb_parser);
259 		sk_psock_start_strp(sk, psock);
260 	}
261 	write_unlock_bh(&sk->sk_callback_lock);
262 	return 0;
263 out_drop:
264 	sk_psock_put(sk, psock);
265 out_progs:
266 	if (msg_parser)
267 		bpf_prog_put(msg_parser);
268 out:
269 	if (skb_progs) {
270 		bpf_prog_put(skb_verdict);
271 		bpf_prog_put(skb_parser);
272 	}
273 	return ret;
274 }
275 
276 static int sock_map_link_no_progs(struct bpf_map *map, struct sock *sk)
277 {
278 	struct sk_psock *psock;
279 	int ret;
280 
281 	psock = sock_map_psock_get_checked(sk);
282 	if (IS_ERR(psock))
283 		return PTR_ERR(psock);
284 
285 	if (!psock) {
286 		psock = sk_psock_init(sk, map->numa_node);
287 		if (!psock)
288 			return -ENOMEM;
289 	}
290 
291 	ret = sock_map_init_proto(sk, psock);
292 	if (ret < 0)
293 		sk_psock_put(sk, psock);
294 	return ret;
295 }
296 
297 static void sock_map_free(struct bpf_map *map)
298 {
299 	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
300 	int i;
301 
302 	/* After the sync no updates or deletes will be in-flight so it
303 	 * is safe to walk map and remove entries without risking a race
304 	 * in EEXIST update case.
305 	 */
306 	synchronize_rcu();
307 	for (i = 0; i < stab->map.max_entries; i++) {
308 		struct sock **psk = &stab->sks[i];
309 		struct sock *sk;
310 
311 		sk = xchg(psk, NULL);
312 		if (sk) {
313 			lock_sock(sk);
314 			rcu_read_lock();
315 			sock_map_unref(sk, psk);
316 			rcu_read_unlock();
317 			release_sock(sk);
318 		}
319 	}
320 
321 	/* wait for psock readers accessing its map link */
322 	synchronize_rcu();
323 
324 	bpf_map_area_free(stab->sks);
325 	kfree(stab);
326 }
327 
328 static void sock_map_release_progs(struct bpf_map *map)
329 {
330 	psock_progs_drop(&container_of(map, struct bpf_stab, map)->progs);
331 }
332 
333 static struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key)
334 {
335 	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
336 
337 	WARN_ON_ONCE(!rcu_read_lock_held());
338 
339 	if (unlikely(key >= map->max_entries))
340 		return NULL;
341 	return READ_ONCE(stab->sks[key]);
342 }
343 
344 static void *sock_map_lookup(struct bpf_map *map, void *key)
345 {
346 	return __sock_map_lookup_elem(map, *(u32 *)key);
347 }
348 
349 static void *sock_map_lookup_sys(struct bpf_map *map, void *key)
350 {
351 	struct sock *sk;
352 
353 	if (map->value_size != sizeof(u64))
354 		return ERR_PTR(-ENOSPC);
355 
356 	sk = __sock_map_lookup_elem(map, *(u32 *)key);
357 	if (!sk)
358 		return ERR_PTR(-ENOENT);
359 
360 	sock_gen_cookie(sk);
361 	return &sk->sk_cookie;
362 }
363 
364 static int __sock_map_delete(struct bpf_stab *stab, struct sock *sk_test,
365 			     struct sock **psk)
366 {
367 	struct sock *sk;
368 	int err = 0;
369 
370 	raw_spin_lock_bh(&stab->lock);
371 	sk = *psk;
372 	if (!sk_test || sk_test == sk)
373 		sk = xchg(psk, NULL);
374 
375 	if (likely(sk))
376 		sock_map_unref(sk, psk);
377 	else
378 		err = -EINVAL;
379 
380 	raw_spin_unlock_bh(&stab->lock);
381 	return err;
382 }
383 
384 static void sock_map_delete_from_link(struct bpf_map *map, struct sock *sk,
385 				      void *link_raw)
386 {
387 	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
388 
389 	__sock_map_delete(stab, sk, link_raw);
390 }
391 
392 static int sock_map_delete_elem(struct bpf_map *map, void *key)
393 {
394 	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
395 	u32 i = *(u32 *)key;
396 	struct sock **psk;
397 
398 	if (unlikely(i >= map->max_entries))
399 		return -EINVAL;
400 
401 	psk = &stab->sks[i];
402 	return __sock_map_delete(stab, NULL, psk);
403 }
404 
405 static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next)
406 {
407 	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
408 	u32 i = key ? *(u32 *)key : U32_MAX;
409 	u32 *key_next = next;
410 
411 	if (i == stab->map.max_entries - 1)
412 		return -ENOENT;
413 	if (i >= stab->map.max_entries)
414 		*key_next = 0;
415 	else
416 		*key_next = i + 1;
417 	return 0;
418 }
419 
420 static bool sock_map_redirect_allowed(const struct sock *sk)
421 {
422 	return sk->sk_state != TCP_LISTEN;
423 }
424 
425 static int sock_map_update_common(struct bpf_map *map, u32 idx,
426 				  struct sock *sk, u64 flags)
427 {
428 	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
429 	struct sk_psock_link *link;
430 	struct sk_psock *psock;
431 	struct sock *osk;
432 	int ret;
433 
434 	WARN_ON_ONCE(!rcu_read_lock_held());
435 	if (unlikely(flags > BPF_EXIST))
436 		return -EINVAL;
437 	if (unlikely(idx >= map->max_entries))
438 		return -E2BIG;
439 	if (inet_csk_has_ulp(sk))
440 		return -EINVAL;
441 
442 	link = sk_psock_init_link();
443 	if (!link)
444 		return -ENOMEM;
445 
446 	/* Only sockets we can redirect into/from in BPF need to hold
447 	 * refs to parser/verdict progs and have their sk_data_ready
448 	 * and sk_write_space callbacks overridden.
449 	 */
450 	if (sock_map_redirect_allowed(sk))
451 		ret = sock_map_link(map, &stab->progs, sk);
452 	else
453 		ret = sock_map_link_no_progs(map, sk);
454 	if (ret < 0)
455 		goto out_free;
456 
457 	psock = sk_psock(sk);
458 	WARN_ON_ONCE(!psock);
459 
460 	raw_spin_lock_bh(&stab->lock);
461 	osk = stab->sks[idx];
462 	if (osk && flags == BPF_NOEXIST) {
463 		ret = -EEXIST;
464 		goto out_unlock;
465 	} else if (!osk && flags == BPF_EXIST) {
466 		ret = -ENOENT;
467 		goto out_unlock;
468 	}
469 
470 	sock_map_add_link(psock, link, map, &stab->sks[idx]);
471 	stab->sks[idx] = sk;
472 	if (osk)
473 		sock_map_unref(osk, &stab->sks[idx]);
474 	raw_spin_unlock_bh(&stab->lock);
475 	return 0;
476 out_unlock:
477 	raw_spin_unlock_bh(&stab->lock);
478 	if (psock)
479 		sk_psock_put(sk, psock);
480 out_free:
481 	sk_psock_free_link(link);
482 	return ret;
483 }
484 
485 static bool sock_map_op_okay(const struct bpf_sock_ops_kern *ops)
486 {
487 	return ops->op == BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB ||
488 	       ops->op == BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB ||
489 	       ops->op == BPF_SOCK_OPS_TCP_LISTEN_CB;
490 }
491 
492 static bool sk_is_tcp(const struct sock *sk)
493 {
494 	return sk->sk_type == SOCK_STREAM &&
495 	       sk->sk_protocol == IPPROTO_TCP;
496 }
497 
498 static bool sk_is_udp(const struct sock *sk)
499 {
500 	return sk->sk_type == SOCK_DGRAM &&
501 	       sk->sk_protocol == IPPROTO_UDP;
502 }
503 
504 static bool sock_map_sk_is_suitable(const struct sock *sk)
505 {
506 	return sk_is_tcp(sk) || sk_is_udp(sk);
507 }
508 
509 static bool sock_map_sk_state_allowed(const struct sock *sk)
510 {
511 	if (sk_is_tcp(sk))
512 		return (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_LISTEN);
513 	else if (sk_is_udp(sk))
514 		return sk_hashed(sk);
515 
516 	return false;
517 }
518 
519 static int sock_map_update_elem(struct bpf_map *map, void *key,
520 				void *value, u64 flags)
521 {
522 	u32 idx = *(u32 *)key;
523 	struct socket *sock;
524 	struct sock *sk;
525 	int ret;
526 	u64 ufd;
527 
528 	if (map->value_size == sizeof(u64))
529 		ufd = *(u64 *)value;
530 	else
531 		ufd = *(u32 *)value;
532 	if (ufd > S32_MAX)
533 		return -EINVAL;
534 
535 	sock = sockfd_lookup(ufd, &ret);
536 	if (!sock)
537 		return ret;
538 	sk = sock->sk;
539 	if (!sk) {
540 		ret = -EINVAL;
541 		goto out;
542 	}
543 	if (!sock_map_sk_is_suitable(sk)) {
544 		ret = -EOPNOTSUPP;
545 		goto out;
546 	}
547 
548 	sock_map_sk_acquire(sk);
549 	if (!sock_map_sk_state_allowed(sk))
550 		ret = -EOPNOTSUPP;
551 	else
552 		ret = sock_map_update_common(map, idx, sk, flags);
553 	sock_map_sk_release(sk);
554 out:
555 	fput(sock->file);
556 	return ret;
557 }
558 
559 BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, sops,
560 	   struct bpf_map *, map, void *, key, u64, flags)
561 {
562 	WARN_ON_ONCE(!rcu_read_lock_held());
563 
564 	if (likely(sock_map_sk_is_suitable(sops->sk) &&
565 		   sock_map_op_okay(sops)))
566 		return sock_map_update_common(map, *(u32 *)key, sops->sk,
567 					      flags);
568 	return -EOPNOTSUPP;
569 }
570 
571 const struct bpf_func_proto bpf_sock_map_update_proto = {
572 	.func		= bpf_sock_map_update,
573 	.gpl_only	= false,
574 	.pkt_access	= true,
575 	.ret_type	= RET_INTEGER,
576 	.arg1_type	= ARG_PTR_TO_CTX,
577 	.arg2_type	= ARG_CONST_MAP_PTR,
578 	.arg3_type	= ARG_PTR_TO_MAP_KEY,
579 	.arg4_type	= ARG_ANYTHING,
580 };
581 
582 BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
583 	   struct bpf_map *, map, u32, key, u64, flags)
584 {
585 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
586 	struct sock *sk;
587 
588 	if (unlikely(flags & ~(BPF_F_INGRESS)))
589 		return SK_DROP;
590 
591 	sk = __sock_map_lookup_elem(map, key);
592 	if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
593 		return SK_DROP;
594 
595 	tcb->bpf.flags = flags;
596 	tcb->bpf.sk_redir = sk;
597 	return SK_PASS;
598 }
599 
600 const struct bpf_func_proto bpf_sk_redirect_map_proto = {
601 	.func           = bpf_sk_redirect_map,
602 	.gpl_only       = false,
603 	.ret_type       = RET_INTEGER,
604 	.arg1_type	= ARG_PTR_TO_CTX,
605 	.arg2_type      = ARG_CONST_MAP_PTR,
606 	.arg3_type      = ARG_ANYTHING,
607 	.arg4_type      = ARG_ANYTHING,
608 };
609 
610 BPF_CALL_4(bpf_msg_redirect_map, struct sk_msg *, msg,
611 	   struct bpf_map *, map, u32, key, u64, flags)
612 {
613 	struct sock *sk;
614 
615 	if (unlikely(flags & ~(BPF_F_INGRESS)))
616 		return SK_DROP;
617 
618 	sk = __sock_map_lookup_elem(map, key);
619 	if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
620 		return SK_DROP;
621 
622 	msg->flags = flags;
623 	msg->sk_redir = sk;
624 	return SK_PASS;
625 }
626 
627 const struct bpf_func_proto bpf_msg_redirect_map_proto = {
628 	.func           = bpf_msg_redirect_map,
629 	.gpl_only       = false,
630 	.ret_type       = RET_INTEGER,
631 	.arg1_type	= ARG_PTR_TO_CTX,
632 	.arg2_type      = ARG_CONST_MAP_PTR,
633 	.arg3_type      = ARG_ANYTHING,
634 	.arg4_type      = ARG_ANYTHING,
635 };
636 
637 const struct bpf_map_ops sock_map_ops = {
638 	.map_alloc		= sock_map_alloc,
639 	.map_free		= sock_map_free,
640 	.map_get_next_key	= sock_map_get_next_key,
641 	.map_lookup_elem_sys_only = sock_map_lookup_sys,
642 	.map_update_elem	= sock_map_update_elem,
643 	.map_delete_elem	= sock_map_delete_elem,
644 	.map_lookup_elem	= sock_map_lookup,
645 	.map_release_uref	= sock_map_release_progs,
646 	.map_check_btf		= map_check_no_btf,
647 };
648 
649 struct bpf_htab_elem {
650 	struct rcu_head rcu;
651 	u32 hash;
652 	struct sock *sk;
653 	struct hlist_node node;
654 	u8 key[];
655 };
656 
657 struct bpf_htab_bucket {
658 	struct hlist_head head;
659 	raw_spinlock_t lock;
660 };
661 
662 struct bpf_htab {
663 	struct bpf_map map;
664 	struct bpf_htab_bucket *buckets;
665 	u32 buckets_num;
666 	u32 elem_size;
667 	struct sk_psock_progs progs;
668 	atomic_t count;
669 };
670 
671 static inline u32 sock_hash_bucket_hash(const void *key, u32 len)
672 {
673 	return jhash(key, len, 0);
674 }
675 
676 static struct bpf_htab_bucket *sock_hash_select_bucket(struct bpf_htab *htab,
677 						       u32 hash)
678 {
679 	return &htab->buckets[hash & (htab->buckets_num - 1)];
680 }
681 
682 static struct bpf_htab_elem *
683 sock_hash_lookup_elem_raw(struct hlist_head *head, u32 hash, void *key,
684 			  u32 key_size)
685 {
686 	struct bpf_htab_elem *elem;
687 
688 	hlist_for_each_entry_rcu(elem, head, node) {
689 		if (elem->hash == hash &&
690 		    !memcmp(&elem->key, key, key_size))
691 			return elem;
692 	}
693 
694 	return NULL;
695 }
696 
697 static struct sock *__sock_hash_lookup_elem(struct bpf_map *map, void *key)
698 {
699 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
700 	u32 key_size = map->key_size, hash;
701 	struct bpf_htab_bucket *bucket;
702 	struct bpf_htab_elem *elem;
703 
704 	WARN_ON_ONCE(!rcu_read_lock_held());
705 
706 	hash = sock_hash_bucket_hash(key, key_size);
707 	bucket = sock_hash_select_bucket(htab, hash);
708 	elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
709 
710 	return elem ? elem->sk : NULL;
711 }
712 
713 static void sock_hash_free_elem(struct bpf_htab *htab,
714 				struct bpf_htab_elem *elem)
715 {
716 	atomic_dec(&htab->count);
717 	kfree_rcu(elem, rcu);
718 }
719 
720 static void sock_hash_delete_from_link(struct bpf_map *map, struct sock *sk,
721 				       void *link_raw)
722 {
723 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
724 	struct bpf_htab_elem *elem_probe, *elem = link_raw;
725 	struct bpf_htab_bucket *bucket;
726 
727 	WARN_ON_ONCE(!rcu_read_lock_held());
728 	bucket = sock_hash_select_bucket(htab, elem->hash);
729 
730 	/* elem may be deleted in parallel from the map, but access here
731 	 * is okay since it's going away only after RCU grace period.
732 	 * However, we need to check whether it's still present.
733 	 */
734 	raw_spin_lock_bh(&bucket->lock);
735 	elem_probe = sock_hash_lookup_elem_raw(&bucket->head, elem->hash,
736 					       elem->key, map->key_size);
737 	if (elem_probe && elem_probe == elem) {
738 		hlist_del_rcu(&elem->node);
739 		sock_map_unref(elem->sk, elem);
740 		sock_hash_free_elem(htab, elem);
741 	}
742 	raw_spin_unlock_bh(&bucket->lock);
743 }
744 
745 static int sock_hash_delete_elem(struct bpf_map *map, void *key)
746 {
747 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
748 	u32 hash, key_size = map->key_size;
749 	struct bpf_htab_bucket *bucket;
750 	struct bpf_htab_elem *elem;
751 	int ret = -ENOENT;
752 
753 	hash = sock_hash_bucket_hash(key, key_size);
754 	bucket = sock_hash_select_bucket(htab, hash);
755 
756 	raw_spin_lock_bh(&bucket->lock);
757 	elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
758 	if (elem) {
759 		hlist_del_rcu(&elem->node);
760 		sock_map_unref(elem->sk, elem);
761 		sock_hash_free_elem(htab, elem);
762 		ret = 0;
763 	}
764 	raw_spin_unlock_bh(&bucket->lock);
765 	return ret;
766 }
767 
768 static struct bpf_htab_elem *sock_hash_alloc_elem(struct bpf_htab *htab,
769 						  void *key, u32 key_size,
770 						  u32 hash, struct sock *sk,
771 						  struct bpf_htab_elem *old)
772 {
773 	struct bpf_htab_elem *new;
774 
775 	if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
776 		if (!old) {
777 			atomic_dec(&htab->count);
778 			return ERR_PTR(-E2BIG);
779 		}
780 	}
781 
782 	new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
783 			   htab->map.numa_node);
784 	if (!new) {
785 		atomic_dec(&htab->count);
786 		return ERR_PTR(-ENOMEM);
787 	}
788 	memcpy(new->key, key, key_size);
789 	new->sk = sk;
790 	new->hash = hash;
791 	return new;
792 }
793 
794 static int sock_hash_update_common(struct bpf_map *map, void *key,
795 				   struct sock *sk, u64 flags)
796 {
797 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
798 	u32 key_size = map->key_size, hash;
799 	struct bpf_htab_elem *elem, *elem_new;
800 	struct bpf_htab_bucket *bucket;
801 	struct sk_psock_link *link;
802 	struct sk_psock *psock;
803 	int ret;
804 
805 	WARN_ON_ONCE(!rcu_read_lock_held());
806 	if (unlikely(flags > BPF_EXIST))
807 		return -EINVAL;
808 	if (inet_csk_has_ulp(sk))
809 		return -EINVAL;
810 
811 	link = sk_psock_init_link();
812 	if (!link)
813 		return -ENOMEM;
814 
815 	/* Only sockets we can redirect into/from in BPF need to hold
816 	 * refs to parser/verdict progs and have their sk_data_ready
817 	 * and sk_write_space callbacks overridden.
818 	 */
819 	if (sock_map_redirect_allowed(sk))
820 		ret = sock_map_link(map, &htab->progs, sk);
821 	else
822 		ret = sock_map_link_no_progs(map, sk);
823 	if (ret < 0)
824 		goto out_free;
825 
826 	psock = sk_psock(sk);
827 	WARN_ON_ONCE(!psock);
828 
829 	hash = sock_hash_bucket_hash(key, key_size);
830 	bucket = sock_hash_select_bucket(htab, hash);
831 
832 	raw_spin_lock_bh(&bucket->lock);
833 	elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
834 	if (elem && flags == BPF_NOEXIST) {
835 		ret = -EEXIST;
836 		goto out_unlock;
837 	} else if (!elem && flags == BPF_EXIST) {
838 		ret = -ENOENT;
839 		goto out_unlock;
840 	}
841 
842 	elem_new = sock_hash_alloc_elem(htab, key, key_size, hash, sk, elem);
843 	if (IS_ERR(elem_new)) {
844 		ret = PTR_ERR(elem_new);
845 		goto out_unlock;
846 	}
847 
848 	sock_map_add_link(psock, link, map, elem_new);
849 	/* Add new element to the head of the list, so that
850 	 * concurrent search will find it before old elem.
851 	 */
852 	hlist_add_head_rcu(&elem_new->node, &bucket->head);
853 	if (elem) {
854 		hlist_del_rcu(&elem->node);
855 		sock_map_unref(elem->sk, elem);
856 		sock_hash_free_elem(htab, elem);
857 	}
858 	raw_spin_unlock_bh(&bucket->lock);
859 	return 0;
860 out_unlock:
861 	raw_spin_unlock_bh(&bucket->lock);
862 	sk_psock_put(sk, psock);
863 out_free:
864 	sk_psock_free_link(link);
865 	return ret;
866 }
867 
868 static int sock_hash_update_elem(struct bpf_map *map, void *key,
869 				 void *value, u64 flags)
870 {
871 	struct socket *sock;
872 	struct sock *sk;
873 	int ret;
874 	u64 ufd;
875 
876 	if (map->value_size == sizeof(u64))
877 		ufd = *(u64 *)value;
878 	else
879 		ufd = *(u32 *)value;
880 	if (ufd > S32_MAX)
881 		return -EINVAL;
882 
883 	sock = sockfd_lookup(ufd, &ret);
884 	if (!sock)
885 		return ret;
886 	sk = sock->sk;
887 	if (!sk) {
888 		ret = -EINVAL;
889 		goto out;
890 	}
891 	if (!sock_map_sk_is_suitable(sk)) {
892 		ret = -EOPNOTSUPP;
893 		goto out;
894 	}
895 
896 	sock_map_sk_acquire(sk);
897 	if (!sock_map_sk_state_allowed(sk))
898 		ret = -EOPNOTSUPP;
899 	else
900 		ret = sock_hash_update_common(map, key, sk, flags);
901 	sock_map_sk_release(sk);
902 out:
903 	fput(sock->file);
904 	return ret;
905 }
906 
907 static int sock_hash_get_next_key(struct bpf_map *map, void *key,
908 				  void *key_next)
909 {
910 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
911 	struct bpf_htab_elem *elem, *elem_next;
912 	u32 hash, key_size = map->key_size;
913 	struct hlist_head *head;
914 	int i = 0;
915 
916 	if (!key)
917 		goto find_first_elem;
918 	hash = sock_hash_bucket_hash(key, key_size);
919 	head = &sock_hash_select_bucket(htab, hash)->head;
920 	elem = sock_hash_lookup_elem_raw(head, hash, key, key_size);
921 	if (!elem)
922 		goto find_first_elem;
923 
924 	elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&elem->node)),
925 				     struct bpf_htab_elem, node);
926 	if (elem_next) {
927 		memcpy(key_next, elem_next->key, key_size);
928 		return 0;
929 	}
930 
931 	i = hash & (htab->buckets_num - 1);
932 	i++;
933 find_first_elem:
934 	for (; i < htab->buckets_num; i++) {
935 		head = &sock_hash_select_bucket(htab, i)->head;
936 		elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
937 					     struct bpf_htab_elem, node);
938 		if (elem_next) {
939 			memcpy(key_next, elem_next->key, key_size);
940 			return 0;
941 		}
942 	}
943 
944 	return -ENOENT;
945 }
946 
947 static struct bpf_map *sock_hash_alloc(union bpf_attr *attr)
948 {
949 	struct bpf_htab *htab;
950 	int i, err;
951 	u64 cost;
952 
953 	if (!capable(CAP_NET_ADMIN))
954 		return ERR_PTR(-EPERM);
955 	if (attr->max_entries == 0 ||
956 	    attr->key_size    == 0 ||
957 	    (attr->value_size != sizeof(u32) &&
958 	     attr->value_size != sizeof(u64)) ||
959 	    attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
960 		return ERR_PTR(-EINVAL);
961 	if (attr->key_size > MAX_BPF_STACK)
962 		return ERR_PTR(-E2BIG);
963 
964 	htab = kzalloc(sizeof(*htab), GFP_USER);
965 	if (!htab)
966 		return ERR_PTR(-ENOMEM);
967 
968 	bpf_map_init_from_attr(&htab->map, attr);
969 
970 	htab->buckets_num = roundup_pow_of_two(htab->map.max_entries);
971 	htab->elem_size = sizeof(struct bpf_htab_elem) +
972 			  round_up(htab->map.key_size, 8);
973 	if (htab->buckets_num == 0 ||
974 	    htab->buckets_num > U32_MAX / sizeof(struct bpf_htab_bucket)) {
975 		err = -EINVAL;
976 		goto free_htab;
977 	}
978 
979 	cost = (u64) htab->buckets_num * sizeof(struct bpf_htab_bucket) +
980 	       (u64) htab->elem_size * htab->map.max_entries;
981 	if (cost >= U32_MAX - PAGE_SIZE) {
982 		err = -EINVAL;
983 		goto free_htab;
984 	}
985 
986 	htab->buckets = bpf_map_area_alloc(htab->buckets_num *
987 					   sizeof(struct bpf_htab_bucket),
988 					   htab->map.numa_node);
989 	if (!htab->buckets) {
990 		err = -ENOMEM;
991 		goto free_htab;
992 	}
993 
994 	for (i = 0; i < htab->buckets_num; i++) {
995 		INIT_HLIST_HEAD(&htab->buckets[i].head);
996 		raw_spin_lock_init(&htab->buckets[i].lock);
997 	}
998 
999 	return &htab->map;
1000 free_htab:
1001 	kfree(htab);
1002 	return ERR_PTR(err);
1003 }
1004 
1005 static void sock_hash_free(struct bpf_map *map)
1006 {
1007 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1008 	struct bpf_htab_bucket *bucket;
1009 	struct bpf_htab_elem *elem;
1010 	struct hlist_node *node;
1011 	int i;
1012 
1013 	/* After the sync no updates or deletes will be in-flight so it
1014 	 * is safe to walk map and remove entries without risking a race
1015 	 * in EEXIST update case.
1016 	 */
1017 	synchronize_rcu();
1018 	for (i = 0; i < htab->buckets_num; i++) {
1019 		bucket = sock_hash_select_bucket(htab, i);
1020 		hlist_for_each_entry_safe(elem, node, &bucket->head, node) {
1021 			hlist_del_rcu(&elem->node);
1022 			lock_sock(elem->sk);
1023 			rcu_read_lock();
1024 			sock_map_unref(elem->sk, elem);
1025 			rcu_read_unlock();
1026 			release_sock(elem->sk);
1027 		}
1028 	}
1029 
1030 	/* wait for psock readers accessing its map link */
1031 	synchronize_rcu();
1032 
1033 	bpf_map_area_free(htab->buckets);
1034 	kfree(htab);
1035 }
1036 
1037 static void *sock_hash_lookup_sys(struct bpf_map *map, void *key)
1038 {
1039 	struct sock *sk;
1040 
1041 	if (map->value_size != sizeof(u64))
1042 		return ERR_PTR(-ENOSPC);
1043 
1044 	sk = __sock_hash_lookup_elem(map, key);
1045 	if (!sk)
1046 		return ERR_PTR(-ENOENT);
1047 
1048 	sock_gen_cookie(sk);
1049 	return &sk->sk_cookie;
1050 }
1051 
1052 static void *sock_hash_lookup(struct bpf_map *map, void *key)
1053 {
1054 	return __sock_hash_lookup_elem(map, key);
1055 }
1056 
1057 static void sock_hash_release_progs(struct bpf_map *map)
1058 {
1059 	psock_progs_drop(&container_of(map, struct bpf_htab, map)->progs);
1060 }
1061 
1062 BPF_CALL_4(bpf_sock_hash_update, struct bpf_sock_ops_kern *, sops,
1063 	   struct bpf_map *, map, void *, key, u64, flags)
1064 {
1065 	WARN_ON_ONCE(!rcu_read_lock_held());
1066 
1067 	if (likely(sock_map_sk_is_suitable(sops->sk) &&
1068 		   sock_map_op_okay(sops)))
1069 		return sock_hash_update_common(map, key, sops->sk, flags);
1070 	return -EOPNOTSUPP;
1071 }
1072 
1073 const struct bpf_func_proto bpf_sock_hash_update_proto = {
1074 	.func		= bpf_sock_hash_update,
1075 	.gpl_only	= false,
1076 	.pkt_access	= true,
1077 	.ret_type	= RET_INTEGER,
1078 	.arg1_type	= ARG_PTR_TO_CTX,
1079 	.arg2_type	= ARG_CONST_MAP_PTR,
1080 	.arg3_type	= ARG_PTR_TO_MAP_KEY,
1081 	.arg4_type	= ARG_ANYTHING,
1082 };
1083 
1084 BPF_CALL_4(bpf_sk_redirect_hash, struct sk_buff *, skb,
1085 	   struct bpf_map *, map, void *, key, u64, flags)
1086 {
1087 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1088 	struct sock *sk;
1089 
1090 	if (unlikely(flags & ~(BPF_F_INGRESS)))
1091 		return SK_DROP;
1092 
1093 	sk = __sock_hash_lookup_elem(map, key);
1094 	if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1095 		return SK_DROP;
1096 
1097 	tcb->bpf.flags = flags;
1098 	tcb->bpf.sk_redir = sk;
1099 	return SK_PASS;
1100 }
1101 
1102 const struct bpf_func_proto bpf_sk_redirect_hash_proto = {
1103 	.func           = bpf_sk_redirect_hash,
1104 	.gpl_only       = false,
1105 	.ret_type       = RET_INTEGER,
1106 	.arg1_type	= ARG_PTR_TO_CTX,
1107 	.arg2_type      = ARG_CONST_MAP_PTR,
1108 	.arg3_type      = ARG_PTR_TO_MAP_KEY,
1109 	.arg4_type      = ARG_ANYTHING,
1110 };
1111 
1112 BPF_CALL_4(bpf_msg_redirect_hash, struct sk_msg *, msg,
1113 	   struct bpf_map *, map, void *, key, u64, flags)
1114 {
1115 	struct sock *sk;
1116 
1117 	if (unlikely(flags & ~(BPF_F_INGRESS)))
1118 		return SK_DROP;
1119 
1120 	sk = __sock_hash_lookup_elem(map, key);
1121 	if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1122 		return SK_DROP;
1123 
1124 	msg->flags = flags;
1125 	msg->sk_redir = sk;
1126 	return SK_PASS;
1127 }
1128 
1129 const struct bpf_func_proto bpf_msg_redirect_hash_proto = {
1130 	.func           = bpf_msg_redirect_hash,
1131 	.gpl_only       = false,
1132 	.ret_type       = RET_INTEGER,
1133 	.arg1_type	= ARG_PTR_TO_CTX,
1134 	.arg2_type      = ARG_CONST_MAP_PTR,
1135 	.arg3_type      = ARG_PTR_TO_MAP_KEY,
1136 	.arg4_type      = ARG_ANYTHING,
1137 };
1138 
1139 const struct bpf_map_ops sock_hash_ops = {
1140 	.map_alloc		= sock_hash_alloc,
1141 	.map_free		= sock_hash_free,
1142 	.map_get_next_key	= sock_hash_get_next_key,
1143 	.map_update_elem	= sock_hash_update_elem,
1144 	.map_delete_elem	= sock_hash_delete_elem,
1145 	.map_lookup_elem	= sock_hash_lookup,
1146 	.map_lookup_elem_sys_only = sock_hash_lookup_sys,
1147 	.map_release_uref	= sock_hash_release_progs,
1148 	.map_check_btf		= map_check_no_btf,
1149 };
1150 
1151 static struct sk_psock_progs *sock_map_progs(struct bpf_map *map)
1152 {
1153 	switch (map->map_type) {
1154 	case BPF_MAP_TYPE_SOCKMAP:
1155 		return &container_of(map, struct bpf_stab, map)->progs;
1156 	case BPF_MAP_TYPE_SOCKHASH:
1157 		return &container_of(map, struct bpf_htab, map)->progs;
1158 	default:
1159 		break;
1160 	}
1161 
1162 	return NULL;
1163 }
1164 
1165 int sock_map_prog_update(struct bpf_map *map, struct bpf_prog *prog,
1166 			 u32 which)
1167 {
1168 	struct sk_psock_progs *progs = sock_map_progs(map);
1169 
1170 	if (!progs)
1171 		return -EOPNOTSUPP;
1172 
1173 	switch (which) {
1174 	case BPF_SK_MSG_VERDICT:
1175 		psock_set_prog(&progs->msg_parser, prog);
1176 		break;
1177 	case BPF_SK_SKB_STREAM_PARSER:
1178 		psock_set_prog(&progs->skb_parser, prog);
1179 		break;
1180 	case BPF_SK_SKB_STREAM_VERDICT:
1181 		psock_set_prog(&progs->skb_verdict, prog);
1182 		break;
1183 	default:
1184 		return -EOPNOTSUPP;
1185 	}
1186 
1187 	return 0;
1188 }
1189 
1190 static void sock_map_unlink(struct sock *sk, struct sk_psock_link *link)
1191 {
1192 	switch (link->map->map_type) {
1193 	case BPF_MAP_TYPE_SOCKMAP:
1194 		return sock_map_delete_from_link(link->map, sk,
1195 						 link->link_raw);
1196 	case BPF_MAP_TYPE_SOCKHASH:
1197 		return sock_hash_delete_from_link(link->map, sk,
1198 						  link->link_raw);
1199 	default:
1200 		break;
1201 	}
1202 }
1203 
1204 static void sock_map_remove_links(struct sock *sk, struct sk_psock *psock)
1205 {
1206 	struct sk_psock_link *link;
1207 
1208 	while ((link = sk_psock_link_pop(psock))) {
1209 		sock_map_unlink(sk, link);
1210 		sk_psock_free_link(link);
1211 	}
1212 }
1213 
1214 void sock_map_unhash(struct sock *sk)
1215 {
1216 	void (*saved_unhash)(struct sock *sk);
1217 	struct sk_psock *psock;
1218 
1219 	rcu_read_lock();
1220 	psock = sk_psock(sk);
1221 	if (unlikely(!psock)) {
1222 		rcu_read_unlock();
1223 		if (sk->sk_prot->unhash)
1224 			sk->sk_prot->unhash(sk);
1225 		return;
1226 	}
1227 
1228 	saved_unhash = psock->saved_unhash;
1229 	sock_map_remove_links(sk, psock);
1230 	rcu_read_unlock();
1231 	saved_unhash(sk);
1232 }
1233 
1234 void sock_map_close(struct sock *sk, long timeout)
1235 {
1236 	void (*saved_close)(struct sock *sk, long timeout);
1237 	struct sk_psock *psock;
1238 
1239 	lock_sock(sk);
1240 	rcu_read_lock();
1241 	psock = sk_psock(sk);
1242 	if (unlikely(!psock)) {
1243 		rcu_read_unlock();
1244 		release_sock(sk);
1245 		return sk->sk_prot->close(sk, timeout);
1246 	}
1247 
1248 	saved_close = psock->saved_close;
1249 	sock_map_remove_links(sk, psock);
1250 	rcu_read_unlock();
1251 	release_sock(sk);
1252 	saved_close(sk, timeout);
1253 }
1254