xref: /linux/net/ipv4/tcp_bpf.c (revision cea0f76a483d1270ac6f6513964e3e75193dda48)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
3 
4 #include <linux/skmsg.h>
5 #include <linux/filter.h>
6 #include <linux/bpf.h>
7 #include <linux/init.h>
8 #include <linux/wait.h>
9 
10 #include <net/inet_common.h>
11 #include <net/tls.h>
12 
13 int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock,
14 		      struct msghdr *msg, int len, int flags)
15 {
16 	struct iov_iter *iter = &msg->msg_iter;
17 	int peek = flags & MSG_PEEK;
18 	int i, ret, copied = 0;
19 	struct sk_msg *msg_rx;
20 
21 	msg_rx = list_first_entry_or_null(&psock->ingress_msg,
22 					  struct sk_msg, list);
23 
24 	while (copied != len) {
25 		struct scatterlist *sge;
26 
27 		if (unlikely(!msg_rx))
28 			break;
29 
30 		i = msg_rx->sg.start;
31 		do {
32 			struct page *page;
33 			int copy;
34 
35 			sge = sk_msg_elem(msg_rx, i);
36 			copy = sge->length;
37 			page = sg_page(sge);
38 			if (copied + copy > len)
39 				copy = len - copied;
40 			ret = copy_page_to_iter(page, sge->offset, copy, iter);
41 			if (ret != copy) {
42 				msg_rx->sg.start = i;
43 				return -EFAULT;
44 			}
45 
46 			copied += copy;
47 			if (likely(!peek)) {
48 				sge->offset += copy;
49 				sge->length -= copy;
50 				sk_mem_uncharge(sk, copy);
51 				msg_rx->sg.size -= copy;
52 
53 				if (!sge->length) {
54 					sk_msg_iter_var_next(i);
55 					if (!msg_rx->skb)
56 						put_page(page);
57 				}
58 			} else {
59 				sk_msg_iter_var_next(i);
60 			}
61 
62 			if (copied == len)
63 				break;
64 		} while (i != msg_rx->sg.end);
65 
66 		if (unlikely(peek)) {
67 			if (msg_rx == list_last_entry(&psock->ingress_msg,
68 						      struct sk_msg, list))
69 				break;
70 			msg_rx = list_next_entry(msg_rx, list);
71 			continue;
72 		}
73 
74 		msg_rx->sg.start = i;
75 		if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
76 			list_del(&msg_rx->list);
77 			if (msg_rx->skb)
78 				consume_skb(msg_rx->skb);
79 			kfree(msg_rx);
80 		}
81 		msg_rx = list_first_entry_or_null(&psock->ingress_msg,
82 						  struct sk_msg, list);
83 	}
84 
85 	return copied;
86 }
87 EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg);
88 
89 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
90 			   struct sk_msg *msg, u32 apply_bytes, int flags)
91 {
92 	bool apply = apply_bytes;
93 	struct scatterlist *sge;
94 	u32 size, copied = 0;
95 	struct sk_msg *tmp;
96 	int i, ret = 0;
97 
98 	tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
99 	if (unlikely(!tmp))
100 		return -ENOMEM;
101 
102 	lock_sock(sk);
103 	tmp->sg.start = msg->sg.start;
104 	i = msg->sg.start;
105 	do {
106 		sge = sk_msg_elem(msg, i);
107 		size = (apply && apply_bytes < sge->length) ?
108 			apply_bytes : sge->length;
109 		if (!sk_wmem_schedule(sk, size)) {
110 			if (!copied)
111 				ret = -ENOMEM;
112 			break;
113 		}
114 
115 		sk_mem_charge(sk, size);
116 		sk_msg_xfer(tmp, msg, i, size);
117 		copied += size;
118 		if (sge->length)
119 			get_page(sk_msg_page(tmp, i));
120 		sk_msg_iter_var_next(i);
121 		tmp->sg.end = i;
122 		if (apply) {
123 			apply_bytes -= size;
124 			if (!apply_bytes)
125 				break;
126 		}
127 	} while (i != msg->sg.end);
128 
129 	if (!ret) {
130 		msg->sg.start = i;
131 		sk_psock_queue_msg(psock, tmp);
132 		sk_psock_data_ready(sk, psock);
133 	} else {
134 		sk_msg_free(sk, tmp);
135 		kfree(tmp);
136 	}
137 
138 	release_sock(sk);
139 	return ret;
140 }
141 
142 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
143 			int flags, bool uncharge)
144 {
145 	bool apply = apply_bytes;
146 	struct scatterlist *sge;
147 	struct page *page;
148 	int size, ret = 0;
149 	u32 off;
150 
151 	while (1) {
152 		bool has_tx_ulp;
153 
154 		sge = sk_msg_elem(msg, msg->sg.start);
155 		size = (apply && apply_bytes < sge->length) ?
156 			apply_bytes : sge->length;
157 		off  = sge->offset;
158 		page = sg_page(sge);
159 
160 		tcp_rate_check_app_limited(sk);
161 retry:
162 		has_tx_ulp = tls_sw_has_ctx_tx(sk);
163 		if (has_tx_ulp) {
164 			flags |= MSG_SENDPAGE_NOPOLICY;
165 			ret = kernel_sendpage_locked(sk,
166 						     page, off, size, flags);
167 		} else {
168 			ret = do_tcp_sendpages(sk, page, off, size, flags);
169 		}
170 
171 		if (ret <= 0)
172 			return ret;
173 		if (apply)
174 			apply_bytes -= ret;
175 		msg->sg.size -= ret;
176 		sge->offset += ret;
177 		sge->length -= ret;
178 		if (uncharge)
179 			sk_mem_uncharge(sk, ret);
180 		if (ret != size) {
181 			size -= ret;
182 			off  += ret;
183 			goto retry;
184 		}
185 		if (!sge->length) {
186 			put_page(page);
187 			sk_msg_iter_next(msg, start);
188 			sg_init_table(sge, 1);
189 			if (msg->sg.start == msg->sg.end)
190 				break;
191 		}
192 		if (apply && !apply_bytes)
193 			break;
194 	}
195 
196 	return 0;
197 }
198 
199 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
200 			       u32 apply_bytes, int flags, bool uncharge)
201 {
202 	int ret;
203 
204 	lock_sock(sk);
205 	ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
206 	release_sock(sk);
207 	return ret;
208 }
209 
210 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
211 			  u32 bytes, int flags)
212 {
213 	bool ingress = sk_msg_to_ingress(msg);
214 	struct sk_psock *psock = sk_psock_get(sk);
215 	int ret;
216 
217 	if (unlikely(!psock)) {
218 		sk_msg_free(sk, msg);
219 		return 0;
220 	}
221 	ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
222 			tcp_bpf_push_locked(sk, msg, bytes, flags, false);
223 	sk_psock_put(sk, psock);
224 	return ret;
225 }
226 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
227 
228 #ifdef CONFIG_BPF_STREAM_PARSER
229 static bool tcp_bpf_stream_read(const struct sock *sk)
230 {
231 	struct sk_psock *psock;
232 	bool empty = true;
233 
234 	rcu_read_lock();
235 	psock = sk_psock(sk);
236 	if (likely(psock))
237 		empty = list_empty(&psock->ingress_msg);
238 	rcu_read_unlock();
239 	return !empty;
240 }
241 
242 static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock,
243 			     int flags, long timeo, int *err)
244 {
245 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
246 	int ret = 0;
247 
248 	if (sk->sk_shutdown & RCV_SHUTDOWN)
249 		return 1;
250 
251 	if (!timeo)
252 		return ret;
253 
254 	add_wait_queue(sk_sleep(sk), &wait);
255 	sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
256 	ret = sk_wait_event(sk, &timeo,
257 			    !list_empty(&psock->ingress_msg) ||
258 			    !skb_queue_empty(&sk->sk_receive_queue), &wait);
259 	sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
260 	remove_wait_queue(sk_sleep(sk), &wait);
261 	return ret;
262 }
263 
264 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
265 		    int nonblock, int flags, int *addr_len)
266 {
267 	struct sk_psock *psock;
268 	int copied, ret;
269 
270 	if (unlikely(flags & MSG_ERRQUEUE))
271 		return inet_recv_error(sk, msg, len, addr_len);
272 
273 	psock = sk_psock_get(sk);
274 	if (unlikely(!psock))
275 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
276 	if (!skb_queue_empty(&sk->sk_receive_queue) &&
277 	    sk_psock_queue_empty(psock)) {
278 		sk_psock_put(sk, psock);
279 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
280 	}
281 	lock_sock(sk);
282 msg_bytes_ready:
283 	copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags);
284 	if (!copied) {
285 		int data, err = 0;
286 		long timeo;
287 
288 		timeo = sock_rcvtimeo(sk, nonblock);
289 		data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err);
290 		if (data) {
291 			if (!sk_psock_queue_empty(psock))
292 				goto msg_bytes_ready;
293 			release_sock(sk);
294 			sk_psock_put(sk, psock);
295 			return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
296 		}
297 		if (err) {
298 			ret = err;
299 			goto out;
300 		}
301 		copied = -EAGAIN;
302 	}
303 	ret = copied;
304 out:
305 	release_sock(sk);
306 	sk_psock_put(sk, psock);
307 	return ret;
308 }
309 
310 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
311 				struct sk_msg *msg, int *copied, int flags)
312 {
313 	bool cork = false, enospc = sk_msg_full(msg);
314 	struct sock *sk_redir;
315 	u32 tosend, delta = 0;
316 	int ret;
317 
318 more_data:
319 	if (psock->eval == __SK_NONE) {
320 		/* Track delta in msg size to add/subtract it on SK_DROP from
321 		 * returned to user copied size. This ensures user doesn't
322 		 * get a positive return code with msg_cut_data and SK_DROP
323 		 * verdict.
324 		 */
325 		delta = msg->sg.size;
326 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
327 		delta -= msg->sg.size;
328 	}
329 
330 	if (msg->cork_bytes &&
331 	    msg->cork_bytes > msg->sg.size && !enospc) {
332 		psock->cork_bytes = msg->cork_bytes - msg->sg.size;
333 		if (!psock->cork) {
334 			psock->cork = kzalloc(sizeof(*psock->cork),
335 					      GFP_ATOMIC | __GFP_NOWARN);
336 			if (!psock->cork)
337 				return -ENOMEM;
338 		}
339 		memcpy(psock->cork, msg, sizeof(*msg));
340 		return 0;
341 	}
342 
343 	tosend = msg->sg.size;
344 	if (psock->apply_bytes && psock->apply_bytes < tosend)
345 		tosend = psock->apply_bytes;
346 
347 	switch (psock->eval) {
348 	case __SK_PASS:
349 		ret = tcp_bpf_push(sk, msg, tosend, flags, true);
350 		if (unlikely(ret)) {
351 			*copied -= sk_msg_free(sk, msg);
352 			break;
353 		}
354 		sk_msg_apply_bytes(psock, tosend);
355 		break;
356 	case __SK_REDIRECT:
357 		sk_redir = psock->sk_redir;
358 		sk_msg_apply_bytes(psock, tosend);
359 		if (psock->cork) {
360 			cork = true;
361 			psock->cork = NULL;
362 		}
363 		sk_msg_return(sk, msg, tosend);
364 		release_sock(sk);
365 		ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
366 		lock_sock(sk);
367 		if (unlikely(ret < 0)) {
368 			int free = sk_msg_free_nocharge(sk, msg);
369 
370 			if (!cork)
371 				*copied -= free;
372 		}
373 		if (cork) {
374 			sk_msg_free(sk, msg);
375 			kfree(msg);
376 			msg = NULL;
377 			ret = 0;
378 		}
379 		break;
380 	case __SK_DROP:
381 	default:
382 		sk_msg_free_partial(sk, msg, tosend);
383 		sk_msg_apply_bytes(psock, tosend);
384 		*copied -= (tosend + delta);
385 		return -EACCES;
386 	}
387 
388 	if (likely(!ret)) {
389 		if (!psock->apply_bytes) {
390 			psock->eval =  __SK_NONE;
391 			if (psock->sk_redir) {
392 				sock_put(psock->sk_redir);
393 				psock->sk_redir = NULL;
394 			}
395 		}
396 		if (msg &&
397 		    msg->sg.data[msg->sg.start].page_link &&
398 		    msg->sg.data[msg->sg.start].length)
399 			goto more_data;
400 	}
401 	return ret;
402 }
403 
404 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
405 {
406 	struct sk_msg tmp, *msg_tx = NULL;
407 	int copied = 0, err = 0;
408 	struct sk_psock *psock;
409 	long timeo;
410 	int flags;
411 
412 	/* Don't let internal do_tcp_sendpages() flags through */
413 	flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
414 	flags |= MSG_NO_SHARED_FRAGS;
415 
416 	psock = sk_psock_get(sk);
417 	if (unlikely(!psock))
418 		return tcp_sendmsg(sk, msg, size);
419 
420 	lock_sock(sk);
421 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
422 	while (msg_data_left(msg)) {
423 		bool enospc = false;
424 		u32 copy, osize;
425 
426 		if (sk->sk_err) {
427 			err = -sk->sk_err;
428 			goto out_err;
429 		}
430 
431 		copy = msg_data_left(msg);
432 		if (!sk_stream_memory_free(sk))
433 			goto wait_for_sndbuf;
434 		if (psock->cork) {
435 			msg_tx = psock->cork;
436 		} else {
437 			msg_tx = &tmp;
438 			sk_msg_init(msg_tx);
439 		}
440 
441 		osize = msg_tx->sg.size;
442 		err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
443 		if (err) {
444 			if (err != -ENOSPC)
445 				goto wait_for_memory;
446 			enospc = true;
447 			copy = msg_tx->sg.size - osize;
448 		}
449 
450 		err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
451 					       copy);
452 		if (err < 0) {
453 			sk_msg_trim(sk, msg_tx, osize);
454 			goto out_err;
455 		}
456 
457 		copied += copy;
458 		if (psock->cork_bytes) {
459 			if (size > psock->cork_bytes)
460 				psock->cork_bytes = 0;
461 			else
462 				psock->cork_bytes -= size;
463 			if (psock->cork_bytes && !enospc)
464 				goto out_err;
465 			/* All cork bytes are accounted, rerun the prog. */
466 			psock->eval = __SK_NONE;
467 			psock->cork_bytes = 0;
468 		}
469 
470 		err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
471 		if (unlikely(err < 0))
472 			goto out_err;
473 		continue;
474 wait_for_sndbuf:
475 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
476 wait_for_memory:
477 		err = sk_stream_wait_memory(sk, &timeo);
478 		if (err) {
479 			if (msg_tx && msg_tx != psock->cork)
480 				sk_msg_free(sk, msg_tx);
481 			goto out_err;
482 		}
483 	}
484 out_err:
485 	if (err < 0)
486 		err = sk_stream_error(sk, msg->msg_flags, err);
487 	release_sock(sk);
488 	sk_psock_put(sk, psock);
489 	return copied ? copied : err;
490 }
491 
492 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
493 			    size_t size, int flags)
494 {
495 	struct sk_msg tmp, *msg = NULL;
496 	int err = 0, copied = 0;
497 	struct sk_psock *psock;
498 	bool enospc = false;
499 
500 	psock = sk_psock_get(sk);
501 	if (unlikely(!psock))
502 		return tcp_sendpage(sk, page, offset, size, flags);
503 
504 	lock_sock(sk);
505 	if (psock->cork) {
506 		msg = psock->cork;
507 	} else {
508 		msg = &tmp;
509 		sk_msg_init(msg);
510 	}
511 
512 	/* Catch case where ring is full and sendpage is stalled. */
513 	if (unlikely(sk_msg_full(msg)))
514 		goto out_err;
515 
516 	sk_msg_page_add(msg, page, size, offset);
517 	sk_mem_charge(sk, size);
518 	copied = size;
519 	if (sk_msg_full(msg))
520 		enospc = true;
521 	if (psock->cork_bytes) {
522 		if (size > psock->cork_bytes)
523 			psock->cork_bytes = 0;
524 		else
525 			psock->cork_bytes -= size;
526 		if (psock->cork_bytes && !enospc)
527 			goto out_err;
528 		/* All cork bytes are accounted, rerun the prog. */
529 		psock->eval = __SK_NONE;
530 		psock->cork_bytes = 0;
531 	}
532 
533 	err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
534 out_err:
535 	release_sock(sk);
536 	sk_psock_put(sk, psock);
537 	return copied ? copied : err;
538 }
539 
540 enum {
541 	TCP_BPF_IPV4,
542 	TCP_BPF_IPV6,
543 	TCP_BPF_NUM_PROTS,
544 };
545 
546 enum {
547 	TCP_BPF_BASE,
548 	TCP_BPF_TX,
549 	TCP_BPF_NUM_CFGS,
550 };
551 
552 static struct proto *tcpv6_prot_saved __read_mostly;
553 static DEFINE_SPINLOCK(tcpv6_prot_lock);
554 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
555 
556 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
557 				   struct proto *base)
558 {
559 	prot[TCP_BPF_BASE]			= *base;
560 	prot[TCP_BPF_BASE].unhash		= sock_map_unhash;
561 	prot[TCP_BPF_BASE].close		= sock_map_close;
562 	prot[TCP_BPF_BASE].recvmsg		= tcp_bpf_recvmsg;
563 	prot[TCP_BPF_BASE].stream_memory_read	= tcp_bpf_stream_read;
564 
565 	prot[TCP_BPF_TX]			= prot[TCP_BPF_BASE];
566 	prot[TCP_BPF_TX].sendmsg		= tcp_bpf_sendmsg;
567 	prot[TCP_BPF_TX].sendpage		= tcp_bpf_sendpage;
568 }
569 
570 static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops)
571 {
572 	if (sk->sk_family == AF_INET6 &&
573 	    unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
574 		spin_lock_bh(&tcpv6_prot_lock);
575 		if (likely(ops != tcpv6_prot_saved)) {
576 			tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
577 			smp_store_release(&tcpv6_prot_saved, ops);
578 		}
579 		spin_unlock_bh(&tcpv6_prot_lock);
580 	}
581 }
582 
583 static int __init tcp_bpf_v4_build_proto(void)
584 {
585 	tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
586 	return 0;
587 }
588 core_initcall(tcp_bpf_v4_build_proto);
589 
590 static int tcp_bpf_assert_proto_ops(struct proto *ops)
591 {
592 	/* In order to avoid retpoline, we make assumptions when we call
593 	 * into ops if e.g. a psock is not present. Make sure they are
594 	 * indeed valid assumptions.
595 	 */
596 	return ops->recvmsg  == tcp_recvmsg &&
597 	       ops->sendmsg  == tcp_sendmsg &&
598 	       ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
599 }
600 
601 struct proto *tcp_bpf_get_proto(struct sock *sk, struct sk_psock *psock)
602 {
603 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
604 	int config = psock->progs.msg_parser   ? TCP_BPF_TX   : TCP_BPF_BASE;
605 
606 	if (!psock->sk_proto) {
607 		struct proto *ops = READ_ONCE(sk->sk_prot);
608 
609 		if (tcp_bpf_assert_proto_ops(ops))
610 			return ERR_PTR(-EINVAL);
611 
612 		tcp_bpf_check_v6_needs_rebuild(sk, ops);
613 	}
614 
615 	return &tcp_bpf_prots[family][config];
616 }
617 
618 /* If a child got cloned from a listening socket that had tcp_bpf
619  * protocol callbacks installed, we need to restore the callbacks to
620  * the default ones because the child does not inherit the psock state
621  * that tcp_bpf callbacks expect.
622  */
623 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
624 {
625 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
626 	struct proto *prot = newsk->sk_prot;
627 
628 	if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
629 		newsk->sk_prot = sk->sk_prot_creator;
630 }
631 #endif /* CONFIG_BPF_STREAM_PARSER */
632