xref: /linux/net/ipv4/tcp_bpf.c (revision 3f0a50f345f78183f6e9b39c2f45ca5dcaa511ca)
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 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
14 			   struct sk_msg *msg, u32 apply_bytes, int flags)
15 {
16 	bool apply = apply_bytes;
17 	struct scatterlist *sge;
18 	u32 size, copied = 0;
19 	struct sk_msg *tmp;
20 	int i, ret = 0;
21 
22 	tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
23 	if (unlikely(!tmp))
24 		return -ENOMEM;
25 
26 	lock_sock(sk);
27 	tmp->sg.start = msg->sg.start;
28 	i = msg->sg.start;
29 	do {
30 		sge = sk_msg_elem(msg, i);
31 		size = (apply && apply_bytes < sge->length) ?
32 			apply_bytes : sge->length;
33 		if (!sk_wmem_schedule(sk, size)) {
34 			if (!copied)
35 				ret = -ENOMEM;
36 			break;
37 		}
38 
39 		sk_mem_charge(sk, size);
40 		sk_msg_xfer(tmp, msg, i, size);
41 		copied += size;
42 		if (sge->length)
43 			get_page(sk_msg_page(tmp, i));
44 		sk_msg_iter_var_next(i);
45 		tmp->sg.end = i;
46 		if (apply) {
47 			apply_bytes -= size;
48 			if (!apply_bytes)
49 				break;
50 		}
51 	} while (i != msg->sg.end);
52 
53 	if (!ret) {
54 		msg->sg.start = i;
55 		sk_psock_queue_msg(psock, tmp);
56 		sk_psock_data_ready(sk, psock);
57 	} else {
58 		sk_msg_free(sk, tmp);
59 		kfree(tmp);
60 	}
61 
62 	release_sock(sk);
63 	return ret;
64 }
65 
66 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
67 			int flags, bool uncharge)
68 {
69 	bool apply = apply_bytes;
70 	struct scatterlist *sge;
71 	struct page *page;
72 	int size, ret = 0;
73 	u32 off;
74 
75 	while (1) {
76 		bool has_tx_ulp;
77 
78 		sge = sk_msg_elem(msg, msg->sg.start);
79 		size = (apply && apply_bytes < sge->length) ?
80 			apply_bytes : sge->length;
81 		off  = sge->offset;
82 		page = sg_page(sge);
83 
84 		tcp_rate_check_app_limited(sk);
85 retry:
86 		has_tx_ulp = tls_sw_has_ctx_tx(sk);
87 		if (has_tx_ulp) {
88 			flags |= MSG_SENDPAGE_NOPOLICY;
89 			ret = kernel_sendpage_locked(sk,
90 						     page, off, size, flags);
91 		} else {
92 			ret = do_tcp_sendpages(sk, page, off, size, flags);
93 		}
94 
95 		if (ret <= 0)
96 			return ret;
97 		if (apply)
98 			apply_bytes -= ret;
99 		msg->sg.size -= ret;
100 		sge->offset += ret;
101 		sge->length -= ret;
102 		if (uncharge)
103 			sk_mem_uncharge(sk, ret);
104 		if (ret != size) {
105 			size -= ret;
106 			off  += ret;
107 			goto retry;
108 		}
109 		if (!sge->length) {
110 			put_page(page);
111 			sk_msg_iter_next(msg, start);
112 			sg_init_table(sge, 1);
113 			if (msg->sg.start == msg->sg.end)
114 				break;
115 		}
116 		if (apply && !apply_bytes)
117 			break;
118 	}
119 
120 	return 0;
121 }
122 
123 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
124 			       u32 apply_bytes, int flags, bool uncharge)
125 {
126 	int ret;
127 
128 	lock_sock(sk);
129 	ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
130 	release_sock(sk);
131 	return ret;
132 }
133 
134 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
135 			  u32 bytes, int flags)
136 {
137 	bool ingress = sk_msg_to_ingress(msg);
138 	struct sk_psock *psock = sk_psock_get(sk);
139 	int ret;
140 
141 	if (unlikely(!psock))
142 		return -EPIPE;
143 
144 	ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
145 			tcp_bpf_push_locked(sk, msg, bytes, flags, false);
146 	sk_psock_put(sk, psock);
147 	return ret;
148 }
149 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
150 
151 #ifdef CONFIG_BPF_SYSCALL
152 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
153 			     long timeo)
154 {
155 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
156 	int ret = 0;
157 
158 	if (sk->sk_shutdown & RCV_SHUTDOWN)
159 		return 1;
160 
161 	if (!timeo)
162 		return ret;
163 
164 	add_wait_queue(sk_sleep(sk), &wait);
165 	sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
166 	ret = sk_wait_event(sk, &timeo,
167 			    !list_empty(&psock->ingress_msg) ||
168 			    !skb_queue_empty(&sk->sk_receive_queue), &wait);
169 	sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
170 	remove_wait_queue(sk_sleep(sk), &wait);
171 	return ret;
172 }
173 
174 static int tcp_bpf_recvmsg_parser(struct sock *sk,
175 				  struct msghdr *msg,
176 				  size_t len,
177 				  int nonblock,
178 				  int flags,
179 				  int *addr_len)
180 {
181 	struct sk_psock *psock;
182 	int copied;
183 
184 	if (unlikely(flags & MSG_ERRQUEUE))
185 		return inet_recv_error(sk, msg, len, addr_len);
186 
187 	psock = sk_psock_get(sk);
188 	if (unlikely(!psock))
189 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
190 
191 	lock_sock(sk);
192 msg_bytes_ready:
193 	copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
194 	if (!copied) {
195 		long timeo;
196 		int data;
197 
198 		if (sock_flag(sk, SOCK_DONE))
199 			goto out;
200 
201 		if (sk->sk_err) {
202 			copied = sock_error(sk);
203 			goto out;
204 		}
205 
206 		if (sk->sk_shutdown & RCV_SHUTDOWN)
207 			goto out;
208 
209 		if (sk->sk_state == TCP_CLOSE) {
210 			copied = -ENOTCONN;
211 			goto out;
212 		}
213 
214 		timeo = sock_rcvtimeo(sk, nonblock);
215 		if (!timeo) {
216 			copied = -EAGAIN;
217 			goto out;
218 		}
219 
220 		if (signal_pending(current)) {
221 			copied = sock_intr_errno(timeo);
222 			goto out;
223 		}
224 
225 		data = tcp_msg_wait_data(sk, psock, timeo);
226 		if (data && !sk_psock_queue_empty(psock))
227 			goto msg_bytes_ready;
228 		copied = -EAGAIN;
229 	}
230 out:
231 	release_sock(sk);
232 	sk_psock_put(sk, psock);
233 	return copied;
234 }
235 
236 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
237 		    int nonblock, int flags, int *addr_len)
238 {
239 	struct sk_psock *psock;
240 	int copied, ret;
241 
242 	if (unlikely(flags & MSG_ERRQUEUE))
243 		return inet_recv_error(sk, msg, len, addr_len);
244 
245 	psock = sk_psock_get(sk);
246 	if (unlikely(!psock))
247 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
248 	if (!skb_queue_empty(&sk->sk_receive_queue) &&
249 	    sk_psock_queue_empty(psock)) {
250 		sk_psock_put(sk, psock);
251 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
252 	}
253 	lock_sock(sk);
254 msg_bytes_ready:
255 	copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
256 	if (!copied) {
257 		long timeo;
258 		int data;
259 
260 		timeo = sock_rcvtimeo(sk, nonblock);
261 		data = tcp_msg_wait_data(sk, psock, timeo);
262 		if (data) {
263 			if (!sk_psock_queue_empty(psock))
264 				goto msg_bytes_ready;
265 			release_sock(sk);
266 			sk_psock_put(sk, psock);
267 			return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
268 		}
269 		copied = -EAGAIN;
270 	}
271 	ret = copied;
272 	release_sock(sk);
273 	sk_psock_put(sk, psock);
274 	return ret;
275 }
276 
277 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
278 				struct sk_msg *msg, int *copied, int flags)
279 {
280 	bool cork = false, enospc = sk_msg_full(msg);
281 	struct sock *sk_redir;
282 	u32 tosend, delta = 0;
283 	u32 eval = __SK_NONE;
284 	int ret;
285 
286 more_data:
287 	if (psock->eval == __SK_NONE) {
288 		/* Track delta in msg size to add/subtract it on SK_DROP from
289 		 * returned to user copied size. This ensures user doesn't
290 		 * get a positive return code with msg_cut_data and SK_DROP
291 		 * verdict.
292 		 */
293 		delta = msg->sg.size;
294 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
295 		delta -= msg->sg.size;
296 	}
297 
298 	if (msg->cork_bytes &&
299 	    msg->cork_bytes > msg->sg.size && !enospc) {
300 		psock->cork_bytes = msg->cork_bytes - msg->sg.size;
301 		if (!psock->cork) {
302 			psock->cork = kzalloc(sizeof(*psock->cork),
303 					      GFP_ATOMIC | __GFP_NOWARN);
304 			if (!psock->cork)
305 				return -ENOMEM;
306 		}
307 		memcpy(psock->cork, msg, sizeof(*msg));
308 		return 0;
309 	}
310 
311 	tosend = msg->sg.size;
312 	if (psock->apply_bytes && psock->apply_bytes < tosend)
313 		tosend = psock->apply_bytes;
314 
315 	switch (psock->eval) {
316 	case __SK_PASS:
317 		ret = tcp_bpf_push(sk, msg, tosend, flags, true);
318 		if (unlikely(ret)) {
319 			*copied -= sk_msg_free(sk, msg);
320 			break;
321 		}
322 		sk_msg_apply_bytes(psock, tosend);
323 		break;
324 	case __SK_REDIRECT:
325 		sk_redir = psock->sk_redir;
326 		sk_msg_apply_bytes(psock, tosend);
327 		if (!psock->apply_bytes) {
328 			/* Clean up before releasing the sock lock. */
329 			eval = psock->eval;
330 			psock->eval = __SK_NONE;
331 			psock->sk_redir = NULL;
332 		}
333 		if (psock->cork) {
334 			cork = true;
335 			psock->cork = NULL;
336 		}
337 		sk_msg_return(sk, msg, msg->sg.size);
338 		release_sock(sk);
339 
340 		ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
341 
342 		if (eval == __SK_REDIRECT)
343 			sock_put(sk_redir);
344 
345 		lock_sock(sk);
346 		if (unlikely(ret < 0)) {
347 			int free = sk_msg_free_nocharge(sk, msg);
348 
349 			if (!cork)
350 				*copied -= free;
351 		}
352 		if (cork) {
353 			sk_msg_free(sk, msg);
354 			kfree(msg);
355 			msg = NULL;
356 			ret = 0;
357 		}
358 		break;
359 	case __SK_DROP:
360 	default:
361 		sk_msg_free_partial(sk, msg, tosend);
362 		sk_msg_apply_bytes(psock, tosend);
363 		*copied -= (tosend + delta);
364 		return -EACCES;
365 	}
366 
367 	if (likely(!ret)) {
368 		if (!psock->apply_bytes) {
369 			psock->eval =  __SK_NONE;
370 			if (psock->sk_redir) {
371 				sock_put(psock->sk_redir);
372 				psock->sk_redir = NULL;
373 			}
374 		}
375 		if (msg &&
376 		    msg->sg.data[msg->sg.start].page_link &&
377 		    msg->sg.data[msg->sg.start].length) {
378 			if (eval == __SK_REDIRECT)
379 				sk_mem_charge(sk, msg->sg.size);
380 			goto more_data;
381 		}
382 	}
383 	return ret;
384 }
385 
386 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
387 {
388 	struct sk_msg tmp, *msg_tx = NULL;
389 	int copied = 0, err = 0;
390 	struct sk_psock *psock;
391 	long timeo;
392 	int flags;
393 
394 	/* Don't let internal do_tcp_sendpages() flags through */
395 	flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
396 	flags |= MSG_NO_SHARED_FRAGS;
397 
398 	psock = sk_psock_get(sk);
399 	if (unlikely(!psock))
400 		return tcp_sendmsg(sk, msg, size);
401 
402 	lock_sock(sk);
403 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
404 	while (msg_data_left(msg)) {
405 		bool enospc = false;
406 		u32 copy, osize;
407 
408 		if (sk->sk_err) {
409 			err = -sk->sk_err;
410 			goto out_err;
411 		}
412 
413 		copy = msg_data_left(msg);
414 		if (!sk_stream_memory_free(sk))
415 			goto wait_for_sndbuf;
416 		if (psock->cork) {
417 			msg_tx = psock->cork;
418 		} else {
419 			msg_tx = &tmp;
420 			sk_msg_init(msg_tx);
421 		}
422 
423 		osize = msg_tx->sg.size;
424 		err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
425 		if (err) {
426 			if (err != -ENOSPC)
427 				goto wait_for_memory;
428 			enospc = true;
429 			copy = msg_tx->sg.size - osize;
430 		}
431 
432 		err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
433 					       copy);
434 		if (err < 0) {
435 			sk_msg_trim(sk, msg_tx, osize);
436 			goto out_err;
437 		}
438 
439 		copied += copy;
440 		if (psock->cork_bytes) {
441 			if (size > psock->cork_bytes)
442 				psock->cork_bytes = 0;
443 			else
444 				psock->cork_bytes -= size;
445 			if (psock->cork_bytes && !enospc)
446 				goto out_err;
447 			/* All cork bytes are accounted, rerun the prog. */
448 			psock->eval = __SK_NONE;
449 			psock->cork_bytes = 0;
450 		}
451 
452 		err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
453 		if (unlikely(err < 0))
454 			goto out_err;
455 		continue;
456 wait_for_sndbuf:
457 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
458 wait_for_memory:
459 		err = sk_stream_wait_memory(sk, &timeo);
460 		if (err) {
461 			if (msg_tx && msg_tx != psock->cork)
462 				sk_msg_free(sk, msg_tx);
463 			goto out_err;
464 		}
465 	}
466 out_err:
467 	if (err < 0)
468 		err = sk_stream_error(sk, msg->msg_flags, err);
469 	release_sock(sk);
470 	sk_psock_put(sk, psock);
471 	return copied ? copied : err;
472 }
473 
474 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
475 			    size_t size, int flags)
476 {
477 	struct sk_msg tmp, *msg = NULL;
478 	int err = 0, copied = 0;
479 	struct sk_psock *psock;
480 	bool enospc = false;
481 
482 	psock = sk_psock_get(sk);
483 	if (unlikely(!psock))
484 		return tcp_sendpage(sk, page, offset, size, flags);
485 
486 	lock_sock(sk);
487 	if (psock->cork) {
488 		msg = psock->cork;
489 	} else {
490 		msg = &tmp;
491 		sk_msg_init(msg);
492 	}
493 
494 	/* Catch case where ring is full and sendpage is stalled. */
495 	if (unlikely(sk_msg_full(msg)))
496 		goto out_err;
497 
498 	sk_msg_page_add(msg, page, size, offset);
499 	sk_mem_charge(sk, size);
500 	copied = size;
501 	if (sk_msg_full(msg))
502 		enospc = true;
503 	if (psock->cork_bytes) {
504 		if (size > psock->cork_bytes)
505 			psock->cork_bytes = 0;
506 		else
507 			psock->cork_bytes -= size;
508 		if (psock->cork_bytes && !enospc)
509 			goto out_err;
510 		/* All cork bytes are accounted, rerun the prog. */
511 		psock->eval = __SK_NONE;
512 		psock->cork_bytes = 0;
513 	}
514 
515 	err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
516 out_err:
517 	release_sock(sk);
518 	sk_psock_put(sk, psock);
519 	return copied ? copied : err;
520 }
521 
522 enum {
523 	TCP_BPF_IPV4,
524 	TCP_BPF_IPV6,
525 	TCP_BPF_NUM_PROTS,
526 };
527 
528 enum {
529 	TCP_BPF_BASE,
530 	TCP_BPF_TX,
531 	TCP_BPF_RX,
532 	TCP_BPF_TXRX,
533 	TCP_BPF_NUM_CFGS,
534 };
535 
536 static struct proto *tcpv6_prot_saved __read_mostly;
537 static DEFINE_SPINLOCK(tcpv6_prot_lock);
538 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
539 
540 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
541 				   struct proto *base)
542 {
543 	prot[TCP_BPF_BASE]			= *base;
544 	prot[TCP_BPF_BASE].close		= sock_map_close;
545 	prot[TCP_BPF_BASE].recvmsg		= tcp_bpf_recvmsg;
546 	prot[TCP_BPF_BASE].sock_is_readable	= sk_msg_is_readable;
547 
548 	prot[TCP_BPF_TX]			= prot[TCP_BPF_BASE];
549 	prot[TCP_BPF_TX].sendmsg		= tcp_bpf_sendmsg;
550 	prot[TCP_BPF_TX].sendpage		= tcp_bpf_sendpage;
551 
552 	prot[TCP_BPF_RX]			= prot[TCP_BPF_BASE];
553 	prot[TCP_BPF_RX].recvmsg		= tcp_bpf_recvmsg_parser;
554 
555 	prot[TCP_BPF_TXRX]			= prot[TCP_BPF_TX];
556 	prot[TCP_BPF_TXRX].recvmsg		= tcp_bpf_recvmsg_parser;
557 }
558 
559 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
560 {
561 	if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
562 		spin_lock_bh(&tcpv6_prot_lock);
563 		if (likely(ops != tcpv6_prot_saved)) {
564 			tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
565 			smp_store_release(&tcpv6_prot_saved, ops);
566 		}
567 		spin_unlock_bh(&tcpv6_prot_lock);
568 	}
569 }
570 
571 static int __init tcp_bpf_v4_build_proto(void)
572 {
573 	tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
574 	return 0;
575 }
576 late_initcall(tcp_bpf_v4_build_proto);
577 
578 static int tcp_bpf_assert_proto_ops(struct proto *ops)
579 {
580 	/* In order to avoid retpoline, we make assumptions when we call
581 	 * into ops if e.g. a psock is not present. Make sure they are
582 	 * indeed valid assumptions.
583 	 */
584 	return ops->recvmsg  == tcp_recvmsg &&
585 	       ops->sendmsg  == tcp_sendmsg &&
586 	       ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
587 }
588 
589 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
590 {
591 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
592 	int config = psock->progs.msg_parser   ? TCP_BPF_TX   : TCP_BPF_BASE;
593 
594 	if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
595 		config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
596 	}
597 
598 	if (restore) {
599 		if (inet_csk_has_ulp(sk)) {
600 			/* TLS does not have an unhash proto in SW cases,
601 			 * but we need to ensure we stop using the sock_map
602 			 * unhash routine because the associated psock is being
603 			 * removed. So use the original unhash handler.
604 			 */
605 			WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
606 			tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
607 		} else {
608 			sk->sk_write_space = psock->saved_write_space;
609 			/* Pairs with lockless read in sk_clone_lock() */
610 			WRITE_ONCE(sk->sk_prot, psock->sk_proto);
611 		}
612 		return 0;
613 	}
614 
615 	if (inet_csk_has_ulp(sk))
616 		return -EINVAL;
617 
618 	if (sk->sk_family == AF_INET6) {
619 		if (tcp_bpf_assert_proto_ops(psock->sk_proto))
620 			return -EINVAL;
621 
622 		tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
623 	}
624 
625 	/* Pairs with lockless read in sk_clone_lock() */
626 	WRITE_ONCE(sk->sk_prot, &tcp_bpf_prots[family][config]);
627 	return 0;
628 }
629 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
630 
631 /* If a child got cloned from a listening socket that had tcp_bpf
632  * protocol callbacks installed, we need to restore the callbacks to
633  * the default ones because the child does not inherit the psock state
634  * that tcp_bpf callbacks expect.
635  */
636 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
637 {
638 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
639 	struct proto *prot = newsk->sk_prot;
640 
641 	if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
642 		newsk->sk_prot = sk->sk_prot_creator;
643 }
644 #endif /* CONFIG_BPF_SYSCALL */
645