xref: /openbmc/linux/net/ipv4/tcp_bpf.c (revision a3b91d8b)
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 			msg_rx = list_next_entry(msg_rx, list);
68 			continue;
69 		}
70 
71 		msg_rx->sg.start = i;
72 		if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
73 			list_del(&msg_rx->list);
74 			if (msg_rx->skb)
75 				consume_skb(msg_rx->skb);
76 			kfree(msg_rx);
77 		}
78 		msg_rx = list_first_entry_or_null(&psock->ingress_msg,
79 						  struct sk_msg, list);
80 	}
81 
82 	return copied;
83 }
84 EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg);
85 
86 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
87 			   struct sk_msg *msg, u32 apply_bytes, int flags)
88 {
89 	bool apply = apply_bytes;
90 	struct scatterlist *sge;
91 	u32 size, copied = 0;
92 	struct sk_msg *tmp;
93 	int i, ret = 0;
94 
95 	tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
96 	if (unlikely(!tmp))
97 		return -ENOMEM;
98 
99 	lock_sock(sk);
100 	tmp->sg.start = msg->sg.start;
101 	i = msg->sg.start;
102 	do {
103 		sge = sk_msg_elem(msg, i);
104 		size = (apply && apply_bytes < sge->length) ?
105 			apply_bytes : sge->length;
106 		if (!sk_wmem_schedule(sk, size)) {
107 			if (!copied)
108 				ret = -ENOMEM;
109 			break;
110 		}
111 
112 		sk_mem_charge(sk, size);
113 		sk_msg_xfer(tmp, msg, i, size);
114 		copied += size;
115 		if (sge->length)
116 			get_page(sk_msg_page(tmp, i));
117 		sk_msg_iter_var_next(i);
118 		tmp->sg.end = i;
119 		if (apply) {
120 			apply_bytes -= size;
121 			if (!apply_bytes)
122 				break;
123 		}
124 	} while (i != msg->sg.end);
125 
126 	if (!ret) {
127 		msg->sg.start = i;
128 		msg->sg.size -= apply_bytes;
129 		sk_psock_queue_msg(psock, tmp);
130 		sk_psock_data_ready(sk, psock);
131 	} else {
132 		sk_msg_free(sk, tmp);
133 		kfree(tmp);
134 	}
135 
136 	release_sock(sk);
137 	return ret;
138 }
139 
140 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
141 			int flags, bool uncharge)
142 {
143 	bool apply = apply_bytes;
144 	struct scatterlist *sge;
145 	struct page *page;
146 	int size, ret = 0;
147 	u32 off;
148 
149 	while (1) {
150 		bool has_tx_ulp;
151 
152 		sge = sk_msg_elem(msg, msg->sg.start);
153 		size = (apply && apply_bytes < sge->length) ?
154 			apply_bytes : sge->length;
155 		off  = sge->offset;
156 		page = sg_page(sge);
157 
158 		tcp_rate_check_app_limited(sk);
159 retry:
160 		has_tx_ulp = tls_sw_has_ctx_tx(sk);
161 		if (has_tx_ulp) {
162 			flags |= MSG_SENDPAGE_NOPOLICY;
163 			ret = kernel_sendpage_locked(sk,
164 						     page, off, size, flags);
165 		} else {
166 			ret = do_tcp_sendpages(sk, page, off, size, flags);
167 		}
168 
169 		if (ret <= 0)
170 			return ret;
171 		if (apply)
172 			apply_bytes -= ret;
173 		msg->sg.size -= ret;
174 		sge->offset += ret;
175 		sge->length -= ret;
176 		if (uncharge)
177 			sk_mem_uncharge(sk, ret);
178 		if (ret != size) {
179 			size -= ret;
180 			off  += ret;
181 			goto retry;
182 		}
183 		if (!sge->length) {
184 			put_page(page);
185 			sk_msg_iter_next(msg, start);
186 			sg_init_table(sge, 1);
187 			if (msg->sg.start == msg->sg.end)
188 				break;
189 		}
190 		if (apply && !apply_bytes)
191 			break;
192 	}
193 
194 	return 0;
195 }
196 
197 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
198 			       u32 apply_bytes, int flags, bool uncharge)
199 {
200 	int ret;
201 
202 	lock_sock(sk);
203 	ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
204 	release_sock(sk);
205 	return ret;
206 }
207 
208 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
209 			  u32 bytes, int flags)
210 {
211 	bool ingress = sk_msg_to_ingress(msg);
212 	struct sk_psock *psock = sk_psock_get(sk);
213 	int ret;
214 
215 	if (unlikely(!psock)) {
216 		sk_msg_free(sk, msg);
217 		return 0;
218 	}
219 	ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
220 			tcp_bpf_push_locked(sk, msg, bytes, flags, false);
221 	sk_psock_put(sk, psock);
222 	return ret;
223 }
224 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
225 
226 #ifdef CONFIG_BPF_STREAM_PARSER
227 static bool tcp_bpf_stream_read(const struct sock *sk)
228 {
229 	struct sk_psock *psock;
230 	bool empty = true;
231 
232 	rcu_read_lock();
233 	psock = sk_psock(sk);
234 	if (likely(psock))
235 		empty = list_empty(&psock->ingress_msg);
236 	rcu_read_unlock();
237 	return !empty;
238 }
239 
240 static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock,
241 			     int flags, long timeo, int *err)
242 {
243 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
244 	int ret = 0;
245 
246 	if (!timeo)
247 		return ret;
248 
249 	add_wait_queue(sk_sleep(sk), &wait);
250 	sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
251 	ret = sk_wait_event(sk, &timeo,
252 			    !list_empty(&psock->ingress_msg) ||
253 			    !skb_queue_empty(&sk->sk_receive_queue), &wait);
254 	sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
255 	remove_wait_queue(sk_sleep(sk), &wait);
256 	return ret;
257 }
258 
259 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
260 		    int nonblock, int flags, int *addr_len)
261 {
262 	struct sk_psock *psock;
263 	int copied, ret;
264 
265 	psock = sk_psock_get(sk);
266 	if (unlikely(!psock))
267 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
268 	if (unlikely(flags & MSG_ERRQUEUE))
269 		return inet_recv_error(sk, msg, len, addr_len);
270 	if (!skb_queue_empty(&sk->sk_receive_queue) &&
271 	    sk_psock_queue_empty(psock))
272 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
273 	lock_sock(sk);
274 msg_bytes_ready:
275 	copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags);
276 	if (!copied) {
277 		int data, err = 0;
278 		long timeo;
279 
280 		timeo = sock_rcvtimeo(sk, nonblock);
281 		data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err);
282 		if (data) {
283 			if (!sk_psock_queue_empty(psock))
284 				goto msg_bytes_ready;
285 			release_sock(sk);
286 			sk_psock_put(sk, psock);
287 			return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
288 		}
289 		if (err) {
290 			ret = err;
291 			goto out;
292 		}
293 		copied = -EAGAIN;
294 	}
295 	ret = copied;
296 out:
297 	release_sock(sk);
298 	sk_psock_put(sk, psock);
299 	return ret;
300 }
301 
302 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
303 				struct sk_msg *msg, int *copied, int flags)
304 {
305 	bool cork = false, enospc = sk_msg_full(msg);
306 	struct sock *sk_redir;
307 	u32 tosend, delta = 0;
308 	int ret;
309 
310 more_data:
311 	if (psock->eval == __SK_NONE) {
312 		/* Track delta in msg size to add/subtract it on SK_DROP from
313 		 * returned to user copied size. This ensures user doesn't
314 		 * get a positive return code with msg_cut_data and SK_DROP
315 		 * verdict.
316 		 */
317 		delta = msg->sg.size;
318 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
319 		delta -= msg->sg.size;
320 	}
321 
322 	if (msg->cork_bytes &&
323 	    msg->cork_bytes > msg->sg.size && !enospc) {
324 		psock->cork_bytes = msg->cork_bytes - msg->sg.size;
325 		if (!psock->cork) {
326 			psock->cork = kzalloc(sizeof(*psock->cork),
327 					      GFP_ATOMIC | __GFP_NOWARN);
328 			if (!psock->cork)
329 				return -ENOMEM;
330 		}
331 		memcpy(psock->cork, msg, sizeof(*msg));
332 		return 0;
333 	}
334 
335 	tosend = msg->sg.size;
336 	if (psock->apply_bytes && psock->apply_bytes < tosend)
337 		tosend = psock->apply_bytes;
338 
339 	switch (psock->eval) {
340 	case __SK_PASS:
341 		ret = tcp_bpf_push(sk, msg, tosend, flags, true);
342 		if (unlikely(ret)) {
343 			*copied -= sk_msg_free(sk, msg);
344 			break;
345 		}
346 		sk_msg_apply_bytes(psock, tosend);
347 		break;
348 	case __SK_REDIRECT:
349 		sk_redir = psock->sk_redir;
350 		sk_msg_apply_bytes(psock, tosend);
351 		if (psock->cork) {
352 			cork = true;
353 			psock->cork = NULL;
354 		}
355 		sk_msg_return(sk, msg, tosend);
356 		release_sock(sk);
357 		ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
358 		lock_sock(sk);
359 		if (unlikely(ret < 0)) {
360 			int free = sk_msg_free_nocharge(sk, msg);
361 
362 			if (!cork)
363 				*copied -= free;
364 		}
365 		if (cork) {
366 			sk_msg_free(sk, msg);
367 			kfree(msg);
368 			msg = NULL;
369 			ret = 0;
370 		}
371 		break;
372 	case __SK_DROP:
373 	default:
374 		sk_msg_free_partial(sk, msg, tosend);
375 		sk_msg_apply_bytes(psock, tosend);
376 		*copied -= (tosend + delta);
377 		return -EACCES;
378 	}
379 
380 	if (likely(!ret)) {
381 		if (!psock->apply_bytes) {
382 			psock->eval =  __SK_NONE;
383 			if (psock->sk_redir) {
384 				sock_put(psock->sk_redir);
385 				psock->sk_redir = NULL;
386 			}
387 		}
388 		if (msg &&
389 		    msg->sg.data[msg->sg.start].page_link &&
390 		    msg->sg.data[msg->sg.start].length)
391 			goto more_data;
392 	}
393 	return ret;
394 }
395 
396 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
397 {
398 	struct sk_msg tmp, *msg_tx = NULL;
399 	int copied = 0, err = 0;
400 	struct sk_psock *psock;
401 	long timeo;
402 	int flags;
403 
404 	/* Don't let internal do_tcp_sendpages() flags through */
405 	flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
406 	flags |= MSG_NO_SHARED_FRAGS;
407 
408 	psock = sk_psock_get(sk);
409 	if (unlikely(!psock))
410 		return tcp_sendmsg(sk, msg, size);
411 
412 	lock_sock(sk);
413 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
414 	while (msg_data_left(msg)) {
415 		bool enospc = false;
416 		u32 copy, osize;
417 
418 		if (sk->sk_err) {
419 			err = -sk->sk_err;
420 			goto out_err;
421 		}
422 
423 		copy = msg_data_left(msg);
424 		if (!sk_stream_memory_free(sk))
425 			goto wait_for_sndbuf;
426 		if (psock->cork) {
427 			msg_tx = psock->cork;
428 		} else {
429 			msg_tx = &tmp;
430 			sk_msg_init(msg_tx);
431 		}
432 
433 		osize = msg_tx->sg.size;
434 		err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
435 		if (err) {
436 			if (err != -ENOSPC)
437 				goto wait_for_memory;
438 			enospc = true;
439 			copy = msg_tx->sg.size - osize;
440 		}
441 
442 		err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
443 					       copy);
444 		if (err < 0) {
445 			sk_msg_trim(sk, msg_tx, osize);
446 			goto out_err;
447 		}
448 
449 		copied += copy;
450 		if (psock->cork_bytes) {
451 			if (size > psock->cork_bytes)
452 				psock->cork_bytes = 0;
453 			else
454 				psock->cork_bytes -= size;
455 			if (psock->cork_bytes && !enospc)
456 				goto out_err;
457 			/* All cork bytes are accounted, rerun the prog. */
458 			psock->eval = __SK_NONE;
459 			psock->cork_bytes = 0;
460 		}
461 
462 		err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
463 		if (unlikely(err < 0))
464 			goto out_err;
465 		continue;
466 wait_for_sndbuf:
467 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
468 wait_for_memory:
469 		err = sk_stream_wait_memory(sk, &timeo);
470 		if (err) {
471 			if (msg_tx && msg_tx != psock->cork)
472 				sk_msg_free(sk, msg_tx);
473 			goto out_err;
474 		}
475 	}
476 out_err:
477 	if (err < 0)
478 		err = sk_stream_error(sk, msg->msg_flags, err);
479 	release_sock(sk);
480 	sk_psock_put(sk, psock);
481 	return copied ? copied : err;
482 }
483 
484 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
485 			    size_t size, int flags)
486 {
487 	struct sk_msg tmp, *msg = NULL;
488 	int err = 0, copied = 0;
489 	struct sk_psock *psock;
490 	bool enospc = false;
491 
492 	psock = sk_psock_get(sk);
493 	if (unlikely(!psock))
494 		return tcp_sendpage(sk, page, offset, size, flags);
495 
496 	lock_sock(sk);
497 	if (psock->cork) {
498 		msg = psock->cork;
499 	} else {
500 		msg = &tmp;
501 		sk_msg_init(msg);
502 	}
503 
504 	/* Catch case where ring is full and sendpage is stalled. */
505 	if (unlikely(sk_msg_full(msg)))
506 		goto out_err;
507 
508 	sk_msg_page_add(msg, page, size, offset);
509 	sk_mem_charge(sk, size);
510 	copied = size;
511 	if (sk_msg_full(msg))
512 		enospc = true;
513 	if (psock->cork_bytes) {
514 		if (size > psock->cork_bytes)
515 			psock->cork_bytes = 0;
516 		else
517 			psock->cork_bytes -= size;
518 		if (psock->cork_bytes && !enospc)
519 			goto out_err;
520 		/* All cork bytes are accounted, rerun the prog. */
521 		psock->eval = __SK_NONE;
522 		psock->cork_bytes = 0;
523 	}
524 
525 	err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
526 out_err:
527 	release_sock(sk);
528 	sk_psock_put(sk, psock);
529 	return copied ? copied : err;
530 }
531 
532 enum {
533 	TCP_BPF_IPV4,
534 	TCP_BPF_IPV6,
535 	TCP_BPF_NUM_PROTS,
536 };
537 
538 enum {
539 	TCP_BPF_BASE,
540 	TCP_BPF_TX,
541 	TCP_BPF_NUM_CFGS,
542 };
543 
544 static struct proto *tcpv6_prot_saved __read_mostly;
545 static DEFINE_SPINLOCK(tcpv6_prot_lock);
546 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
547 
548 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
549 				   struct proto *base)
550 {
551 	prot[TCP_BPF_BASE]			= *base;
552 	prot[TCP_BPF_BASE].unhash		= sock_map_unhash;
553 	prot[TCP_BPF_BASE].close		= sock_map_close;
554 	prot[TCP_BPF_BASE].recvmsg		= tcp_bpf_recvmsg;
555 	prot[TCP_BPF_BASE].stream_memory_read	= tcp_bpf_stream_read;
556 
557 	prot[TCP_BPF_TX]			= prot[TCP_BPF_BASE];
558 	prot[TCP_BPF_TX].sendmsg		= tcp_bpf_sendmsg;
559 	prot[TCP_BPF_TX].sendpage		= tcp_bpf_sendpage;
560 }
561 
562 static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops)
563 {
564 	if (sk->sk_family == AF_INET6 &&
565 	    unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
566 		spin_lock_bh(&tcpv6_prot_lock);
567 		if (likely(ops != tcpv6_prot_saved)) {
568 			tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
569 			smp_store_release(&tcpv6_prot_saved, ops);
570 		}
571 		spin_unlock_bh(&tcpv6_prot_lock);
572 	}
573 }
574 
575 static int __init tcp_bpf_v4_build_proto(void)
576 {
577 	tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
578 	return 0;
579 }
580 core_initcall(tcp_bpf_v4_build_proto);
581 
582 static int tcp_bpf_assert_proto_ops(struct proto *ops)
583 {
584 	/* In order to avoid retpoline, we make assumptions when we call
585 	 * into ops if e.g. a psock is not present. Make sure they are
586 	 * indeed valid assumptions.
587 	 */
588 	return ops->recvmsg  == tcp_recvmsg &&
589 	       ops->sendmsg  == tcp_sendmsg &&
590 	       ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
591 }
592 
593 struct proto *tcp_bpf_get_proto(struct sock *sk, struct sk_psock *psock)
594 {
595 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
596 	int config = psock->progs.msg_parser   ? TCP_BPF_TX   : TCP_BPF_BASE;
597 
598 	if (!psock->sk_proto) {
599 		struct proto *ops = READ_ONCE(sk->sk_prot);
600 
601 		if (tcp_bpf_assert_proto_ops(ops))
602 			return ERR_PTR(-EINVAL);
603 
604 		tcp_bpf_check_v6_needs_rebuild(sk, ops);
605 	}
606 
607 	return &tcp_bpf_prots[family][config];
608 }
609 
610 /* If a child got cloned from a listening socket that had tcp_bpf
611  * protocol callbacks installed, we need to restore the callbacks to
612  * the default ones because the child does not inherit the psock state
613  * that tcp_bpf callbacks expect.
614  */
615 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
616 {
617 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
618 	struct proto *prot = newsk->sk_prot;
619 
620 	if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
621 		newsk->sk_prot = sk->sk_prot_creator;
622 }
623 #endif /* CONFIG_BPF_STREAM_PARSER */
624