xref: /openbmc/linux/net/ipv4/tcp_bpf.c (revision 842ed298)
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 	struct sk_msg *msg_rx;
19 	int i, copied = 0;
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 			copy = copy_page_to_iter(page, sge->offset, copy, iter);
41 			if (!copy)
42 				return copied ? copied : -EFAULT;
43 
44 			copied += copy;
45 			if (likely(!peek)) {
46 				sge->offset += copy;
47 				sge->length -= copy;
48 				if (!msg_rx->skb)
49 					sk_mem_uncharge(sk, copy);
50 				msg_rx->sg.size -= copy;
51 
52 				if (!sge->length) {
53 					sk_msg_iter_var_next(i);
54 					if (!msg_rx->skb)
55 						put_page(page);
56 				}
57 			} else {
58 				/* Lets not optimize peek case if copy_page_to_iter
59 				 * didn't copy the entire length lets just break.
60 				 */
61 				if (copy != sge->length)
62 					return copied;
63 				sk_msg_iter_var_next(i);
64 			}
65 
66 			if (copied == len)
67 				break;
68 		} while (i != msg_rx->sg.end);
69 
70 		if (unlikely(peek)) {
71 			if (msg_rx == list_last_entry(&psock->ingress_msg,
72 						      struct sk_msg, list))
73 				break;
74 			msg_rx = list_next_entry(msg_rx, list);
75 			continue;
76 		}
77 
78 		msg_rx->sg.start = i;
79 		if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
80 			list_del(&msg_rx->list);
81 			if (msg_rx->skb)
82 				consume_skb(msg_rx->skb);
83 			kfree(msg_rx);
84 		}
85 		msg_rx = list_first_entry_or_null(&psock->ingress_msg,
86 						  struct sk_msg, list);
87 	}
88 
89 	return copied;
90 }
91 EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg);
92 
93 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
94 			   struct sk_msg *msg, u32 apply_bytes, int flags)
95 {
96 	bool apply = apply_bytes;
97 	struct scatterlist *sge;
98 	u32 size, copied = 0;
99 	struct sk_msg *tmp;
100 	int i, ret = 0;
101 
102 	tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
103 	if (unlikely(!tmp))
104 		return -ENOMEM;
105 
106 	lock_sock(sk);
107 	tmp->sg.start = msg->sg.start;
108 	i = msg->sg.start;
109 	do {
110 		sge = sk_msg_elem(msg, i);
111 		size = (apply && apply_bytes < sge->length) ?
112 			apply_bytes : sge->length;
113 		if (!sk_wmem_schedule(sk, size)) {
114 			if (!copied)
115 				ret = -ENOMEM;
116 			break;
117 		}
118 
119 		sk_mem_charge(sk, size);
120 		sk_msg_xfer(tmp, msg, i, size);
121 		copied += size;
122 		if (sge->length)
123 			get_page(sk_msg_page(tmp, i));
124 		sk_msg_iter_var_next(i);
125 		tmp->sg.end = i;
126 		if (apply) {
127 			apply_bytes -= size;
128 			if (!apply_bytes)
129 				break;
130 		}
131 	} while (i != msg->sg.end);
132 
133 	if (!ret) {
134 		msg->sg.start = i;
135 		sk_psock_queue_msg(psock, tmp);
136 		sk_psock_data_ready(sk, psock);
137 	} else {
138 		sk_msg_free(sk, tmp);
139 		kfree(tmp);
140 	}
141 
142 	release_sock(sk);
143 	return ret;
144 }
145 
146 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
147 			int flags, bool uncharge)
148 {
149 	bool apply = apply_bytes;
150 	struct scatterlist *sge;
151 	struct page *page;
152 	int size, ret = 0;
153 	u32 off;
154 
155 	while (1) {
156 		bool has_tx_ulp;
157 
158 		sge = sk_msg_elem(msg, msg->sg.start);
159 		size = (apply && apply_bytes < sge->length) ?
160 			apply_bytes : sge->length;
161 		off  = sge->offset;
162 		page = sg_page(sge);
163 
164 		tcp_rate_check_app_limited(sk);
165 retry:
166 		has_tx_ulp = tls_sw_has_ctx_tx(sk);
167 		if (has_tx_ulp) {
168 			flags |= MSG_SENDPAGE_NOPOLICY;
169 			ret = kernel_sendpage_locked(sk,
170 						     page, off, size, flags);
171 		} else {
172 			ret = do_tcp_sendpages(sk, page, off, size, flags);
173 		}
174 
175 		if (ret <= 0)
176 			return ret;
177 		if (apply)
178 			apply_bytes -= ret;
179 		msg->sg.size -= ret;
180 		sge->offset += ret;
181 		sge->length -= ret;
182 		if (uncharge)
183 			sk_mem_uncharge(sk, ret);
184 		if (ret != size) {
185 			size -= ret;
186 			off  += ret;
187 			goto retry;
188 		}
189 		if (!sge->length) {
190 			put_page(page);
191 			sk_msg_iter_next(msg, start);
192 			sg_init_table(sge, 1);
193 			if (msg->sg.start == msg->sg.end)
194 				break;
195 		}
196 		if (apply && !apply_bytes)
197 			break;
198 	}
199 
200 	return 0;
201 }
202 
203 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
204 			       u32 apply_bytes, int flags, bool uncharge)
205 {
206 	int ret;
207 
208 	lock_sock(sk);
209 	ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
210 	release_sock(sk);
211 	return ret;
212 }
213 
214 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
215 			  u32 bytes, int flags)
216 {
217 	bool ingress = sk_msg_to_ingress(msg);
218 	struct sk_psock *psock = sk_psock_get(sk);
219 	int ret;
220 
221 	if (unlikely(!psock)) {
222 		sk_msg_free(sk, msg);
223 		return 0;
224 	}
225 	ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
226 			tcp_bpf_push_locked(sk, msg, bytes, flags, false);
227 	sk_psock_put(sk, psock);
228 	return ret;
229 }
230 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
231 
232 #ifdef CONFIG_BPF_STREAM_PARSER
233 static bool tcp_bpf_stream_read(const struct sock *sk)
234 {
235 	struct sk_psock *psock;
236 	bool empty = true;
237 
238 	rcu_read_lock();
239 	psock = sk_psock(sk);
240 	if (likely(psock))
241 		empty = list_empty(&psock->ingress_msg);
242 	rcu_read_unlock();
243 	return !empty;
244 }
245 
246 static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock,
247 			     int flags, long timeo, int *err)
248 {
249 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
250 	int ret = 0;
251 
252 	if (sk->sk_shutdown & RCV_SHUTDOWN)
253 		return 1;
254 
255 	if (!timeo)
256 		return ret;
257 
258 	add_wait_queue(sk_sleep(sk), &wait);
259 	sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
260 	ret = sk_wait_event(sk, &timeo,
261 			    !list_empty(&psock->ingress_msg) ||
262 			    !skb_queue_empty(&sk->sk_receive_queue), &wait);
263 	sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
264 	remove_wait_queue(sk_sleep(sk), &wait);
265 	return ret;
266 }
267 
268 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
269 		    int nonblock, int flags, int *addr_len)
270 {
271 	struct sk_psock *psock;
272 	int copied, ret;
273 
274 	if (unlikely(flags & MSG_ERRQUEUE))
275 		return inet_recv_error(sk, msg, len, addr_len);
276 
277 	psock = sk_psock_get(sk);
278 	if (unlikely(!psock))
279 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
280 	if (!skb_queue_empty(&sk->sk_receive_queue) &&
281 	    sk_psock_queue_empty(psock)) {
282 		sk_psock_put(sk, psock);
283 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
284 	}
285 	lock_sock(sk);
286 msg_bytes_ready:
287 	copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags);
288 	if (!copied) {
289 		int data, err = 0;
290 		long timeo;
291 
292 		timeo = sock_rcvtimeo(sk, nonblock);
293 		data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err);
294 		if (data) {
295 			if (!sk_psock_queue_empty(psock))
296 				goto msg_bytes_ready;
297 			release_sock(sk);
298 			sk_psock_put(sk, psock);
299 			return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
300 		}
301 		if (err) {
302 			ret = err;
303 			goto out;
304 		}
305 		copied = -EAGAIN;
306 	}
307 	ret = copied;
308 out:
309 	release_sock(sk);
310 	sk_psock_put(sk, psock);
311 	return ret;
312 }
313 
314 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
315 				struct sk_msg *msg, int *copied, int flags)
316 {
317 	bool cork = false, enospc = sk_msg_full(msg);
318 	struct sock *sk_redir;
319 	u32 tosend, delta = 0;
320 	int ret;
321 
322 more_data:
323 	if (psock->eval == __SK_NONE) {
324 		/* Track delta in msg size to add/subtract it on SK_DROP from
325 		 * returned to user copied size. This ensures user doesn't
326 		 * get a positive return code with msg_cut_data and SK_DROP
327 		 * verdict.
328 		 */
329 		delta = msg->sg.size;
330 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
331 		delta -= msg->sg.size;
332 	}
333 
334 	if (msg->cork_bytes &&
335 	    msg->cork_bytes > msg->sg.size && !enospc) {
336 		psock->cork_bytes = msg->cork_bytes - msg->sg.size;
337 		if (!psock->cork) {
338 			psock->cork = kzalloc(sizeof(*psock->cork),
339 					      GFP_ATOMIC | __GFP_NOWARN);
340 			if (!psock->cork)
341 				return -ENOMEM;
342 		}
343 		memcpy(psock->cork, msg, sizeof(*msg));
344 		return 0;
345 	}
346 
347 	tosend = msg->sg.size;
348 	if (psock->apply_bytes && psock->apply_bytes < tosend)
349 		tosend = psock->apply_bytes;
350 
351 	switch (psock->eval) {
352 	case __SK_PASS:
353 		ret = tcp_bpf_push(sk, msg, tosend, flags, true);
354 		if (unlikely(ret)) {
355 			*copied -= sk_msg_free(sk, msg);
356 			break;
357 		}
358 		sk_msg_apply_bytes(psock, tosend);
359 		break;
360 	case __SK_REDIRECT:
361 		sk_redir = psock->sk_redir;
362 		sk_msg_apply_bytes(psock, tosend);
363 		if (psock->cork) {
364 			cork = true;
365 			psock->cork = NULL;
366 		}
367 		sk_msg_return(sk, msg, tosend);
368 		release_sock(sk);
369 		ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
370 		lock_sock(sk);
371 		if (unlikely(ret < 0)) {
372 			int free = sk_msg_free_nocharge(sk, msg);
373 
374 			if (!cork)
375 				*copied -= free;
376 		}
377 		if (cork) {
378 			sk_msg_free(sk, msg);
379 			kfree(msg);
380 			msg = NULL;
381 			ret = 0;
382 		}
383 		break;
384 	case __SK_DROP:
385 	default:
386 		sk_msg_free_partial(sk, msg, tosend);
387 		sk_msg_apply_bytes(psock, tosend);
388 		*copied -= (tosend + delta);
389 		return -EACCES;
390 	}
391 
392 	if (likely(!ret)) {
393 		if (!psock->apply_bytes) {
394 			psock->eval =  __SK_NONE;
395 			if (psock->sk_redir) {
396 				sock_put(psock->sk_redir);
397 				psock->sk_redir = NULL;
398 			}
399 		}
400 		if (msg &&
401 		    msg->sg.data[msg->sg.start].page_link &&
402 		    msg->sg.data[msg->sg.start].length)
403 			goto more_data;
404 	}
405 	return ret;
406 }
407 
408 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
409 {
410 	struct sk_msg tmp, *msg_tx = NULL;
411 	int copied = 0, err = 0;
412 	struct sk_psock *psock;
413 	long timeo;
414 	int flags;
415 
416 	/* Don't let internal do_tcp_sendpages() flags through */
417 	flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
418 	flags |= MSG_NO_SHARED_FRAGS;
419 
420 	psock = sk_psock_get(sk);
421 	if (unlikely(!psock))
422 		return tcp_sendmsg(sk, msg, size);
423 
424 	lock_sock(sk);
425 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
426 	while (msg_data_left(msg)) {
427 		bool enospc = false;
428 		u32 copy, osize;
429 
430 		if (sk->sk_err) {
431 			err = -sk->sk_err;
432 			goto out_err;
433 		}
434 
435 		copy = msg_data_left(msg);
436 		if (!sk_stream_memory_free(sk))
437 			goto wait_for_sndbuf;
438 		if (psock->cork) {
439 			msg_tx = psock->cork;
440 		} else {
441 			msg_tx = &tmp;
442 			sk_msg_init(msg_tx);
443 		}
444 
445 		osize = msg_tx->sg.size;
446 		err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
447 		if (err) {
448 			if (err != -ENOSPC)
449 				goto wait_for_memory;
450 			enospc = true;
451 			copy = msg_tx->sg.size - osize;
452 		}
453 
454 		err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
455 					       copy);
456 		if (err < 0) {
457 			sk_msg_trim(sk, msg_tx, osize);
458 			goto out_err;
459 		}
460 
461 		copied += copy;
462 		if (psock->cork_bytes) {
463 			if (size > psock->cork_bytes)
464 				psock->cork_bytes = 0;
465 			else
466 				psock->cork_bytes -= size;
467 			if (psock->cork_bytes && !enospc)
468 				goto out_err;
469 			/* All cork bytes are accounted, rerun the prog. */
470 			psock->eval = __SK_NONE;
471 			psock->cork_bytes = 0;
472 		}
473 
474 		err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
475 		if (unlikely(err < 0))
476 			goto out_err;
477 		continue;
478 wait_for_sndbuf:
479 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
480 wait_for_memory:
481 		err = sk_stream_wait_memory(sk, &timeo);
482 		if (err) {
483 			if (msg_tx && msg_tx != psock->cork)
484 				sk_msg_free(sk, msg_tx);
485 			goto out_err;
486 		}
487 	}
488 out_err:
489 	if (err < 0)
490 		err = sk_stream_error(sk, msg->msg_flags, err);
491 	release_sock(sk);
492 	sk_psock_put(sk, psock);
493 	return copied ? copied : err;
494 }
495 
496 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
497 			    size_t size, int flags)
498 {
499 	struct sk_msg tmp, *msg = NULL;
500 	int err = 0, copied = 0;
501 	struct sk_psock *psock;
502 	bool enospc = false;
503 
504 	psock = sk_psock_get(sk);
505 	if (unlikely(!psock))
506 		return tcp_sendpage(sk, page, offset, size, flags);
507 
508 	lock_sock(sk);
509 	if (psock->cork) {
510 		msg = psock->cork;
511 	} else {
512 		msg = &tmp;
513 		sk_msg_init(msg);
514 	}
515 
516 	/* Catch case where ring is full and sendpage is stalled. */
517 	if (unlikely(sk_msg_full(msg)))
518 		goto out_err;
519 
520 	sk_msg_page_add(msg, page, size, offset);
521 	sk_mem_charge(sk, size);
522 	copied = size;
523 	if (sk_msg_full(msg))
524 		enospc = true;
525 	if (psock->cork_bytes) {
526 		if (size > psock->cork_bytes)
527 			psock->cork_bytes = 0;
528 		else
529 			psock->cork_bytes -= size;
530 		if (psock->cork_bytes && !enospc)
531 			goto out_err;
532 		/* All cork bytes are accounted, rerun the prog. */
533 		psock->eval = __SK_NONE;
534 		psock->cork_bytes = 0;
535 	}
536 
537 	err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
538 out_err:
539 	release_sock(sk);
540 	sk_psock_put(sk, psock);
541 	return copied ? copied : err;
542 }
543 
544 enum {
545 	TCP_BPF_IPV4,
546 	TCP_BPF_IPV6,
547 	TCP_BPF_NUM_PROTS,
548 };
549 
550 enum {
551 	TCP_BPF_BASE,
552 	TCP_BPF_TX,
553 	TCP_BPF_NUM_CFGS,
554 };
555 
556 static struct proto *tcpv6_prot_saved __read_mostly;
557 static DEFINE_SPINLOCK(tcpv6_prot_lock);
558 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
559 
560 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
561 				   struct proto *base)
562 {
563 	prot[TCP_BPF_BASE]			= *base;
564 	prot[TCP_BPF_BASE].unhash		= sock_map_unhash;
565 	prot[TCP_BPF_BASE].close		= sock_map_close;
566 	prot[TCP_BPF_BASE].recvmsg		= tcp_bpf_recvmsg;
567 	prot[TCP_BPF_BASE].stream_memory_read	= tcp_bpf_stream_read;
568 
569 	prot[TCP_BPF_TX]			= prot[TCP_BPF_BASE];
570 	prot[TCP_BPF_TX].sendmsg		= tcp_bpf_sendmsg;
571 	prot[TCP_BPF_TX].sendpage		= tcp_bpf_sendpage;
572 }
573 
574 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
575 {
576 	if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
577 		spin_lock_bh(&tcpv6_prot_lock);
578 		if (likely(ops != tcpv6_prot_saved)) {
579 			tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
580 			smp_store_release(&tcpv6_prot_saved, ops);
581 		}
582 		spin_unlock_bh(&tcpv6_prot_lock);
583 	}
584 }
585 
586 static int __init tcp_bpf_v4_build_proto(void)
587 {
588 	tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
589 	return 0;
590 }
591 core_initcall(tcp_bpf_v4_build_proto);
592 
593 static int tcp_bpf_assert_proto_ops(struct proto *ops)
594 {
595 	/* In order to avoid retpoline, we make assumptions when we call
596 	 * into ops if e.g. a psock is not present. Make sure they are
597 	 * indeed valid assumptions.
598 	 */
599 	return ops->recvmsg  == tcp_recvmsg &&
600 	       ops->sendmsg  == tcp_sendmsg &&
601 	       ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
602 }
603 
604 struct proto *tcp_bpf_get_proto(struct sock *sk, struct sk_psock *psock)
605 {
606 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
607 	int config = psock->progs.msg_parser   ? TCP_BPF_TX   : TCP_BPF_BASE;
608 
609 	if (sk->sk_family == AF_INET6) {
610 		if (tcp_bpf_assert_proto_ops(psock->sk_proto))
611 			return ERR_PTR(-EINVAL);
612 
613 		tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
614 	}
615 
616 	return &tcp_bpf_prots[family][config];
617 }
618 
619 /* If a child got cloned from a listening socket that had tcp_bpf
620  * protocol callbacks installed, we need to restore the callbacks to
621  * the default ones because the child does not inherit the psock state
622  * that tcp_bpf callbacks expect.
623  */
624 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
625 {
626 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
627 	struct proto *prot = newsk->sk_prot;
628 
629 	if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
630 		newsk->sk_prot = sk->sk_prot_creator;
631 }
632 #endif /* CONFIG_BPF_STREAM_PARSER */
633