xref: /openbmc/linux/net/ipv4/tcp_bpf.c (revision c29b9772)
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 #include <linux/util_macros.h>
10 
11 #include <net/inet_common.h>
12 #include <net/tls.h>
13 
14 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
15 			   struct sk_msg *msg, u32 apply_bytes, int flags)
16 {
17 	bool apply = apply_bytes;
18 	struct scatterlist *sge;
19 	u32 size, copied = 0;
20 	struct sk_msg *tmp;
21 	int i, ret = 0;
22 
23 	tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
24 	if (unlikely(!tmp))
25 		return -ENOMEM;
26 
27 	lock_sock(sk);
28 	tmp->sg.start = msg->sg.start;
29 	i = msg->sg.start;
30 	do {
31 		sge = sk_msg_elem(msg, i);
32 		size = (apply && apply_bytes < sge->length) ?
33 			apply_bytes : sge->length;
34 		if (!sk_wmem_schedule(sk, size)) {
35 			if (!copied)
36 				ret = -ENOMEM;
37 			break;
38 		}
39 
40 		sk_mem_charge(sk, size);
41 		sk_msg_xfer(tmp, msg, i, size);
42 		copied += size;
43 		if (sge->length)
44 			get_page(sk_msg_page(tmp, i));
45 		sk_msg_iter_var_next(i);
46 		tmp->sg.end = i;
47 		if (apply) {
48 			apply_bytes -= size;
49 			if (!apply_bytes) {
50 				if (sge->length)
51 					sk_msg_iter_var_prev(i);
52 				break;
53 			}
54 		}
55 	} while (i != msg->sg.end);
56 
57 	if (!ret) {
58 		msg->sg.start = i;
59 		sk_psock_queue_msg(psock, tmp);
60 		sk_psock_data_ready(sk, psock);
61 	} else {
62 		sk_msg_free(sk, tmp);
63 		kfree(tmp);
64 	}
65 
66 	release_sock(sk);
67 	return ret;
68 }
69 
70 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
71 			int flags, bool uncharge)
72 {
73 	bool apply = apply_bytes;
74 	struct scatterlist *sge;
75 	struct page *page;
76 	int size, ret = 0;
77 	u32 off;
78 
79 	while (1) {
80 		bool has_tx_ulp;
81 
82 		sge = sk_msg_elem(msg, msg->sg.start);
83 		size = (apply && apply_bytes < sge->length) ?
84 			apply_bytes : sge->length;
85 		off  = sge->offset;
86 		page = sg_page(sge);
87 
88 		tcp_rate_check_app_limited(sk);
89 retry:
90 		has_tx_ulp = tls_sw_has_ctx_tx(sk);
91 		if (has_tx_ulp) {
92 			flags |= MSG_SENDPAGE_NOPOLICY;
93 			ret = kernel_sendpage_locked(sk,
94 						     page, off, size, flags);
95 		} else {
96 			ret = do_tcp_sendpages(sk, page, off, size, flags);
97 		}
98 
99 		if (ret <= 0)
100 			return ret;
101 		if (apply)
102 			apply_bytes -= ret;
103 		msg->sg.size -= ret;
104 		sge->offset += ret;
105 		sge->length -= ret;
106 		if (uncharge)
107 			sk_mem_uncharge(sk, ret);
108 		if (ret != size) {
109 			size -= ret;
110 			off  += ret;
111 			goto retry;
112 		}
113 		if (!sge->length) {
114 			put_page(page);
115 			sk_msg_iter_next(msg, start);
116 			sg_init_table(sge, 1);
117 			if (msg->sg.start == msg->sg.end)
118 				break;
119 		}
120 		if (apply && !apply_bytes)
121 			break;
122 	}
123 
124 	return 0;
125 }
126 
127 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
128 			       u32 apply_bytes, int flags, bool uncharge)
129 {
130 	int ret;
131 
132 	lock_sock(sk);
133 	ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
134 	release_sock(sk);
135 	return ret;
136 }
137 
138 int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress,
139 			  struct sk_msg *msg, u32 bytes, int flags)
140 {
141 	struct sk_psock *psock = sk_psock_get(sk);
142 	int ret;
143 
144 	if (unlikely(!psock))
145 		return -EPIPE;
146 
147 	ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
148 			tcp_bpf_push_locked(sk, msg, bytes, flags, false);
149 	sk_psock_put(sk, psock);
150 	return ret;
151 }
152 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
153 
154 #ifdef CONFIG_BPF_SYSCALL
155 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
156 			     long timeo)
157 {
158 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
159 	int ret = 0;
160 
161 	if (sk->sk_shutdown & RCV_SHUTDOWN)
162 		return 1;
163 
164 	if (!timeo)
165 		return ret;
166 
167 	add_wait_queue(sk_sleep(sk), &wait);
168 	sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
169 	ret = sk_wait_event(sk, &timeo,
170 			    !list_empty(&psock->ingress_msg) ||
171 			    !skb_queue_empty(&sk->sk_receive_queue), &wait);
172 	sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
173 	remove_wait_queue(sk_sleep(sk), &wait);
174 	return ret;
175 }
176 
177 static int tcp_bpf_recvmsg_parser(struct sock *sk,
178 				  struct msghdr *msg,
179 				  size_t len,
180 				  int flags,
181 				  int *addr_len)
182 {
183 	struct sk_psock *psock;
184 	int copied;
185 
186 	if (unlikely(flags & MSG_ERRQUEUE))
187 		return inet_recv_error(sk, msg, len, addr_len);
188 
189 	psock = sk_psock_get(sk);
190 	if (unlikely(!psock))
191 		return tcp_recvmsg(sk, msg, len, flags, addr_len);
192 
193 	lock_sock(sk);
194 msg_bytes_ready:
195 	copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
196 	if (!copied) {
197 		long timeo;
198 		int data;
199 
200 		if (sock_flag(sk, SOCK_DONE))
201 			goto out;
202 
203 		if (sk->sk_err) {
204 			copied = sock_error(sk);
205 			goto out;
206 		}
207 
208 		if (sk->sk_shutdown & RCV_SHUTDOWN)
209 			goto out;
210 
211 		if (sk->sk_state == TCP_CLOSE) {
212 			copied = -ENOTCONN;
213 			goto out;
214 		}
215 
216 		timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
217 		if (!timeo) {
218 			copied = -EAGAIN;
219 			goto out;
220 		}
221 
222 		if (signal_pending(current)) {
223 			copied = sock_intr_errno(timeo);
224 			goto out;
225 		}
226 
227 		data = tcp_msg_wait_data(sk, psock, timeo);
228 		if (data && !sk_psock_queue_empty(psock))
229 			goto msg_bytes_ready;
230 		copied = -EAGAIN;
231 	}
232 out:
233 	release_sock(sk);
234 	sk_psock_put(sk, psock);
235 	return copied;
236 }
237 
238 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
239 			   int flags, int *addr_len)
240 {
241 	struct sk_psock *psock;
242 	int copied, ret;
243 
244 	if (unlikely(flags & MSG_ERRQUEUE))
245 		return inet_recv_error(sk, msg, len, addr_len);
246 
247 	psock = sk_psock_get(sk);
248 	if (unlikely(!psock))
249 		return tcp_recvmsg(sk, msg, len, flags, addr_len);
250 	if (!skb_queue_empty(&sk->sk_receive_queue) &&
251 	    sk_psock_queue_empty(psock)) {
252 		sk_psock_put(sk, psock);
253 		return tcp_recvmsg(sk, msg, len, flags, addr_len);
254 	}
255 	lock_sock(sk);
256 msg_bytes_ready:
257 	copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
258 	if (!copied) {
259 		long timeo;
260 		int data;
261 
262 		timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
263 		data = tcp_msg_wait_data(sk, psock, timeo);
264 		if (data) {
265 			if (!sk_psock_queue_empty(psock))
266 				goto msg_bytes_ready;
267 			release_sock(sk);
268 			sk_psock_put(sk, psock);
269 			return tcp_recvmsg(sk, msg, len, flags, addr_len);
270 		}
271 		copied = -EAGAIN;
272 	}
273 	ret = copied;
274 	release_sock(sk);
275 	sk_psock_put(sk, psock);
276 	return ret;
277 }
278 
279 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
280 				struct sk_msg *msg, int *copied, int flags)
281 {
282 	bool cork = false, enospc = sk_msg_full(msg), redir_ingress;
283 	struct sock *sk_redir;
284 	u32 tosend, origsize, sent, delta = 0;
285 	u32 eval;
286 	int ret;
287 
288 more_data:
289 	if (psock->eval == __SK_NONE) {
290 		/* Track delta in msg size to add/subtract it on SK_DROP from
291 		 * returned to user copied size. This ensures user doesn't
292 		 * get a positive return code with msg_cut_data and SK_DROP
293 		 * verdict.
294 		 */
295 		delta = msg->sg.size;
296 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
297 		delta -= msg->sg.size;
298 	}
299 
300 	if (msg->cork_bytes &&
301 	    msg->cork_bytes > msg->sg.size && !enospc) {
302 		psock->cork_bytes = msg->cork_bytes - msg->sg.size;
303 		if (!psock->cork) {
304 			psock->cork = kzalloc(sizeof(*psock->cork),
305 					      GFP_ATOMIC | __GFP_NOWARN);
306 			if (!psock->cork)
307 				return -ENOMEM;
308 		}
309 		memcpy(psock->cork, msg, sizeof(*msg));
310 		return 0;
311 	}
312 
313 	tosend = msg->sg.size;
314 	if (psock->apply_bytes && psock->apply_bytes < tosend)
315 		tosend = psock->apply_bytes;
316 	eval = __SK_NONE;
317 
318 	switch (psock->eval) {
319 	case __SK_PASS:
320 		ret = tcp_bpf_push(sk, msg, tosend, flags, true);
321 		if (unlikely(ret)) {
322 			*copied -= sk_msg_free(sk, msg);
323 			break;
324 		}
325 		sk_msg_apply_bytes(psock, tosend);
326 		break;
327 	case __SK_REDIRECT:
328 		redir_ingress = psock->redir_ingress;
329 		sk_redir = psock->sk_redir;
330 		sk_msg_apply_bytes(psock, tosend);
331 		if (!psock->apply_bytes) {
332 			/* Clean up before releasing the sock lock. */
333 			eval = psock->eval;
334 			psock->eval = __SK_NONE;
335 			psock->sk_redir = NULL;
336 		}
337 		if (psock->cork) {
338 			cork = true;
339 			psock->cork = NULL;
340 		}
341 		sk_msg_return(sk, msg, tosend);
342 		release_sock(sk);
343 
344 		origsize = msg->sg.size;
345 		ret = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
346 					    msg, tosend, flags);
347 		sent = origsize - msg->sg.size;
348 
349 		if (eval == __SK_REDIRECT)
350 			sock_put(sk_redir);
351 
352 		lock_sock(sk);
353 		if (unlikely(ret < 0)) {
354 			int free = sk_msg_free_nocharge(sk, msg);
355 
356 			if (!cork)
357 				*copied -= free;
358 		}
359 		if (cork) {
360 			sk_msg_free(sk, msg);
361 			kfree(msg);
362 			msg = NULL;
363 			ret = 0;
364 		}
365 		break;
366 	case __SK_DROP:
367 	default:
368 		sk_msg_free_partial(sk, msg, tosend);
369 		sk_msg_apply_bytes(psock, tosend);
370 		*copied -= (tosend + delta);
371 		return -EACCES;
372 	}
373 
374 	if (likely(!ret)) {
375 		if (!psock->apply_bytes) {
376 			psock->eval =  __SK_NONE;
377 			if (psock->sk_redir) {
378 				sock_put(psock->sk_redir);
379 				psock->sk_redir = NULL;
380 			}
381 		}
382 		if (msg &&
383 		    msg->sg.data[msg->sg.start].page_link &&
384 		    msg->sg.data[msg->sg.start].length) {
385 			if (eval == __SK_REDIRECT)
386 				sk_mem_charge(sk, tosend - sent);
387 			goto more_data;
388 		}
389 	}
390 	return ret;
391 }
392 
393 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
394 {
395 	struct sk_msg tmp, *msg_tx = NULL;
396 	int copied = 0, err = 0;
397 	struct sk_psock *psock;
398 	long timeo;
399 	int flags;
400 
401 	/* Don't let internal do_tcp_sendpages() flags through */
402 	flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
403 	flags |= MSG_NO_SHARED_FRAGS;
404 
405 	psock = sk_psock_get(sk);
406 	if (unlikely(!psock))
407 		return tcp_sendmsg(sk, msg, size);
408 
409 	lock_sock(sk);
410 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
411 	while (msg_data_left(msg)) {
412 		bool enospc = false;
413 		u32 copy, osize;
414 
415 		if (sk->sk_err) {
416 			err = -sk->sk_err;
417 			goto out_err;
418 		}
419 
420 		copy = msg_data_left(msg);
421 		if (!sk_stream_memory_free(sk))
422 			goto wait_for_sndbuf;
423 		if (psock->cork) {
424 			msg_tx = psock->cork;
425 		} else {
426 			msg_tx = &tmp;
427 			sk_msg_init(msg_tx);
428 		}
429 
430 		osize = msg_tx->sg.size;
431 		err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
432 		if (err) {
433 			if (err != -ENOSPC)
434 				goto wait_for_memory;
435 			enospc = true;
436 			copy = msg_tx->sg.size - osize;
437 		}
438 
439 		err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
440 					       copy);
441 		if (err < 0) {
442 			sk_msg_trim(sk, msg_tx, osize);
443 			goto out_err;
444 		}
445 
446 		copied += copy;
447 		if (psock->cork_bytes) {
448 			if (size > psock->cork_bytes)
449 				psock->cork_bytes = 0;
450 			else
451 				psock->cork_bytes -= size;
452 			if (psock->cork_bytes && !enospc)
453 				goto out_err;
454 			/* All cork bytes are accounted, rerun the prog. */
455 			psock->eval = __SK_NONE;
456 			psock->cork_bytes = 0;
457 		}
458 
459 		err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
460 		if (unlikely(err < 0))
461 			goto out_err;
462 		continue;
463 wait_for_sndbuf:
464 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
465 wait_for_memory:
466 		err = sk_stream_wait_memory(sk, &timeo);
467 		if (err) {
468 			if (msg_tx && msg_tx != psock->cork)
469 				sk_msg_free(sk, msg_tx);
470 			goto out_err;
471 		}
472 	}
473 out_err:
474 	if (err < 0)
475 		err = sk_stream_error(sk, msg->msg_flags, err);
476 	release_sock(sk);
477 	sk_psock_put(sk, psock);
478 	return copied ? copied : err;
479 }
480 
481 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
482 			    size_t size, int flags)
483 {
484 	struct sk_msg tmp, *msg = NULL;
485 	int err = 0, copied = 0;
486 	struct sk_psock *psock;
487 	bool enospc = false;
488 
489 	psock = sk_psock_get(sk);
490 	if (unlikely(!psock))
491 		return tcp_sendpage(sk, page, offset, size, flags);
492 
493 	lock_sock(sk);
494 	if (psock->cork) {
495 		msg = psock->cork;
496 	} else {
497 		msg = &tmp;
498 		sk_msg_init(msg);
499 	}
500 
501 	/* Catch case where ring is full and sendpage is stalled. */
502 	if (unlikely(sk_msg_full(msg)))
503 		goto out_err;
504 
505 	sk_msg_page_add(msg, page, size, offset);
506 	sk_mem_charge(sk, size);
507 	copied = size;
508 	if (sk_msg_full(msg))
509 		enospc = true;
510 	if (psock->cork_bytes) {
511 		if (size > psock->cork_bytes)
512 			psock->cork_bytes = 0;
513 		else
514 			psock->cork_bytes -= size;
515 		if (psock->cork_bytes && !enospc)
516 			goto out_err;
517 		/* All cork bytes are accounted, rerun the prog. */
518 		psock->eval = __SK_NONE;
519 		psock->cork_bytes = 0;
520 	}
521 
522 	err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
523 out_err:
524 	release_sock(sk);
525 	sk_psock_put(sk, psock);
526 	return copied ? copied : err;
527 }
528 
529 enum {
530 	TCP_BPF_IPV4,
531 	TCP_BPF_IPV6,
532 	TCP_BPF_NUM_PROTS,
533 };
534 
535 enum {
536 	TCP_BPF_BASE,
537 	TCP_BPF_TX,
538 	TCP_BPF_RX,
539 	TCP_BPF_TXRX,
540 	TCP_BPF_NUM_CFGS,
541 };
542 
543 static struct proto *tcpv6_prot_saved __read_mostly;
544 static DEFINE_SPINLOCK(tcpv6_prot_lock);
545 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
546 
547 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
548 				   struct proto *base)
549 {
550 	prot[TCP_BPF_BASE]			= *base;
551 	prot[TCP_BPF_BASE].destroy		= sock_map_destroy;
552 	prot[TCP_BPF_BASE].close		= sock_map_close;
553 	prot[TCP_BPF_BASE].recvmsg		= tcp_bpf_recvmsg;
554 	prot[TCP_BPF_BASE].sock_is_readable	= sk_msg_is_readable;
555 
556 	prot[TCP_BPF_TX]			= prot[TCP_BPF_BASE];
557 	prot[TCP_BPF_TX].sendmsg		= tcp_bpf_sendmsg;
558 	prot[TCP_BPF_TX].sendpage		= tcp_bpf_sendpage;
559 
560 	prot[TCP_BPF_RX]			= prot[TCP_BPF_BASE];
561 	prot[TCP_BPF_RX].recvmsg		= tcp_bpf_recvmsg_parser;
562 
563 	prot[TCP_BPF_TXRX]			= prot[TCP_BPF_TX];
564 	prot[TCP_BPF_TXRX].recvmsg		= tcp_bpf_recvmsg_parser;
565 }
566 
567 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
568 {
569 	if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
570 		spin_lock_bh(&tcpv6_prot_lock);
571 		if (likely(ops != tcpv6_prot_saved)) {
572 			tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
573 			smp_store_release(&tcpv6_prot_saved, ops);
574 		}
575 		spin_unlock_bh(&tcpv6_prot_lock);
576 	}
577 }
578 
579 static int __init tcp_bpf_v4_build_proto(void)
580 {
581 	tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
582 	return 0;
583 }
584 late_initcall(tcp_bpf_v4_build_proto);
585 
586 static int tcp_bpf_assert_proto_ops(struct proto *ops)
587 {
588 	/* In order to avoid retpoline, we make assumptions when we call
589 	 * into ops if e.g. a psock is not present. Make sure they are
590 	 * indeed valid assumptions.
591 	 */
592 	return ops->recvmsg  == tcp_recvmsg &&
593 	       ops->sendmsg  == tcp_sendmsg &&
594 	       ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
595 }
596 
597 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
598 {
599 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
600 	int config = psock->progs.msg_parser   ? TCP_BPF_TX   : TCP_BPF_BASE;
601 
602 	if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
603 		config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
604 	}
605 
606 	if (restore) {
607 		if (inet_csk_has_ulp(sk)) {
608 			/* TLS does not have an unhash proto in SW cases,
609 			 * but we need to ensure we stop using the sock_map
610 			 * unhash routine because the associated psock is being
611 			 * removed. So use the original unhash handler.
612 			 */
613 			WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
614 			tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
615 		} else {
616 			sk->sk_write_space = psock->saved_write_space;
617 			/* Pairs with lockless read in sk_clone_lock() */
618 			sock_replace_proto(sk, psock->sk_proto);
619 		}
620 		return 0;
621 	}
622 
623 	if (sk->sk_family == AF_INET6) {
624 		if (tcp_bpf_assert_proto_ops(psock->sk_proto))
625 			return -EINVAL;
626 
627 		tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
628 	}
629 
630 	/* Pairs with lockless read in sk_clone_lock() */
631 	sock_replace_proto(sk, &tcp_bpf_prots[family][config]);
632 	return 0;
633 }
634 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
635 
636 /* If a child got cloned from a listening socket that had tcp_bpf
637  * protocol callbacks installed, we need to restore the callbacks to
638  * the default ones because the child does not inherit the psock state
639  * that tcp_bpf callbacks expect.
640  */
641 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
642 {
643 	struct proto *prot = newsk->sk_prot;
644 
645 	if (is_insidevar(prot, tcp_bpf_prots))
646 		newsk->sk_prot = sk->sk_prot_creator;
647 }
648 #endif /* CONFIG_BPF_SYSCALL */
649