xref: /openbmc/linux/net/ipv4/tcp_bpf.c (revision 8931ddd8)
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 		sk_msg_free(sk, msg);
143 		return 0;
144 	}
145 	ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
146 			tcp_bpf_push_locked(sk, msg, bytes, flags, false);
147 	sk_psock_put(sk, psock);
148 	return ret;
149 }
150 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
151 
152 #ifdef CONFIG_BPF_SYSCALL
153 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
154 			     long timeo)
155 {
156 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
157 	int ret = 0;
158 
159 	if (sk->sk_shutdown & RCV_SHUTDOWN)
160 		return 1;
161 
162 	if (!timeo)
163 		return ret;
164 
165 	add_wait_queue(sk_sleep(sk), &wait);
166 	sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
167 	ret = sk_wait_event(sk, &timeo,
168 			    !list_empty(&psock->ingress_msg) ||
169 			    !skb_queue_empty(&sk->sk_receive_queue), &wait);
170 	sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
171 	remove_wait_queue(sk_sleep(sk), &wait);
172 	return ret;
173 }
174 
175 static int tcp_bpf_recvmsg_parser(struct sock *sk,
176 				  struct msghdr *msg,
177 				  size_t len,
178 				  int nonblock,
179 				  int flags,
180 				  int *addr_len)
181 {
182 	struct sk_psock *psock;
183 	int copied;
184 
185 	if (unlikely(flags & MSG_ERRQUEUE))
186 		return inet_recv_error(sk, msg, len, addr_len);
187 
188 	psock = sk_psock_get(sk);
189 	if (unlikely(!psock))
190 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
191 
192 	lock_sock(sk);
193 msg_bytes_ready:
194 	copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
195 	if (!copied) {
196 		long timeo;
197 		int data;
198 
199 		if (sock_flag(sk, SOCK_DONE))
200 			goto out;
201 
202 		if (sk->sk_err) {
203 			copied = sock_error(sk);
204 			goto out;
205 		}
206 
207 		if (sk->sk_shutdown & RCV_SHUTDOWN)
208 			goto out;
209 
210 		if (sk->sk_state == TCP_CLOSE) {
211 			copied = -ENOTCONN;
212 			goto out;
213 		}
214 
215 		timeo = sock_rcvtimeo(sk, nonblock);
216 		if (!timeo) {
217 			copied = -EAGAIN;
218 			goto out;
219 		}
220 
221 		if (signal_pending(current)) {
222 			copied = sock_intr_errno(timeo);
223 			goto out;
224 		}
225 
226 		data = tcp_msg_wait_data(sk, psock, timeo);
227 		if (data && !sk_psock_queue_empty(psock))
228 			goto msg_bytes_ready;
229 		copied = -EAGAIN;
230 	}
231 out:
232 	release_sock(sk);
233 	sk_psock_put(sk, psock);
234 	return copied;
235 }
236 
237 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
238 		    int nonblock, int flags, int *addr_len)
239 {
240 	struct sk_psock *psock;
241 	int copied, ret;
242 
243 	if (unlikely(flags & MSG_ERRQUEUE))
244 		return inet_recv_error(sk, msg, len, addr_len);
245 
246 	psock = sk_psock_get(sk);
247 	if (unlikely(!psock))
248 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
249 	if (!skb_queue_empty(&sk->sk_receive_queue) &&
250 	    sk_psock_queue_empty(psock)) {
251 		sk_psock_put(sk, psock);
252 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
253 	}
254 	lock_sock(sk);
255 msg_bytes_ready:
256 	copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
257 	if (!copied) {
258 		long timeo;
259 		int data;
260 
261 		timeo = sock_rcvtimeo(sk, nonblock);
262 		data = tcp_msg_wait_data(sk, psock, timeo);
263 		if (data) {
264 			if (!sk_psock_queue_empty(psock))
265 				goto msg_bytes_ready;
266 			release_sock(sk);
267 			sk_psock_put(sk, psock);
268 			return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
269 		}
270 		copied = -EAGAIN;
271 	}
272 	ret = copied;
273 	release_sock(sk);
274 	sk_psock_put(sk, psock);
275 	return ret;
276 }
277 
278 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
279 				struct sk_msg *msg, int *copied, int flags)
280 {
281 	bool cork = false, enospc = sk_msg_full(msg);
282 	struct sock *sk_redir;
283 	u32 tosend, delta = 0;
284 	u32 eval = __SK_NONE;
285 	int ret;
286 
287 more_data:
288 	if (psock->eval == __SK_NONE) {
289 		/* Track delta in msg size to add/subtract it on SK_DROP from
290 		 * returned to user copied size. This ensures user doesn't
291 		 * get a positive return code with msg_cut_data and SK_DROP
292 		 * verdict.
293 		 */
294 		delta = msg->sg.size;
295 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
296 		delta -= msg->sg.size;
297 	}
298 
299 	if (msg->cork_bytes &&
300 	    msg->cork_bytes > msg->sg.size && !enospc) {
301 		psock->cork_bytes = msg->cork_bytes - msg->sg.size;
302 		if (!psock->cork) {
303 			psock->cork = kzalloc(sizeof(*psock->cork),
304 					      GFP_ATOMIC | __GFP_NOWARN);
305 			if (!psock->cork)
306 				return -ENOMEM;
307 		}
308 		memcpy(psock->cork, msg, sizeof(*msg));
309 		return 0;
310 	}
311 
312 	tosend = msg->sg.size;
313 	if (psock->apply_bytes && psock->apply_bytes < tosend)
314 		tosend = psock->apply_bytes;
315 
316 	switch (psock->eval) {
317 	case __SK_PASS:
318 		ret = tcp_bpf_push(sk, msg, tosend, flags, true);
319 		if (unlikely(ret)) {
320 			*copied -= sk_msg_free(sk, msg);
321 			break;
322 		}
323 		sk_msg_apply_bytes(psock, tosend);
324 		break;
325 	case __SK_REDIRECT:
326 		sk_redir = psock->sk_redir;
327 		sk_msg_apply_bytes(psock, tosend);
328 		if (!psock->apply_bytes) {
329 			/* Clean up before releasing the sock lock. */
330 			eval = psock->eval;
331 			psock->eval = __SK_NONE;
332 			psock->sk_redir = NULL;
333 		}
334 		if (psock->cork) {
335 			cork = true;
336 			psock->cork = NULL;
337 		}
338 		sk_msg_return(sk, msg, tosend);
339 		release_sock(sk);
340 
341 		ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
342 
343 		if (eval == __SK_REDIRECT)
344 			sock_put(sk_redir);
345 
346 		lock_sock(sk);
347 		if (unlikely(ret < 0)) {
348 			int free = sk_msg_free_nocharge(sk, msg);
349 
350 			if (!cork)
351 				*copied -= free;
352 		}
353 		if (cork) {
354 			sk_msg_free(sk, msg);
355 			kfree(msg);
356 			msg = NULL;
357 			ret = 0;
358 		}
359 		break;
360 	case __SK_DROP:
361 	default:
362 		sk_msg_free_partial(sk, msg, tosend);
363 		sk_msg_apply_bytes(psock, tosend);
364 		*copied -= (tosend + delta);
365 		return -EACCES;
366 	}
367 
368 	if (likely(!ret)) {
369 		if (!psock->apply_bytes) {
370 			psock->eval =  __SK_NONE;
371 			if (psock->sk_redir) {
372 				sock_put(psock->sk_redir);
373 				psock->sk_redir = NULL;
374 			}
375 		}
376 		if (msg &&
377 		    msg->sg.data[msg->sg.start].page_link &&
378 		    msg->sg.data[msg->sg.start].length)
379 			goto more_data;
380 	}
381 	return ret;
382 }
383 
384 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
385 {
386 	struct sk_msg tmp, *msg_tx = NULL;
387 	int copied = 0, err = 0;
388 	struct sk_psock *psock;
389 	long timeo;
390 	int flags;
391 
392 	/* Don't let internal do_tcp_sendpages() flags through */
393 	flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
394 	flags |= MSG_NO_SHARED_FRAGS;
395 
396 	psock = sk_psock_get(sk);
397 	if (unlikely(!psock))
398 		return tcp_sendmsg(sk, msg, size);
399 
400 	lock_sock(sk);
401 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
402 	while (msg_data_left(msg)) {
403 		bool enospc = false;
404 		u32 copy, osize;
405 
406 		if (sk->sk_err) {
407 			err = -sk->sk_err;
408 			goto out_err;
409 		}
410 
411 		copy = msg_data_left(msg);
412 		if (!sk_stream_memory_free(sk))
413 			goto wait_for_sndbuf;
414 		if (psock->cork) {
415 			msg_tx = psock->cork;
416 		} else {
417 			msg_tx = &tmp;
418 			sk_msg_init(msg_tx);
419 		}
420 
421 		osize = msg_tx->sg.size;
422 		err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
423 		if (err) {
424 			if (err != -ENOSPC)
425 				goto wait_for_memory;
426 			enospc = true;
427 			copy = msg_tx->sg.size - osize;
428 		}
429 
430 		err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
431 					       copy);
432 		if (err < 0) {
433 			sk_msg_trim(sk, msg_tx, osize);
434 			goto out_err;
435 		}
436 
437 		copied += copy;
438 		if (psock->cork_bytes) {
439 			if (size > psock->cork_bytes)
440 				psock->cork_bytes = 0;
441 			else
442 				psock->cork_bytes -= size;
443 			if (psock->cork_bytes && !enospc)
444 				goto out_err;
445 			/* All cork bytes are accounted, rerun the prog. */
446 			psock->eval = __SK_NONE;
447 			psock->cork_bytes = 0;
448 		}
449 
450 		err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
451 		if (unlikely(err < 0))
452 			goto out_err;
453 		continue;
454 wait_for_sndbuf:
455 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
456 wait_for_memory:
457 		err = sk_stream_wait_memory(sk, &timeo);
458 		if (err) {
459 			if (msg_tx && msg_tx != psock->cork)
460 				sk_msg_free(sk, msg_tx);
461 			goto out_err;
462 		}
463 	}
464 out_err:
465 	if (err < 0)
466 		err = sk_stream_error(sk, msg->msg_flags, err);
467 	release_sock(sk);
468 	sk_psock_put(sk, psock);
469 	return copied ? copied : err;
470 }
471 
472 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
473 			    size_t size, int flags)
474 {
475 	struct sk_msg tmp, *msg = NULL;
476 	int err = 0, copied = 0;
477 	struct sk_psock *psock;
478 	bool enospc = false;
479 
480 	psock = sk_psock_get(sk);
481 	if (unlikely(!psock))
482 		return tcp_sendpage(sk, page, offset, size, flags);
483 
484 	lock_sock(sk);
485 	if (psock->cork) {
486 		msg = psock->cork;
487 	} else {
488 		msg = &tmp;
489 		sk_msg_init(msg);
490 	}
491 
492 	/* Catch case where ring is full and sendpage is stalled. */
493 	if (unlikely(sk_msg_full(msg)))
494 		goto out_err;
495 
496 	sk_msg_page_add(msg, page, size, offset);
497 	sk_mem_charge(sk, size);
498 	copied = size;
499 	if (sk_msg_full(msg))
500 		enospc = true;
501 	if (psock->cork_bytes) {
502 		if (size > psock->cork_bytes)
503 			psock->cork_bytes = 0;
504 		else
505 			psock->cork_bytes -= size;
506 		if (psock->cork_bytes && !enospc)
507 			goto out_err;
508 		/* All cork bytes are accounted, rerun the prog. */
509 		psock->eval = __SK_NONE;
510 		psock->cork_bytes = 0;
511 	}
512 
513 	err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
514 out_err:
515 	release_sock(sk);
516 	sk_psock_put(sk, psock);
517 	return copied ? copied : err;
518 }
519 
520 enum {
521 	TCP_BPF_IPV4,
522 	TCP_BPF_IPV6,
523 	TCP_BPF_NUM_PROTS,
524 };
525 
526 enum {
527 	TCP_BPF_BASE,
528 	TCP_BPF_TX,
529 	TCP_BPF_RX,
530 	TCP_BPF_TXRX,
531 	TCP_BPF_NUM_CFGS,
532 };
533 
534 static struct proto *tcpv6_prot_saved __read_mostly;
535 static DEFINE_SPINLOCK(tcpv6_prot_lock);
536 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
537 
538 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
539 				   struct proto *base)
540 {
541 	prot[TCP_BPF_BASE]			= *base;
542 	prot[TCP_BPF_BASE].close		= sock_map_close;
543 	prot[TCP_BPF_BASE].recvmsg		= tcp_bpf_recvmsg;
544 	prot[TCP_BPF_BASE].sock_is_readable	= sk_msg_is_readable;
545 
546 	prot[TCP_BPF_TX]			= prot[TCP_BPF_BASE];
547 	prot[TCP_BPF_TX].sendmsg		= tcp_bpf_sendmsg;
548 	prot[TCP_BPF_TX].sendpage		= tcp_bpf_sendpage;
549 
550 	prot[TCP_BPF_RX]			= prot[TCP_BPF_BASE];
551 	prot[TCP_BPF_RX].recvmsg		= tcp_bpf_recvmsg_parser;
552 
553 	prot[TCP_BPF_TXRX]			= prot[TCP_BPF_TX];
554 	prot[TCP_BPF_TXRX].recvmsg		= tcp_bpf_recvmsg_parser;
555 }
556 
557 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
558 {
559 	if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
560 		spin_lock_bh(&tcpv6_prot_lock);
561 		if (likely(ops != tcpv6_prot_saved)) {
562 			tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
563 			smp_store_release(&tcpv6_prot_saved, ops);
564 		}
565 		spin_unlock_bh(&tcpv6_prot_lock);
566 	}
567 }
568 
569 static int __init tcp_bpf_v4_build_proto(void)
570 {
571 	tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
572 	return 0;
573 }
574 late_initcall(tcp_bpf_v4_build_proto);
575 
576 static int tcp_bpf_assert_proto_ops(struct proto *ops)
577 {
578 	/* In order to avoid retpoline, we make assumptions when we call
579 	 * into ops if e.g. a psock is not present. Make sure they are
580 	 * indeed valid assumptions.
581 	 */
582 	return ops->recvmsg  == tcp_recvmsg &&
583 	       ops->sendmsg  == tcp_sendmsg &&
584 	       ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
585 }
586 
587 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
588 {
589 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
590 	int config = psock->progs.msg_parser   ? TCP_BPF_TX   : TCP_BPF_BASE;
591 
592 	if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
593 		config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
594 	}
595 
596 	if (restore) {
597 		if (inet_csk_has_ulp(sk)) {
598 			/* TLS does not have an unhash proto in SW cases,
599 			 * but we need to ensure we stop using the sock_map
600 			 * unhash routine because the associated psock is being
601 			 * removed. So use the original unhash handler.
602 			 */
603 			WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
604 			tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
605 		} else {
606 			sk->sk_write_space = psock->saved_write_space;
607 			/* Pairs with lockless read in sk_clone_lock() */
608 			WRITE_ONCE(sk->sk_prot, psock->sk_proto);
609 		}
610 		return 0;
611 	}
612 
613 	if (inet_csk_has_ulp(sk))
614 		return -EINVAL;
615 
616 	if (sk->sk_family == AF_INET6) {
617 		if (tcp_bpf_assert_proto_ops(psock->sk_proto))
618 			return -EINVAL;
619 
620 		tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
621 	}
622 
623 	/* Pairs with lockless read in sk_clone_lock() */
624 	WRITE_ONCE(sk->sk_prot, &tcp_bpf_prots[family][config]);
625 	return 0;
626 }
627 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
628 
629 /* If a child got cloned from a listening socket that had tcp_bpf
630  * protocol callbacks installed, we need to restore the callbacks to
631  * the default ones because the child does not inherit the psock state
632  * that tcp_bpf callbacks expect.
633  */
634 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
635 {
636 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
637 	struct proto *prot = newsk->sk_prot;
638 
639 	if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
640 		newsk->sk_prot = sk->sk_prot_creator;
641 }
642 #endif /* CONFIG_BPF_SYSCALL */
643