xref: /openbmc/linux/net/ipv4/tcp_bpf.c (revision 63705da3)
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 		timeo = sock_rcvtimeo(sk, nonblock);
200 		data = tcp_msg_wait_data(sk, psock, timeo);
201 		if (data && !sk_psock_queue_empty(psock))
202 			goto msg_bytes_ready;
203 		copied = -EAGAIN;
204 	}
205 	release_sock(sk);
206 	sk_psock_put(sk, psock);
207 	return copied;
208 }
209 
210 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
211 		    int nonblock, int flags, int *addr_len)
212 {
213 	struct sk_psock *psock;
214 	int copied, ret;
215 
216 	if (unlikely(flags & MSG_ERRQUEUE))
217 		return inet_recv_error(sk, msg, len, addr_len);
218 
219 	psock = sk_psock_get(sk);
220 	if (unlikely(!psock))
221 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
222 	if (!skb_queue_empty(&sk->sk_receive_queue) &&
223 	    sk_psock_queue_empty(psock)) {
224 		sk_psock_put(sk, psock);
225 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
226 	}
227 	lock_sock(sk);
228 msg_bytes_ready:
229 	copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
230 	if (!copied) {
231 		long timeo;
232 		int data;
233 
234 		timeo = sock_rcvtimeo(sk, nonblock);
235 		data = tcp_msg_wait_data(sk, psock, timeo);
236 		if (data) {
237 			if (!sk_psock_queue_empty(psock))
238 				goto msg_bytes_ready;
239 			release_sock(sk);
240 			sk_psock_put(sk, psock);
241 			return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
242 		}
243 		copied = -EAGAIN;
244 	}
245 	ret = copied;
246 	release_sock(sk);
247 	sk_psock_put(sk, psock);
248 	return ret;
249 }
250 
251 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
252 				struct sk_msg *msg, int *copied, int flags)
253 {
254 	bool cork = false, enospc = sk_msg_full(msg);
255 	struct sock *sk_redir;
256 	u32 tosend, delta = 0;
257 	u32 eval = __SK_NONE;
258 	int ret;
259 
260 more_data:
261 	if (psock->eval == __SK_NONE) {
262 		/* Track delta in msg size to add/subtract it on SK_DROP from
263 		 * returned to user copied size. This ensures user doesn't
264 		 * get a positive return code with msg_cut_data and SK_DROP
265 		 * verdict.
266 		 */
267 		delta = msg->sg.size;
268 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
269 		delta -= msg->sg.size;
270 	}
271 
272 	if (msg->cork_bytes &&
273 	    msg->cork_bytes > msg->sg.size && !enospc) {
274 		psock->cork_bytes = msg->cork_bytes - msg->sg.size;
275 		if (!psock->cork) {
276 			psock->cork = kzalloc(sizeof(*psock->cork),
277 					      GFP_ATOMIC | __GFP_NOWARN);
278 			if (!psock->cork)
279 				return -ENOMEM;
280 		}
281 		memcpy(psock->cork, msg, sizeof(*msg));
282 		return 0;
283 	}
284 
285 	tosend = msg->sg.size;
286 	if (psock->apply_bytes && psock->apply_bytes < tosend)
287 		tosend = psock->apply_bytes;
288 
289 	switch (psock->eval) {
290 	case __SK_PASS:
291 		ret = tcp_bpf_push(sk, msg, tosend, flags, true);
292 		if (unlikely(ret)) {
293 			*copied -= sk_msg_free(sk, msg);
294 			break;
295 		}
296 		sk_msg_apply_bytes(psock, tosend);
297 		break;
298 	case __SK_REDIRECT:
299 		sk_redir = psock->sk_redir;
300 		sk_msg_apply_bytes(psock, tosend);
301 		if (!psock->apply_bytes) {
302 			/* Clean up before releasing the sock lock. */
303 			eval = psock->eval;
304 			psock->eval = __SK_NONE;
305 			psock->sk_redir = NULL;
306 		}
307 		if (psock->cork) {
308 			cork = true;
309 			psock->cork = NULL;
310 		}
311 		sk_msg_return(sk, msg, tosend);
312 		release_sock(sk);
313 
314 		ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
315 
316 		if (eval == __SK_REDIRECT)
317 			sock_put(sk_redir);
318 
319 		lock_sock(sk);
320 		if (unlikely(ret < 0)) {
321 			int free = sk_msg_free_nocharge(sk, msg);
322 
323 			if (!cork)
324 				*copied -= free;
325 		}
326 		if (cork) {
327 			sk_msg_free(sk, msg);
328 			kfree(msg);
329 			msg = NULL;
330 			ret = 0;
331 		}
332 		break;
333 	case __SK_DROP:
334 	default:
335 		sk_msg_free_partial(sk, msg, tosend);
336 		sk_msg_apply_bytes(psock, tosend);
337 		*copied -= (tosend + delta);
338 		return -EACCES;
339 	}
340 
341 	if (likely(!ret)) {
342 		if (!psock->apply_bytes) {
343 			psock->eval =  __SK_NONE;
344 			if (psock->sk_redir) {
345 				sock_put(psock->sk_redir);
346 				psock->sk_redir = NULL;
347 			}
348 		}
349 		if (msg &&
350 		    msg->sg.data[msg->sg.start].page_link &&
351 		    msg->sg.data[msg->sg.start].length)
352 			goto more_data;
353 	}
354 	return ret;
355 }
356 
357 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
358 {
359 	struct sk_msg tmp, *msg_tx = NULL;
360 	int copied = 0, err = 0;
361 	struct sk_psock *psock;
362 	long timeo;
363 	int flags;
364 
365 	/* Don't let internal do_tcp_sendpages() flags through */
366 	flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
367 	flags |= MSG_NO_SHARED_FRAGS;
368 
369 	psock = sk_psock_get(sk);
370 	if (unlikely(!psock))
371 		return tcp_sendmsg(sk, msg, size);
372 
373 	lock_sock(sk);
374 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
375 	while (msg_data_left(msg)) {
376 		bool enospc = false;
377 		u32 copy, osize;
378 
379 		if (sk->sk_err) {
380 			err = -sk->sk_err;
381 			goto out_err;
382 		}
383 
384 		copy = msg_data_left(msg);
385 		if (!sk_stream_memory_free(sk))
386 			goto wait_for_sndbuf;
387 		if (psock->cork) {
388 			msg_tx = psock->cork;
389 		} else {
390 			msg_tx = &tmp;
391 			sk_msg_init(msg_tx);
392 		}
393 
394 		osize = msg_tx->sg.size;
395 		err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
396 		if (err) {
397 			if (err != -ENOSPC)
398 				goto wait_for_memory;
399 			enospc = true;
400 			copy = msg_tx->sg.size - osize;
401 		}
402 
403 		err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
404 					       copy);
405 		if (err < 0) {
406 			sk_msg_trim(sk, msg_tx, osize);
407 			goto out_err;
408 		}
409 
410 		copied += copy;
411 		if (psock->cork_bytes) {
412 			if (size > psock->cork_bytes)
413 				psock->cork_bytes = 0;
414 			else
415 				psock->cork_bytes -= size;
416 			if (psock->cork_bytes && !enospc)
417 				goto out_err;
418 			/* All cork bytes are accounted, rerun the prog. */
419 			psock->eval = __SK_NONE;
420 			psock->cork_bytes = 0;
421 		}
422 
423 		err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
424 		if (unlikely(err < 0))
425 			goto out_err;
426 		continue;
427 wait_for_sndbuf:
428 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
429 wait_for_memory:
430 		err = sk_stream_wait_memory(sk, &timeo);
431 		if (err) {
432 			if (msg_tx && msg_tx != psock->cork)
433 				sk_msg_free(sk, msg_tx);
434 			goto out_err;
435 		}
436 	}
437 out_err:
438 	if (err < 0)
439 		err = sk_stream_error(sk, msg->msg_flags, err);
440 	release_sock(sk);
441 	sk_psock_put(sk, psock);
442 	return copied ? copied : err;
443 }
444 
445 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
446 			    size_t size, int flags)
447 {
448 	struct sk_msg tmp, *msg = NULL;
449 	int err = 0, copied = 0;
450 	struct sk_psock *psock;
451 	bool enospc = false;
452 
453 	psock = sk_psock_get(sk);
454 	if (unlikely(!psock))
455 		return tcp_sendpage(sk, page, offset, size, flags);
456 
457 	lock_sock(sk);
458 	if (psock->cork) {
459 		msg = psock->cork;
460 	} else {
461 		msg = &tmp;
462 		sk_msg_init(msg);
463 	}
464 
465 	/* Catch case where ring is full and sendpage is stalled. */
466 	if (unlikely(sk_msg_full(msg)))
467 		goto out_err;
468 
469 	sk_msg_page_add(msg, page, size, offset);
470 	sk_mem_charge(sk, size);
471 	copied = size;
472 	if (sk_msg_full(msg))
473 		enospc = true;
474 	if (psock->cork_bytes) {
475 		if (size > psock->cork_bytes)
476 			psock->cork_bytes = 0;
477 		else
478 			psock->cork_bytes -= size;
479 		if (psock->cork_bytes && !enospc)
480 			goto out_err;
481 		/* All cork bytes are accounted, rerun the prog. */
482 		psock->eval = __SK_NONE;
483 		psock->cork_bytes = 0;
484 	}
485 
486 	err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
487 out_err:
488 	release_sock(sk);
489 	sk_psock_put(sk, psock);
490 	return copied ? copied : err;
491 }
492 
493 enum {
494 	TCP_BPF_IPV4,
495 	TCP_BPF_IPV6,
496 	TCP_BPF_NUM_PROTS,
497 };
498 
499 enum {
500 	TCP_BPF_BASE,
501 	TCP_BPF_TX,
502 	TCP_BPF_RX,
503 	TCP_BPF_TXRX,
504 	TCP_BPF_NUM_CFGS,
505 };
506 
507 static struct proto *tcpv6_prot_saved __read_mostly;
508 static DEFINE_SPINLOCK(tcpv6_prot_lock);
509 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
510 
511 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
512 				   struct proto *base)
513 {
514 	prot[TCP_BPF_BASE]			= *base;
515 	prot[TCP_BPF_BASE].close		= sock_map_close;
516 	prot[TCP_BPF_BASE].recvmsg		= tcp_bpf_recvmsg;
517 	prot[TCP_BPF_BASE].sock_is_readable	= sk_msg_is_readable;
518 
519 	prot[TCP_BPF_TX]			= prot[TCP_BPF_BASE];
520 	prot[TCP_BPF_TX].sendmsg		= tcp_bpf_sendmsg;
521 	prot[TCP_BPF_TX].sendpage		= tcp_bpf_sendpage;
522 
523 	prot[TCP_BPF_RX]			= prot[TCP_BPF_BASE];
524 	prot[TCP_BPF_RX].recvmsg		= tcp_bpf_recvmsg_parser;
525 
526 	prot[TCP_BPF_TXRX]			= prot[TCP_BPF_TX];
527 	prot[TCP_BPF_TXRX].recvmsg		= tcp_bpf_recvmsg_parser;
528 }
529 
530 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
531 {
532 	if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
533 		spin_lock_bh(&tcpv6_prot_lock);
534 		if (likely(ops != tcpv6_prot_saved)) {
535 			tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
536 			smp_store_release(&tcpv6_prot_saved, ops);
537 		}
538 		spin_unlock_bh(&tcpv6_prot_lock);
539 	}
540 }
541 
542 static int __init tcp_bpf_v4_build_proto(void)
543 {
544 	tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
545 	return 0;
546 }
547 late_initcall(tcp_bpf_v4_build_proto);
548 
549 static int tcp_bpf_assert_proto_ops(struct proto *ops)
550 {
551 	/* In order to avoid retpoline, we make assumptions when we call
552 	 * into ops if e.g. a psock is not present. Make sure they are
553 	 * indeed valid assumptions.
554 	 */
555 	return ops->recvmsg  == tcp_recvmsg &&
556 	       ops->sendmsg  == tcp_sendmsg &&
557 	       ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
558 }
559 
560 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
561 {
562 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
563 	int config = psock->progs.msg_parser   ? TCP_BPF_TX   : TCP_BPF_BASE;
564 
565 	if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
566 		config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
567 	}
568 
569 	if (restore) {
570 		if (inet_csk_has_ulp(sk)) {
571 			/* TLS does not have an unhash proto in SW cases,
572 			 * but we need to ensure we stop using the sock_map
573 			 * unhash routine because the associated psock is being
574 			 * removed. So use the original unhash handler.
575 			 */
576 			WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
577 			tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
578 		} else {
579 			sk->sk_write_space = psock->saved_write_space;
580 			/* Pairs with lockless read in sk_clone_lock() */
581 			WRITE_ONCE(sk->sk_prot, psock->sk_proto);
582 		}
583 		return 0;
584 	}
585 
586 	if (inet_csk_has_ulp(sk))
587 		return -EINVAL;
588 
589 	if (sk->sk_family == AF_INET6) {
590 		if (tcp_bpf_assert_proto_ops(psock->sk_proto))
591 			return -EINVAL;
592 
593 		tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
594 	}
595 
596 	/* Pairs with lockless read in sk_clone_lock() */
597 	WRITE_ONCE(sk->sk_prot, &tcp_bpf_prots[family][config]);
598 	return 0;
599 }
600 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
601 
602 /* If a child got cloned from a listening socket that had tcp_bpf
603  * protocol callbacks installed, we need to restore the callbacks to
604  * the default ones because the child does not inherit the psock state
605  * that tcp_bpf callbacks expect.
606  */
607 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
608 {
609 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
610 	struct proto *prot = newsk->sk_prot;
611 
612 	if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
613 		newsk->sk_prot = sk->sk_prot_creator;
614 }
615 #endif /* CONFIG_BPF_SYSCALL */
616