xref: /openbmc/linux/net/ipv4/tcp_bpf.c (revision 97e6ea6d) 
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 bool tcp_bpf_stream_read(const struct sock *sk)
154 {
155 	struct sk_psock *psock;
156 	bool empty = true;
157 
158 	rcu_read_lock();
159 	psock = sk_psock(sk);
160 	if (likely(psock))
161 		empty = list_empty(&psock->ingress_msg);
162 	rcu_read_unlock();
163 	return !empty;
164 }
165 
166 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
167 			     long timeo)
168 {
169 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
170 	int ret = 0;
171 
172 	if (sk->sk_shutdown & RCV_SHUTDOWN)
173 		return 1;
174 
175 	if (!timeo)
176 		return ret;
177 
178 	add_wait_queue(sk_sleep(sk), &wait);
179 	sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
180 	ret = sk_wait_event(sk, &timeo,
181 			    !list_empty(&psock->ingress_msg) ||
182 			    !skb_queue_empty(&sk->sk_receive_queue), &wait);
183 	sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
184 	remove_wait_queue(sk_sleep(sk), &wait);
185 	return ret;
186 }
187 
188 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
189 		    int nonblock, int flags, int *addr_len)
190 {
191 	struct sk_psock *psock;
192 	int copied, ret;
193 
194 	if (unlikely(flags & MSG_ERRQUEUE))
195 		return inet_recv_error(sk, msg, len, addr_len);
196 
197 	psock = sk_psock_get(sk);
198 	if (unlikely(!psock))
199 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
200 	if (!skb_queue_empty(&sk->sk_receive_queue) &&
201 	    sk_psock_queue_empty(psock)) {
202 		sk_psock_put(sk, psock);
203 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
204 	}
205 	lock_sock(sk);
206 msg_bytes_ready:
207 	copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
208 	if (!copied) {
209 		long timeo;
210 		int data;
211 
212 		timeo = sock_rcvtimeo(sk, nonblock);
213 		data = tcp_msg_wait_data(sk, psock, timeo);
214 		if (data) {
215 			if (!sk_psock_queue_empty(psock))
216 				goto msg_bytes_ready;
217 			release_sock(sk);
218 			sk_psock_put(sk, psock);
219 			return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
220 		}
221 		copied = -EAGAIN;
222 	}
223 	ret = copied;
224 	release_sock(sk);
225 	sk_psock_put(sk, psock);
226 	return ret;
227 }
228 
229 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
230 				struct sk_msg *msg, int *copied, int flags)
231 {
232 	bool cork = false, enospc = sk_msg_full(msg);
233 	struct sock *sk_redir;
234 	u32 tosend, delta = 0;
235 	int ret;
236 
237 more_data:
238 	if (psock->eval == __SK_NONE) {
239 		/* Track delta in msg size to add/subtract it on SK_DROP from
240 		 * returned to user copied size. This ensures user doesn't
241 		 * get a positive return code with msg_cut_data and SK_DROP
242 		 * verdict.
243 		 */
244 		delta = msg->sg.size;
245 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
246 		delta -= msg->sg.size;
247 	}
248 
249 	if (msg->cork_bytes &&
250 	    msg->cork_bytes > msg->sg.size && !enospc) {
251 		psock->cork_bytes = msg->cork_bytes - msg->sg.size;
252 		if (!psock->cork) {
253 			psock->cork = kzalloc(sizeof(*psock->cork),
254 					      GFP_ATOMIC | __GFP_NOWARN);
255 			if (!psock->cork)
256 				return -ENOMEM;
257 		}
258 		memcpy(psock->cork, msg, sizeof(*msg));
259 		return 0;
260 	}
261 
262 	tosend = msg->sg.size;
263 	if (psock->apply_bytes && psock->apply_bytes < tosend)
264 		tosend = psock->apply_bytes;
265 
266 	switch (psock->eval) {
267 	case __SK_PASS:
268 		ret = tcp_bpf_push(sk, msg, tosend, flags, true);
269 		if (unlikely(ret)) {
270 			*copied -= sk_msg_free(sk, msg);
271 			break;
272 		}
273 		sk_msg_apply_bytes(psock, tosend);
274 		break;
275 	case __SK_REDIRECT:
276 		sk_redir = psock->sk_redir;
277 		sk_msg_apply_bytes(psock, tosend);
278 		if (psock->cork) {
279 			cork = true;
280 			psock->cork = NULL;
281 		}
282 		sk_msg_return(sk, msg, tosend);
283 		release_sock(sk);
284 		ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
285 		lock_sock(sk);
286 		if (unlikely(ret < 0)) {
287 			int free = sk_msg_free_nocharge(sk, msg);
288 
289 			if (!cork)
290 				*copied -= free;
291 		}
292 		if (cork) {
293 			sk_msg_free(sk, msg);
294 			kfree(msg);
295 			msg = NULL;
296 			ret = 0;
297 		}
298 		break;
299 	case __SK_DROP:
300 	default:
301 		sk_msg_free_partial(sk, msg, tosend);
302 		sk_msg_apply_bytes(psock, tosend);
303 		*copied -= (tosend + delta);
304 		return -EACCES;
305 	}
306 
307 	if (likely(!ret)) {
308 		if (!psock->apply_bytes) {
309 			psock->eval =  __SK_NONE;
310 			if (psock->sk_redir) {
311 				sock_put(psock->sk_redir);
312 				psock->sk_redir = NULL;
313 			}
314 		}
315 		if (msg &&
316 		    msg->sg.data[msg->sg.start].page_link &&
317 		    msg->sg.data[msg->sg.start].length)
318 			goto more_data;
319 	}
320 	return ret;
321 }
322 
323 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
324 {
325 	struct sk_msg tmp, *msg_tx = NULL;
326 	int copied = 0, err = 0;
327 	struct sk_psock *psock;
328 	long timeo;
329 	int flags;
330 
331 	/* Don't let internal do_tcp_sendpages() flags through */
332 	flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
333 	flags |= MSG_NO_SHARED_FRAGS;
334 
335 	psock = sk_psock_get(sk);
336 	if (unlikely(!psock))
337 		return tcp_sendmsg(sk, msg, size);
338 
339 	lock_sock(sk);
340 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
341 	while (msg_data_left(msg)) {
342 		bool enospc = false;
343 		u32 copy, osize;
344 
345 		if (sk->sk_err) {
346 			err = -sk->sk_err;
347 			goto out_err;
348 		}
349 
350 		copy = msg_data_left(msg);
351 		if (!sk_stream_memory_free(sk))
352 			goto wait_for_sndbuf;
353 		if (psock->cork) {
354 			msg_tx = psock->cork;
355 		} else {
356 			msg_tx = &tmp;
357 			sk_msg_init(msg_tx);
358 		}
359 
360 		osize = msg_tx->sg.size;
361 		err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
362 		if (err) {
363 			if (err != -ENOSPC)
364 				goto wait_for_memory;
365 			enospc = true;
366 			copy = msg_tx->sg.size - osize;
367 		}
368 
369 		err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
370 					       copy);
371 		if (err < 0) {
372 			sk_msg_trim(sk, msg_tx, osize);
373 			goto out_err;
374 		}
375 
376 		copied += copy;
377 		if (psock->cork_bytes) {
378 			if (size > psock->cork_bytes)
379 				psock->cork_bytes = 0;
380 			else
381 				psock->cork_bytes -= size;
382 			if (psock->cork_bytes && !enospc)
383 				goto out_err;
384 			/* All cork bytes are accounted, rerun the prog. */
385 			psock->eval = __SK_NONE;
386 			psock->cork_bytes = 0;
387 		}
388 
389 		err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
390 		if (unlikely(err < 0))
391 			goto out_err;
392 		continue;
393 wait_for_sndbuf:
394 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
395 wait_for_memory:
396 		err = sk_stream_wait_memory(sk, &timeo);
397 		if (err) {
398 			if (msg_tx && msg_tx != psock->cork)
399 				sk_msg_free(sk, msg_tx);
400 			goto out_err;
401 		}
402 	}
403 out_err:
404 	if (err < 0)
405 		err = sk_stream_error(sk, msg->msg_flags, err);
406 	release_sock(sk);
407 	sk_psock_put(sk, psock);
408 	return copied ? copied : err;
409 }
410 
411 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
412 			    size_t size, int flags)
413 {
414 	struct sk_msg tmp, *msg = NULL;
415 	int err = 0, copied = 0;
416 	struct sk_psock *psock;
417 	bool enospc = false;
418 
419 	psock = sk_psock_get(sk);
420 	if (unlikely(!psock))
421 		return tcp_sendpage(sk, page, offset, size, flags);
422 
423 	lock_sock(sk);
424 	if (psock->cork) {
425 		msg = psock->cork;
426 	} else {
427 		msg = &tmp;
428 		sk_msg_init(msg);
429 	}
430 
431 	/* Catch case where ring is full and sendpage is stalled. */
432 	if (unlikely(sk_msg_full(msg)))
433 		goto out_err;
434 
435 	sk_msg_page_add(msg, page, size, offset);
436 	sk_mem_charge(sk, size);
437 	copied = size;
438 	if (sk_msg_full(msg))
439 		enospc = true;
440 	if (psock->cork_bytes) {
441 		if (size > psock->cork_bytes)
442 			psock->cork_bytes = 0;
443 		else
444 			psock->cork_bytes -= size;
445 		if (psock->cork_bytes && !enospc)
446 			goto out_err;
447 		/* All cork bytes are accounted, rerun the prog. */
448 		psock->eval = __SK_NONE;
449 		psock->cork_bytes = 0;
450 	}
451 
452 	err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
453 out_err:
454 	release_sock(sk);
455 	sk_psock_put(sk, psock);
456 	return copied ? copied : err;
457 }
458 
459 enum {
460 	TCP_BPF_IPV4,
461 	TCP_BPF_IPV6,
462 	TCP_BPF_NUM_PROTS,
463 };
464 
465 enum {
466 	TCP_BPF_BASE,
467 	TCP_BPF_TX,
468 	TCP_BPF_NUM_CFGS,
469 };
470 
471 static struct proto *tcpv6_prot_saved __read_mostly;
472 static DEFINE_SPINLOCK(tcpv6_prot_lock);
473 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
474 
475 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
476 				   struct proto *base)
477 {
478 	prot[TCP_BPF_BASE]			= *base;
479 	prot[TCP_BPF_BASE].unhash		= sock_map_unhash;
480 	prot[TCP_BPF_BASE].close		= sock_map_close;
481 	prot[TCP_BPF_BASE].recvmsg		= tcp_bpf_recvmsg;
482 	prot[TCP_BPF_BASE].stream_memory_read	= tcp_bpf_stream_read;
483 
484 	prot[TCP_BPF_TX]			= prot[TCP_BPF_BASE];
485 	prot[TCP_BPF_TX].sendmsg		= tcp_bpf_sendmsg;
486 	prot[TCP_BPF_TX].sendpage		= tcp_bpf_sendpage;
487 }
488 
489 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
490 {
491 	if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
492 		spin_lock_bh(&tcpv6_prot_lock);
493 		if (likely(ops != tcpv6_prot_saved)) {
494 			tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
495 			smp_store_release(&tcpv6_prot_saved, ops);
496 		}
497 		spin_unlock_bh(&tcpv6_prot_lock);
498 	}
499 }
500 
501 static int __init tcp_bpf_v4_build_proto(void)
502 {
503 	tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
504 	return 0;
505 }
506 late_initcall(tcp_bpf_v4_build_proto);
507 
508 static int tcp_bpf_assert_proto_ops(struct proto *ops)
509 {
510 	/* In order to avoid retpoline, we make assumptions when we call
511 	 * into ops if e.g. a psock is not present. Make sure they are
512 	 * indeed valid assumptions.
513 	 */
514 	return ops->recvmsg  == tcp_recvmsg &&
515 	       ops->sendmsg  == tcp_sendmsg &&
516 	       ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
517 }
518 
519 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
520 {
521 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
522 	int config = psock->progs.msg_parser   ? TCP_BPF_TX   : TCP_BPF_BASE;
523 
524 	if (restore) {
525 		if (inet_csk_has_ulp(sk)) {
526 			/* TLS does not have an unhash proto in SW cases,
527 			 * but we need to ensure we stop using the sock_map
528 			 * unhash routine because the associated psock is being
529 			 * removed. So use the original unhash handler.
530 			 */
531 			WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
532 			tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
533 		} else {
534 			sk->sk_write_space = psock->saved_write_space;
535 			/* Pairs with lockless read in sk_clone_lock() */
536 			WRITE_ONCE(sk->sk_prot, psock->sk_proto);
537 		}
538 		return 0;
539 	}
540 
541 	if (inet_csk_has_ulp(sk))
542 		return -EINVAL;
543 
544 	if (sk->sk_family == AF_INET6) {
545 		if (tcp_bpf_assert_proto_ops(psock->sk_proto))
546 			return -EINVAL;
547 
548 		tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
549 	}
550 
551 	/* Pairs with lockless read in sk_clone_lock() */
552 	WRITE_ONCE(sk->sk_prot, &tcp_bpf_prots[family][config]);
553 	return 0;
554 }
555 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
556 
557 /* If a child got cloned from a listening socket that had tcp_bpf
558  * protocol callbacks installed, we need to restore the callbacks to
559  * the default ones because the child does not inherit the psock state
560  * that tcp_bpf callbacks expect.
561  */
562 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
563 {
564 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
565 	struct proto *prot = newsk->sk_prot;
566 
567 	if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
568 		newsk->sk_prot = sk->sk_prot_creator;
569 }
570 #endif /* CONFIG_BPF_SYSCALL */
571