xref: /linux/net/ipv4/tcp_bpf.c (revision d0f482bb06f9447d44d2cae0386a0bd768c3cc16)
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_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
167 		    int nonblock, int flags, int *addr_len)
168 {
169 	struct sk_psock *psock;
170 	int copied, ret;
171 
172 	if (unlikely(flags & MSG_ERRQUEUE))
173 		return inet_recv_error(sk, msg, len, addr_len);
174 
175 	psock = sk_psock_get(sk);
176 	if (unlikely(!psock))
177 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
178 	if (!skb_queue_empty(&sk->sk_receive_queue) &&
179 	    sk_psock_queue_empty(psock)) {
180 		sk_psock_put(sk, psock);
181 		return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
182 	}
183 	lock_sock(sk);
184 msg_bytes_ready:
185 	copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
186 	if (!copied) {
187 		long timeo;
188 		int data;
189 
190 		timeo = sock_rcvtimeo(sk, nonblock);
191 		data = sk_msg_wait_data(sk, psock, timeo);
192 		if (data) {
193 			if (!sk_psock_queue_empty(psock))
194 				goto msg_bytes_ready;
195 			release_sock(sk);
196 			sk_psock_put(sk, psock);
197 			return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
198 		}
199 		copied = -EAGAIN;
200 	}
201 	ret = copied;
202 	release_sock(sk);
203 	sk_psock_put(sk, psock);
204 	return ret;
205 }
206 
207 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
208 				struct sk_msg *msg, int *copied, int flags)
209 {
210 	bool cork = false, enospc = sk_msg_full(msg);
211 	struct sock *sk_redir;
212 	u32 tosend, delta = 0;
213 	int ret;
214 
215 more_data:
216 	if (psock->eval == __SK_NONE) {
217 		/* Track delta in msg size to add/subtract it on SK_DROP from
218 		 * returned to user copied size. This ensures user doesn't
219 		 * get a positive return code with msg_cut_data and SK_DROP
220 		 * verdict.
221 		 */
222 		delta = msg->sg.size;
223 		psock->eval = sk_psock_msg_verdict(sk, psock, msg);
224 		delta -= msg->sg.size;
225 	}
226 
227 	if (msg->cork_bytes &&
228 	    msg->cork_bytes > msg->sg.size && !enospc) {
229 		psock->cork_bytes = msg->cork_bytes - msg->sg.size;
230 		if (!psock->cork) {
231 			psock->cork = kzalloc(sizeof(*psock->cork),
232 					      GFP_ATOMIC | __GFP_NOWARN);
233 			if (!psock->cork)
234 				return -ENOMEM;
235 		}
236 		memcpy(psock->cork, msg, sizeof(*msg));
237 		return 0;
238 	}
239 
240 	tosend = msg->sg.size;
241 	if (psock->apply_bytes && psock->apply_bytes < tosend)
242 		tosend = psock->apply_bytes;
243 
244 	switch (psock->eval) {
245 	case __SK_PASS:
246 		ret = tcp_bpf_push(sk, msg, tosend, flags, true);
247 		if (unlikely(ret)) {
248 			*copied -= sk_msg_free(sk, msg);
249 			break;
250 		}
251 		sk_msg_apply_bytes(psock, tosend);
252 		break;
253 	case __SK_REDIRECT:
254 		sk_redir = psock->sk_redir;
255 		sk_msg_apply_bytes(psock, tosend);
256 		if (psock->cork) {
257 			cork = true;
258 			psock->cork = NULL;
259 		}
260 		sk_msg_return(sk, msg, tosend);
261 		release_sock(sk);
262 		ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
263 		lock_sock(sk);
264 		if (unlikely(ret < 0)) {
265 			int free = sk_msg_free_nocharge(sk, msg);
266 
267 			if (!cork)
268 				*copied -= free;
269 		}
270 		if (cork) {
271 			sk_msg_free(sk, msg);
272 			kfree(msg);
273 			msg = NULL;
274 			ret = 0;
275 		}
276 		break;
277 	case __SK_DROP:
278 	default:
279 		sk_msg_free_partial(sk, msg, tosend);
280 		sk_msg_apply_bytes(psock, tosend);
281 		*copied -= (tosend + delta);
282 		return -EACCES;
283 	}
284 
285 	if (likely(!ret)) {
286 		if (!psock->apply_bytes) {
287 			psock->eval =  __SK_NONE;
288 			if (psock->sk_redir) {
289 				sock_put(psock->sk_redir);
290 				psock->sk_redir = NULL;
291 			}
292 		}
293 		if (msg &&
294 		    msg->sg.data[msg->sg.start].page_link &&
295 		    msg->sg.data[msg->sg.start].length)
296 			goto more_data;
297 	}
298 	return ret;
299 }
300 
301 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
302 {
303 	struct sk_msg tmp, *msg_tx = NULL;
304 	int copied = 0, err = 0;
305 	struct sk_psock *psock;
306 	long timeo;
307 	int flags;
308 
309 	/* Don't let internal do_tcp_sendpages() flags through */
310 	flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
311 	flags |= MSG_NO_SHARED_FRAGS;
312 
313 	psock = sk_psock_get(sk);
314 	if (unlikely(!psock))
315 		return tcp_sendmsg(sk, msg, size);
316 
317 	lock_sock(sk);
318 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
319 	while (msg_data_left(msg)) {
320 		bool enospc = false;
321 		u32 copy, osize;
322 
323 		if (sk->sk_err) {
324 			err = -sk->sk_err;
325 			goto out_err;
326 		}
327 
328 		copy = msg_data_left(msg);
329 		if (!sk_stream_memory_free(sk))
330 			goto wait_for_sndbuf;
331 		if (psock->cork) {
332 			msg_tx = psock->cork;
333 		} else {
334 			msg_tx = &tmp;
335 			sk_msg_init(msg_tx);
336 		}
337 
338 		osize = msg_tx->sg.size;
339 		err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
340 		if (err) {
341 			if (err != -ENOSPC)
342 				goto wait_for_memory;
343 			enospc = true;
344 			copy = msg_tx->sg.size - osize;
345 		}
346 
347 		err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
348 					       copy);
349 		if (err < 0) {
350 			sk_msg_trim(sk, msg_tx, osize);
351 			goto out_err;
352 		}
353 
354 		copied += copy;
355 		if (psock->cork_bytes) {
356 			if (size > psock->cork_bytes)
357 				psock->cork_bytes = 0;
358 			else
359 				psock->cork_bytes -= size;
360 			if (psock->cork_bytes && !enospc)
361 				goto out_err;
362 			/* All cork bytes are accounted, rerun the prog. */
363 			psock->eval = __SK_NONE;
364 			psock->cork_bytes = 0;
365 		}
366 
367 		err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
368 		if (unlikely(err < 0))
369 			goto out_err;
370 		continue;
371 wait_for_sndbuf:
372 		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
373 wait_for_memory:
374 		err = sk_stream_wait_memory(sk, &timeo);
375 		if (err) {
376 			if (msg_tx && msg_tx != psock->cork)
377 				sk_msg_free(sk, msg_tx);
378 			goto out_err;
379 		}
380 	}
381 out_err:
382 	if (err < 0)
383 		err = sk_stream_error(sk, msg->msg_flags, err);
384 	release_sock(sk);
385 	sk_psock_put(sk, psock);
386 	return copied ? copied : err;
387 }
388 
389 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
390 			    size_t size, int flags)
391 {
392 	struct sk_msg tmp, *msg = NULL;
393 	int err = 0, copied = 0;
394 	struct sk_psock *psock;
395 	bool enospc = false;
396 
397 	psock = sk_psock_get(sk);
398 	if (unlikely(!psock))
399 		return tcp_sendpage(sk, page, offset, size, flags);
400 
401 	lock_sock(sk);
402 	if (psock->cork) {
403 		msg = psock->cork;
404 	} else {
405 		msg = &tmp;
406 		sk_msg_init(msg);
407 	}
408 
409 	/* Catch case where ring is full and sendpage is stalled. */
410 	if (unlikely(sk_msg_full(msg)))
411 		goto out_err;
412 
413 	sk_msg_page_add(msg, page, size, offset);
414 	sk_mem_charge(sk, size);
415 	copied = size;
416 	if (sk_msg_full(msg))
417 		enospc = true;
418 	if (psock->cork_bytes) {
419 		if (size > psock->cork_bytes)
420 			psock->cork_bytes = 0;
421 		else
422 			psock->cork_bytes -= size;
423 		if (psock->cork_bytes && !enospc)
424 			goto out_err;
425 		/* All cork bytes are accounted, rerun the prog. */
426 		psock->eval = __SK_NONE;
427 		psock->cork_bytes = 0;
428 	}
429 
430 	err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
431 out_err:
432 	release_sock(sk);
433 	sk_psock_put(sk, psock);
434 	return copied ? copied : err;
435 }
436 
437 enum {
438 	TCP_BPF_IPV4,
439 	TCP_BPF_IPV6,
440 	TCP_BPF_NUM_PROTS,
441 };
442 
443 enum {
444 	TCP_BPF_BASE,
445 	TCP_BPF_TX,
446 	TCP_BPF_NUM_CFGS,
447 };
448 
449 static struct proto *tcpv6_prot_saved __read_mostly;
450 static DEFINE_SPINLOCK(tcpv6_prot_lock);
451 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
452 
453 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
454 				   struct proto *base)
455 {
456 	prot[TCP_BPF_BASE]			= *base;
457 	prot[TCP_BPF_BASE].unhash		= sock_map_unhash;
458 	prot[TCP_BPF_BASE].close		= sock_map_close;
459 	prot[TCP_BPF_BASE].recvmsg		= tcp_bpf_recvmsg;
460 	prot[TCP_BPF_BASE].stream_memory_read	= tcp_bpf_stream_read;
461 
462 	prot[TCP_BPF_TX]			= prot[TCP_BPF_BASE];
463 	prot[TCP_BPF_TX].sendmsg		= tcp_bpf_sendmsg;
464 	prot[TCP_BPF_TX].sendpage		= tcp_bpf_sendpage;
465 }
466 
467 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
468 {
469 	if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
470 		spin_lock_bh(&tcpv6_prot_lock);
471 		if (likely(ops != tcpv6_prot_saved)) {
472 			tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
473 			smp_store_release(&tcpv6_prot_saved, ops);
474 		}
475 		spin_unlock_bh(&tcpv6_prot_lock);
476 	}
477 }
478 
479 static int __init tcp_bpf_v4_build_proto(void)
480 {
481 	tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
482 	return 0;
483 }
484 core_initcall(tcp_bpf_v4_build_proto);
485 
486 static int tcp_bpf_assert_proto_ops(struct proto *ops)
487 {
488 	/* In order to avoid retpoline, we make assumptions when we call
489 	 * into ops if e.g. a psock is not present. Make sure they are
490 	 * indeed valid assumptions.
491 	 */
492 	return ops->recvmsg  == tcp_recvmsg &&
493 	       ops->sendmsg  == tcp_sendmsg &&
494 	       ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
495 }
496 
497 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
498 {
499 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
500 	int config = psock->progs.msg_parser   ? TCP_BPF_TX   : TCP_BPF_BASE;
501 
502 	if (restore) {
503 		if (inet_csk_has_ulp(sk)) {
504 			/* TLS does not have an unhash proto in SW cases,
505 			 * but we need to ensure we stop using the sock_map
506 			 * unhash routine because the associated psock is being
507 			 * removed. So use the original unhash handler.
508 			 */
509 			WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
510 			tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
511 		} else {
512 			sk->sk_write_space = psock->saved_write_space;
513 			/* Pairs with lockless read in sk_clone_lock() */
514 			WRITE_ONCE(sk->sk_prot, psock->sk_proto);
515 		}
516 		return 0;
517 	}
518 
519 	if (inet_csk_has_ulp(sk))
520 		return -EINVAL;
521 
522 	if (sk->sk_family == AF_INET6) {
523 		if (tcp_bpf_assert_proto_ops(psock->sk_proto))
524 			return -EINVAL;
525 
526 		tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
527 	}
528 
529 	/* Pairs with lockless read in sk_clone_lock() */
530 	WRITE_ONCE(sk->sk_prot, &tcp_bpf_prots[family][config]);
531 	return 0;
532 }
533 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
534 
535 /* If a child got cloned from a listening socket that had tcp_bpf
536  * protocol callbacks installed, we need to restore the callbacks to
537  * the default ones because the child does not inherit the psock state
538  * that tcp_bpf callbacks expect.
539  */
540 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
541 {
542 	int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
543 	struct proto *prot = newsk->sk_prot;
544 
545 	if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
546 		newsk->sk_prot = sk->sk_prot_creator;
547 }
548 #endif /* CONFIG_BPF_SYSCALL */
549