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