Loading...
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#include <linux/util_macros.h>
10
11#include <net/inet_common.h>
12#include <net/tls.h>
13
14void tcp_eat_skb(struct sock *sk, struct sk_buff *skb)
15{
16 struct tcp_sock *tcp;
17 int copied;
18
19 if (!skb || !skb->len || !sk_is_tcp(sk))
20 return;
21
22 if (skb_bpf_strparser(skb))
23 return;
24
25 tcp = tcp_sk(sk);
26 copied = tcp->copied_seq + skb->len;
27 WRITE_ONCE(tcp->copied_seq, copied);
28 tcp_rcv_space_adjust(sk);
29 __tcp_cleanup_rbuf(sk, skb->len);
30}
31
32static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
33 struct sk_msg *msg, u32 apply_bytes)
34{
35 bool apply = apply_bytes;
36 struct scatterlist *sge;
37 u32 size, copied = 0;
38 struct sk_msg *tmp;
39 int i, ret = 0;
40
41 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
42 if (unlikely(!tmp))
43 return -ENOMEM;
44
45 lock_sock(sk);
46 tmp->sg.start = msg->sg.start;
47 i = msg->sg.start;
48 do {
49 sge = sk_msg_elem(msg, i);
50 size = (apply && apply_bytes < sge->length) ?
51 apply_bytes : sge->length;
52 if (!__sk_rmem_schedule(sk, size, false)) {
53 if (!copied)
54 ret = -ENOMEM;
55 break;
56 }
57
58 sk_mem_charge(sk, size);
59 atomic_add(size, &sk->sk_rmem_alloc);
60 sk_msg_xfer(tmp, msg, i, size);
61 copied += size;
62 if (sge->length)
63 get_page(sk_msg_page(tmp, i));
64 sk_msg_iter_var_next(i);
65 tmp->sg.end = i;
66 if (apply) {
67 apply_bytes -= size;
68 if (!apply_bytes) {
69 if (sge->length)
70 sk_msg_iter_var_prev(i);
71 break;
72 }
73 }
74 } while (i != msg->sg.end);
75
76 if (!ret) {
77 msg->sg.start = i;
78 if (!sk_psock_queue_msg(psock, tmp))
79 atomic_sub(copied, &sk->sk_rmem_alloc);
80 sk_psock_data_ready(sk, psock);
81 } else {
82 sk_msg_free(sk, tmp);
83 kfree(tmp);
84 }
85
86 release_sock(sk);
87 return ret;
88}
89
90static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
91 int flags, bool uncharge)
92{
93 struct msghdr msghdr = {};
94 bool apply = apply_bytes;
95 struct scatterlist *sge;
96 struct page *page;
97 int size, ret = 0;
98 u32 off;
99
100 while (1) {
101 struct bio_vec bvec;
102 bool has_tx_ulp;
103
104 sge = sk_msg_elem(msg, msg->sg.start);
105 size = (apply && apply_bytes < sge->length) ?
106 apply_bytes : sge->length;
107 off = sge->offset;
108 page = sg_page(sge);
109
110 tcp_rate_check_app_limited(sk);
111retry:
112 msghdr.msg_flags = flags | MSG_SPLICE_PAGES;
113 has_tx_ulp = tls_sw_has_ctx_tx(sk);
114 if (has_tx_ulp)
115 msghdr.msg_flags |= MSG_SENDPAGE_NOPOLICY;
116
117 if (size < sge->length && msg->sg.start != msg->sg.end)
118 msghdr.msg_flags |= MSG_MORE;
119
120 bvec_set_page(&bvec, page, size, off);
121 iov_iter_bvec(&msghdr.msg_iter, ITER_SOURCE, &bvec, 1, size);
122 ret = tcp_sendmsg_locked(sk, &msghdr, size);
123 if (ret <= 0)
124 return ret;
125
126 if (apply)
127 apply_bytes -= ret;
128 msg->sg.size -= ret;
129 sge->offset += ret;
130 sge->length -= ret;
131 if (uncharge)
132 sk_mem_uncharge(sk, ret);
133 if (ret != size) {
134 size -= ret;
135 off += ret;
136 goto retry;
137 }
138 if (!sge->length) {
139 put_page(page);
140 sk_msg_iter_next(msg, start);
141 sg_init_table(sge, 1);
142 if (msg->sg.start == msg->sg.end)
143 break;
144 }
145 if (apply && !apply_bytes)
146 break;
147 }
148
149 return 0;
150}
151
152static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
153 u32 apply_bytes, int flags, bool uncharge)
154{
155 int ret;
156
157 lock_sock(sk);
158 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
159 release_sock(sk);
160 return ret;
161}
162
163int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress,
164 struct sk_msg *msg, u32 bytes, int flags)
165{
166 struct sk_psock *psock = sk_psock_get(sk);
167 int ret;
168
169 if (unlikely(!psock))
170 return -EPIPE;
171
172 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes) :
173 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
174 sk_psock_put(sk, psock);
175 return ret;
176}
177EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
178
179#ifdef CONFIG_BPF_SYSCALL
180static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
181 long timeo)
182{
183 DEFINE_WAIT_FUNC(wait, woken_wake_function);
184 int ret = 0;
185
186 if (sk->sk_shutdown & RCV_SHUTDOWN)
187 return 1;
188
189 if (!timeo)
190 return ret;
191
192 add_wait_queue(sk_sleep(sk), &wait);
193 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
194 ret = sk_wait_event(sk, &timeo,
195 !list_empty(&psock->ingress_msg) ||
196 !skb_queue_empty_lockless(&sk->sk_receive_queue), &wait);
197 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
198 remove_wait_queue(sk_sleep(sk), &wait);
199 return ret;
200}
201
202static bool is_next_msg_fin(struct sk_psock *psock)
203{
204 struct scatterlist *sge;
205 struct sk_msg *msg_rx;
206 int i;
207
208 msg_rx = sk_psock_peek_msg(psock);
209 i = msg_rx->sg.start;
210 sge = sk_msg_elem(msg_rx, i);
211 if (!sge->length) {
212 struct sk_buff *skb = msg_rx->skb;
213
214 if (skb && TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
215 return true;
216 }
217 return false;
218}
219
220static int tcp_bpf_recvmsg_parser(struct sock *sk,
221 struct msghdr *msg,
222 size_t len,
223 int flags,
224 int *addr_len)
225{
226 int peek = flags & MSG_PEEK;
227 struct sk_psock *psock;
228 struct tcp_sock *tcp;
229 int copied = 0;
230 u32 seq;
231
232 if (unlikely(flags & MSG_ERRQUEUE))
233 return inet_recv_error(sk, msg, len, addr_len);
234
235 if (!len)
236 return 0;
237
238 psock = sk_psock_get(sk);
239 if (unlikely(!psock))
240 return tcp_recvmsg(sk, msg, len, flags, addr_len);
241
242 lock_sock(sk);
243 tcp = tcp_sk(sk);
244 seq = tcp->copied_seq;
245 /* We may have received data on the sk_receive_queue pre-accept and
246 * then we can not use read_skb in this context because we haven't
247 * assigned a sk_socket yet so have no link to the ops. The work-around
248 * is to check the sk_receive_queue and in these cases read skbs off
249 * queue again. The read_skb hook is not running at this point because
250 * of lock_sock so we avoid having multiple runners in read_skb.
251 */
252 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) {
253 tcp_data_ready(sk);
254 /* This handles the ENOMEM errors if we both receive data
255 * pre accept and are already under memory pressure. At least
256 * let user know to retry.
257 */
258 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) {
259 copied = -EAGAIN;
260 goto out;
261 }
262 }
263
264msg_bytes_ready:
265 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
266 /* The typical case for EFAULT is the socket was gracefully
267 * shutdown with a FIN pkt. So check here the other case is
268 * some error on copy_page_to_iter which would be unexpected.
269 * On fin return correct return code to zero.
270 */
271 if (copied == -EFAULT) {
272 bool is_fin = is_next_msg_fin(psock);
273
274 if (is_fin) {
275 copied = 0;
276 seq++;
277 goto out;
278 }
279 }
280 seq += copied;
281 if (!copied) {
282 long timeo;
283 int data;
284
285 if (sock_flag(sk, SOCK_DONE))
286 goto out;
287
288 if (sk->sk_err) {
289 copied = sock_error(sk);
290 goto out;
291 }
292
293 if (sk->sk_shutdown & RCV_SHUTDOWN)
294 goto out;
295
296 if (sk->sk_state == TCP_CLOSE) {
297 copied = -ENOTCONN;
298 goto out;
299 }
300
301 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
302 if (!timeo) {
303 copied = -EAGAIN;
304 goto out;
305 }
306
307 if (signal_pending(current)) {
308 copied = sock_intr_errno(timeo);
309 goto out;
310 }
311
312 data = tcp_msg_wait_data(sk, psock, timeo);
313 if (data < 0) {
314 copied = data;
315 goto unlock;
316 }
317 if (data && !sk_psock_queue_empty(psock))
318 goto msg_bytes_ready;
319 copied = -EAGAIN;
320 }
321out:
322 if (!peek)
323 WRITE_ONCE(tcp->copied_seq, seq);
324 tcp_rcv_space_adjust(sk);
325 if (copied > 0)
326 __tcp_cleanup_rbuf(sk, copied);
327
328unlock:
329 release_sock(sk);
330 sk_psock_put(sk, psock);
331 return copied;
332}
333
334static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
335 int flags, int *addr_len)
336{
337 struct sk_psock *psock;
338 int copied, ret;
339
340 if (unlikely(flags & MSG_ERRQUEUE))
341 return inet_recv_error(sk, msg, len, addr_len);
342
343 if (!len)
344 return 0;
345
346 psock = sk_psock_get(sk);
347 if (unlikely(!psock))
348 return tcp_recvmsg(sk, msg, len, flags, addr_len);
349 if (!skb_queue_empty(&sk->sk_receive_queue) &&
350 sk_psock_queue_empty(psock)) {
351 sk_psock_put(sk, psock);
352 return tcp_recvmsg(sk, msg, len, flags, addr_len);
353 }
354 lock_sock(sk);
355msg_bytes_ready:
356 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
357 if (!copied) {
358 long timeo;
359 int data;
360
361 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
362 data = tcp_msg_wait_data(sk, psock, timeo);
363 if (data < 0) {
364 ret = data;
365 goto unlock;
366 }
367 if (data) {
368 if (!sk_psock_queue_empty(psock))
369 goto msg_bytes_ready;
370 release_sock(sk);
371 sk_psock_put(sk, psock);
372 return tcp_recvmsg(sk, msg, len, flags, addr_len);
373 }
374 copied = -EAGAIN;
375 }
376 ret = copied;
377
378unlock:
379 release_sock(sk);
380 sk_psock_put(sk, psock);
381 return ret;
382}
383
384static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
385 struct sk_msg *msg, int *copied, int flags)
386{
387 bool cork = false, enospc = sk_msg_full(msg), redir_ingress;
388 struct sock *sk_redir;
389 u32 tosend, origsize, sent, delta = 0;
390 u32 eval;
391 int ret;
392
393more_data:
394 if (psock->eval == __SK_NONE) {
395 /* Track delta in msg size to add/subtract it on SK_DROP from
396 * returned to user copied size. This ensures user doesn't
397 * get a positive return code with msg_cut_data and SK_DROP
398 * verdict.
399 */
400 delta = msg->sg.size;
401 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
402 delta -= msg->sg.size;
403 }
404
405 if (msg->cork_bytes &&
406 msg->cork_bytes > msg->sg.size && !enospc) {
407 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
408 if (!psock->cork) {
409 psock->cork = kzalloc(sizeof(*psock->cork),
410 GFP_ATOMIC | __GFP_NOWARN);
411 if (!psock->cork)
412 return -ENOMEM;
413 }
414 memcpy(psock->cork, msg, sizeof(*msg));
415 return 0;
416 }
417
418 tosend = msg->sg.size;
419 if (psock->apply_bytes && psock->apply_bytes < tosend)
420 tosend = psock->apply_bytes;
421 eval = __SK_NONE;
422
423 switch (psock->eval) {
424 case __SK_PASS:
425 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
426 if (unlikely(ret)) {
427 *copied -= sk_msg_free(sk, msg);
428 break;
429 }
430 sk_msg_apply_bytes(psock, tosend);
431 break;
432 case __SK_REDIRECT:
433 redir_ingress = psock->redir_ingress;
434 sk_redir = psock->sk_redir;
435 sk_msg_apply_bytes(psock, tosend);
436 if (!psock->apply_bytes) {
437 /* Clean up before releasing the sock lock. */
438 eval = psock->eval;
439 psock->eval = __SK_NONE;
440 psock->sk_redir = NULL;
441 }
442 if (psock->cork) {
443 cork = true;
444 psock->cork = NULL;
445 }
446 release_sock(sk);
447
448 origsize = msg->sg.size;
449 ret = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
450 msg, tosend, flags);
451 sent = origsize - msg->sg.size;
452
453 if (eval == __SK_REDIRECT)
454 sock_put(sk_redir);
455
456 lock_sock(sk);
457 sk_mem_uncharge(sk, sent);
458 if (unlikely(ret < 0)) {
459 int free = sk_msg_free(sk, msg);
460
461 if (!cork)
462 *copied -= free;
463 }
464 if (cork) {
465 sk_msg_free(sk, msg);
466 kfree(msg);
467 msg = NULL;
468 ret = 0;
469 }
470 break;
471 case __SK_DROP:
472 default:
473 sk_msg_free(sk, msg);
474 sk_msg_apply_bytes(psock, tosend);
475 *copied -= (tosend + delta);
476 return -EACCES;
477 }
478
479 if (likely(!ret)) {
480 if (!psock->apply_bytes) {
481 psock->eval = __SK_NONE;
482 if (psock->sk_redir) {
483 sock_put(psock->sk_redir);
484 psock->sk_redir = NULL;
485 }
486 }
487 if (msg &&
488 msg->sg.data[msg->sg.start].page_link &&
489 msg->sg.data[msg->sg.start].length)
490 goto more_data;
491 }
492 return ret;
493}
494
495static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
496{
497 struct sk_msg tmp, *msg_tx = NULL;
498 int copied = 0, err = 0, ret = 0;
499 struct sk_psock *psock;
500 long timeo;
501 int flags;
502
503 /* Don't let internal flags through */
504 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
505 flags |= MSG_NO_SHARED_FRAGS;
506
507 psock = sk_psock_get(sk);
508 if (unlikely(!psock))
509 return tcp_sendmsg(sk, msg, size);
510
511 lock_sock(sk);
512 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
513 while (msg_data_left(msg)) {
514 bool enospc = false;
515 u32 copy, osize;
516
517 if (sk->sk_err) {
518 err = -sk->sk_err;
519 goto out_err;
520 }
521
522 copy = msg_data_left(msg);
523 if (!sk_stream_memory_free(sk))
524 goto wait_for_sndbuf;
525 if (psock->cork) {
526 msg_tx = psock->cork;
527 } else {
528 msg_tx = &tmp;
529 sk_msg_init(msg_tx);
530 }
531
532 osize = msg_tx->sg.size;
533 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
534 if (err) {
535 if (err != -ENOSPC)
536 goto wait_for_memory;
537 enospc = true;
538 copy = msg_tx->sg.size - osize;
539 }
540
541 ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
542 copy);
543 if (ret < 0) {
544 sk_msg_trim(sk, msg_tx, osize);
545 goto out_err;
546 }
547
548 copied += ret;
549 if (psock->cork_bytes) {
550 if (size > psock->cork_bytes)
551 psock->cork_bytes = 0;
552 else
553 psock->cork_bytes -= size;
554 if (psock->cork_bytes && !enospc)
555 goto out_err;
556 /* All cork bytes are accounted, rerun the prog. */
557 psock->eval = __SK_NONE;
558 psock->cork_bytes = 0;
559 }
560
561 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
562 if (unlikely(err < 0))
563 goto out_err;
564 continue;
565wait_for_sndbuf:
566 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
567wait_for_memory:
568 err = sk_stream_wait_memory(sk, &timeo);
569 if (err) {
570 if (msg_tx && msg_tx != psock->cork)
571 sk_msg_free(sk, msg_tx);
572 goto out_err;
573 }
574 }
575out_err:
576 if (err < 0)
577 err = sk_stream_error(sk, msg->msg_flags, err);
578 release_sock(sk);
579 sk_psock_put(sk, psock);
580 return copied > 0 ? copied : err;
581}
582
583enum {
584 TCP_BPF_IPV4,
585 TCP_BPF_IPV6,
586 TCP_BPF_NUM_PROTS,
587};
588
589enum {
590 TCP_BPF_BASE,
591 TCP_BPF_TX,
592 TCP_BPF_RX,
593 TCP_BPF_TXRX,
594 TCP_BPF_NUM_CFGS,
595};
596
597static struct proto *tcpv6_prot_saved __read_mostly;
598static DEFINE_SPINLOCK(tcpv6_prot_lock);
599static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
600
601static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
602 struct proto *base)
603{
604 prot[TCP_BPF_BASE] = *base;
605 prot[TCP_BPF_BASE].destroy = sock_map_destroy;
606 prot[TCP_BPF_BASE].close = sock_map_close;
607 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
608 prot[TCP_BPF_BASE].sock_is_readable = sk_msg_is_readable;
609
610 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
611 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
612
613 prot[TCP_BPF_RX] = prot[TCP_BPF_BASE];
614 prot[TCP_BPF_RX].recvmsg = tcp_bpf_recvmsg_parser;
615
616 prot[TCP_BPF_TXRX] = prot[TCP_BPF_TX];
617 prot[TCP_BPF_TXRX].recvmsg = tcp_bpf_recvmsg_parser;
618}
619
620static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
621{
622 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
623 spin_lock_bh(&tcpv6_prot_lock);
624 if (likely(ops != tcpv6_prot_saved)) {
625 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
626 smp_store_release(&tcpv6_prot_saved, ops);
627 }
628 spin_unlock_bh(&tcpv6_prot_lock);
629 }
630}
631
632static int __init tcp_bpf_v4_build_proto(void)
633{
634 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
635 return 0;
636}
637late_initcall(tcp_bpf_v4_build_proto);
638
639static int tcp_bpf_assert_proto_ops(struct proto *ops)
640{
641 /* In order to avoid retpoline, we make assumptions when we call
642 * into ops if e.g. a psock is not present. Make sure they are
643 * indeed valid assumptions.
644 */
645 return ops->recvmsg == tcp_recvmsg &&
646 ops->sendmsg == tcp_sendmsg ? 0 : -ENOTSUPP;
647}
648
649int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
650{
651 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
652 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
653
654 if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
655 config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
656 }
657
658 if (restore) {
659 if (inet_csk_has_ulp(sk)) {
660 /* TLS does not have an unhash proto in SW cases,
661 * but we need to ensure we stop using the sock_map
662 * unhash routine because the associated psock is being
663 * removed. So use the original unhash handler.
664 */
665 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
666 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
667 } else {
668 sk->sk_write_space = psock->saved_write_space;
669 /* Pairs with lockless read in sk_clone_lock() */
670 sock_replace_proto(sk, psock->sk_proto);
671 }
672 return 0;
673 }
674
675 if (sk->sk_family == AF_INET6) {
676 if (tcp_bpf_assert_proto_ops(psock->sk_proto))
677 return -EINVAL;
678
679 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
680 }
681
682 /* Pairs with lockless read in sk_clone_lock() */
683 sock_replace_proto(sk, &tcp_bpf_prots[family][config]);
684 return 0;
685}
686EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
687
688/* If a child got cloned from a listening socket that had tcp_bpf
689 * protocol callbacks installed, we need to restore the callbacks to
690 * the default ones because the child does not inherit the psock state
691 * that tcp_bpf callbacks expect.
692 */
693void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
694{
695 struct proto *prot = newsk->sk_prot;
696
697 if (is_insidevar(prot, tcp_bpf_prots))
698 newsk->sk_prot = sk->sk_prot_creator;
699}
700#endif /* CONFIG_BPF_SYSCALL */
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
13int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock,
14 struct msghdr *msg, int len, int flags)
15{
16 struct iov_iter *iter = &msg->msg_iter;
17 int peek = flags & MSG_PEEK;
18 int i, ret, copied = 0;
19 struct sk_msg *msg_rx;
20
21 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
22 struct sk_msg, list);
23
24 while (copied != len) {
25 struct scatterlist *sge;
26
27 if (unlikely(!msg_rx))
28 break;
29
30 i = msg_rx->sg.start;
31 do {
32 struct page *page;
33 int copy;
34
35 sge = sk_msg_elem(msg_rx, i);
36 copy = sge->length;
37 page = sg_page(sge);
38 if (copied + copy > len)
39 copy = len - copied;
40 ret = copy_page_to_iter(page, sge->offset, copy, iter);
41 if (ret != copy) {
42 msg_rx->sg.start = i;
43 return -EFAULT;
44 }
45
46 copied += copy;
47 if (likely(!peek)) {
48 sge->offset += copy;
49 sge->length -= copy;
50 sk_mem_uncharge(sk, copy);
51 msg_rx->sg.size -= copy;
52
53 if (!sge->length) {
54 sk_msg_iter_var_next(i);
55 if (!msg_rx->skb)
56 put_page(page);
57 }
58 } else {
59 sk_msg_iter_var_next(i);
60 }
61
62 if (copied == len)
63 break;
64 } while (i != msg_rx->sg.end);
65
66 if (unlikely(peek)) {
67 if (msg_rx == list_last_entry(&psock->ingress_msg,
68 struct sk_msg, list))
69 break;
70 msg_rx = list_next_entry(msg_rx, list);
71 continue;
72 }
73
74 msg_rx->sg.start = i;
75 if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
76 list_del(&msg_rx->list);
77 if (msg_rx->skb)
78 consume_skb(msg_rx->skb);
79 kfree(msg_rx);
80 }
81 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
82 struct sk_msg, list);
83 }
84
85 return copied;
86}
87EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg);
88
89static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
90 struct sk_msg *msg, u32 apply_bytes, int flags)
91{
92 bool apply = apply_bytes;
93 struct scatterlist *sge;
94 u32 size, copied = 0;
95 struct sk_msg *tmp;
96 int i, ret = 0;
97
98 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
99 if (unlikely(!tmp))
100 return -ENOMEM;
101
102 lock_sock(sk);
103 tmp->sg.start = msg->sg.start;
104 i = msg->sg.start;
105 do {
106 sge = sk_msg_elem(msg, i);
107 size = (apply && apply_bytes < sge->length) ?
108 apply_bytes : sge->length;
109 if (!sk_wmem_schedule(sk, size)) {
110 if (!copied)
111 ret = -ENOMEM;
112 break;
113 }
114
115 sk_mem_charge(sk, size);
116 sk_msg_xfer(tmp, msg, i, size);
117 copied += size;
118 if (sge->length)
119 get_page(sk_msg_page(tmp, i));
120 sk_msg_iter_var_next(i);
121 tmp->sg.end = i;
122 if (apply) {
123 apply_bytes -= size;
124 if (!apply_bytes)
125 break;
126 }
127 } while (i != msg->sg.end);
128
129 if (!ret) {
130 msg->sg.start = i;
131 sk_psock_queue_msg(psock, tmp);
132 sk_psock_data_ready(sk, psock);
133 } else {
134 sk_msg_free(sk, tmp);
135 kfree(tmp);
136 }
137
138 release_sock(sk);
139 return ret;
140}
141
142static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
143 int flags, bool uncharge)
144{
145 bool apply = apply_bytes;
146 struct scatterlist *sge;
147 struct page *page;
148 int size, ret = 0;
149 u32 off;
150
151 while (1) {
152 bool has_tx_ulp;
153
154 sge = sk_msg_elem(msg, msg->sg.start);
155 size = (apply && apply_bytes < sge->length) ?
156 apply_bytes : sge->length;
157 off = sge->offset;
158 page = sg_page(sge);
159
160 tcp_rate_check_app_limited(sk);
161retry:
162 has_tx_ulp = tls_sw_has_ctx_tx(sk);
163 if (has_tx_ulp) {
164 flags |= MSG_SENDPAGE_NOPOLICY;
165 ret = kernel_sendpage_locked(sk,
166 page, off, size, flags);
167 } else {
168 ret = do_tcp_sendpages(sk, page, off, size, flags);
169 }
170
171 if (ret <= 0)
172 return ret;
173 if (apply)
174 apply_bytes -= ret;
175 msg->sg.size -= ret;
176 sge->offset += ret;
177 sge->length -= ret;
178 if (uncharge)
179 sk_mem_uncharge(sk, ret);
180 if (ret != size) {
181 size -= ret;
182 off += ret;
183 goto retry;
184 }
185 if (!sge->length) {
186 put_page(page);
187 sk_msg_iter_next(msg, start);
188 sg_init_table(sge, 1);
189 if (msg->sg.start == msg->sg.end)
190 break;
191 }
192 if (apply && !apply_bytes)
193 break;
194 }
195
196 return 0;
197}
198
199static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
200 u32 apply_bytes, int flags, bool uncharge)
201{
202 int ret;
203
204 lock_sock(sk);
205 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
206 release_sock(sk);
207 return ret;
208}
209
210int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
211 u32 bytes, int flags)
212{
213 bool ingress = sk_msg_to_ingress(msg);
214 struct sk_psock *psock = sk_psock_get(sk);
215 int ret;
216
217 if (unlikely(!psock)) {
218 sk_msg_free(sk, msg);
219 return 0;
220 }
221 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
222 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
223 sk_psock_put(sk, psock);
224 return ret;
225}
226EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
227
228#ifdef CONFIG_BPF_STREAM_PARSER
229static bool tcp_bpf_stream_read(const struct sock *sk)
230{
231 struct sk_psock *psock;
232 bool empty = true;
233
234 rcu_read_lock();
235 psock = sk_psock(sk);
236 if (likely(psock))
237 empty = list_empty(&psock->ingress_msg);
238 rcu_read_unlock();
239 return !empty;
240}
241
242static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock,
243 int flags, long timeo, int *err)
244{
245 DEFINE_WAIT_FUNC(wait, woken_wake_function);
246 int ret = 0;
247
248 if (sk->sk_shutdown & RCV_SHUTDOWN)
249 return 1;
250
251 if (!timeo)
252 return ret;
253
254 add_wait_queue(sk_sleep(sk), &wait);
255 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
256 ret = sk_wait_event(sk, &timeo,
257 !list_empty(&psock->ingress_msg) ||
258 !skb_queue_empty(&sk->sk_receive_queue), &wait);
259 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
260 remove_wait_queue(sk_sleep(sk), &wait);
261 return ret;
262}
263
264static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
265 int nonblock, int flags, int *addr_len)
266{
267 struct sk_psock *psock;
268 int copied, ret;
269
270 if (unlikely(flags & MSG_ERRQUEUE))
271 return inet_recv_error(sk, msg, len, addr_len);
272
273 psock = sk_psock_get(sk);
274 if (unlikely(!psock))
275 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
276 if (!skb_queue_empty(&sk->sk_receive_queue) &&
277 sk_psock_queue_empty(psock)) {
278 sk_psock_put(sk, psock);
279 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
280 }
281 lock_sock(sk);
282msg_bytes_ready:
283 copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags);
284 if (!copied) {
285 int data, err = 0;
286 long timeo;
287
288 timeo = sock_rcvtimeo(sk, nonblock);
289 data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err);
290 if (data) {
291 if (!sk_psock_queue_empty(psock))
292 goto msg_bytes_ready;
293 release_sock(sk);
294 sk_psock_put(sk, psock);
295 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
296 }
297 if (err) {
298 ret = err;
299 goto out;
300 }
301 copied = -EAGAIN;
302 }
303 ret = copied;
304out:
305 release_sock(sk);
306 sk_psock_put(sk, psock);
307 return ret;
308}
309
310static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
311 struct sk_msg *msg, int *copied, int flags)
312{
313 bool cork = false, enospc = sk_msg_full(msg);
314 struct sock *sk_redir;
315 u32 tosend, delta = 0;
316 int ret;
317
318more_data:
319 if (psock->eval == __SK_NONE) {
320 /* Track delta in msg size to add/subtract it on SK_DROP from
321 * returned to user copied size. This ensures user doesn't
322 * get a positive return code with msg_cut_data and SK_DROP
323 * verdict.
324 */
325 delta = msg->sg.size;
326 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
327 delta -= msg->sg.size;
328 }
329
330 if (msg->cork_bytes &&
331 msg->cork_bytes > msg->sg.size && !enospc) {
332 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
333 if (!psock->cork) {
334 psock->cork = kzalloc(sizeof(*psock->cork),
335 GFP_ATOMIC | __GFP_NOWARN);
336 if (!psock->cork)
337 return -ENOMEM;
338 }
339 memcpy(psock->cork, msg, sizeof(*msg));
340 return 0;
341 }
342
343 tosend = msg->sg.size;
344 if (psock->apply_bytes && psock->apply_bytes < tosend)
345 tosend = psock->apply_bytes;
346
347 switch (psock->eval) {
348 case __SK_PASS:
349 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
350 if (unlikely(ret)) {
351 *copied -= sk_msg_free(sk, msg);
352 break;
353 }
354 sk_msg_apply_bytes(psock, tosend);
355 break;
356 case __SK_REDIRECT:
357 sk_redir = psock->sk_redir;
358 sk_msg_apply_bytes(psock, tosend);
359 if (psock->cork) {
360 cork = true;
361 psock->cork = NULL;
362 }
363 sk_msg_return(sk, msg, tosend);
364 release_sock(sk);
365 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
366 lock_sock(sk);
367 if (unlikely(ret < 0)) {
368 int free = sk_msg_free_nocharge(sk, msg);
369
370 if (!cork)
371 *copied -= free;
372 }
373 if (cork) {
374 sk_msg_free(sk, msg);
375 kfree(msg);
376 msg = NULL;
377 ret = 0;
378 }
379 break;
380 case __SK_DROP:
381 default:
382 sk_msg_free_partial(sk, msg, tosend);
383 sk_msg_apply_bytes(psock, tosend);
384 *copied -= (tosend + delta);
385 return -EACCES;
386 }
387
388 if (likely(!ret)) {
389 if (!psock->apply_bytes) {
390 psock->eval = __SK_NONE;
391 if (psock->sk_redir) {
392 sock_put(psock->sk_redir);
393 psock->sk_redir = NULL;
394 }
395 }
396 if (msg &&
397 msg->sg.data[msg->sg.start].page_link &&
398 msg->sg.data[msg->sg.start].length)
399 goto more_data;
400 }
401 return ret;
402}
403
404static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
405{
406 struct sk_msg tmp, *msg_tx = NULL;
407 int copied = 0, err = 0;
408 struct sk_psock *psock;
409 long timeo;
410 int flags;
411
412 /* Don't let internal do_tcp_sendpages() flags through */
413 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
414 flags |= MSG_NO_SHARED_FRAGS;
415
416 psock = sk_psock_get(sk);
417 if (unlikely(!psock))
418 return tcp_sendmsg(sk, msg, size);
419
420 lock_sock(sk);
421 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
422 while (msg_data_left(msg)) {
423 bool enospc = false;
424 u32 copy, osize;
425
426 if (sk->sk_err) {
427 err = -sk->sk_err;
428 goto out_err;
429 }
430
431 copy = msg_data_left(msg);
432 if (!sk_stream_memory_free(sk))
433 goto wait_for_sndbuf;
434 if (psock->cork) {
435 msg_tx = psock->cork;
436 } else {
437 msg_tx = &tmp;
438 sk_msg_init(msg_tx);
439 }
440
441 osize = msg_tx->sg.size;
442 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
443 if (err) {
444 if (err != -ENOSPC)
445 goto wait_for_memory;
446 enospc = true;
447 copy = msg_tx->sg.size - osize;
448 }
449
450 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
451 copy);
452 if (err < 0) {
453 sk_msg_trim(sk, msg_tx, osize);
454 goto out_err;
455 }
456
457 copied += copy;
458 if (psock->cork_bytes) {
459 if (size > psock->cork_bytes)
460 psock->cork_bytes = 0;
461 else
462 psock->cork_bytes -= size;
463 if (psock->cork_bytes && !enospc)
464 goto out_err;
465 /* All cork bytes are accounted, rerun the prog. */
466 psock->eval = __SK_NONE;
467 psock->cork_bytes = 0;
468 }
469
470 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
471 if (unlikely(err < 0))
472 goto out_err;
473 continue;
474wait_for_sndbuf:
475 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
476wait_for_memory:
477 err = sk_stream_wait_memory(sk, &timeo);
478 if (err) {
479 if (msg_tx && msg_tx != psock->cork)
480 sk_msg_free(sk, msg_tx);
481 goto out_err;
482 }
483 }
484out_err:
485 if (err < 0)
486 err = sk_stream_error(sk, msg->msg_flags, err);
487 release_sock(sk);
488 sk_psock_put(sk, psock);
489 return copied ? copied : err;
490}
491
492static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
493 size_t size, int flags)
494{
495 struct sk_msg tmp, *msg = NULL;
496 int err = 0, copied = 0;
497 struct sk_psock *psock;
498 bool enospc = false;
499
500 psock = sk_psock_get(sk);
501 if (unlikely(!psock))
502 return tcp_sendpage(sk, page, offset, size, flags);
503
504 lock_sock(sk);
505 if (psock->cork) {
506 msg = psock->cork;
507 } else {
508 msg = &tmp;
509 sk_msg_init(msg);
510 }
511
512 /* Catch case where ring is full and sendpage is stalled. */
513 if (unlikely(sk_msg_full(msg)))
514 goto out_err;
515
516 sk_msg_page_add(msg, page, size, offset);
517 sk_mem_charge(sk, size);
518 copied = size;
519 if (sk_msg_full(msg))
520 enospc = true;
521 if (psock->cork_bytes) {
522 if (size > psock->cork_bytes)
523 psock->cork_bytes = 0;
524 else
525 psock->cork_bytes -= size;
526 if (psock->cork_bytes && !enospc)
527 goto out_err;
528 /* All cork bytes are accounted, rerun the prog. */
529 psock->eval = __SK_NONE;
530 psock->cork_bytes = 0;
531 }
532
533 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
534out_err:
535 release_sock(sk);
536 sk_psock_put(sk, psock);
537 return copied ? copied : err;
538}
539
540enum {
541 TCP_BPF_IPV4,
542 TCP_BPF_IPV6,
543 TCP_BPF_NUM_PROTS,
544};
545
546enum {
547 TCP_BPF_BASE,
548 TCP_BPF_TX,
549 TCP_BPF_NUM_CFGS,
550};
551
552static struct proto *tcpv6_prot_saved __read_mostly;
553static DEFINE_SPINLOCK(tcpv6_prot_lock);
554static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
555
556static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
557 struct proto *base)
558{
559 prot[TCP_BPF_BASE] = *base;
560 prot[TCP_BPF_BASE].unhash = sock_map_unhash;
561 prot[TCP_BPF_BASE].close = sock_map_close;
562 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
563 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read;
564
565 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
566 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
567 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
568}
569
570static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops)
571{
572 if (sk->sk_family == AF_INET6 &&
573 unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
574 spin_lock_bh(&tcpv6_prot_lock);
575 if (likely(ops != tcpv6_prot_saved)) {
576 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
577 smp_store_release(&tcpv6_prot_saved, ops);
578 }
579 spin_unlock_bh(&tcpv6_prot_lock);
580 }
581}
582
583static int __init tcp_bpf_v4_build_proto(void)
584{
585 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
586 return 0;
587}
588core_initcall(tcp_bpf_v4_build_proto);
589
590static int tcp_bpf_assert_proto_ops(struct proto *ops)
591{
592 /* In order to avoid retpoline, we make assumptions when we call
593 * into ops if e.g. a psock is not present. Make sure they are
594 * indeed valid assumptions.
595 */
596 return ops->recvmsg == tcp_recvmsg &&
597 ops->sendmsg == tcp_sendmsg &&
598 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
599}
600
601struct proto *tcp_bpf_get_proto(struct sock *sk, struct sk_psock *psock)
602{
603 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
604 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
605
606 if (!psock->sk_proto) {
607 struct proto *ops = READ_ONCE(sk->sk_prot);
608
609 if (tcp_bpf_assert_proto_ops(ops))
610 return ERR_PTR(-EINVAL);
611
612 tcp_bpf_check_v6_needs_rebuild(sk, ops);
613 }
614
615 return &tcp_bpf_prots[family][config];
616}
617
618/* If a child got cloned from a listening socket that had tcp_bpf
619 * protocol callbacks installed, we need to restore the callbacks to
620 * the default ones because the child does not inherit the psock state
621 * that tcp_bpf callbacks expect.
622 */
623void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
624{
625 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
626 struct proto *prot = newsk->sk_prot;
627
628 if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
629 newsk->sk_prot = sk->sk_prot_creator;
630}
631#endif /* CONFIG_BPF_STREAM_PARSER */