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