1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
4 #include <linux/skmsg.h>
5 #include <linux/filter.h>
7 #include <linux/init.h>
8 #include <linux/wait.h>
9 #include <linux/util_macros.h>
11 #include <net/inet_common.h>
14 void tcp_eat_skb(struct sock *sk, struct sk_buff *skb)
19 if (!skb || !skb->len || !sk_is_tcp(sk))
22 if (skb_bpf_strparser(skb))
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);
32 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
33 struct sk_msg *msg, u32 apply_bytes, int flags)
35 bool apply = apply_bytes;
36 struct scatterlist *sge;
41 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
46 tmp->sg.start = msg->sg.start;
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)) {
58 sk_mem_charge(sk, size);
59 sk_msg_xfer(tmp, msg, i, size);
62 get_page(sk_msg_page(tmp, i));
63 sk_msg_iter_var_next(i);
69 sk_msg_iter_var_prev(i);
73 } while (i != msg->sg.end);
77 sk_psock_queue_msg(psock, tmp);
78 sk_psock_data_ready(sk, psock);
88 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
89 int flags, bool uncharge)
91 bool apply = apply_bytes;
92 struct scatterlist *sge;
100 sge = sk_msg_elem(msg, msg->sg.start);
101 size = (apply && apply_bytes < sge->length) ?
102 apply_bytes : sge->length;
106 tcp_rate_check_app_limited(sk);
108 has_tx_ulp = tls_sw_has_ctx_tx(sk);
110 flags |= MSG_SENDPAGE_NOPOLICY;
111 ret = kernel_sendpage_locked(sk,
112 page, off, size, flags);
114 ret = do_tcp_sendpages(sk, page, off, size, flags);
125 sk_mem_uncharge(sk, ret);
133 sk_msg_iter_next(msg, start);
134 sg_init_table(sge, 1);
135 if (msg->sg.start == msg->sg.end)
138 if (apply && !apply_bytes)
145 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
146 u32 apply_bytes, int flags, bool uncharge)
151 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
156 int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress,
157 struct sk_msg *msg, u32 bytes, int flags)
159 struct sk_psock *psock = sk_psock_get(sk);
162 if (unlikely(!psock))
165 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
166 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
167 sk_psock_put(sk, psock);
170 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
172 #ifdef CONFIG_BPF_SYSCALL
173 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
176 DEFINE_WAIT_FUNC(wait, woken_wake_function);
179 if (sk->sk_shutdown & RCV_SHUTDOWN)
185 add_wait_queue(sk_sleep(sk), &wait);
186 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
187 ret = sk_wait_event(sk, &timeo,
188 !list_empty(&psock->ingress_msg) ||
189 !skb_queue_empty_lockless(&sk->sk_receive_queue), &wait);
190 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
191 remove_wait_queue(sk_sleep(sk), &wait);
195 static bool is_next_msg_fin(struct sk_psock *psock)
197 struct scatterlist *sge;
198 struct sk_msg *msg_rx;
201 msg_rx = sk_psock_peek_msg(psock);
202 i = msg_rx->sg.start;
203 sge = sk_msg_elem(msg_rx, i);
205 struct sk_buff *skb = msg_rx->skb;
207 if (skb && TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
213 static int tcp_bpf_recvmsg_parser(struct sock *sk,
219 struct tcp_sock *tcp = tcp_sk(sk);
220 int peek = flags & MSG_PEEK;
221 u32 seq = tcp->copied_seq;
222 struct sk_psock *psock;
225 if (unlikely(flags & MSG_ERRQUEUE))
226 return inet_recv_error(sk, msg, len, addr_len);
231 psock = sk_psock_get(sk);
232 if (unlikely(!psock))
233 return tcp_recvmsg(sk, msg, len, flags, addr_len);
237 /* We may have received data on the sk_receive_queue pre-accept and
238 * then we can not use read_skb in this context because we haven't
239 * assigned a sk_socket yet so have no link to the ops. The work-around
240 * is to check the sk_receive_queue and in these cases read skbs off
241 * queue again. The read_skb hook is not running at this point because
242 * of lock_sock so we avoid having multiple runners in read_skb.
244 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) {
246 /* This handles the ENOMEM errors if we both receive data
247 * pre accept and are already under memory pressure. At least
248 * let user know to retry.
250 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) {
257 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
258 /* The typical case for EFAULT is the socket was gracefully
259 * shutdown with a FIN pkt. So check here the other case is
260 * some error on copy_page_to_iter which would be unexpected.
261 * On fin return correct return code to zero.
263 if (copied == -EFAULT) {
264 bool is_fin = is_next_msg_fin(psock);
277 if (sock_flag(sk, SOCK_DONE))
281 copied = sock_error(sk);
285 if (sk->sk_shutdown & RCV_SHUTDOWN)
288 if (sk->sk_state == TCP_CLOSE) {
293 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
299 if (signal_pending(current)) {
300 copied = sock_intr_errno(timeo);
304 data = tcp_msg_wait_data(sk, psock, timeo);
309 if (data && !sk_psock_queue_empty(psock))
310 goto msg_bytes_ready;
315 WRITE_ONCE(tcp->copied_seq, seq);
316 tcp_rcv_space_adjust(sk);
318 __tcp_cleanup_rbuf(sk, copied);
322 sk_psock_put(sk, psock);
326 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
327 int flags, int *addr_len)
329 struct sk_psock *psock;
332 if (unlikely(flags & MSG_ERRQUEUE))
333 return inet_recv_error(sk, msg, len, addr_len);
338 psock = sk_psock_get(sk);
339 if (unlikely(!psock))
340 return tcp_recvmsg(sk, msg, len, flags, addr_len);
341 if (!skb_queue_empty(&sk->sk_receive_queue) &&
342 sk_psock_queue_empty(psock)) {
343 sk_psock_put(sk, psock);
344 return tcp_recvmsg(sk, msg, len, flags, addr_len);
348 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
353 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
354 data = tcp_msg_wait_data(sk, psock, timeo);
360 if (!sk_psock_queue_empty(psock))
361 goto msg_bytes_ready;
363 sk_psock_put(sk, psock);
364 return tcp_recvmsg(sk, msg, len, flags, addr_len);
372 sk_psock_put(sk, psock);
376 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
377 struct sk_msg *msg, int *copied, int flags)
379 bool cork = false, enospc = sk_msg_full(msg), redir_ingress;
380 struct sock *sk_redir;
381 u32 tosend, origsize, sent, delta = 0;
386 if (psock->eval == __SK_NONE) {
387 /* Track delta in msg size to add/subtract it on SK_DROP from
388 * returned to user copied size. This ensures user doesn't
389 * get a positive return code with msg_cut_data and SK_DROP
392 delta = msg->sg.size;
393 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
394 delta -= msg->sg.size;
397 if (msg->cork_bytes &&
398 msg->cork_bytes > msg->sg.size && !enospc) {
399 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
401 psock->cork = kzalloc(sizeof(*psock->cork),
402 GFP_ATOMIC | __GFP_NOWARN);
406 memcpy(psock->cork, msg, sizeof(*msg));
410 tosend = msg->sg.size;
411 if (psock->apply_bytes && psock->apply_bytes < tosend)
412 tosend = psock->apply_bytes;
415 switch (psock->eval) {
417 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
419 *copied -= sk_msg_free(sk, msg);
422 sk_msg_apply_bytes(psock, tosend);
425 redir_ingress = psock->redir_ingress;
426 sk_redir = psock->sk_redir;
427 sk_msg_apply_bytes(psock, tosend);
428 if (!psock->apply_bytes) {
429 /* Clean up before releasing the sock lock. */
431 psock->eval = __SK_NONE;
432 psock->sk_redir = NULL;
438 sk_msg_return(sk, msg, tosend);
441 origsize = msg->sg.size;
442 ret = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
444 sent = origsize - msg->sg.size;
446 if (eval == __SK_REDIRECT)
450 if (unlikely(ret < 0)) {
451 int free = sk_msg_free_nocharge(sk, msg);
457 sk_msg_free(sk, msg);
465 sk_msg_free_partial(sk, msg, tosend);
466 sk_msg_apply_bytes(psock, tosend);
467 *copied -= (tosend + delta);
472 if (!psock->apply_bytes) {
473 psock->eval = __SK_NONE;
474 if (psock->sk_redir) {
475 sock_put(psock->sk_redir);
476 psock->sk_redir = NULL;
480 msg->sg.data[msg->sg.start].page_link &&
481 msg->sg.data[msg->sg.start].length) {
482 if (eval == __SK_REDIRECT)
483 sk_mem_charge(sk, tosend - sent);
490 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
492 struct sk_msg tmp, *msg_tx = NULL;
493 int copied = 0, err = 0;
494 struct sk_psock *psock;
498 /* Don't let internal do_tcp_sendpages() flags through */
499 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
500 flags |= MSG_NO_SHARED_FRAGS;
502 psock = sk_psock_get(sk);
503 if (unlikely(!psock))
504 return tcp_sendmsg(sk, msg, size);
507 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
508 while (msg_data_left(msg)) {
517 copy = msg_data_left(msg);
518 if (!sk_stream_memory_free(sk))
519 goto wait_for_sndbuf;
521 msg_tx = psock->cork;
527 osize = msg_tx->sg.size;
528 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
531 goto wait_for_memory;
533 copy = msg_tx->sg.size - osize;
536 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
539 sk_msg_trim(sk, msg_tx, osize);
544 if (psock->cork_bytes) {
545 if (size > psock->cork_bytes)
546 psock->cork_bytes = 0;
548 psock->cork_bytes -= size;
549 if (psock->cork_bytes && !enospc)
551 /* All cork bytes are accounted, rerun the prog. */
552 psock->eval = __SK_NONE;
553 psock->cork_bytes = 0;
556 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
557 if (unlikely(err < 0))
561 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
563 err = sk_stream_wait_memory(sk, &timeo);
565 if (msg_tx && msg_tx != psock->cork)
566 sk_msg_free(sk, msg_tx);
572 err = sk_stream_error(sk, msg->msg_flags, err);
574 sk_psock_put(sk, psock);
575 return copied ? copied : err;
578 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
579 size_t size, int flags)
581 struct sk_msg tmp, *msg = NULL;
582 int err = 0, copied = 0;
583 struct sk_psock *psock;
586 psock = sk_psock_get(sk);
587 if (unlikely(!psock))
588 return tcp_sendpage(sk, page, offset, size, flags);
598 /* Catch case where ring is full and sendpage is stalled. */
599 if (unlikely(sk_msg_full(msg)))
602 sk_msg_page_add(msg, page, size, offset);
603 sk_mem_charge(sk, size);
605 if (sk_msg_full(msg))
607 if (psock->cork_bytes) {
608 if (size > psock->cork_bytes)
609 psock->cork_bytes = 0;
611 psock->cork_bytes -= size;
612 if (psock->cork_bytes && !enospc)
614 /* All cork bytes are accounted, rerun the prog. */
615 psock->eval = __SK_NONE;
616 psock->cork_bytes = 0;
619 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
622 sk_psock_put(sk, psock);
623 return copied ? copied : err;
640 static struct proto *tcpv6_prot_saved __read_mostly;
641 static DEFINE_SPINLOCK(tcpv6_prot_lock);
642 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
644 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
647 prot[TCP_BPF_BASE] = *base;
648 prot[TCP_BPF_BASE].destroy = sock_map_destroy;
649 prot[TCP_BPF_BASE].close = sock_map_close;
650 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
651 prot[TCP_BPF_BASE].sock_is_readable = sk_msg_is_readable;
653 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
654 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
655 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
657 prot[TCP_BPF_RX] = prot[TCP_BPF_BASE];
658 prot[TCP_BPF_RX].recvmsg = tcp_bpf_recvmsg_parser;
660 prot[TCP_BPF_TXRX] = prot[TCP_BPF_TX];
661 prot[TCP_BPF_TXRX].recvmsg = tcp_bpf_recvmsg_parser;
664 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
666 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
667 spin_lock_bh(&tcpv6_prot_lock);
668 if (likely(ops != tcpv6_prot_saved)) {
669 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
670 smp_store_release(&tcpv6_prot_saved, ops);
672 spin_unlock_bh(&tcpv6_prot_lock);
676 static int __init tcp_bpf_v4_build_proto(void)
678 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
681 late_initcall(tcp_bpf_v4_build_proto);
683 static int tcp_bpf_assert_proto_ops(struct proto *ops)
685 /* In order to avoid retpoline, we make assumptions when we call
686 * into ops if e.g. a psock is not present. Make sure they are
687 * indeed valid assumptions.
689 return ops->recvmsg == tcp_recvmsg &&
690 ops->sendmsg == tcp_sendmsg &&
691 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
694 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
696 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
697 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
699 if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
700 config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
704 if (inet_csk_has_ulp(sk)) {
705 /* TLS does not have an unhash proto in SW cases,
706 * but we need to ensure we stop using the sock_map
707 * unhash routine because the associated psock is being
708 * removed. So use the original unhash handler.
710 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
711 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
713 sk->sk_write_space = psock->saved_write_space;
714 /* Pairs with lockless read in sk_clone_lock() */
715 sock_replace_proto(sk, psock->sk_proto);
720 if (sk->sk_family == AF_INET6) {
721 if (tcp_bpf_assert_proto_ops(psock->sk_proto))
724 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
727 /* Pairs with lockless read in sk_clone_lock() */
728 sock_replace_proto(sk, &tcp_bpf_prots[family][config]);
731 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
733 /* If a child got cloned from a listening socket that had tcp_bpf
734 * protocol callbacks installed, we need to restore the callbacks to
735 * the default ones because the child does not inherit the psock state
736 * that tcp_bpf callbacks expect.
738 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
740 struct proto *prot = newsk->sk_prot;
742 if (is_insidevar(prot, tcp_bpf_prots))
743 newsk->sk_prot = sk->sk_prot_creator;
745 #endif /* CONFIG_BPF_SYSCALL */