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 struct msghdr msghdr = {};
92 bool apply = apply_bytes;
93 struct scatterlist *sge;
102 sge = sk_msg_elem(msg, msg->sg.start);
103 size = (apply && apply_bytes < sge->length) ?
104 apply_bytes : sge->length;
108 tcp_rate_check_app_limited(sk);
110 msghdr.msg_flags = flags | MSG_SPLICE_PAGES;
111 has_tx_ulp = tls_sw_has_ctx_tx(sk);
113 msghdr.msg_flags |= MSG_SENDPAGE_NOPOLICY;
115 if (size < sge->length && msg->sg.start != msg->sg.end)
116 msghdr.msg_flags |= MSG_MORE;
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);
130 sk_mem_uncharge(sk, ret);
138 sk_msg_iter_next(msg, start);
139 sg_init_table(sge, 1);
140 if (msg->sg.start == msg->sg.end)
143 if (apply && !apply_bytes)
150 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
151 u32 apply_bytes, int flags, bool uncharge)
156 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
161 int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress,
162 struct sk_msg *msg, u32 bytes, int flags)
164 struct sk_psock *psock = sk_psock_get(sk);
167 if (unlikely(!psock))
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);
175 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
177 #ifdef CONFIG_BPF_SYSCALL
178 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
181 DEFINE_WAIT_FUNC(wait, woken_wake_function);
184 if (sk->sk_shutdown & RCV_SHUTDOWN)
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);
200 static bool is_next_msg_fin(struct sk_psock *psock)
202 struct scatterlist *sge;
203 struct sk_msg *msg_rx;
206 msg_rx = sk_psock_peek_msg(psock);
207 i = msg_rx->sg.start;
208 sge = sk_msg_elem(msg_rx, i);
210 struct sk_buff *skb = msg_rx->skb;
212 if (skb && TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
218 static int tcp_bpf_recvmsg_parser(struct sock *sk,
224 struct tcp_sock *tcp = tcp_sk(sk);
225 u32 seq = tcp->copied_seq;
226 struct sk_psock *psock;
229 if (unlikely(flags & MSG_ERRQUEUE))
230 return inet_recv_error(sk, msg, len, addr_len);
235 psock = sk_psock_get(sk);
236 if (unlikely(!psock))
237 return tcp_recvmsg(sk, msg, len, flags, addr_len);
241 /* We may have received data on the sk_receive_queue pre-accept and
242 * then we can not use read_skb in this context because we haven't
243 * assigned a sk_socket yet so have no link to the ops. The work-around
244 * is to check the sk_receive_queue and in these cases read skbs off
245 * queue again. The read_skb hook is not running at this point because
246 * of lock_sock so we avoid having multiple runners in read_skb.
248 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) {
250 /* This handles the ENOMEM errors if we both receive data
251 * pre accept and are already under memory pressure. At least
252 * let user know to retry.
254 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) {
261 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
262 /* The typical case for EFAULT is the socket was gracefully
263 * shutdown with a FIN pkt. So check here the other case is
264 * some error on copy_page_to_iter which would be unexpected.
265 * On fin return correct return code to zero.
267 if (copied == -EFAULT) {
268 bool is_fin = is_next_msg_fin(psock);
281 if (sock_flag(sk, SOCK_DONE))
285 copied = sock_error(sk);
289 if (sk->sk_shutdown & RCV_SHUTDOWN)
292 if (sk->sk_state == TCP_CLOSE) {
297 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
303 if (signal_pending(current)) {
304 copied = sock_intr_errno(timeo);
308 data = tcp_msg_wait_data(sk, psock, timeo);
309 if (data && !sk_psock_queue_empty(psock))
310 goto msg_bytes_ready;
314 WRITE_ONCE(tcp->copied_seq, seq);
315 tcp_rcv_space_adjust(sk);
317 __tcp_cleanup_rbuf(sk, copied);
319 sk_psock_put(sk, psock);
323 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
324 int flags, int *addr_len)
326 struct sk_psock *psock;
329 if (unlikely(flags & MSG_ERRQUEUE))
330 return inet_recv_error(sk, msg, len, addr_len);
335 psock = sk_psock_get(sk);
336 if (unlikely(!psock))
337 return tcp_recvmsg(sk, msg, len, flags, addr_len);
338 if (!skb_queue_empty(&sk->sk_receive_queue) &&
339 sk_psock_queue_empty(psock)) {
340 sk_psock_put(sk, psock);
341 return tcp_recvmsg(sk, msg, len, flags, addr_len);
345 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
350 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
351 data = tcp_msg_wait_data(sk, psock, timeo);
353 if (!sk_psock_queue_empty(psock))
354 goto msg_bytes_ready;
356 sk_psock_put(sk, psock);
357 return tcp_recvmsg(sk, msg, len, flags, addr_len);
363 sk_psock_put(sk, psock);
367 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
368 struct sk_msg *msg, int *copied, int flags)
370 bool cork = false, enospc = sk_msg_full(msg), redir_ingress;
371 struct sock *sk_redir;
372 u32 tosend, origsize, sent, delta = 0;
377 if (psock->eval == __SK_NONE) {
378 /* Track delta in msg size to add/subtract it on SK_DROP from
379 * returned to user copied size. This ensures user doesn't
380 * get a positive return code with msg_cut_data and SK_DROP
383 delta = msg->sg.size;
384 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
385 delta -= msg->sg.size;
388 if (msg->cork_bytes &&
389 msg->cork_bytes > msg->sg.size && !enospc) {
390 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
392 psock->cork = kzalloc(sizeof(*psock->cork),
393 GFP_ATOMIC | __GFP_NOWARN);
397 memcpy(psock->cork, msg, sizeof(*msg));
401 tosend = msg->sg.size;
402 if (psock->apply_bytes && psock->apply_bytes < tosend)
403 tosend = psock->apply_bytes;
406 switch (psock->eval) {
408 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
410 *copied -= sk_msg_free(sk, msg);
413 sk_msg_apply_bytes(psock, tosend);
416 redir_ingress = psock->redir_ingress;
417 sk_redir = psock->sk_redir;
418 sk_msg_apply_bytes(psock, tosend);
419 if (!psock->apply_bytes) {
420 /* Clean up before releasing the sock lock. */
422 psock->eval = __SK_NONE;
423 psock->sk_redir = NULL;
429 sk_msg_return(sk, msg, tosend);
432 origsize = msg->sg.size;
433 ret = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
435 sent = origsize - msg->sg.size;
437 if (eval == __SK_REDIRECT)
441 if (unlikely(ret < 0)) {
442 int free = sk_msg_free_nocharge(sk, msg);
448 sk_msg_free(sk, msg);
456 sk_msg_free_partial(sk, msg, tosend);
457 sk_msg_apply_bytes(psock, tosend);
458 *copied -= (tosend + delta);
463 if (!psock->apply_bytes) {
464 psock->eval = __SK_NONE;
465 if (psock->sk_redir) {
466 sock_put(psock->sk_redir);
467 psock->sk_redir = NULL;
471 msg->sg.data[msg->sg.start].page_link &&
472 msg->sg.data[msg->sg.start].length) {
473 if (eval == __SK_REDIRECT)
474 sk_mem_charge(sk, tosend - sent);
481 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
483 struct sk_msg tmp, *msg_tx = NULL;
484 int copied = 0, err = 0;
485 struct sk_psock *psock;
489 /* Don't let internal flags through */
490 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
491 flags |= MSG_NO_SHARED_FRAGS;
493 psock = sk_psock_get(sk);
494 if (unlikely(!psock))
495 return tcp_sendmsg(sk, msg, size);
498 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
499 while (msg_data_left(msg)) {
508 copy = msg_data_left(msg);
509 if (!sk_stream_memory_free(sk))
510 goto wait_for_sndbuf;
512 msg_tx = psock->cork;
518 osize = msg_tx->sg.size;
519 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
522 goto wait_for_memory;
524 copy = msg_tx->sg.size - osize;
527 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
530 sk_msg_trim(sk, msg_tx, osize);
535 if (psock->cork_bytes) {
536 if (size > psock->cork_bytes)
537 psock->cork_bytes = 0;
539 psock->cork_bytes -= size;
540 if (psock->cork_bytes && !enospc)
542 /* All cork bytes are accounted, rerun the prog. */
543 psock->eval = __SK_NONE;
544 psock->cork_bytes = 0;
547 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
548 if (unlikely(err < 0))
552 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
554 err = sk_stream_wait_memory(sk, &timeo);
556 if (msg_tx && msg_tx != psock->cork)
557 sk_msg_free(sk, msg_tx);
563 err = sk_stream_error(sk, msg->msg_flags, err);
565 sk_psock_put(sk, psock);
566 return copied ? copied : err;
583 static struct proto *tcpv6_prot_saved __read_mostly;
584 static DEFINE_SPINLOCK(tcpv6_prot_lock);
585 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
587 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
590 prot[TCP_BPF_BASE] = *base;
591 prot[TCP_BPF_BASE].destroy = sock_map_destroy;
592 prot[TCP_BPF_BASE].close = sock_map_close;
593 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
594 prot[TCP_BPF_BASE].sock_is_readable = sk_msg_is_readable;
596 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
597 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
599 prot[TCP_BPF_RX] = prot[TCP_BPF_BASE];
600 prot[TCP_BPF_RX].recvmsg = tcp_bpf_recvmsg_parser;
602 prot[TCP_BPF_TXRX] = prot[TCP_BPF_TX];
603 prot[TCP_BPF_TXRX].recvmsg = tcp_bpf_recvmsg_parser;
606 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
608 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
609 spin_lock_bh(&tcpv6_prot_lock);
610 if (likely(ops != tcpv6_prot_saved)) {
611 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
612 smp_store_release(&tcpv6_prot_saved, ops);
614 spin_unlock_bh(&tcpv6_prot_lock);
618 static int __init tcp_bpf_v4_build_proto(void)
620 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
623 late_initcall(tcp_bpf_v4_build_proto);
625 static int tcp_bpf_assert_proto_ops(struct proto *ops)
627 /* In order to avoid retpoline, we make assumptions when we call
628 * into ops if e.g. a psock is not present. Make sure they are
629 * indeed valid assumptions.
631 return ops->recvmsg == tcp_recvmsg &&
632 ops->sendmsg == tcp_sendmsg ? 0 : -ENOTSUPP;
635 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
637 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
638 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
640 if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
641 config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
645 if (inet_csk_has_ulp(sk)) {
646 /* TLS does not have an unhash proto in SW cases,
647 * but we need to ensure we stop using the sock_map
648 * unhash routine because the associated psock is being
649 * removed. So use the original unhash handler.
651 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
652 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
654 sk->sk_write_space = psock->saved_write_space;
655 /* Pairs with lockless read in sk_clone_lock() */
656 sock_replace_proto(sk, psock->sk_proto);
661 if (sk->sk_family == AF_INET6) {
662 if (tcp_bpf_assert_proto_ops(psock->sk_proto))
665 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
668 /* Pairs with lockless read in sk_clone_lock() */
669 sock_replace_proto(sk, &tcp_bpf_prots[family][config]);
672 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
674 /* If a child got cloned from a listening socket that had tcp_bpf
675 * protocol callbacks installed, we need to restore the callbacks to
676 * the default ones because the child does not inherit the psock state
677 * that tcp_bpf callbacks expect.
679 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
681 struct proto *prot = newsk->sk_prot;
683 if (is_insidevar(prot, tcp_bpf_prots))
684 newsk->sk_prot = sk->sk_prot_creator;
686 #endif /* CONFIG_BPF_SYSCALL */