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 u32 seq = tcp->copied_seq;
221 struct sk_psock *psock;
224 if (unlikely(flags & MSG_ERRQUEUE))
225 return inet_recv_error(sk, msg, len, addr_len);
230 psock = sk_psock_get(sk);
231 if (unlikely(!psock))
232 return tcp_recvmsg(sk, msg, len, flags, addr_len);
236 /* We may have received data on the sk_receive_queue pre-accept and
237 * then we can not use read_skb in this context because we haven't
238 * assigned a sk_socket yet so have no link to the ops. The work-around
239 * is to check the sk_receive_queue and in these cases read skbs off
240 * queue again. The read_skb hook is not running at this point because
241 * of lock_sock so we avoid having multiple runners in read_skb.
243 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) {
245 /* This handles the ENOMEM errors if we both receive data
246 * pre accept and are already under memory pressure. At least
247 * let user know to retry.
249 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) {
256 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
257 /* The typical case for EFAULT is the socket was gracefully
258 * shutdown with a FIN pkt. So check here the other case is
259 * some error on copy_page_to_iter which would be unexpected.
260 * On fin return correct return code to zero.
262 if (copied == -EFAULT) {
263 bool is_fin = is_next_msg_fin(psock);
276 if (sock_flag(sk, SOCK_DONE))
280 copied = sock_error(sk);
284 if (sk->sk_shutdown & RCV_SHUTDOWN)
287 if (sk->sk_state == TCP_CLOSE) {
292 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
298 if (signal_pending(current)) {
299 copied = sock_intr_errno(timeo);
303 data = tcp_msg_wait_data(sk, psock, timeo);
304 if (data && !sk_psock_queue_empty(psock))
305 goto msg_bytes_ready;
309 WRITE_ONCE(tcp->copied_seq, seq);
310 tcp_rcv_space_adjust(sk);
312 __tcp_cleanup_rbuf(sk, copied);
314 sk_psock_put(sk, psock);
318 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
319 int flags, int *addr_len)
321 struct sk_psock *psock;
324 if (unlikely(flags & MSG_ERRQUEUE))
325 return inet_recv_error(sk, msg, len, addr_len);
330 psock = sk_psock_get(sk);
331 if (unlikely(!psock))
332 return tcp_recvmsg(sk, msg, len, flags, addr_len);
333 if (!skb_queue_empty(&sk->sk_receive_queue) &&
334 sk_psock_queue_empty(psock)) {
335 sk_psock_put(sk, psock);
336 return tcp_recvmsg(sk, msg, len, flags, addr_len);
340 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
345 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
346 data = tcp_msg_wait_data(sk, psock, timeo);
348 if (!sk_psock_queue_empty(psock))
349 goto msg_bytes_ready;
351 sk_psock_put(sk, psock);
352 return tcp_recvmsg(sk, msg, len, flags, addr_len);
358 sk_psock_put(sk, psock);
362 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
363 struct sk_msg *msg, int *copied, int flags)
365 bool cork = false, enospc = sk_msg_full(msg), redir_ingress;
366 struct sock *sk_redir;
367 u32 tosend, origsize, sent, delta = 0;
372 if (psock->eval == __SK_NONE) {
373 /* Track delta in msg size to add/subtract it on SK_DROP from
374 * returned to user copied size. This ensures user doesn't
375 * get a positive return code with msg_cut_data and SK_DROP
378 delta = msg->sg.size;
379 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
380 delta -= msg->sg.size;
383 if (msg->cork_bytes &&
384 msg->cork_bytes > msg->sg.size && !enospc) {
385 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
387 psock->cork = kzalloc(sizeof(*psock->cork),
388 GFP_ATOMIC | __GFP_NOWARN);
392 memcpy(psock->cork, msg, sizeof(*msg));
396 tosend = msg->sg.size;
397 if (psock->apply_bytes && psock->apply_bytes < tosend)
398 tosend = psock->apply_bytes;
401 switch (psock->eval) {
403 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
405 *copied -= sk_msg_free(sk, msg);
408 sk_msg_apply_bytes(psock, tosend);
411 redir_ingress = psock->redir_ingress;
412 sk_redir = psock->sk_redir;
413 sk_msg_apply_bytes(psock, tosend);
414 if (!psock->apply_bytes) {
415 /* Clean up before releasing the sock lock. */
417 psock->eval = __SK_NONE;
418 psock->sk_redir = NULL;
424 sk_msg_return(sk, msg, tosend);
427 origsize = msg->sg.size;
428 ret = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
430 sent = origsize - msg->sg.size;
432 if (eval == __SK_REDIRECT)
436 if (unlikely(ret < 0)) {
437 int free = sk_msg_free_nocharge(sk, msg);
443 sk_msg_free(sk, msg);
451 sk_msg_free_partial(sk, msg, tosend);
452 sk_msg_apply_bytes(psock, tosend);
453 *copied -= (tosend + delta);
458 if (!psock->apply_bytes) {
459 psock->eval = __SK_NONE;
460 if (psock->sk_redir) {
461 sock_put(psock->sk_redir);
462 psock->sk_redir = NULL;
466 msg->sg.data[msg->sg.start].page_link &&
467 msg->sg.data[msg->sg.start].length) {
468 if (eval == __SK_REDIRECT)
469 sk_mem_charge(sk, tosend - sent);
476 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
478 struct sk_msg tmp, *msg_tx = NULL;
479 int copied = 0, err = 0;
480 struct sk_psock *psock;
484 /* Don't let internal do_tcp_sendpages() flags through */
485 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
486 flags |= MSG_NO_SHARED_FRAGS;
488 psock = sk_psock_get(sk);
489 if (unlikely(!psock))
490 return tcp_sendmsg(sk, msg, size);
493 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
494 while (msg_data_left(msg)) {
503 copy = msg_data_left(msg);
504 if (!sk_stream_memory_free(sk))
505 goto wait_for_sndbuf;
507 msg_tx = psock->cork;
513 osize = msg_tx->sg.size;
514 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
517 goto wait_for_memory;
519 copy = msg_tx->sg.size - osize;
522 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
525 sk_msg_trim(sk, msg_tx, osize);
530 if (psock->cork_bytes) {
531 if (size > psock->cork_bytes)
532 psock->cork_bytes = 0;
534 psock->cork_bytes -= size;
535 if (psock->cork_bytes && !enospc)
537 /* All cork bytes are accounted, rerun the prog. */
538 psock->eval = __SK_NONE;
539 psock->cork_bytes = 0;
542 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
543 if (unlikely(err < 0))
547 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
549 err = sk_stream_wait_memory(sk, &timeo);
551 if (msg_tx && msg_tx != psock->cork)
552 sk_msg_free(sk, msg_tx);
558 err = sk_stream_error(sk, msg->msg_flags, err);
560 sk_psock_put(sk, psock);
561 return copied ? copied : err;
564 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
565 size_t size, int flags)
567 struct sk_msg tmp, *msg = NULL;
568 int err = 0, copied = 0;
569 struct sk_psock *psock;
572 psock = sk_psock_get(sk);
573 if (unlikely(!psock))
574 return tcp_sendpage(sk, page, offset, size, flags);
584 /* Catch case where ring is full and sendpage is stalled. */
585 if (unlikely(sk_msg_full(msg)))
588 sk_msg_page_add(msg, page, size, offset);
589 sk_mem_charge(sk, size);
591 if (sk_msg_full(msg))
593 if (psock->cork_bytes) {
594 if (size > psock->cork_bytes)
595 psock->cork_bytes = 0;
597 psock->cork_bytes -= size;
598 if (psock->cork_bytes && !enospc)
600 /* All cork bytes are accounted, rerun the prog. */
601 psock->eval = __SK_NONE;
602 psock->cork_bytes = 0;
605 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
608 sk_psock_put(sk, psock);
609 return copied ? copied : err;
626 static struct proto *tcpv6_prot_saved __read_mostly;
627 static DEFINE_SPINLOCK(tcpv6_prot_lock);
628 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
630 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
633 prot[TCP_BPF_BASE] = *base;
634 prot[TCP_BPF_BASE].destroy = sock_map_destroy;
635 prot[TCP_BPF_BASE].close = sock_map_close;
636 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
637 prot[TCP_BPF_BASE].sock_is_readable = sk_msg_is_readable;
639 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
640 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
641 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
643 prot[TCP_BPF_RX] = prot[TCP_BPF_BASE];
644 prot[TCP_BPF_RX].recvmsg = tcp_bpf_recvmsg_parser;
646 prot[TCP_BPF_TXRX] = prot[TCP_BPF_TX];
647 prot[TCP_BPF_TXRX].recvmsg = tcp_bpf_recvmsg_parser;
650 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
652 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
653 spin_lock_bh(&tcpv6_prot_lock);
654 if (likely(ops != tcpv6_prot_saved)) {
655 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
656 smp_store_release(&tcpv6_prot_saved, ops);
658 spin_unlock_bh(&tcpv6_prot_lock);
662 static int __init tcp_bpf_v4_build_proto(void)
664 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
667 late_initcall(tcp_bpf_v4_build_proto);
669 static int tcp_bpf_assert_proto_ops(struct proto *ops)
671 /* In order to avoid retpoline, we make assumptions when we call
672 * into ops if e.g. a psock is not present. Make sure they are
673 * indeed valid assumptions.
675 return ops->recvmsg == tcp_recvmsg &&
676 ops->sendmsg == tcp_sendmsg &&
677 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
680 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
682 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
683 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
685 if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
686 config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
690 if (inet_csk_has_ulp(sk)) {
691 /* TLS does not have an unhash proto in SW cases,
692 * but we need to ensure we stop using the sock_map
693 * unhash routine because the associated psock is being
694 * removed. So use the original unhash handler.
696 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
697 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
699 sk->sk_write_space = psock->saved_write_space;
700 /* Pairs with lockless read in sk_clone_lock() */
701 sock_replace_proto(sk, psock->sk_proto);
706 if (sk->sk_family == AF_INET6) {
707 if (tcp_bpf_assert_proto_ops(psock->sk_proto))
710 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
713 /* Pairs with lockless read in sk_clone_lock() */
714 sock_replace_proto(sk, &tcp_bpf_prots[family][config]);
717 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
719 /* If a child got cloned from a listening socket that had tcp_bpf
720 * protocol callbacks installed, we need to restore the callbacks to
721 * the default ones because the child does not inherit the psock state
722 * that tcp_bpf callbacks expect.
724 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
726 struct proto *prot = newsk->sk_prot;
728 if (is_insidevar(prot, tcp_bpf_prots))
729 newsk->sk_prot = sk->sk_prot_creator;
731 #endif /* CONFIG_BPF_SYSCALL */