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 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
15 struct sk_msg *msg, u32 apply_bytes, int flags)
17 bool apply = apply_bytes;
18 struct scatterlist *sge;
23 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
28 tmp->sg.start = msg->sg.start;
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)) {
40 sk_mem_charge(sk, size);
41 sk_msg_xfer(tmp, msg, i, size);
44 get_page(sk_msg_page(tmp, i));
45 sk_msg_iter_var_next(i);
51 sk_msg_iter_var_prev(i);
55 } while (i != msg->sg.end);
59 sk_psock_queue_msg(psock, tmp);
60 sk_psock_data_ready(sk, psock);
70 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
71 int flags, bool uncharge)
73 bool apply = apply_bytes;
74 struct scatterlist *sge;
82 sge = sk_msg_elem(msg, msg->sg.start);
83 size = (apply && apply_bytes < sge->length) ?
84 apply_bytes : sge->length;
88 tcp_rate_check_app_limited(sk);
90 has_tx_ulp = tls_sw_has_ctx_tx(sk);
92 flags |= MSG_SENDPAGE_NOPOLICY;
93 ret = kernel_sendpage_locked(sk,
94 page, off, size, flags);
96 ret = do_tcp_sendpages(sk, page, off, size, flags);
107 sk_mem_uncharge(sk, ret);
115 sk_msg_iter_next(msg, start);
116 sg_init_table(sge, 1);
117 if (msg->sg.start == msg->sg.end)
120 if (apply && !apply_bytes)
127 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
128 u32 apply_bytes, int flags, bool uncharge)
133 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
138 int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress,
139 struct sk_msg *msg, u32 bytes, int flags)
141 struct sk_psock *psock = sk_psock_get(sk);
144 if (unlikely(!psock))
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);
152 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
154 #ifdef CONFIG_BPF_SYSCALL
155 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
158 DEFINE_WAIT_FUNC(wait, woken_wake_function);
161 if (sk->sk_shutdown & RCV_SHUTDOWN)
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_lockless(&sk->sk_receive_queue), &wait);
172 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
173 remove_wait_queue(sk_sleep(sk), &wait);
177 static bool is_next_msg_fin(struct sk_psock *psock)
179 struct scatterlist *sge;
180 struct sk_msg *msg_rx;
183 msg_rx = sk_psock_peek_msg(psock);
184 i = msg_rx->sg.start;
185 sge = sk_msg_elem(msg_rx, i);
187 struct sk_buff *skb = msg_rx->skb;
189 if (skb && TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
195 static int tcp_bpf_recvmsg_parser(struct sock *sk,
201 struct sk_psock *psock;
204 if (unlikely(flags & MSG_ERRQUEUE))
205 return inet_recv_error(sk, msg, len, addr_len);
210 psock = sk_psock_get(sk);
211 if (unlikely(!psock))
212 return tcp_recvmsg(sk, msg, len, flags, addr_len);
216 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
217 /* The typical case for EFAULT is the socket was gracefully
218 * shutdown with a FIN pkt. So check here the other case is
219 * some error on copy_page_to_iter which would be unexpected.
220 * On fin return correct return code to zero.
222 if (copied == -EFAULT) {
223 bool is_fin = is_next_msg_fin(psock);
234 if (sock_flag(sk, SOCK_DONE))
238 copied = sock_error(sk);
242 if (sk->sk_shutdown & RCV_SHUTDOWN)
245 if (sk->sk_state == TCP_CLOSE) {
250 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
256 if (signal_pending(current)) {
257 copied = sock_intr_errno(timeo);
261 data = tcp_msg_wait_data(sk, psock, timeo);
262 if (data && !sk_psock_queue_empty(psock))
263 goto msg_bytes_ready;
268 sk_psock_put(sk, psock);
272 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
273 int flags, int *addr_len)
275 struct sk_psock *psock;
278 if (unlikely(flags & MSG_ERRQUEUE))
279 return inet_recv_error(sk, msg, len, addr_len);
284 psock = sk_psock_get(sk);
285 if (unlikely(!psock))
286 return tcp_recvmsg(sk, msg, len, flags, addr_len);
287 if (!skb_queue_empty(&sk->sk_receive_queue) &&
288 sk_psock_queue_empty(psock)) {
289 sk_psock_put(sk, psock);
290 return tcp_recvmsg(sk, msg, len, flags, addr_len);
294 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
299 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
300 data = tcp_msg_wait_data(sk, psock, timeo);
302 if (!sk_psock_queue_empty(psock))
303 goto msg_bytes_ready;
305 sk_psock_put(sk, psock);
306 return tcp_recvmsg(sk, msg, len, flags, addr_len);
312 sk_psock_put(sk, psock);
316 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
317 struct sk_msg *msg, int *copied, int flags)
319 bool cork = false, enospc = sk_msg_full(msg), redir_ingress;
320 struct sock *sk_redir;
321 u32 tosend, origsize, sent, delta = 0;
326 if (psock->eval == __SK_NONE) {
327 /* Track delta in msg size to add/subtract it on SK_DROP from
328 * returned to user copied size. This ensures user doesn't
329 * get a positive return code with msg_cut_data and SK_DROP
332 delta = msg->sg.size;
333 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
334 delta -= msg->sg.size;
337 if (msg->cork_bytes &&
338 msg->cork_bytes > msg->sg.size && !enospc) {
339 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
341 psock->cork = kzalloc(sizeof(*psock->cork),
342 GFP_ATOMIC | __GFP_NOWARN);
346 memcpy(psock->cork, msg, sizeof(*msg));
350 tosend = msg->sg.size;
351 if (psock->apply_bytes && psock->apply_bytes < tosend)
352 tosend = psock->apply_bytes;
355 switch (psock->eval) {
357 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
359 *copied -= sk_msg_free(sk, msg);
362 sk_msg_apply_bytes(psock, tosend);
365 redir_ingress = psock->redir_ingress;
366 sk_redir = psock->sk_redir;
367 sk_msg_apply_bytes(psock, tosend);
368 if (!psock->apply_bytes) {
369 /* Clean up before releasing the sock lock. */
371 psock->eval = __SK_NONE;
372 psock->sk_redir = NULL;
378 sk_msg_return(sk, msg, tosend);
381 origsize = msg->sg.size;
382 ret = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
384 sent = origsize - msg->sg.size;
386 if (eval == __SK_REDIRECT)
390 if (unlikely(ret < 0)) {
391 int free = sk_msg_free_nocharge(sk, msg);
397 sk_msg_free(sk, msg);
405 sk_msg_free_partial(sk, msg, tosend);
406 sk_msg_apply_bytes(psock, tosend);
407 *copied -= (tosend + delta);
412 if (!psock->apply_bytes) {
413 psock->eval = __SK_NONE;
414 if (psock->sk_redir) {
415 sock_put(psock->sk_redir);
416 psock->sk_redir = NULL;
420 msg->sg.data[msg->sg.start].page_link &&
421 msg->sg.data[msg->sg.start].length) {
422 if (eval == __SK_REDIRECT)
423 sk_mem_charge(sk, tosend - sent);
430 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
432 struct sk_msg tmp, *msg_tx = NULL;
433 int copied = 0, err = 0;
434 struct sk_psock *psock;
438 /* Don't let internal do_tcp_sendpages() flags through */
439 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
440 flags |= MSG_NO_SHARED_FRAGS;
442 psock = sk_psock_get(sk);
443 if (unlikely(!psock))
444 return tcp_sendmsg(sk, msg, size);
447 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
448 while (msg_data_left(msg)) {
457 copy = msg_data_left(msg);
458 if (!sk_stream_memory_free(sk))
459 goto wait_for_sndbuf;
461 msg_tx = psock->cork;
467 osize = msg_tx->sg.size;
468 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
471 goto wait_for_memory;
473 copy = msg_tx->sg.size - osize;
476 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
479 sk_msg_trim(sk, msg_tx, osize);
484 if (psock->cork_bytes) {
485 if (size > psock->cork_bytes)
486 psock->cork_bytes = 0;
488 psock->cork_bytes -= size;
489 if (psock->cork_bytes && !enospc)
491 /* All cork bytes are accounted, rerun the prog. */
492 psock->eval = __SK_NONE;
493 psock->cork_bytes = 0;
496 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
497 if (unlikely(err < 0))
501 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
503 err = sk_stream_wait_memory(sk, &timeo);
505 if (msg_tx && msg_tx != psock->cork)
506 sk_msg_free(sk, msg_tx);
512 err = sk_stream_error(sk, msg->msg_flags, err);
514 sk_psock_put(sk, psock);
515 return copied ? copied : err;
518 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
519 size_t size, int flags)
521 struct sk_msg tmp, *msg = NULL;
522 int err = 0, copied = 0;
523 struct sk_psock *psock;
526 psock = sk_psock_get(sk);
527 if (unlikely(!psock))
528 return tcp_sendpage(sk, page, offset, size, flags);
538 /* Catch case where ring is full and sendpage is stalled. */
539 if (unlikely(sk_msg_full(msg)))
542 sk_msg_page_add(msg, page, size, offset);
543 sk_mem_charge(sk, size);
545 if (sk_msg_full(msg))
547 if (psock->cork_bytes) {
548 if (size > psock->cork_bytes)
549 psock->cork_bytes = 0;
551 psock->cork_bytes -= size;
552 if (psock->cork_bytes && !enospc)
554 /* All cork bytes are accounted, rerun the prog. */
555 psock->eval = __SK_NONE;
556 psock->cork_bytes = 0;
559 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
562 sk_psock_put(sk, psock);
563 return copied ? copied : err;
580 static struct proto *tcpv6_prot_saved __read_mostly;
581 static DEFINE_SPINLOCK(tcpv6_prot_lock);
582 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
584 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
587 prot[TCP_BPF_BASE] = *base;
588 prot[TCP_BPF_BASE].destroy = sock_map_destroy;
589 prot[TCP_BPF_BASE].close = sock_map_close;
590 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
591 prot[TCP_BPF_BASE].sock_is_readable = sk_msg_is_readable;
593 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
594 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
595 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
597 prot[TCP_BPF_RX] = prot[TCP_BPF_BASE];
598 prot[TCP_BPF_RX].recvmsg = tcp_bpf_recvmsg_parser;
600 prot[TCP_BPF_TXRX] = prot[TCP_BPF_TX];
601 prot[TCP_BPF_TXRX].recvmsg = tcp_bpf_recvmsg_parser;
604 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
606 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
607 spin_lock_bh(&tcpv6_prot_lock);
608 if (likely(ops != tcpv6_prot_saved)) {
609 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
610 smp_store_release(&tcpv6_prot_saved, ops);
612 spin_unlock_bh(&tcpv6_prot_lock);
616 static int __init tcp_bpf_v4_build_proto(void)
618 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
621 late_initcall(tcp_bpf_v4_build_proto);
623 static int tcp_bpf_assert_proto_ops(struct proto *ops)
625 /* In order to avoid retpoline, we make assumptions when we call
626 * into ops if e.g. a psock is not present. Make sure they are
627 * indeed valid assumptions.
629 return ops->recvmsg == tcp_recvmsg &&
630 ops->sendmsg == tcp_sendmsg &&
631 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
634 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
636 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
637 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
639 if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
640 config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
644 if (inet_csk_has_ulp(sk)) {
645 /* TLS does not have an unhash proto in SW cases,
646 * but we need to ensure we stop using the sock_map
647 * unhash routine because the associated psock is being
648 * removed. So use the original unhash handler.
650 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
651 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
653 sk->sk_write_space = psock->saved_write_space;
654 /* Pairs with lockless read in sk_clone_lock() */
655 sock_replace_proto(sk, psock->sk_proto);
660 if (sk->sk_family == AF_INET6) {
661 if (tcp_bpf_assert_proto_ops(psock->sk_proto))
664 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
667 /* Pairs with lockless read in sk_clone_lock() */
668 sock_replace_proto(sk, &tcp_bpf_prots[family][config]);
671 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
673 /* If a child got cloned from a listening socket that had tcp_bpf
674 * protocol callbacks installed, we need to restore the callbacks to
675 * the default ones because the child does not inherit the psock state
676 * that tcp_bpf callbacks expect.
678 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
680 struct proto *prot = newsk->sk_prot;
682 if (is_insidevar(prot, tcp_bpf_prots))
683 newsk->sk_prot = sk->sk_prot_creator;
685 #endif /* CONFIG_BPF_SYSCALL */