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>
10 #include <net/inet_common.h>
12 static bool tcp_bpf_stream_read(const struct sock *sk)
14 struct sk_psock *psock;
20 empty = list_empty(&psock->ingress_msg);
25 static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock,
26 int flags, long timeo, int *err)
28 DEFINE_WAIT_FUNC(wait, woken_wake_function);
31 add_wait_queue(sk_sleep(sk), &wait);
32 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
33 ret = sk_wait_event(sk, &timeo,
34 !list_empty(&psock->ingress_msg) ||
35 !skb_queue_empty(&sk->sk_receive_queue), &wait);
36 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
37 remove_wait_queue(sk_sleep(sk), &wait);
41 int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock,
42 struct msghdr *msg, int len, int flags)
44 struct iov_iter *iter = &msg->msg_iter;
45 int peek = flags & MSG_PEEK;
46 int i, ret, copied = 0;
47 struct sk_msg *msg_rx;
49 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
52 while (copied != len) {
53 struct scatterlist *sge;
55 if (unlikely(!msg_rx))
63 sge = sk_msg_elem(msg_rx, i);
66 if (copied + copy > len)
68 ret = copy_page_to_iter(page, sge->offset, copy, iter);
78 sk_mem_uncharge(sk, copy);
79 msg_rx->sg.size -= copy;
82 sk_msg_iter_var_next(i);
87 sk_msg_iter_var_next(i);
92 } while (i != msg_rx->sg.end);
95 msg_rx = list_next_entry(msg_rx, list);
100 if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
101 list_del(&msg_rx->list);
103 consume_skb(msg_rx->skb);
106 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
107 struct sk_msg, list);
112 EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg);
114 int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
115 int nonblock, int flags, int *addr_len)
117 struct sk_psock *psock;
120 if (unlikely(flags & MSG_ERRQUEUE))
121 return inet_recv_error(sk, msg, len, addr_len);
122 if (!skb_queue_empty(&sk->sk_receive_queue))
123 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
125 psock = sk_psock_get(sk);
126 if (unlikely(!psock))
127 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
130 copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags);
135 timeo = sock_rcvtimeo(sk, nonblock);
136 data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err);
138 if (skb_queue_empty(&sk->sk_receive_queue))
139 goto msg_bytes_ready;
141 sk_psock_put(sk, psock);
142 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
152 sk_psock_put(sk, psock);
156 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
157 struct sk_msg *msg, u32 apply_bytes, int flags)
159 bool apply = apply_bytes;
160 struct scatterlist *sge;
161 u32 size, copied = 0;
165 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
170 tmp->sg.start = msg->sg.start;
173 sge = sk_msg_elem(msg, i);
174 size = (apply && apply_bytes < sge->length) ?
175 apply_bytes : sge->length;
176 if (!sk_wmem_schedule(sk, size)) {
182 sk_mem_charge(sk, size);
183 sk_msg_xfer(tmp, msg, i, size);
186 get_page(sk_msg_page(tmp, i));
187 sk_msg_iter_var_next(i);
194 } while (i != msg->sg.end);
198 msg->sg.size -= apply_bytes;
199 sk_psock_queue_msg(psock, tmp);
200 sk->sk_data_ready(sk);
202 sk_msg_free(sk, tmp);
210 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
211 int flags, bool uncharge)
213 bool apply = apply_bytes;
214 struct scatterlist *sge;
220 sge = sk_msg_elem(msg, msg->sg.start);
221 size = (apply && apply_bytes < sge->length) ?
222 apply_bytes : sge->length;
226 tcp_rate_check_app_limited(sk);
228 ret = do_tcp_sendpages(sk, page, off, size, flags);
237 sk_mem_uncharge(sk, ret);
245 sk_msg_iter_next(msg, start);
246 sg_init_table(sge, 1);
247 if (msg->sg.start == msg->sg.end)
250 if (apply && !apply_bytes)
257 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
258 u32 apply_bytes, int flags, bool uncharge)
263 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
268 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
269 u32 bytes, int flags)
271 bool ingress = sk_msg_to_ingress(msg);
272 struct sk_psock *psock = sk_psock_get(sk);
275 if (unlikely(!psock)) {
276 sk_msg_free(sk, msg);
279 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
280 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
281 sk_psock_put(sk, psock);
284 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
286 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
287 struct sk_msg *msg, int *copied, int flags)
289 bool cork = false, enospc = msg->sg.start == msg->sg.end;
290 struct sock *sk_redir;
295 if (psock->eval == __SK_NONE)
296 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
298 if (msg->cork_bytes &&
299 msg->cork_bytes > msg->sg.size && !enospc) {
300 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
302 psock->cork = kzalloc(sizeof(*psock->cork),
303 GFP_ATOMIC | __GFP_NOWARN);
307 memcpy(psock->cork, msg, sizeof(*msg));
311 tosend = msg->sg.size;
312 if (psock->apply_bytes && psock->apply_bytes < tosend)
313 tosend = psock->apply_bytes;
315 switch (psock->eval) {
317 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
319 *copied -= sk_msg_free(sk, msg);
322 sk_msg_apply_bytes(psock, tosend);
325 sk_redir = psock->sk_redir;
326 sk_msg_apply_bytes(psock, tosend);
331 sk_msg_return(sk, msg, tosend);
333 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
335 if (unlikely(ret < 0)) {
336 int free = sk_msg_free_nocharge(sk, msg);
342 sk_msg_free(sk, msg);
350 sk_msg_free_partial(sk, msg, tosend);
351 sk_msg_apply_bytes(psock, tosend);
357 if (!psock->apply_bytes) {
358 psock->eval = __SK_NONE;
359 if (psock->sk_redir) {
360 sock_put(psock->sk_redir);
361 psock->sk_redir = NULL;
365 msg->sg.data[msg->sg.start].page_link &&
366 msg->sg.data[msg->sg.start].length)
372 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
374 struct sk_msg tmp, *msg_tx = NULL;
375 int flags = msg->msg_flags | MSG_NO_SHARED_FRAGS;
376 int copied = 0, err = 0;
377 struct sk_psock *psock;
380 psock = sk_psock_get(sk);
381 if (unlikely(!psock))
382 return tcp_sendmsg(sk, msg, size);
385 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
386 while (msg_data_left(msg)) {
395 copy = msg_data_left(msg);
396 if (!sk_stream_memory_free(sk))
397 goto wait_for_sndbuf;
399 msg_tx = psock->cork;
405 osize = msg_tx->sg.size;
406 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
409 goto wait_for_memory;
411 copy = msg_tx->sg.size - osize;
414 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
417 sk_msg_trim(sk, msg_tx, osize);
422 if (psock->cork_bytes) {
423 if (size > psock->cork_bytes)
424 psock->cork_bytes = 0;
426 psock->cork_bytes -= size;
427 if (psock->cork_bytes && !enospc)
429 /* All cork bytes are accounted, rerun the prog. */
430 psock->eval = __SK_NONE;
431 psock->cork_bytes = 0;
434 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
435 if (unlikely(err < 0))
439 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
441 err = sk_stream_wait_memory(sk, &timeo);
443 if (msg_tx && msg_tx != psock->cork)
444 sk_msg_free(sk, msg_tx);
450 err = sk_stream_error(sk, msg->msg_flags, err);
452 sk_psock_put(sk, psock);
453 return copied ? copied : err;
456 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
457 size_t size, int flags)
459 struct sk_msg tmp, *msg = NULL;
460 int err = 0, copied = 0;
461 struct sk_psock *psock;
464 psock = sk_psock_get(sk);
465 if (unlikely(!psock))
466 return tcp_sendpage(sk, page, offset, size, flags);
476 /* Catch case where ring is full and sendpage is stalled. */
477 if (unlikely(sk_msg_full(msg)))
480 sk_msg_page_add(msg, page, size, offset);
481 sk_mem_charge(sk, size);
483 if (sk_msg_full(msg))
485 if (psock->cork_bytes) {
486 if (size > psock->cork_bytes)
487 psock->cork_bytes = 0;
489 psock->cork_bytes -= size;
490 if (psock->cork_bytes && !enospc)
492 /* All cork bytes are accounted, rerun the prog. */
493 psock->eval = __SK_NONE;
494 psock->cork_bytes = 0;
497 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
500 sk_psock_put(sk, psock);
501 return copied ? copied : err;
504 static void tcp_bpf_remove(struct sock *sk, struct sk_psock *psock)
506 struct sk_psock_link *link;
508 sk_psock_cork_free(psock);
509 __sk_psock_purge_ingress_msg(psock);
510 while ((link = sk_psock_link_pop(psock))) {
511 sk_psock_unlink(sk, link);
512 sk_psock_free_link(link);
516 static void tcp_bpf_unhash(struct sock *sk)
518 void (*saved_unhash)(struct sock *sk);
519 struct sk_psock *psock;
522 psock = sk_psock(sk);
523 if (unlikely(!psock)) {
525 if (sk->sk_prot->unhash)
526 sk->sk_prot->unhash(sk);
530 saved_unhash = psock->saved_unhash;
531 tcp_bpf_remove(sk, psock);
536 static void tcp_bpf_close(struct sock *sk, long timeout)
538 void (*saved_close)(struct sock *sk, long timeout);
539 struct sk_psock *psock;
543 psock = sk_psock(sk);
544 if (unlikely(!psock)) {
547 return sk->sk_prot->close(sk, timeout);
550 saved_close = psock->saved_close;
551 tcp_bpf_remove(sk, psock);
554 saved_close(sk, timeout);
569 static struct proto *tcpv6_prot_saved __read_mostly;
570 static DEFINE_SPINLOCK(tcpv6_prot_lock);
571 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
573 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
576 prot[TCP_BPF_BASE] = *base;
577 prot[TCP_BPF_BASE].unhash = tcp_bpf_unhash;
578 prot[TCP_BPF_BASE].close = tcp_bpf_close;
579 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
580 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read;
582 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
583 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
584 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
587 static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops)
589 if (sk->sk_family == AF_INET6 &&
590 unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
591 spin_lock_bh(&tcpv6_prot_lock);
592 if (likely(ops != tcpv6_prot_saved)) {
593 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
594 smp_store_release(&tcpv6_prot_saved, ops);
596 spin_unlock_bh(&tcpv6_prot_lock);
600 static int __init tcp_bpf_v4_build_proto(void)
602 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
605 core_initcall(tcp_bpf_v4_build_proto);
607 static void tcp_bpf_update_sk_prot(struct sock *sk, struct sk_psock *psock)
609 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
610 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
612 sk_psock_update_proto(sk, psock, &tcp_bpf_prots[family][config]);
615 static void tcp_bpf_reinit_sk_prot(struct sock *sk, struct sk_psock *psock)
617 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
618 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
620 /* Reinit occurs when program types change e.g. TCP_BPF_TX is removed
621 * or added requiring sk_prot hook updates. We keep original saved
622 * hooks in this case.
624 sk->sk_prot = &tcp_bpf_prots[family][config];
627 static int tcp_bpf_assert_proto_ops(struct proto *ops)
629 /* In order to avoid retpoline, we make assumptions when we call
630 * into ops if e.g. a psock is not present. Make sure they are
631 * indeed valid assumptions.
633 return ops->recvmsg == tcp_recvmsg &&
634 ops->sendmsg == tcp_sendmsg &&
635 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
638 void tcp_bpf_reinit(struct sock *sk)
640 struct sk_psock *psock;
642 sock_owned_by_me(sk);
645 psock = sk_psock(sk);
646 tcp_bpf_reinit_sk_prot(sk, psock);
650 int tcp_bpf_init(struct sock *sk)
652 struct proto *ops = READ_ONCE(sk->sk_prot);
653 struct sk_psock *psock;
655 sock_owned_by_me(sk);
658 psock = sk_psock(sk);
659 if (unlikely(!psock || psock->sk_proto ||
660 tcp_bpf_assert_proto_ops(ops))) {
664 tcp_bpf_check_v6_needs_rebuild(sk, ops);
665 tcp_bpf_update_sk_prot(sk, psock);