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);
153 sk_psock_put(sk, psock);
157 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
158 struct sk_msg *msg, u32 apply_bytes, int flags)
160 bool apply = apply_bytes;
161 struct scatterlist *sge;
162 u32 size, copied = 0;
166 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
171 tmp->sg.start = msg->sg.start;
174 sge = sk_msg_elem(msg, i);
175 size = (apply && apply_bytes < sge->length) ?
176 apply_bytes : sge->length;
177 if (!sk_wmem_schedule(sk, size)) {
183 sk_mem_charge(sk, size);
184 sk_msg_xfer(tmp, msg, i, size);
187 get_page(sk_msg_page(tmp, i));
188 sk_msg_iter_var_next(i);
195 } while (i != msg->sg.end);
199 msg->sg.size -= apply_bytes;
200 sk_psock_queue_msg(psock, tmp);
201 sk->sk_data_ready(sk);
203 sk_msg_free(sk, tmp);
211 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
212 int flags, bool uncharge)
214 bool apply = apply_bytes;
215 struct scatterlist *sge;
221 sge = sk_msg_elem(msg, msg->sg.start);
222 size = (apply && apply_bytes < sge->length) ?
223 apply_bytes : sge->length;
227 tcp_rate_check_app_limited(sk);
229 ret = do_tcp_sendpages(sk, page, off, size, flags);
238 sk_mem_uncharge(sk, ret);
246 sk_msg_iter_next(msg, start);
247 sg_init_table(sge, 1);
248 if (msg->sg.start == msg->sg.end)
251 if (apply && !apply_bytes)
258 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
259 u32 apply_bytes, int flags, bool uncharge)
264 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
269 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
270 u32 bytes, int flags)
272 bool ingress = sk_msg_to_ingress(msg);
273 struct sk_psock *psock = sk_psock_get(sk);
276 if (unlikely(!psock)) {
277 sk_msg_free(sk, msg);
280 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
281 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
282 sk_psock_put(sk, psock);
285 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
287 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
288 struct sk_msg *msg, int *copied, int flags)
290 bool cork = false, enospc = msg->sg.start == msg->sg.end;
291 struct sock *sk_redir;
292 u32 tosend, delta = 0;
296 if (psock->eval == __SK_NONE) {
297 /* Track delta in msg size to add/subtract it on SK_DROP from
298 * returned to user copied size. This ensures user doesn't
299 * get a positive return code with msg_cut_data and SK_DROP
302 delta = msg->sg.size;
303 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
304 if (msg->sg.size < delta)
305 delta -= msg->sg.size;
310 if (msg->cork_bytes &&
311 msg->cork_bytes > msg->sg.size && !enospc) {
312 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
314 psock->cork = kzalloc(sizeof(*psock->cork),
315 GFP_ATOMIC | __GFP_NOWARN);
319 memcpy(psock->cork, msg, sizeof(*msg));
323 tosend = msg->sg.size;
324 if (psock->apply_bytes && psock->apply_bytes < tosend)
325 tosend = psock->apply_bytes;
327 switch (psock->eval) {
329 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
331 *copied -= sk_msg_free(sk, msg);
334 sk_msg_apply_bytes(psock, tosend);
337 sk_redir = psock->sk_redir;
338 sk_msg_apply_bytes(psock, tosend);
343 sk_msg_return(sk, msg, tosend);
345 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
347 if (unlikely(ret < 0)) {
348 int free = sk_msg_free_nocharge(sk, msg);
354 sk_msg_free(sk, msg);
362 sk_msg_free_partial(sk, msg, tosend);
363 sk_msg_apply_bytes(psock, tosend);
364 *copied -= (tosend + delta);
369 if (!psock->apply_bytes) {
370 psock->eval = __SK_NONE;
371 if (psock->sk_redir) {
372 sock_put(psock->sk_redir);
373 psock->sk_redir = NULL;
377 msg->sg.data[msg->sg.start].page_link &&
378 msg->sg.data[msg->sg.start].length)
384 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
386 struct sk_msg tmp, *msg_tx = NULL;
387 int flags = msg->msg_flags | MSG_NO_SHARED_FRAGS;
388 int copied = 0, err = 0;
389 struct sk_psock *psock;
392 psock = sk_psock_get(sk);
393 if (unlikely(!psock))
394 return tcp_sendmsg(sk, msg, size);
397 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
398 while (msg_data_left(msg)) {
407 copy = msg_data_left(msg);
408 if (!sk_stream_memory_free(sk))
409 goto wait_for_sndbuf;
411 msg_tx = psock->cork;
417 osize = msg_tx->sg.size;
418 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
421 goto wait_for_memory;
423 copy = msg_tx->sg.size - osize;
426 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
429 sk_msg_trim(sk, msg_tx, osize);
434 if (psock->cork_bytes) {
435 if (size > psock->cork_bytes)
436 psock->cork_bytes = 0;
438 psock->cork_bytes -= size;
439 if (psock->cork_bytes && !enospc)
441 /* All cork bytes are accounted, rerun the prog. */
442 psock->eval = __SK_NONE;
443 psock->cork_bytes = 0;
446 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
447 if (unlikely(err < 0))
451 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
453 err = sk_stream_wait_memory(sk, &timeo);
455 if (msg_tx && msg_tx != psock->cork)
456 sk_msg_free(sk, msg_tx);
462 err = sk_stream_error(sk, msg->msg_flags, err);
464 sk_psock_put(sk, psock);
465 return copied ? copied : err;
468 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
469 size_t size, int flags)
471 struct sk_msg tmp, *msg = NULL;
472 int err = 0, copied = 0;
473 struct sk_psock *psock;
476 psock = sk_psock_get(sk);
477 if (unlikely(!psock))
478 return tcp_sendpage(sk, page, offset, size, flags);
488 /* Catch case where ring is full and sendpage is stalled. */
489 if (unlikely(sk_msg_full(msg)))
492 sk_msg_page_add(msg, page, size, offset);
493 sk_mem_charge(sk, size);
495 if (sk_msg_full(msg))
497 if (psock->cork_bytes) {
498 if (size > psock->cork_bytes)
499 psock->cork_bytes = 0;
501 psock->cork_bytes -= size;
502 if (psock->cork_bytes && !enospc)
504 /* All cork bytes are accounted, rerun the prog. */
505 psock->eval = __SK_NONE;
506 psock->cork_bytes = 0;
509 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
512 sk_psock_put(sk, psock);
513 return copied ? copied : err;
516 static void tcp_bpf_remove(struct sock *sk, struct sk_psock *psock)
518 struct sk_psock_link *link;
520 sk_psock_cork_free(psock);
521 __sk_psock_purge_ingress_msg(psock);
522 while ((link = sk_psock_link_pop(psock))) {
523 sk_psock_unlink(sk, link);
524 sk_psock_free_link(link);
528 static void tcp_bpf_unhash(struct sock *sk)
530 void (*saved_unhash)(struct sock *sk);
531 struct sk_psock *psock;
534 psock = sk_psock(sk);
535 if (unlikely(!psock)) {
537 if (sk->sk_prot->unhash)
538 sk->sk_prot->unhash(sk);
542 saved_unhash = psock->saved_unhash;
543 tcp_bpf_remove(sk, psock);
548 static void tcp_bpf_close(struct sock *sk, long timeout)
550 void (*saved_close)(struct sock *sk, long timeout);
551 struct sk_psock *psock;
555 psock = sk_psock(sk);
556 if (unlikely(!psock)) {
559 return sk->sk_prot->close(sk, timeout);
562 saved_close = psock->saved_close;
563 tcp_bpf_remove(sk, psock);
566 saved_close(sk, timeout);
581 static struct proto *tcpv6_prot_saved __read_mostly;
582 static DEFINE_SPINLOCK(tcpv6_prot_lock);
583 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
585 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
588 prot[TCP_BPF_BASE] = *base;
589 prot[TCP_BPF_BASE].unhash = tcp_bpf_unhash;
590 prot[TCP_BPF_BASE].close = tcp_bpf_close;
591 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
592 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read;
594 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
595 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
596 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
599 static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops)
601 if (sk->sk_family == AF_INET6 &&
602 unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
603 spin_lock_bh(&tcpv6_prot_lock);
604 if (likely(ops != tcpv6_prot_saved)) {
605 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
606 smp_store_release(&tcpv6_prot_saved, ops);
608 spin_unlock_bh(&tcpv6_prot_lock);
612 static int __init tcp_bpf_v4_build_proto(void)
614 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
617 core_initcall(tcp_bpf_v4_build_proto);
619 static void tcp_bpf_update_sk_prot(struct sock *sk, struct sk_psock *psock)
621 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
622 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
624 sk_psock_update_proto(sk, psock, &tcp_bpf_prots[family][config]);
627 static void tcp_bpf_reinit_sk_prot(struct sock *sk, struct sk_psock *psock)
629 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
630 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
632 /* Reinit occurs when program types change e.g. TCP_BPF_TX is removed
633 * or added requiring sk_prot hook updates. We keep original saved
634 * hooks in this case.
636 sk->sk_prot = &tcp_bpf_prots[family][config];
639 static int tcp_bpf_assert_proto_ops(struct proto *ops)
641 /* In order to avoid retpoline, we make assumptions when we call
642 * into ops if e.g. a psock is not present. Make sure they are
643 * indeed valid assumptions.
645 return ops->recvmsg == tcp_recvmsg &&
646 ops->sendmsg == tcp_sendmsg &&
647 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
650 void tcp_bpf_reinit(struct sock *sk)
652 struct sk_psock *psock;
654 sock_owned_by_me(sk);
657 psock = sk_psock(sk);
658 tcp_bpf_reinit_sk_prot(sk, psock);
662 int tcp_bpf_init(struct sock *sk)
664 struct proto *ops = READ_ONCE(sk->sk_prot);
665 struct sk_psock *psock;
667 sock_owned_by_me(sk);
670 psock = sk_psock(sk);
671 if (unlikely(!psock || psock->sk_proto ||
672 tcp_bpf_assert_proto_ops(ops))) {
676 tcp_bpf_check_v6_needs_rebuild(sk, ops);
677 tcp_bpf_update_sk_prot(sk, psock);