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>
13 int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock,
14 struct msghdr *msg, int len, int flags)
16 struct iov_iter *iter = &msg->msg_iter;
17 int peek = flags & MSG_PEEK;
18 int i, ret, copied = 0;
19 struct sk_msg *msg_rx;
21 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
24 while (copied != len) {
25 struct scatterlist *sge;
27 if (unlikely(!msg_rx))
35 sge = sk_msg_elem(msg_rx, i);
38 if (copied + copy > len)
40 ret = copy_page_to_iter(page, sge->offset, copy, iter);
50 sk_mem_uncharge(sk, copy);
51 msg_rx->sg.size -= copy;
54 sk_msg_iter_var_next(i);
59 sk_msg_iter_var_next(i);
64 } while (i != msg_rx->sg.end);
67 if (msg_rx == list_last_entry(&psock->ingress_msg,
70 msg_rx = list_next_entry(msg_rx, list);
75 if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
76 list_del(&msg_rx->list);
78 consume_skb(msg_rx->skb);
81 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
87 EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg);
89 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
90 struct sk_msg *msg, u32 apply_bytes, int flags)
92 bool apply = apply_bytes;
93 struct scatterlist *sge;
98 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
103 tmp->sg.start = msg->sg.start;
106 sge = sk_msg_elem(msg, i);
107 size = (apply && apply_bytes < sge->length) ?
108 apply_bytes : sge->length;
109 if (!sk_wmem_schedule(sk, size)) {
115 sk_mem_charge(sk, size);
116 sk_msg_xfer(tmp, msg, i, size);
119 get_page(sk_msg_page(tmp, i));
120 sk_msg_iter_var_next(i);
127 } while (i != msg->sg.end);
131 sk_psock_queue_msg(psock, tmp);
132 sk_psock_data_ready(sk, psock);
134 sk_msg_free(sk, tmp);
142 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
143 int flags, bool uncharge)
145 bool apply = apply_bytes;
146 struct scatterlist *sge;
154 sge = sk_msg_elem(msg, msg->sg.start);
155 size = (apply && apply_bytes < sge->length) ?
156 apply_bytes : sge->length;
160 tcp_rate_check_app_limited(sk);
162 has_tx_ulp = tls_sw_has_ctx_tx(sk);
164 flags |= MSG_SENDPAGE_NOPOLICY;
165 ret = kernel_sendpage_locked(sk,
166 page, off, size, flags);
168 ret = do_tcp_sendpages(sk, page, off, size, flags);
179 sk_mem_uncharge(sk, ret);
187 sk_msg_iter_next(msg, start);
188 sg_init_table(sge, 1);
189 if (msg->sg.start == msg->sg.end)
192 if (apply && !apply_bytes)
199 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
200 u32 apply_bytes, int flags, bool uncharge)
205 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
210 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
211 u32 bytes, int flags)
213 bool ingress = sk_msg_to_ingress(msg);
214 struct sk_psock *psock = sk_psock_get(sk);
217 if (unlikely(!psock)) {
218 sk_msg_free(sk, msg);
221 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
222 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
223 sk_psock_put(sk, psock);
226 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
228 #ifdef CONFIG_BPF_STREAM_PARSER
229 static bool tcp_bpf_stream_read(const struct sock *sk)
231 struct sk_psock *psock;
235 psock = sk_psock(sk);
237 empty = list_empty(&psock->ingress_msg);
242 static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock,
243 int flags, long timeo, int *err)
245 DEFINE_WAIT_FUNC(wait, woken_wake_function);
248 if (sk->sk_shutdown & RCV_SHUTDOWN)
254 add_wait_queue(sk_sleep(sk), &wait);
255 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
256 ret = sk_wait_event(sk, &timeo,
257 !list_empty(&psock->ingress_msg) ||
258 !skb_queue_empty(&sk->sk_receive_queue), &wait);
259 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
260 remove_wait_queue(sk_sleep(sk), &wait);
264 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
265 int nonblock, int flags, int *addr_len)
267 struct sk_psock *psock;
270 if (unlikely(flags & MSG_ERRQUEUE))
271 return inet_recv_error(sk, msg, len, addr_len);
273 psock = sk_psock_get(sk);
274 if (unlikely(!psock))
275 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
276 if (!skb_queue_empty(&sk->sk_receive_queue) &&
277 sk_psock_queue_empty(psock)) {
278 sk_psock_put(sk, psock);
279 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
283 copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags);
288 timeo = sock_rcvtimeo(sk, nonblock);
289 data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err);
291 if (!sk_psock_queue_empty(psock))
292 goto msg_bytes_ready;
294 sk_psock_put(sk, psock);
295 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
306 sk_psock_put(sk, psock);
310 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
311 struct sk_msg *msg, int *copied, int flags)
313 bool cork = false, enospc = sk_msg_full(msg);
314 struct sock *sk_redir;
315 u32 tosend, delta = 0;
319 if (psock->eval == __SK_NONE) {
320 /* Track delta in msg size to add/subtract it on SK_DROP from
321 * returned to user copied size. This ensures user doesn't
322 * get a positive return code with msg_cut_data and SK_DROP
325 delta = msg->sg.size;
326 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
327 delta -= msg->sg.size;
330 if (msg->cork_bytes &&
331 msg->cork_bytes > msg->sg.size && !enospc) {
332 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
334 psock->cork = kzalloc(sizeof(*psock->cork),
335 GFP_ATOMIC | __GFP_NOWARN);
339 memcpy(psock->cork, msg, sizeof(*msg));
343 tosend = msg->sg.size;
344 if (psock->apply_bytes && psock->apply_bytes < tosend)
345 tosend = psock->apply_bytes;
347 switch (psock->eval) {
349 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
351 *copied -= sk_msg_free(sk, msg);
354 sk_msg_apply_bytes(psock, tosend);
357 sk_redir = psock->sk_redir;
358 sk_msg_apply_bytes(psock, tosend);
363 sk_msg_return(sk, msg, tosend);
365 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
367 if (unlikely(ret < 0)) {
368 int free = sk_msg_free_nocharge(sk, msg);
374 sk_msg_free(sk, msg);
382 sk_msg_free_partial(sk, msg, tosend);
383 sk_msg_apply_bytes(psock, tosend);
384 *copied -= (tosend + delta);
389 if (!psock->apply_bytes) {
390 psock->eval = __SK_NONE;
391 if (psock->sk_redir) {
392 sock_put(psock->sk_redir);
393 psock->sk_redir = NULL;
397 msg->sg.data[msg->sg.start].page_link &&
398 msg->sg.data[msg->sg.start].length)
404 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
406 struct sk_msg tmp, *msg_tx = NULL;
407 int copied = 0, err = 0;
408 struct sk_psock *psock;
412 /* Don't let internal do_tcp_sendpages() flags through */
413 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
414 flags |= MSG_NO_SHARED_FRAGS;
416 psock = sk_psock_get(sk);
417 if (unlikely(!psock))
418 return tcp_sendmsg(sk, msg, size);
421 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
422 while (msg_data_left(msg)) {
431 copy = msg_data_left(msg);
432 if (!sk_stream_memory_free(sk))
433 goto wait_for_sndbuf;
435 msg_tx = psock->cork;
441 osize = msg_tx->sg.size;
442 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
445 goto wait_for_memory;
447 copy = msg_tx->sg.size - osize;
450 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
453 sk_msg_trim(sk, msg_tx, osize);
458 if (psock->cork_bytes) {
459 if (size > psock->cork_bytes)
460 psock->cork_bytes = 0;
462 psock->cork_bytes -= size;
463 if (psock->cork_bytes && !enospc)
465 /* All cork bytes are accounted, rerun the prog. */
466 psock->eval = __SK_NONE;
467 psock->cork_bytes = 0;
470 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
471 if (unlikely(err < 0))
475 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
477 err = sk_stream_wait_memory(sk, &timeo);
479 if (msg_tx && msg_tx != psock->cork)
480 sk_msg_free(sk, msg_tx);
486 err = sk_stream_error(sk, msg->msg_flags, err);
488 sk_psock_put(sk, psock);
489 return copied ? copied : err;
492 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
493 size_t size, int flags)
495 struct sk_msg tmp, *msg = NULL;
496 int err = 0, copied = 0;
497 struct sk_psock *psock;
500 psock = sk_psock_get(sk);
501 if (unlikely(!psock))
502 return tcp_sendpage(sk, page, offset, size, flags);
512 /* Catch case where ring is full and sendpage is stalled. */
513 if (unlikely(sk_msg_full(msg)))
516 sk_msg_page_add(msg, page, size, offset);
517 sk_mem_charge(sk, size);
519 if (sk_msg_full(msg))
521 if (psock->cork_bytes) {
522 if (size > psock->cork_bytes)
523 psock->cork_bytes = 0;
525 psock->cork_bytes -= size;
526 if (psock->cork_bytes && !enospc)
528 /* All cork bytes are accounted, rerun the prog. */
529 psock->eval = __SK_NONE;
530 psock->cork_bytes = 0;
533 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
536 sk_psock_put(sk, psock);
537 return copied ? copied : err;
552 static struct proto *tcpv6_prot_saved __read_mostly;
553 static DEFINE_SPINLOCK(tcpv6_prot_lock);
554 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
556 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
559 prot[TCP_BPF_BASE] = *base;
560 prot[TCP_BPF_BASE].unhash = sock_map_unhash;
561 prot[TCP_BPF_BASE].close = sock_map_close;
562 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
563 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read;
565 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
566 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
567 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
570 static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops)
572 if (sk->sk_family == AF_INET6 &&
573 unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
574 spin_lock_bh(&tcpv6_prot_lock);
575 if (likely(ops != tcpv6_prot_saved)) {
576 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
577 smp_store_release(&tcpv6_prot_saved, ops);
579 spin_unlock_bh(&tcpv6_prot_lock);
583 static int __init tcp_bpf_v4_build_proto(void)
585 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
588 core_initcall(tcp_bpf_v4_build_proto);
590 static int tcp_bpf_assert_proto_ops(struct proto *ops)
592 /* In order to avoid retpoline, we make assumptions when we call
593 * into ops if e.g. a psock is not present. Make sure they are
594 * indeed valid assumptions.
596 return ops->recvmsg == tcp_recvmsg &&
597 ops->sendmsg == tcp_sendmsg &&
598 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
601 struct proto *tcp_bpf_get_proto(struct sock *sk, struct sk_psock *psock)
603 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
604 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
606 if (!psock->sk_proto) {
607 struct proto *ops = READ_ONCE(sk->sk_prot);
609 if (tcp_bpf_assert_proto_ops(ops))
610 return ERR_PTR(-EINVAL);
612 tcp_bpf_check_v6_needs_rebuild(sk, ops);
615 return &tcp_bpf_prots[family][config];
618 /* If a child got cloned from a listening socket that had tcp_bpf
619 * protocol callbacks installed, we need to restore the callbacks to
620 * the default ones because the child does not inherit the psock state
621 * that tcp_bpf callbacks expect.
623 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
625 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
626 struct proto *prot = newsk->sk_prot;
628 if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
629 newsk->sk_prot = sk->sk_prot_creator;
631 #endif /* CONFIG_BPF_STREAM_PARSER */