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 msg_rx = list_next_entry(msg_rx, list);
72 if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
73 list_del(&msg_rx->list);
75 consume_skb(msg_rx->skb);
78 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
84 EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg);
86 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
87 struct sk_msg *msg, u32 apply_bytes, int flags)
89 bool apply = apply_bytes;
90 struct scatterlist *sge;
95 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
100 tmp->sg.start = msg->sg.start;
103 sge = sk_msg_elem(msg, i);
104 size = (apply && apply_bytes < sge->length) ?
105 apply_bytes : sge->length;
106 if (!sk_wmem_schedule(sk, size)) {
112 sk_mem_charge(sk, size);
113 sk_msg_xfer(tmp, msg, i, size);
116 get_page(sk_msg_page(tmp, i));
117 sk_msg_iter_var_next(i);
124 } while (i != msg->sg.end);
128 sk_psock_queue_msg(psock, tmp);
129 sk_psock_data_ready(sk, psock);
131 sk_msg_free(sk, tmp);
139 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
140 int flags, bool uncharge)
142 bool apply = apply_bytes;
143 struct scatterlist *sge;
151 sge = sk_msg_elem(msg, msg->sg.start);
152 size = (apply && apply_bytes < sge->length) ?
153 apply_bytes : sge->length;
157 tcp_rate_check_app_limited(sk);
159 has_tx_ulp = tls_sw_has_ctx_tx(sk);
161 flags |= MSG_SENDPAGE_NOPOLICY;
162 ret = kernel_sendpage_locked(sk,
163 page, off, size, flags);
165 ret = do_tcp_sendpages(sk, page, off, size, flags);
176 sk_mem_uncharge(sk, ret);
184 sk_msg_iter_next(msg, start);
185 sg_init_table(sge, 1);
186 if (msg->sg.start == msg->sg.end)
189 if (apply && !apply_bytes)
196 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
197 u32 apply_bytes, int flags, bool uncharge)
202 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
207 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
208 u32 bytes, int flags)
210 bool ingress = sk_msg_to_ingress(msg);
211 struct sk_psock *psock = sk_psock_get(sk);
214 if (unlikely(!psock)) {
215 sk_msg_free(sk, msg);
218 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
219 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
220 sk_psock_put(sk, psock);
223 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
225 #ifdef CONFIG_BPF_STREAM_PARSER
226 static bool tcp_bpf_stream_read(const struct sock *sk)
228 struct sk_psock *psock;
232 psock = sk_psock(sk);
234 empty = list_empty(&psock->ingress_msg);
239 static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock,
240 int flags, long timeo, int *err)
242 DEFINE_WAIT_FUNC(wait, woken_wake_function);
248 add_wait_queue(sk_sleep(sk), &wait);
249 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
250 ret = sk_wait_event(sk, &timeo,
251 !list_empty(&psock->ingress_msg) ||
252 !skb_queue_empty(&sk->sk_receive_queue), &wait);
253 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
254 remove_wait_queue(sk_sleep(sk), &wait);
258 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
259 int nonblock, int flags, int *addr_len)
261 struct sk_psock *psock;
264 if (unlikely(flags & MSG_ERRQUEUE))
265 return inet_recv_error(sk, msg, len, addr_len);
267 psock = sk_psock_get(sk);
268 if (unlikely(!psock))
269 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
270 if (!skb_queue_empty(&sk->sk_receive_queue) &&
271 sk_psock_queue_empty(psock)) {
272 sk_psock_put(sk, psock);
273 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
277 copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags);
282 timeo = sock_rcvtimeo(sk, nonblock);
283 data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err);
285 if (!sk_psock_queue_empty(psock))
286 goto msg_bytes_ready;
288 sk_psock_put(sk, psock);
289 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
300 sk_psock_put(sk, psock);
304 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
305 struct sk_msg *msg, int *copied, int flags)
307 bool cork = false, enospc = sk_msg_full(msg);
308 struct sock *sk_redir;
309 u32 tosend, delta = 0;
313 if (psock->eval == __SK_NONE) {
314 /* Track delta in msg size to add/subtract it on SK_DROP from
315 * returned to user copied size. This ensures user doesn't
316 * get a positive return code with msg_cut_data and SK_DROP
319 delta = msg->sg.size;
320 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
321 delta -= msg->sg.size;
324 if (msg->cork_bytes &&
325 msg->cork_bytes > msg->sg.size && !enospc) {
326 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
328 psock->cork = kzalloc(sizeof(*psock->cork),
329 GFP_ATOMIC | __GFP_NOWARN);
333 memcpy(psock->cork, msg, sizeof(*msg));
337 tosend = msg->sg.size;
338 if (psock->apply_bytes && psock->apply_bytes < tosend)
339 tosend = psock->apply_bytes;
341 switch (psock->eval) {
343 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
345 *copied -= sk_msg_free(sk, msg);
348 sk_msg_apply_bytes(psock, tosend);
351 sk_redir = psock->sk_redir;
352 sk_msg_apply_bytes(psock, tosend);
357 sk_msg_return(sk, msg, tosend);
359 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
361 if (unlikely(ret < 0)) {
362 int free = sk_msg_free_nocharge(sk, msg);
368 sk_msg_free(sk, msg);
376 sk_msg_free_partial(sk, msg, tosend);
377 sk_msg_apply_bytes(psock, tosend);
378 *copied -= (tosend + delta);
383 if (!psock->apply_bytes) {
384 psock->eval = __SK_NONE;
385 if (psock->sk_redir) {
386 sock_put(psock->sk_redir);
387 psock->sk_redir = NULL;
391 msg->sg.data[msg->sg.start].page_link &&
392 msg->sg.data[msg->sg.start].length)
398 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
400 struct sk_msg tmp, *msg_tx = NULL;
401 int copied = 0, err = 0;
402 struct sk_psock *psock;
406 /* Don't let internal do_tcp_sendpages() flags through */
407 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
408 flags |= MSG_NO_SHARED_FRAGS;
410 psock = sk_psock_get(sk);
411 if (unlikely(!psock))
412 return tcp_sendmsg(sk, msg, size);
415 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
416 while (msg_data_left(msg)) {
425 copy = msg_data_left(msg);
426 if (!sk_stream_memory_free(sk))
427 goto wait_for_sndbuf;
429 msg_tx = psock->cork;
435 osize = msg_tx->sg.size;
436 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
439 goto wait_for_memory;
441 copy = msg_tx->sg.size - osize;
444 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
447 sk_msg_trim(sk, msg_tx, osize);
452 if (psock->cork_bytes) {
453 if (size > psock->cork_bytes)
454 psock->cork_bytes = 0;
456 psock->cork_bytes -= size;
457 if (psock->cork_bytes && !enospc)
459 /* All cork bytes are accounted, rerun the prog. */
460 psock->eval = __SK_NONE;
461 psock->cork_bytes = 0;
464 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
465 if (unlikely(err < 0))
469 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
471 err = sk_stream_wait_memory(sk, &timeo);
473 if (msg_tx && msg_tx != psock->cork)
474 sk_msg_free(sk, msg_tx);
480 err = sk_stream_error(sk, msg->msg_flags, err);
482 sk_psock_put(sk, psock);
483 return copied ? copied : err;
486 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
487 size_t size, int flags)
489 struct sk_msg tmp, *msg = NULL;
490 int err = 0, copied = 0;
491 struct sk_psock *psock;
494 psock = sk_psock_get(sk);
495 if (unlikely(!psock))
496 return tcp_sendpage(sk, page, offset, size, flags);
506 /* Catch case where ring is full and sendpage is stalled. */
507 if (unlikely(sk_msg_full(msg)))
510 sk_msg_page_add(msg, page, size, offset);
511 sk_mem_charge(sk, size);
513 if (sk_msg_full(msg))
515 if (psock->cork_bytes) {
516 if (size > psock->cork_bytes)
517 psock->cork_bytes = 0;
519 psock->cork_bytes -= size;
520 if (psock->cork_bytes && !enospc)
522 /* All cork bytes are accounted, rerun the prog. */
523 psock->eval = __SK_NONE;
524 psock->cork_bytes = 0;
527 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
530 sk_psock_put(sk, psock);
531 return copied ? copied : err;
546 static struct proto *tcpv6_prot_saved __read_mostly;
547 static DEFINE_SPINLOCK(tcpv6_prot_lock);
548 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
550 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
553 prot[TCP_BPF_BASE] = *base;
554 prot[TCP_BPF_BASE].unhash = sock_map_unhash;
555 prot[TCP_BPF_BASE].close = sock_map_close;
556 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
557 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read;
559 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
560 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
561 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
564 static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops)
566 if (sk->sk_family == AF_INET6 &&
567 unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
568 spin_lock_bh(&tcpv6_prot_lock);
569 if (likely(ops != tcpv6_prot_saved)) {
570 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
571 smp_store_release(&tcpv6_prot_saved, ops);
573 spin_unlock_bh(&tcpv6_prot_lock);
577 static int __init tcp_bpf_v4_build_proto(void)
579 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
582 core_initcall(tcp_bpf_v4_build_proto);
584 static int tcp_bpf_assert_proto_ops(struct proto *ops)
586 /* In order to avoid retpoline, we make assumptions when we call
587 * into ops if e.g. a psock is not present. Make sure they are
588 * indeed valid assumptions.
590 return ops->recvmsg == tcp_recvmsg &&
591 ops->sendmsg == tcp_sendmsg &&
592 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
595 struct proto *tcp_bpf_get_proto(struct sock *sk, struct sk_psock *psock)
597 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
598 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
600 if (!psock->sk_proto) {
601 struct proto *ops = READ_ONCE(sk->sk_prot);
603 if (tcp_bpf_assert_proto_ops(ops))
604 return ERR_PTR(-EINVAL);
606 tcp_bpf_check_v6_needs_rebuild(sk, ops);
609 return &tcp_bpf_prots[family][config];
612 /* If a child got cloned from a listening socket that had tcp_bpf
613 * protocol callbacks installed, we need to restore the callbacks to
614 * the default ones because the child does not inherit the psock state
615 * that tcp_bpf callbacks expect.
617 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
619 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
620 struct proto *prot = newsk->sk_prot;
622 if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
623 newsk->sk_prot = sk->sk_prot_creator;
625 #endif /* CONFIG_BPF_STREAM_PARSER */