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 static bool tcp_bpf_stream_read(const struct sock *sk)
15 struct sk_psock *psock;
21 empty = list_empty(&psock->ingress_msg);
26 static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock,
27 int flags, long timeo, int *err)
29 DEFINE_WAIT_FUNC(wait, woken_wake_function);
35 add_wait_queue(sk_sleep(sk), &wait);
36 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
37 ret = sk_wait_event(sk, &timeo,
38 !list_empty(&psock->ingress_msg) ||
39 !skb_queue_empty(&sk->sk_receive_queue), &wait);
40 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
41 remove_wait_queue(sk_sleep(sk), &wait);
45 int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock,
46 struct msghdr *msg, int len, int flags)
48 struct iov_iter *iter = &msg->msg_iter;
49 int peek = flags & MSG_PEEK;
50 struct sk_msg *msg_rx;
53 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
56 while (copied != len) {
57 struct scatterlist *sge;
59 if (unlikely(!msg_rx))
67 sge = sk_msg_elem(msg_rx, i);
70 if (copied + copy > len)
72 copy = copy_page_to_iter(page, sge->offset, copy, iter);
74 return copied ? copied : -EFAULT;
81 sk_mem_uncharge(sk, copy);
82 msg_rx->sg.size -= copy;
85 sk_msg_iter_var_next(i);
90 /* Lets not optimize peek case if copy_page_to_iter
91 * didn't copy the entire length lets just break.
93 if (copy != sge->length)
95 sk_msg_iter_var_next(i);
100 } while (i != msg_rx->sg.end);
102 if (unlikely(peek)) {
103 if (msg_rx == list_last_entry(&psock->ingress_msg,
104 struct sk_msg, list))
106 msg_rx = list_next_entry(msg_rx, list);
110 msg_rx->sg.start = i;
111 if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
112 list_del(&msg_rx->list);
114 consume_skb(msg_rx->skb);
117 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
118 struct sk_msg, list);
123 EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg);
125 int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
126 int nonblock, int flags, int *addr_len)
128 struct sk_psock *psock;
131 if (unlikely(flags & MSG_ERRQUEUE))
132 return inet_recv_error(sk, msg, len, addr_len);
134 psock = sk_psock_get(sk);
135 if (unlikely(!psock))
136 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
137 if (!skb_queue_empty(&sk->sk_receive_queue) &&
138 sk_psock_queue_empty(psock)) {
139 sk_psock_put(sk, psock);
140 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
144 copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags);
149 timeo = sock_rcvtimeo(sk, nonblock);
150 data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err);
152 if (!sk_psock_queue_empty(psock))
153 goto msg_bytes_ready;
155 sk_psock_put(sk, psock);
156 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
167 sk_psock_put(sk, psock);
171 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
172 struct sk_msg *msg, u32 apply_bytes, int flags)
174 bool apply = apply_bytes;
175 struct scatterlist *sge;
176 u32 size, copied = 0;
180 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
185 tmp->sg.start = msg->sg.start;
188 sge = sk_msg_elem(msg, i);
189 size = (apply && apply_bytes < sge->length) ?
190 apply_bytes : sge->length;
191 if (!sk_wmem_schedule(sk, size)) {
197 sk_mem_charge(sk, size);
198 sk_msg_xfer(tmp, msg, i, size);
201 get_page(sk_msg_page(tmp, i));
202 sk_msg_iter_var_next(i);
208 sk_msg_iter_var_prev(i);
212 } while (i != msg->sg.end);
216 sk_psock_queue_msg(psock, tmp);
217 sk_psock_data_ready(sk, psock);
219 sk_msg_free(sk, tmp);
227 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
228 int flags, bool uncharge)
230 bool apply = apply_bytes;
231 struct scatterlist *sge;
239 sge = sk_msg_elem(msg, msg->sg.start);
240 size = (apply && apply_bytes < sge->length) ?
241 apply_bytes : sge->length;
245 tcp_rate_check_app_limited(sk);
247 has_tx_ulp = tls_sw_has_ctx_tx(sk);
249 flags |= MSG_SENDPAGE_NOPOLICY;
250 ret = kernel_sendpage_locked(sk,
251 page, off, size, flags);
253 ret = do_tcp_sendpages(sk, page, off, size, flags);
264 sk_mem_uncharge(sk, ret);
272 sk_msg_iter_next(msg, start);
273 sg_init_table(sge, 1);
274 if (msg->sg.start == msg->sg.end)
277 if (apply && !apply_bytes)
284 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
285 u32 apply_bytes, int flags, bool uncharge)
290 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
295 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
296 u32 bytes, int flags)
298 bool ingress = sk_msg_to_ingress(msg);
299 struct sk_psock *psock = sk_psock_get(sk);
302 if (unlikely(!psock))
305 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
306 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
307 sk_psock_put(sk, psock);
310 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
312 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
313 struct sk_msg *msg, int *copied, int flags)
315 bool cork = false, enospc = sk_msg_full(msg);
316 struct sock *sk_redir;
317 u32 tosend, origsize, sent, delta = 0;
322 if (psock->eval == __SK_NONE) {
323 /* Track delta in msg size to add/subtract it on SK_DROP from
324 * returned to user copied size. This ensures user doesn't
325 * get a positive return code with msg_cut_data and SK_DROP
328 delta = msg->sg.size;
329 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
330 delta -= msg->sg.size;
333 if (msg->cork_bytes &&
334 msg->cork_bytes > msg->sg.size && !enospc) {
335 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
337 psock->cork = kzalloc(sizeof(*psock->cork),
338 GFP_ATOMIC | __GFP_NOWARN);
342 memcpy(psock->cork, msg, sizeof(*msg));
346 tosend = msg->sg.size;
347 if (psock->apply_bytes && psock->apply_bytes < tosend)
348 tosend = psock->apply_bytes;
351 switch (psock->eval) {
353 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
355 *copied -= sk_msg_free(sk, msg);
358 sk_msg_apply_bytes(psock, tosend);
361 sk_redir = psock->sk_redir;
362 sk_msg_apply_bytes(psock, tosend);
363 if (!psock->apply_bytes) {
364 /* Clean up before releasing the sock lock. */
366 psock->eval = __SK_NONE;
367 psock->sk_redir = NULL;
373 sk_msg_return(sk, msg, tosend);
376 origsize = msg->sg.size;
377 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
378 sent = origsize - msg->sg.size;
380 if (eval == __SK_REDIRECT)
384 if (unlikely(ret < 0)) {
385 int free = sk_msg_free_nocharge(sk, msg);
391 sk_msg_free(sk, msg);
399 sk_msg_free_partial(sk, msg, tosend);
400 sk_msg_apply_bytes(psock, tosend);
401 *copied -= (tosend + delta);
406 if (!psock->apply_bytes) {
407 psock->eval = __SK_NONE;
408 if (psock->sk_redir) {
409 sock_put(psock->sk_redir);
410 psock->sk_redir = NULL;
414 msg->sg.data[msg->sg.start].page_link &&
415 msg->sg.data[msg->sg.start].length) {
416 if (eval == __SK_REDIRECT)
417 sk_mem_charge(sk, tosend - sent);
424 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
426 struct sk_msg tmp, *msg_tx = NULL;
427 int copied = 0, err = 0;
428 struct sk_psock *psock;
432 /* Don't let internal do_tcp_sendpages() flags through */
433 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
434 flags |= MSG_NO_SHARED_FRAGS;
436 psock = sk_psock_get(sk);
437 if (unlikely(!psock))
438 return tcp_sendmsg(sk, msg, size);
441 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
442 while (msg_data_left(msg)) {
451 copy = msg_data_left(msg);
452 if (!sk_stream_memory_free(sk))
453 goto wait_for_sndbuf;
455 msg_tx = psock->cork;
461 osize = msg_tx->sg.size;
462 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
465 goto wait_for_memory;
467 copy = msg_tx->sg.size - osize;
470 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
473 sk_msg_trim(sk, msg_tx, osize);
478 if (psock->cork_bytes) {
479 if (size > psock->cork_bytes)
480 psock->cork_bytes = 0;
482 psock->cork_bytes -= size;
483 if (psock->cork_bytes && !enospc)
485 /* All cork bytes are accounted, rerun the prog. */
486 psock->eval = __SK_NONE;
487 psock->cork_bytes = 0;
490 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
491 if (unlikely(err < 0))
495 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
497 err = sk_stream_wait_memory(sk, &timeo);
499 if (msg_tx && msg_tx != psock->cork)
500 sk_msg_free(sk, msg_tx);
506 err = sk_stream_error(sk, msg->msg_flags, err);
508 sk_psock_put(sk, psock);
509 return copied ? copied : err;
512 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
513 size_t size, int flags)
515 struct sk_msg tmp, *msg = NULL;
516 int err = 0, copied = 0;
517 struct sk_psock *psock;
520 psock = sk_psock_get(sk);
521 if (unlikely(!psock))
522 return tcp_sendpage(sk, page, offset, size, flags);
532 /* Catch case where ring is full and sendpage is stalled. */
533 if (unlikely(sk_msg_full(msg)))
536 sk_msg_page_add(msg, page, size, offset);
537 sk_mem_charge(sk, size);
539 if (sk_msg_full(msg))
541 if (psock->cork_bytes) {
542 if (size > psock->cork_bytes)
543 psock->cork_bytes = 0;
545 psock->cork_bytes -= size;
546 if (psock->cork_bytes && !enospc)
548 /* All cork bytes are accounted, rerun the prog. */
549 psock->eval = __SK_NONE;
550 psock->cork_bytes = 0;
553 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
556 sk_psock_put(sk, psock);
557 return copied ? copied : err;
560 static void tcp_bpf_remove(struct sock *sk, struct sk_psock *psock)
562 struct sk_psock_link *link;
564 while ((link = sk_psock_link_pop(psock))) {
565 sk_psock_unlink(sk, link);
566 sk_psock_free_link(link);
570 static void tcp_bpf_unhash(struct sock *sk)
572 void (*saved_unhash)(struct sock *sk);
573 struct sk_psock *psock;
576 psock = sk_psock(sk);
577 if (unlikely(!psock)) {
579 if (sk->sk_prot->unhash)
580 sk->sk_prot->unhash(sk);
584 saved_unhash = psock->saved_unhash;
585 tcp_bpf_remove(sk, psock);
590 static void tcp_bpf_close(struct sock *sk, long timeout)
592 void (*saved_close)(struct sock *sk, long timeout);
593 struct sk_psock *psock;
597 psock = sk_psock(sk);
598 if (unlikely(!psock)) {
601 return sk->sk_prot->close(sk, timeout);
604 saved_close = psock->saved_close;
605 tcp_bpf_remove(sk, psock);
608 saved_close(sk, timeout);
623 static struct proto *tcpv6_prot_saved __read_mostly;
624 static DEFINE_SPINLOCK(tcpv6_prot_lock);
625 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
627 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
630 prot[TCP_BPF_BASE] = *base;
631 prot[TCP_BPF_BASE].unhash = tcp_bpf_unhash;
632 prot[TCP_BPF_BASE].close = tcp_bpf_close;
633 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
634 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read;
636 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
637 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
638 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
641 static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops)
643 if (sk->sk_family == AF_INET6 &&
644 unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
645 spin_lock_bh(&tcpv6_prot_lock);
646 if (likely(ops != tcpv6_prot_saved)) {
647 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
648 smp_store_release(&tcpv6_prot_saved, ops);
650 spin_unlock_bh(&tcpv6_prot_lock);
654 static int __init tcp_bpf_v4_build_proto(void)
656 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
659 late_initcall(tcp_bpf_v4_build_proto);
661 static void tcp_bpf_update_sk_prot(struct sock *sk, struct sk_psock *psock)
663 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
664 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
666 sk_psock_update_proto(sk, psock, &tcp_bpf_prots[family][config]);
669 static void tcp_bpf_reinit_sk_prot(struct sock *sk, struct sk_psock *psock)
671 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
672 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
674 /* Reinit occurs when program types change e.g. TCP_BPF_TX is removed
675 * or added requiring sk_prot hook updates. We keep original saved
676 * hooks in this case.
678 sk->sk_prot = &tcp_bpf_prots[family][config];
681 static int tcp_bpf_assert_proto_ops(struct proto *ops)
683 /* In order to avoid retpoline, we make assumptions when we call
684 * into ops if e.g. a psock is not present. Make sure they are
685 * indeed valid assumptions.
687 return ops->recvmsg == tcp_recvmsg &&
688 ops->sendmsg == tcp_sendmsg &&
689 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
692 void tcp_bpf_reinit(struct sock *sk)
694 struct sk_psock *psock;
696 sock_owned_by_me(sk);
699 psock = sk_psock(sk);
700 tcp_bpf_reinit_sk_prot(sk, psock);
704 int tcp_bpf_init(struct sock *sk)
706 struct proto *ops = READ_ONCE(sk->sk_prot);
707 struct sk_psock *psock;
709 sock_owned_by_me(sk);
712 psock = sk_psock(sk);
713 if (unlikely(!psock || psock->sk_proto ||
714 tcp_bpf_assert_proto_ops(ops))) {
718 tcp_bpf_check_v6_needs_rebuild(sk, ops);
719 tcp_bpf_update_sk_prot(sk, psock);