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)
44 struct iov_iter *iter = &msg->msg_iter;
45 int i, ret, copied = 0;
47 while (copied != len) {
48 struct scatterlist *sge;
49 struct sk_msg *msg_rx;
51 msg_rx = list_first_entry_or_null(&psock->ingress_msg,
53 if (unlikely(!msg_rx))
61 sge = sk_msg_elem(msg_rx, i);
64 if (copied + copy > len)
66 ret = copy_page_to_iter(page, sge->offset, copy, iter);
75 sk_mem_uncharge(sk, copy);
78 if (i == MAX_SKB_FRAGS)
86 } while (i != msg_rx->sg.end);
89 if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) {
90 list_del(&msg_rx->list);
92 consume_skb(msg_rx->skb);
99 EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg);
101 int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
102 int nonblock, int flags, int *addr_len)
104 struct sk_psock *psock;
107 if (unlikely(flags & MSG_ERRQUEUE))
108 return inet_recv_error(sk, msg, len, addr_len);
109 if (!skb_queue_empty(&sk->sk_receive_queue))
110 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
112 psock = sk_psock_get(sk);
113 if (unlikely(!psock))
114 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
117 copied = __tcp_bpf_recvmsg(sk, psock, msg, len);
122 timeo = sock_rcvtimeo(sk, nonblock);
123 data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err);
125 if (skb_queue_empty(&sk->sk_receive_queue))
126 goto msg_bytes_ready;
128 sk_psock_put(sk, psock);
129 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
139 sk_psock_put(sk, psock);
143 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
144 struct sk_msg *msg, u32 apply_bytes, int flags)
146 bool apply = apply_bytes;
147 struct scatterlist *sge;
148 u32 size, copied = 0;
152 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
157 tmp->sg.start = msg->sg.start;
160 sge = sk_msg_elem(msg, i);
161 size = (apply && apply_bytes < sge->length) ?
162 apply_bytes : sge->length;
163 if (!sk_wmem_schedule(sk, size)) {
169 sk_mem_charge(sk, size);
170 sk_msg_xfer(tmp, msg, i, size);
173 get_page(sk_msg_page(tmp, i));
174 sk_msg_iter_var_next(i);
181 } while (i != msg->sg.end);
185 msg->sg.size -= apply_bytes;
186 sk_psock_queue_msg(psock, tmp);
187 sk->sk_data_ready(sk);
189 sk_msg_free(sk, tmp);
197 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
198 int flags, bool uncharge)
200 bool apply = apply_bytes;
201 struct scatterlist *sge;
207 sge = sk_msg_elem(msg, msg->sg.start);
208 size = (apply && apply_bytes < sge->length) ?
209 apply_bytes : sge->length;
213 tcp_rate_check_app_limited(sk);
215 ret = do_tcp_sendpages(sk, page, off, size, flags);
224 sk_mem_uncharge(sk, ret);
232 sk_msg_iter_next(msg, start);
233 sg_init_table(sge, 1);
234 if (msg->sg.start == msg->sg.end)
237 if (apply && !apply_bytes)
244 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
245 u32 apply_bytes, int flags, bool uncharge)
250 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
255 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
256 u32 bytes, int flags)
258 bool ingress = sk_msg_to_ingress(msg);
259 struct sk_psock *psock = sk_psock_get(sk);
262 if (unlikely(!psock)) {
263 sk_msg_free(sk, msg);
266 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
267 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
268 sk_psock_put(sk, psock);
271 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
273 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
274 struct sk_msg *msg, int *copied, int flags)
276 bool cork = false, enospc = msg->sg.start == msg->sg.end;
277 struct sock *sk_redir;
282 if (psock->eval == __SK_NONE)
283 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
285 if (msg->cork_bytes &&
286 msg->cork_bytes > msg->sg.size && !enospc) {
287 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
289 psock->cork = kzalloc(sizeof(*psock->cork),
290 GFP_ATOMIC | __GFP_NOWARN);
294 memcpy(psock->cork, msg, sizeof(*msg));
298 tosend = msg->sg.size;
299 if (psock->apply_bytes && psock->apply_bytes < tosend)
300 tosend = psock->apply_bytes;
302 switch (psock->eval) {
304 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
306 *copied -= sk_msg_free(sk, msg);
309 sk_msg_apply_bytes(psock, tosend);
312 sk_redir = psock->sk_redir;
313 sk_msg_apply_bytes(psock, tosend);
318 sk_msg_return(sk, msg, tosend);
320 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
322 if (unlikely(ret < 0)) {
323 int free = sk_msg_free_nocharge(sk, msg);
329 sk_msg_free(sk, msg);
337 sk_msg_free_partial(sk, msg, tosend);
338 sk_msg_apply_bytes(psock, tosend);
344 if (!psock->apply_bytes) {
345 psock->eval = __SK_NONE;
346 if (psock->sk_redir) {
347 sock_put(psock->sk_redir);
348 psock->sk_redir = NULL;
352 msg->sg.data[msg->sg.start].page_link &&
353 msg->sg.data[msg->sg.start].length)
359 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
361 struct sk_msg tmp, *msg_tx = NULL;
362 int flags = msg->msg_flags | MSG_NO_SHARED_FRAGS;
363 int copied = 0, err = 0;
364 struct sk_psock *psock;
367 psock = sk_psock_get(sk);
368 if (unlikely(!psock))
369 return tcp_sendmsg(sk, msg, size);
372 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
373 while (msg_data_left(msg)) {
382 copy = msg_data_left(msg);
383 if (!sk_stream_memory_free(sk))
384 goto wait_for_sndbuf;
386 msg_tx = psock->cork;
392 osize = msg_tx->sg.size;
393 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
396 goto wait_for_memory;
398 copy = msg_tx->sg.size - osize;
401 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
404 sk_msg_trim(sk, msg_tx, osize);
409 if (psock->cork_bytes) {
410 if (size > psock->cork_bytes)
411 psock->cork_bytes = 0;
413 psock->cork_bytes -= size;
414 if (psock->cork_bytes && !enospc)
416 /* All cork bytes are accounted, rerun the prog. */
417 psock->eval = __SK_NONE;
418 psock->cork_bytes = 0;
421 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
422 if (unlikely(err < 0))
426 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
428 err = sk_stream_wait_memory(sk, &timeo);
430 if (msg_tx && msg_tx != psock->cork)
431 sk_msg_free(sk, msg_tx);
437 err = sk_stream_error(sk, msg->msg_flags, err);
439 sk_psock_put(sk, psock);
440 return copied ? copied : err;
443 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
444 size_t size, int flags)
446 struct sk_msg tmp, *msg = NULL;
447 int err = 0, copied = 0;
448 struct sk_psock *psock;
451 psock = sk_psock_get(sk);
452 if (unlikely(!psock))
453 return tcp_sendpage(sk, page, offset, size, flags);
463 /* Catch case where ring is full and sendpage is stalled. */
464 if (unlikely(sk_msg_full(msg)))
467 sk_msg_page_add(msg, page, size, offset);
468 sk_mem_charge(sk, size);
470 if (sk_msg_full(msg))
472 if (psock->cork_bytes) {
473 if (size > psock->cork_bytes)
474 psock->cork_bytes = 0;
476 psock->cork_bytes -= size;
477 if (psock->cork_bytes && !enospc)
479 /* All cork bytes are accounted, rerun the prog. */
480 psock->eval = __SK_NONE;
481 psock->cork_bytes = 0;
484 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
487 sk_psock_put(sk, psock);
488 return copied ? copied : err;
491 static void tcp_bpf_remove(struct sock *sk, struct sk_psock *psock)
493 struct sk_psock_link *link;
495 sk_psock_cork_free(psock);
496 __sk_psock_purge_ingress_msg(psock);
497 while ((link = sk_psock_link_pop(psock))) {
498 sk_psock_unlink(sk, link);
499 sk_psock_free_link(link);
503 static void tcp_bpf_unhash(struct sock *sk)
505 void (*saved_unhash)(struct sock *sk);
506 struct sk_psock *psock;
509 psock = sk_psock(sk);
510 if (unlikely(!psock)) {
512 if (sk->sk_prot->unhash)
513 sk->sk_prot->unhash(sk);
517 saved_unhash = psock->saved_unhash;
518 tcp_bpf_remove(sk, psock);
523 static void tcp_bpf_close(struct sock *sk, long timeout)
525 void (*saved_close)(struct sock *sk, long timeout);
526 struct sk_psock *psock;
530 psock = sk_psock(sk);
531 if (unlikely(!psock)) {
534 return sk->sk_prot->close(sk, timeout);
537 saved_close = psock->saved_close;
538 tcp_bpf_remove(sk, psock);
541 saved_close(sk, timeout);
556 static struct proto *tcpv6_prot_saved __read_mostly;
557 static DEFINE_SPINLOCK(tcpv6_prot_lock);
558 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
560 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
563 prot[TCP_BPF_BASE] = *base;
564 prot[TCP_BPF_BASE].unhash = tcp_bpf_unhash;
565 prot[TCP_BPF_BASE].close = tcp_bpf_close;
566 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
567 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read;
569 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
570 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
571 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
574 static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops)
576 if (sk->sk_family == AF_INET6 &&
577 unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
578 spin_lock_bh(&tcpv6_prot_lock);
579 if (likely(ops != tcpv6_prot_saved)) {
580 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
581 smp_store_release(&tcpv6_prot_saved, ops);
583 spin_unlock_bh(&tcpv6_prot_lock);
587 static int __init tcp_bpf_v4_build_proto(void)
589 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
592 core_initcall(tcp_bpf_v4_build_proto);
594 static void tcp_bpf_update_sk_prot(struct sock *sk, struct sk_psock *psock)
596 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
597 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
599 sk_psock_update_proto(sk, psock, &tcp_bpf_prots[family][config]);
602 static void tcp_bpf_reinit_sk_prot(struct sock *sk, struct sk_psock *psock)
604 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
605 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
607 /* Reinit occurs when program types change e.g. TCP_BPF_TX is removed
608 * or added requiring sk_prot hook updates. We keep original saved
609 * hooks in this case.
611 sk->sk_prot = &tcp_bpf_prots[family][config];
614 static int tcp_bpf_assert_proto_ops(struct proto *ops)
616 /* In order to avoid retpoline, we make assumptions when we call
617 * into ops if e.g. a psock is not present. Make sure they are
618 * indeed valid assumptions.
620 return ops->recvmsg == tcp_recvmsg &&
621 ops->sendmsg == tcp_sendmsg &&
622 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
625 void tcp_bpf_reinit(struct sock *sk)
627 struct sk_psock *psock;
629 sock_owned_by_me(sk);
632 psock = sk_psock(sk);
633 tcp_bpf_reinit_sk_prot(sk, psock);
637 int tcp_bpf_init(struct sock *sk)
639 struct proto *ops = READ_ONCE(sk->sk_prot);
640 struct sk_psock *psock;
642 sock_owned_by_me(sk);
645 psock = sk_psock(sk);
646 if (unlikely(!psock || psock->sk_proto ||
647 tcp_bpf_assert_proto_ops(ops))) {
651 tcp_bpf_check_v6_needs_rebuild(sk, ops);
652 tcp_bpf_update_sk_prot(sk, psock);