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 int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock,
14 struct sk_msg *msg, u32 apply_bytes, int flags)
16 bool apply = apply_bytes;
17 struct scatterlist *sge;
22 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL);
27 tmp->sg.start = msg->sg.start;
30 sge = sk_msg_elem(msg, i);
31 size = (apply && apply_bytes < sge->length) ?
32 apply_bytes : sge->length;
33 if (!sk_wmem_schedule(sk, size)) {
39 sk_mem_charge(sk, size);
40 sk_msg_xfer(tmp, msg, i, size);
43 get_page(sk_msg_page(tmp, i));
44 sk_msg_iter_var_next(i);
51 } while (i != msg->sg.end);
55 sk_psock_queue_msg(psock, tmp);
56 sk_psock_data_ready(sk, psock);
66 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
67 int flags, bool uncharge)
69 bool apply = apply_bytes;
70 struct scatterlist *sge;
78 sge = sk_msg_elem(msg, msg->sg.start);
79 size = (apply && apply_bytes < sge->length) ?
80 apply_bytes : sge->length;
84 tcp_rate_check_app_limited(sk);
86 has_tx_ulp = tls_sw_has_ctx_tx(sk);
88 flags |= MSG_SENDPAGE_NOPOLICY;
89 ret = kernel_sendpage_locked(sk,
90 page, off, size, flags);
92 ret = do_tcp_sendpages(sk, page, off, size, flags);
103 sk_mem_uncharge(sk, ret);
111 sk_msg_iter_next(msg, start);
112 sg_init_table(sge, 1);
113 if (msg->sg.start == msg->sg.end)
116 if (apply && !apply_bytes)
123 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
124 u32 apply_bytes, int flags, bool uncharge)
129 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
134 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg,
135 u32 bytes, int flags)
137 bool ingress = sk_msg_to_ingress(msg);
138 struct sk_psock *psock = sk_psock_get(sk);
141 if (unlikely(!psock))
144 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
145 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
146 sk_psock_put(sk, psock);
149 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
151 #ifdef CONFIG_BPF_SYSCALL
152 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
155 DEFINE_WAIT_FUNC(wait, woken_wake_function);
158 if (sk->sk_shutdown & RCV_SHUTDOWN)
164 add_wait_queue(sk_sleep(sk), &wait);
165 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
166 ret = sk_wait_event(sk, &timeo,
167 !list_empty(&psock->ingress_msg) ||
168 !skb_queue_empty(&sk->sk_receive_queue), &wait);
169 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
170 remove_wait_queue(sk_sleep(sk), &wait);
174 static int tcp_bpf_recvmsg_parser(struct sock *sk,
180 struct sk_psock *psock;
183 if (unlikely(flags & MSG_ERRQUEUE))
184 return inet_recv_error(sk, msg, len, addr_len);
186 psock = sk_psock_get(sk);
187 if (unlikely(!psock))
188 return tcp_recvmsg(sk, msg, len, flags, addr_len);
192 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
197 if (sock_flag(sk, SOCK_DONE))
201 copied = sock_error(sk);
205 if (sk->sk_shutdown & RCV_SHUTDOWN)
208 if (sk->sk_state == TCP_CLOSE) {
213 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
219 if (signal_pending(current)) {
220 copied = sock_intr_errno(timeo);
224 data = tcp_msg_wait_data(sk, psock, timeo);
225 if (data && !sk_psock_queue_empty(psock))
226 goto msg_bytes_ready;
231 sk_psock_put(sk, psock);
235 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
236 int flags, int *addr_len)
238 struct sk_psock *psock;
241 if (unlikely(flags & MSG_ERRQUEUE))
242 return inet_recv_error(sk, msg, len, addr_len);
244 psock = sk_psock_get(sk);
245 if (unlikely(!psock))
246 return tcp_recvmsg(sk, msg, len, flags, addr_len);
247 if (!skb_queue_empty(&sk->sk_receive_queue) &&
248 sk_psock_queue_empty(psock)) {
249 sk_psock_put(sk, psock);
250 return tcp_recvmsg(sk, msg, len, flags, addr_len);
254 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
259 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
260 data = tcp_msg_wait_data(sk, psock, timeo);
262 if (!sk_psock_queue_empty(psock))
263 goto msg_bytes_ready;
265 sk_psock_put(sk, psock);
266 return tcp_recvmsg(sk, msg, len, flags, addr_len);
272 sk_psock_put(sk, psock);
276 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
277 struct sk_msg *msg, int *copied, int flags)
279 bool cork = false, enospc = sk_msg_full(msg);
280 struct sock *sk_redir;
281 u32 tosend, delta = 0;
282 u32 eval = __SK_NONE;
286 if (psock->eval == __SK_NONE) {
287 /* Track delta in msg size to add/subtract it on SK_DROP from
288 * returned to user copied size. This ensures user doesn't
289 * get a positive return code with msg_cut_data and SK_DROP
292 delta = msg->sg.size;
293 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
294 delta -= msg->sg.size;
297 if (msg->cork_bytes &&
298 msg->cork_bytes > msg->sg.size && !enospc) {
299 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
301 psock->cork = kzalloc(sizeof(*psock->cork),
302 GFP_ATOMIC | __GFP_NOWARN);
306 memcpy(psock->cork, msg, sizeof(*msg));
310 tosend = msg->sg.size;
311 if (psock->apply_bytes && psock->apply_bytes < tosend)
312 tosend = psock->apply_bytes;
314 switch (psock->eval) {
316 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
318 *copied -= sk_msg_free(sk, msg);
321 sk_msg_apply_bytes(psock, tosend);
324 sk_redir = psock->sk_redir;
325 sk_msg_apply_bytes(psock, tosend);
326 if (!psock->apply_bytes) {
327 /* Clean up before releasing the sock lock. */
329 psock->eval = __SK_NONE;
330 psock->sk_redir = NULL;
336 sk_msg_return(sk, msg, msg->sg.size);
339 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
341 if (eval == __SK_REDIRECT)
345 if (unlikely(ret < 0)) {
346 int free = sk_msg_free_nocharge(sk, msg);
352 sk_msg_free(sk, msg);
360 sk_msg_free_partial(sk, msg, tosend);
361 sk_msg_apply_bytes(psock, tosend);
362 *copied -= (tosend + delta);
367 if (!psock->apply_bytes) {
368 psock->eval = __SK_NONE;
369 if (psock->sk_redir) {
370 sock_put(psock->sk_redir);
371 psock->sk_redir = NULL;
375 msg->sg.data[msg->sg.start].page_link &&
376 msg->sg.data[msg->sg.start].length) {
377 if (eval == __SK_REDIRECT)
378 sk_mem_charge(sk, msg->sg.size);
385 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
387 struct sk_msg tmp, *msg_tx = NULL;
388 int copied = 0, err = 0;
389 struct sk_psock *psock;
393 /* Don't let internal do_tcp_sendpages() flags through */
394 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
395 flags |= MSG_NO_SHARED_FRAGS;
397 psock = sk_psock_get(sk);
398 if (unlikely(!psock))
399 return tcp_sendmsg(sk, msg, size);
402 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
403 while (msg_data_left(msg)) {
412 copy = msg_data_left(msg);
413 if (!sk_stream_memory_free(sk))
414 goto wait_for_sndbuf;
416 msg_tx = psock->cork;
422 osize = msg_tx->sg.size;
423 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
426 goto wait_for_memory;
428 copy = msg_tx->sg.size - osize;
431 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
434 sk_msg_trim(sk, msg_tx, osize);
439 if (psock->cork_bytes) {
440 if (size > psock->cork_bytes)
441 psock->cork_bytes = 0;
443 psock->cork_bytes -= size;
444 if (psock->cork_bytes && !enospc)
446 /* All cork bytes are accounted, rerun the prog. */
447 psock->eval = __SK_NONE;
448 psock->cork_bytes = 0;
451 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
452 if (unlikely(err < 0))
456 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
458 err = sk_stream_wait_memory(sk, &timeo);
460 if (msg_tx && msg_tx != psock->cork)
461 sk_msg_free(sk, msg_tx);
467 err = sk_stream_error(sk, msg->msg_flags, err);
469 sk_psock_put(sk, psock);
470 return copied ? copied : err;
473 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
474 size_t size, int flags)
476 struct sk_msg tmp, *msg = NULL;
477 int err = 0, copied = 0;
478 struct sk_psock *psock;
481 psock = sk_psock_get(sk);
482 if (unlikely(!psock))
483 return tcp_sendpage(sk, page, offset, size, flags);
493 /* Catch case where ring is full and sendpage is stalled. */
494 if (unlikely(sk_msg_full(msg)))
497 sk_msg_page_add(msg, page, size, offset);
498 sk_mem_charge(sk, size);
500 if (sk_msg_full(msg))
502 if (psock->cork_bytes) {
503 if (size > psock->cork_bytes)
504 psock->cork_bytes = 0;
506 psock->cork_bytes -= size;
507 if (psock->cork_bytes && !enospc)
509 /* All cork bytes are accounted, rerun the prog. */
510 psock->eval = __SK_NONE;
511 psock->cork_bytes = 0;
514 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
517 sk_psock_put(sk, psock);
518 return copied ? copied : err;
535 static struct proto *tcpv6_prot_saved __read_mostly;
536 static DEFINE_SPINLOCK(tcpv6_prot_lock);
537 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
539 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
542 prot[TCP_BPF_BASE] = *base;
543 prot[TCP_BPF_BASE].close = sock_map_close;
544 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
545 prot[TCP_BPF_BASE].sock_is_readable = sk_msg_is_readable;
547 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
548 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
549 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
551 prot[TCP_BPF_RX] = prot[TCP_BPF_BASE];
552 prot[TCP_BPF_RX].recvmsg = tcp_bpf_recvmsg_parser;
554 prot[TCP_BPF_TXRX] = prot[TCP_BPF_TX];
555 prot[TCP_BPF_TXRX].recvmsg = tcp_bpf_recvmsg_parser;
558 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
560 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
561 spin_lock_bh(&tcpv6_prot_lock);
562 if (likely(ops != tcpv6_prot_saved)) {
563 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
564 smp_store_release(&tcpv6_prot_saved, ops);
566 spin_unlock_bh(&tcpv6_prot_lock);
570 static int __init tcp_bpf_v4_build_proto(void)
572 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
575 late_initcall(tcp_bpf_v4_build_proto);
577 static int tcp_bpf_assert_proto_ops(struct proto *ops)
579 /* In order to avoid retpoline, we make assumptions when we call
580 * into ops if e.g. a psock is not present. Make sure they are
581 * indeed valid assumptions.
583 return ops->recvmsg == tcp_recvmsg &&
584 ops->sendmsg == tcp_sendmsg &&
585 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
588 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
590 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
591 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
593 if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
594 config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
598 if (inet_csk_has_ulp(sk)) {
599 /* TLS does not have an unhash proto in SW cases,
600 * but we need to ensure we stop using the sock_map
601 * unhash routine because the associated psock is being
602 * removed. So use the original unhash handler.
604 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
605 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
607 sk->sk_write_space = psock->saved_write_space;
608 /* Pairs with lockless read in sk_clone_lock() */
609 WRITE_ONCE(sk->sk_prot, psock->sk_proto);
614 if (inet_csk_has_ulp(sk))
617 if (sk->sk_family == AF_INET6) {
618 if (tcp_bpf_assert_proto_ops(psock->sk_proto))
621 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
624 /* Pairs with lockless read in sk_clone_lock() */
625 WRITE_ONCE(sk->sk_prot, &tcp_bpf_prots[family][config]);
628 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
630 /* If a child got cloned from a listening socket that had tcp_bpf
631 * protocol callbacks installed, we need to restore the callbacks to
632 * the default ones because the child does not inherit the psock state
633 * that tcp_bpf callbacks expect.
635 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
637 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
638 struct proto *prot = newsk->sk_prot;
640 if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
641 newsk->sk_prot = sk->sk_prot_creator;
643 #endif /* CONFIG_BPF_SYSCALL */