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);
50 sk_msg_iter_var_prev(i);
54 } while (i != msg->sg.end);
58 sk_psock_queue_msg(psock, tmp);
59 sk_psock_data_ready(sk, psock);
69 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes,
70 int flags, bool uncharge)
72 bool apply = apply_bytes;
73 struct scatterlist *sge;
81 sge = sk_msg_elem(msg, msg->sg.start);
82 size = (apply && apply_bytes < sge->length) ?
83 apply_bytes : sge->length;
87 tcp_rate_check_app_limited(sk);
89 has_tx_ulp = tls_sw_has_ctx_tx(sk);
91 flags |= MSG_SENDPAGE_NOPOLICY;
92 ret = kernel_sendpage_locked(sk,
93 page, off, size, flags);
95 ret = do_tcp_sendpages(sk, page, off, size, flags);
106 sk_mem_uncharge(sk, ret);
114 sk_msg_iter_next(msg, start);
115 sg_init_table(sge, 1);
116 if (msg->sg.start == msg->sg.end)
119 if (apply && !apply_bytes)
126 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg,
127 u32 apply_bytes, int flags, bool uncharge)
132 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge);
137 int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress,
138 struct sk_msg *msg, u32 bytes, int flags)
140 struct sk_psock *psock = sk_psock_get(sk);
143 if (unlikely(!psock))
146 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
147 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
148 sk_psock_put(sk, psock);
151 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
153 #ifdef CONFIG_BPF_SYSCALL
154 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
157 DEFINE_WAIT_FUNC(wait, woken_wake_function);
160 if (sk->sk_shutdown & RCV_SHUTDOWN)
166 add_wait_queue(sk_sleep(sk), &wait);
167 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
168 ret = sk_wait_event(sk, &timeo,
169 !list_empty(&psock->ingress_msg) ||
170 !skb_queue_empty(&sk->sk_receive_queue), &wait);
171 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
172 remove_wait_queue(sk_sleep(sk), &wait);
176 static int tcp_bpf_recvmsg_parser(struct sock *sk,
182 struct sk_psock *psock;
185 if (unlikely(flags & MSG_ERRQUEUE))
186 return inet_recv_error(sk, msg, len, addr_len);
188 psock = sk_psock_get(sk);
189 if (unlikely(!psock))
190 return tcp_recvmsg(sk, msg, len, flags, addr_len);
194 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
199 if (sock_flag(sk, SOCK_DONE))
203 copied = sock_error(sk);
207 if (sk->sk_shutdown & RCV_SHUTDOWN)
210 if (sk->sk_state == TCP_CLOSE) {
215 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
221 if (signal_pending(current)) {
222 copied = sock_intr_errno(timeo);
226 data = tcp_msg_wait_data(sk, psock, timeo);
227 if (data && !sk_psock_queue_empty(psock))
228 goto msg_bytes_ready;
233 sk_psock_put(sk, psock);
237 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
238 int flags, int *addr_len)
240 struct sk_psock *psock;
243 if (unlikely(flags & MSG_ERRQUEUE))
244 return inet_recv_error(sk, msg, len, addr_len);
246 psock = sk_psock_get(sk);
247 if (unlikely(!psock))
248 return tcp_recvmsg(sk, msg, len, flags, addr_len);
249 if (!skb_queue_empty(&sk->sk_receive_queue) &&
250 sk_psock_queue_empty(psock)) {
251 sk_psock_put(sk, psock);
252 return tcp_recvmsg(sk, msg, len, flags, addr_len);
256 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
261 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
262 data = tcp_msg_wait_data(sk, psock, timeo);
264 if (!sk_psock_queue_empty(psock))
265 goto msg_bytes_ready;
267 sk_psock_put(sk, psock);
268 return tcp_recvmsg(sk, msg, len, flags, addr_len);
274 sk_psock_put(sk, psock);
278 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
279 struct sk_msg *msg, int *copied, int flags)
281 bool cork = false, enospc = sk_msg_full(msg), redir_ingress;
282 struct sock *sk_redir;
283 u32 tosend, origsize, sent, delta = 0;
288 if (psock->eval == __SK_NONE) {
289 /* Track delta in msg size to add/subtract it on SK_DROP from
290 * returned to user copied size. This ensures user doesn't
291 * get a positive return code with msg_cut_data and SK_DROP
294 delta = msg->sg.size;
295 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
296 delta -= msg->sg.size;
299 if (msg->cork_bytes &&
300 msg->cork_bytes > msg->sg.size && !enospc) {
301 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
303 psock->cork = kzalloc(sizeof(*psock->cork),
304 GFP_ATOMIC | __GFP_NOWARN);
308 memcpy(psock->cork, msg, sizeof(*msg));
312 tosend = msg->sg.size;
313 if (psock->apply_bytes && psock->apply_bytes < tosend)
314 tosend = psock->apply_bytes;
317 switch (psock->eval) {
319 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
321 *copied -= sk_msg_free(sk, msg);
324 sk_msg_apply_bytes(psock, tosend);
327 redir_ingress = psock->redir_ingress;
328 sk_redir = psock->sk_redir;
329 sk_msg_apply_bytes(psock, tosend);
330 if (!psock->apply_bytes) {
331 /* Clean up before releasing the sock lock. */
333 psock->eval = __SK_NONE;
334 psock->sk_redir = NULL;
340 sk_msg_return(sk, msg, tosend);
343 origsize = msg->sg.size;
344 ret = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
346 sent = origsize - msg->sg.size;
348 if (eval == __SK_REDIRECT)
352 if (unlikely(ret < 0)) {
353 int free = sk_msg_free_nocharge(sk, msg);
359 sk_msg_free(sk, msg);
367 sk_msg_free_partial(sk, msg, tosend);
368 sk_msg_apply_bytes(psock, tosend);
369 *copied -= (tosend + delta);
374 if (!psock->apply_bytes) {
375 psock->eval = __SK_NONE;
376 if (psock->sk_redir) {
377 sock_put(psock->sk_redir);
378 psock->sk_redir = NULL;
382 msg->sg.data[msg->sg.start].page_link &&
383 msg->sg.data[msg->sg.start].length) {
384 if (eval == __SK_REDIRECT)
385 sk_mem_charge(sk, tosend - sent);
392 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
394 struct sk_msg tmp, *msg_tx = NULL;
395 int copied = 0, err = 0;
396 struct sk_psock *psock;
400 /* Don't let internal do_tcp_sendpages() flags through */
401 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
402 flags |= MSG_NO_SHARED_FRAGS;
404 psock = sk_psock_get(sk);
405 if (unlikely(!psock))
406 return tcp_sendmsg(sk, msg, size);
409 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
410 while (msg_data_left(msg)) {
419 copy = msg_data_left(msg);
420 if (!sk_stream_memory_free(sk))
421 goto wait_for_sndbuf;
423 msg_tx = psock->cork;
429 osize = msg_tx->sg.size;
430 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
433 goto wait_for_memory;
435 copy = msg_tx->sg.size - osize;
438 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
441 sk_msg_trim(sk, msg_tx, osize);
446 if (psock->cork_bytes) {
447 if (size > psock->cork_bytes)
448 psock->cork_bytes = 0;
450 psock->cork_bytes -= size;
451 if (psock->cork_bytes && !enospc)
453 /* All cork bytes are accounted, rerun the prog. */
454 psock->eval = __SK_NONE;
455 psock->cork_bytes = 0;
458 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
459 if (unlikely(err < 0))
463 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
465 err = sk_stream_wait_memory(sk, &timeo);
467 if (msg_tx && msg_tx != psock->cork)
468 sk_msg_free(sk, msg_tx);
474 err = sk_stream_error(sk, msg->msg_flags, err);
476 sk_psock_put(sk, psock);
477 return copied ? copied : err;
480 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
481 size_t size, int flags)
483 struct sk_msg tmp, *msg = NULL;
484 int err = 0, copied = 0;
485 struct sk_psock *psock;
488 psock = sk_psock_get(sk);
489 if (unlikely(!psock))
490 return tcp_sendpage(sk, page, offset, size, flags);
500 /* Catch case where ring is full and sendpage is stalled. */
501 if (unlikely(sk_msg_full(msg)))
504 sk_msg_page_add(msg, page, size, offset);
505 sk_mem_charge(sk, size);
507 if (sk_msg_full(msg))
509 if (psock->cork_bytes) {
510 if (size > psock->cork_bytes)
511 psock->cork_bytes = 0;
513 psock->cork_bytes -= size;
514 if (psock->cork_bytes && !enospc)
516 /* All cork bytes are accounted, rerun the prog. */
517 psock->eval = __SK_NONE;
518 psock->cork_bytes = 0;
521 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags);
524 sk_psock_put(sk, psock);
525 return copied ? copied : err;
542 static struct proto *tcpv6_prot_saved __read_mostly;
543 static DEFINE_SPINLOCK(tcpv6_prot_lock);
544 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
546 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
549 prot[TCP_BPF_BASE] = *base;
550 prot[TCP_BPF_BASE].destroy = sock_map_destroy;
551 prot[TCP_BPF_BASE].close = sock_map_close;
552 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
553 prot[TCP_BPF_BASE].sock_is_readable = sk_msg_is_readable;
555 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
556 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
557 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
559 prot[TCP_BPF_RX] = prot[TCP_BPF_BASE];
560 prot[TCP_BPF_RX].recvmsg = tcp_bpf_recvmsg_parser;
562 prot[TCP_BPF_TXRX] = prot[TCP_BPF_TX];
563 prot[TCP_BPF_TXRX].recvmsg = tcp_bpf_recvmsg_parser;
566 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
568 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
569 spin_lock_bh(&tcpv6_prot_lock);
570 if (likely(ops != tcpv6_prot_saved)) {
571 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
572 smp_store_release(&tcpv6_prot_saved, ops);
574 spin_unlock_bh(&tcpv6_prot_lock);
578 static int __init tcp_bpf_v4_build_proto(void)
580 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
583 late_initcall(tcp_bpf_v4_build_proto);
585 static int tcp_bpf_assert_proto_ops(struct proto *ops)
587 /* In order to avoid retpoline, we make assumptions when we call
588 * into ops if e.g. a psock is not present. Make sure they are
589 * indeed valid assumptions.
591 return ops->recvmsg == tcp_recvmsg &&
592 ops->sendmsg == tcp_sendmsg &&
593 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
596 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
598 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
599 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
601 if (psock->progs.stream_verdict || psock->progs.skb_verdict) {
602 config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX;
606 if (inet_csk_has_ulp(sk)) {
607 /* TLS does not have an unhash proto in SW cases,
608 * but we need to ensure we stop using the sock_map
609 * unhash routine because the associated psock is being
610 * removed. So use the original unhash handler.
612 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
613 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
615 sk->sk_write_space = psock->saved_write_space;
616 /* Pairs with lockless read in sk_clone_lock() */
617 sock_replace_proto(sk, psock->sk_proto);
622 if (sk->sk_family == AF_INET6) {
623 if (tcp_bpf_assert_proto_ops(psock->sk_proto))
626 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
629 /* Pairs with lockless read in sk_clone_lock() */
630 sock_replace_proto(sk, &tcp_bpf_prots[family][config]);
633 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
635 /* If a child got cloned from a listening socket that had tcp_bpf
636 * protocol callbacks installed, we need to restore the callbacks to
637 * the default ones because the child does not inherit the psock state
638 * that tcp_bpf callbacks expect.
640 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
642 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
643 struct proto *prot = newsk->sk_prot;
645 if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
646 newsk->sk_prot = sk->sk_prot_creator;
648 #endif /* CONFIG_BPF_SYSCALL */