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)) {
142 sk_msg_free(sk, msg);
145 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) :
146 tcp_bpf_push_locked(sk, msg, bytes, flags, false);
147 sk_psock_put(sk, psock);
150 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir);
152 #ifdef CONFIG_BPF_SYSCALL
153 static bool tcp_bpf_stream_read(const struct sock *sk)
155 struct sk_psock *psock;
159 psock = sk_psock(sk);
161 empty = list_empty(&psock->ingress_msg);
166 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock,
169 DEFINE_WAIT_FUNC(wait, woken_wake_function);
172 if (sk->sk_shutdown & RCV_SHUTDOWN)
178 add_wait_queue(sk_sleep(sk), &wait);
179 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
180 ret = sk_wait_event(sk, &timeo,
181 !list_empty(&psock->ingress_msg) ||
182 !skb_queue_empty(&sk->sk_receive_queue), &wait);
183 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
184 remove_wait_queue(sk_sleep(sk), &wait);
188 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
189 int nonblock, int flags, int *addr_len)
191 struct sk_psock *psock;
194 if (unlikely(flags & MSG_ERRQUEUE))
195 return inet_recv_error(sk, msg, len, addr_len);
197 psock = sk_psock_get(sk);
198 if (unlikely(!psock))
199 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
200 if (!skb_queue_empty(&sk->sk_receive_queue) &&
201 sk_psock_queue_empty(psock)) {
202 sk_psock_put(sk, psock);
203 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
207 copied = sk_msg_recvmsg(sk, psock, msg, len, flags);
212 timeo = sock_rcvtimeo(sk, nonblock);
213 data = tcp_msg_wait_data(sk, psock, timeo);
215 if (!sk_psock_queue_empty(psock))
216 goto msg_bytes_ready;
218 sk_psock_put(sk, psock);
219 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
225 sk_psock_put(sk, psock);
229 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock,
230 struct sk_msg *msg, int *copied, int flags)
232 bool cork = false, enospc = sk_msg_full(msg);
233 struct sock *sk_redir;
234 u32 tosend, delta = 0;
235 u32 eval = __SK_NONE;
239 if (psock->eval == __SK_NONE) {
240 /* Track delta in msg size to add/subtract it on SK_DROP from
241 * returned to user copied size. This ensures user doesn't
242 * get a positive return code with msg_cut_data and SK_DROP
245 delta = msg->sg.size;
246 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
247 delta -= msg->sg.size;
250 if (msg->cork_bytes &&
251 msg->cork_bytes > msg->sg.size && !enospc) {
252 psock->cork_bytes = msg->cork_bytes - msg->sg.size;
254 psock->cork = kzalloc(sizeof(*psock->cork),
255 GFP_ATOMIC | __GFP_NOWARN);
259 memcpy(psock->cork, msg, sizeof(*msg));
263 tosend = msg->sg.size;
264 if (psock->apply_bytes && psock->apply_bytes < tosend)
265 tosend = psock->apply_bytes;
267 switch (psock->eval) {
269 ret = tcp_bpf_push(sk, msg, tosend, flags, true);
271 *copied -= sk_msg_free(sk, msg);
274 sk_msg_apply_bytes(psock, tosend);
277 sk_redir = psock->sk_redir;
278 sk_msg_apply_bytes(psock, tosend);
279 if (!psock->apply_bytes) {
280 /* Clean up before releasing the sock lock. */
282 psock->eval = __SK_NONE;
283 psock->sk_redir = NULL;
289 sk_msg_return(sk, msg, tosend);
292 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags);
294 if (eval == __SK_REDIRECT)
298 if (unlikely(ret < 0)) {
299 int free = sk_msg_free_nocharge(sk, msg);
305 sk_msg_free(sk, msg);
313 sk_msg_free_partial(sk, msg, tosend);
314 sk_msg_apply_bytes(psock, tosend);
315 *copied -= (tosend + delta);
320 if (!psock->apply_bytes) {
321 psock->eval = __SK_NONE;
322 if (psock->sk_redir) {
323 sock_put(psock->sk_redir);
324 psock->sk_redir = NULL;
328 msg->sg.data[msg->sg.start].page_link &&
329 msg->sg.data[msg->sg.start].length)
335 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
337 struct sk_msg tmp, *msg_tx = NULL;
338 int copied = 0, err = 0;
339 struct sk_psock *psock;
343 /* Don't let internal do_tcp_sendpages() flags through */
344 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED);
345 flags |= MSG_NO_SHARED_FRAGS;
347 psock = sk_psock_get(sk);
348 if (unlikely(!psock))
349 return tcp_sendmsg(sk, msg, size);
352 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
353 while (msg_data_left(msg)) {
362 copy = msg_data_left(msg);
363 if (!sk_stream_memory_free(sk))
364 goto wait_for_sndbuf;
366 msg_tx = psock->cork;
372 osize = msg_tx->sg.size;
373 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1);
376 goto wait_for_memory;
378 copy = msg_tx->sg.size - osize;
381 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx,
384 sk_msg_trim(sk, msg_tx, osize);
389 if (psock->cork_bytes) {
390 if (size > psock->cork_bytes)
391 psock->cork_bytes = 0;
393 psock->cork_bytes -= size;
394 if (psock->cork_bytes && !enospc)
396 /* All cork bytes are accounted, rerun the prog. */
397 psock->eval = __SK_NONE;
398 psock->cork_bytes = 0;
401 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags);
402 if (unlikely(err < 0))
406 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
408 err = sk_stream_wait_memory(sk, &timeo);
410 if (msg_tx && msg_tx != psock->cork)
411 sk_msg_free(sk, msg_tx);
417 err = sk_stream_error(sk, msg->msg_flags, err);
419 sk_psock_put(sk, psock);
420 return copied ? copied : err;
423 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset,
424 size_t size, int flags)
426 struct sk_msg tmp, *msg = NULL;
427 int err = 0, copied = 0;
428 struct sk_psock *psock;
431 psock = sk_psock_get(sk);
432 if (unlikely(!psock))
433 return tcp_sendpage(sk, page, offset, size, flags);
443 /* Catch case where ring is full and sendpage is stalled. */
444 if (unlikely(sk_msg_full(msg)))
447 sk_msg_page_add(msg, page, size, offset);
448 sk_mem_charge(sk, size);
450 if (sk_msg_full(msg))
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, &copied, flags);
467 sk_psock_put(sk, psock);
468 return copied ? copied : err;
483 static struct proto *tcpv6_prot_saved __read_mostly;
484 static DEFINE_SPINLOCK(tcpv6_prot_lock);
485 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS];
487 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS],
490 prot[TCP_BPF_BASE] = *base;
491 prot[TCP_BPF_BASE].unhash = sock_map_unhash;
492 prot[TCP_BPF_BASE].close = sock_map_close;
493 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg;
494 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read;
496 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE];
497 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg;
498 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage;
501 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops)
503 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) {
504 spin_lock_bh(&tcpv6_prot_lock);
505 if (likely(ops != tcpv6_prot_saved)) {
506 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops);
507 smp_store_release(&tcpv6_prot_saved, ops);
509 spin_unlock_bh(&tcpv6_prot_lock);
513 static int __init tcp_bpf_v4_build_proto(void)
515 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot);
518 late_initcall(tcp_bpf_v4_build_proto);
520 static int tcp_bpf_assert_proto_ops(struct proto *ops)
522 /* In order to avoid retpoline, we make assumptions when we call
523 * into ops if e.g. a psock is not present. Make sure they are
524 * indeed valid assumptions.
526 return ops->recvmsg == tcp_recvmsg &&
527 ops->sendmsg == tcp_sendmsg &&
528 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP;
531 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore)
533 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
534 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE;
537 if (inet_csk_has_ulp(sk)) {
538 /* TLS does not have an unhash proto in SW cases,
539 * but we need to ensure we stop using the sock_map
540 * unhash routine because the associated psock is being
541 * removed. So use the original unhash handler.
543 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash);
544 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space);
546 sk->sk_write_space = psock->saved_write_space;
547 /* Pairs with lockless read in sk_clone_lock() */
548 WRITE_ONCE(sk->sk_prot, psock->sk_proto);
553 if (inet_csk_has_ulp(sk))
556 if (sk->sk_family == AF_INET6) {
557 if (tcp_bpf_assert_proto_ops(psock->sk_proto))
560 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto);
563 /* Pairs with lockless read in sk_clone_lock() */
564 WRITE_ONCE(sk->sk_prot, &tcp_bpf_prots[family][config]);
567 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto);
569 /* If a child got cloned from a listening socket that had tcp_bpf
570 * protocol callbacks installed, we need to restore the callbacks to
571 * the default ones because the child does not inherit the psock state
572 * that tcp_bpf callbacks expect.
574 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk)
576 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4;
577 struct proto *prot = newsk->sk_prot;
579 if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE])
580 newsk->sk_prot = sk->sk_prot_creator;
582 #endif /* CONFIG_BPF_SYSCALL */