1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */ 3 4 #include <linux/skmsg.h> 5 #include <linux/filter.h> 6 #include <linux/bpf.h> 7 #include <linux/init.h> 8 #include <linux/wait.h> 9 #include <linux/util_macros.h> 10 11 #include <net/inet_common.h> 12 #include <net/tls.h> 13 14 void tcp_eat_skb(struct sock *sk, struct sk_buff *skb) 15 { 16 struct tcp_sock *tcp; 17 int copied; 18 19 if (!skb || !skb->len || !sk_is_tcp(sk)) 20 return; 21 22 if (skb_bpf_strparser(skb)) 23 return; 24 25 tcp = tcp_sk(sk); 26 copied = tcp->copied_seq + skb->len; 27 WRITE_ONCE(tcp->copied_seq, copied); 28 tcp_rcv_space_adjust(sk); 29 __tcp_cleanup_rbuf(sk, skb->len); 30 } 31 32 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock, 33 struct sk_msg *msg, u32 apply_bytes, int flags) 34 { 35 bool apply = apply_bytes; 36 struct scatterlist *sge; 37 u32 size, copied = 0; 38 struct sk_msg *tmp; 39 int i, ret = 0; 40 41 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL); 42 if (unlikely(!tmp)) 43 return -ENOMEM; 44 45 lock_sock(sk); 46 tmp->sg.start = msg->sg.start; 47 i = msg->sg.start; 48 do { 49 sge = sk_msg_elem(msg, i); 50 size = (apply && apply_bytes < sge->length) ? 51 apply_bytes : sge->length; 52 if (!sk_wmem_schedule(sk, size)) { 53 if (!copied) 54 ret = -ENOMEM; 55 break; 56 } 57 58 sk_mem_charge(sk, size); 59 sk_msg_xfer(tmp, msg, i, size); 60 copied += size; 61 if (sge->length) 62 get_page(sk_msg_page(tmp, i)); 63 sk_msg_iter_var_next(i); 64 tmp->sg.end = i; 65 if (apply) { 66 apply_bytes -= size; 67 if (!apply_bytes) { 68 if (sge->length) 69 sk_msg_iter_var_prev(i); 70 break; 71 } 72 } 73 } while (i != msg->sg.end); 74 75 if (!ret) { 76 msg->sg.start = i; 77 sk_psock_queue_msg(psock, tmp); 78 sk_psock_data_ready(sk, psock); 79 } else { 80 sk_msg_free(sk, tmp); 81 kfree(tmp); 82 } 83 84 release_sock(sk); 85 return ret; 86 } 87 88 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes, 89 int flags, bool uncharge) 90 { 91 struct msghdr msghdr = {}; 92 bool apply = apply_bytes; 93 struct scatterlist *sge; 94 struct page *page; 95 int size, ret = 0; 96 u32 off; 97 98 while (1) { 99 struct bio_vec bvec; 100 bool has_tx_ulp; 101 102 sge = sk_msg_elem(msg, msg->sg.start); 103 size = (apply && apply_bytes < sge->length) ? 104 apply_bytes : sge->length; 105 off = sge->offset; 106 page = sg_page(sge); 107 108 tcp_rate_check_app_limited(sk); 109 retry: 110 msghdr.msg_flags = flags | MSG_SPLICE_PAGES; 111 has_tx_ulp = tls_sw_has_ctx_tx(sk); 112 if (has_tx_ulp) 113 msghdr.msg_flags |= MSG_SENDPAGE_NOPOLICY; 114 115 if (size < sge->length && msg->sg.start != msg->sg.end) 116 msghdr.msg_flags |= MSG_MORE; 117 118 bvec_set_page(&bvec, page, size, off); 119 iov_iter_bvec(&msghdr.msg_iter, ITER_SOURCE, &bvec, 1, size); 120 ret = tcp_sendmsg_locked(sk, &msghdr, size); 121 if (ret <= 0) 122 return ret; 123 124 if (apply) 125 apply_bytes -= ret; 126 msg->sg.size -= ret; 127 sge->offset += ret; 128 sge->length -= ret; 129 if (uncharge) 130 sk_mem_uncharge(sk, ret); 131 if (ret != size) { 132 size -= ret; 133 off += ret; 134 goto retry; 135 } 136 if (!sge->length) { 137 put_page(page); 138 sk_msg_iter_next(msg, start); 139 sg_init_table(sge, 1); 140 if (msg->sg.start == msg->sg.end) 141 break; 142 } 143 if (apply && !apply_bytes) 144 break; 145 } 146 147 return 0; 148 } 149 150 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg, 151 u32 apply_bytes, int flags, bool uncharge) 152 { 153 int ret; 154 155 lock_sock(sk); 156 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge); 157 release_sock(sk); 158 return ret; 159 } 160 161 int tcp_bpf_sendmsg_redir(struct sock *sk, bool ingress, 162 struct sk_msg *msg, u32 bytes, int flags) 163 { 164 struct sk_psock *psock = sk_psock_get(sk); 165 int ret; 166 167 if (unlikely(!psock)) 168 return -EPIPE; 169 170 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) : 171 tcp_bpf_push_locked(sk, msg, bytes, flags, false); 172 sk_psock_put(sk, psock); 173 return ret; 174 } 175 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir); 176 177 #ifdef CONFIG_BPF_SYSCALL 178 static int tcp_msg_wait_data(struct sock *sk, struct sk_psock *psock, 179 long timeo) 180 { 181 DEFINE_WAIT_FUNC(wait, woken_wake_function); 182 int ret = 0; 183 184 if (sk->sk_shutdown & RCV_SHUTDOWN) 185 return 1; 186 187 if (!timeo) 188 return ret; 189 190 add_wait_queue(sk_sleep(sk), &wait); 191 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk); 192 ret = sk_wait_event(sk, &timeo, 193 !list_empty(&psock->ingress_msg) || 194 !skb_queue_empty_lockless(&sk->sk_receive_queue), &wait); 195 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk); 196 remove_wait_queue(sk_sleep(sk), &wait); 197 return ret; 198 } 199 200 static bool is_next_msg_fin(struct sk_psock *psock) 201 { 202 struct scatterlist *sge; 203 struct sk_msg *msg_rx; 204 int i; 205 206 msg_rx = sk_psock_peek_msg(psock); 207 i = msg_rx->sg.start; 208 sge = sk_msg_elem(msg_rx, i); 209 if (!sge->length) { 210 struct sk_buff *skb = msg_rx->skb; 211 212 if (skb && TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 213 return true; 214 } 215 return false; 216 } 217 218 static int tcp_bpf_recvmsg_parser(struct sock *sk, 219 struct msghdr *msg, 220 size_t len, 221 int flags, 222 int *addr_len) 223 { 224 struct tcp_sock *tcp = tcp_sk(sk); 225 int peek = flags & MSG_PEEK; 226 u32 seq = tcp->copied_seq; 227 struct sk_psock *psock; 228 int copied = 0; 229 230 if (unlikely(flags & MSG_ERRQUEUE)) 231 return inet_recv_error(sk, msg, len, addr_len); 232 233 if (!len) 234 return 0; 235 236 psock = sk_psock_get(sk); 237 if (unlikely(!psock)) 238 return tcp_recvmsg(sk, msg, len, flags, addr_len); 239 240 lock_sock(sk); 241 242 /* We may have received data on the sk_receive_queue pre-accept and 243 * then we can not use read_skb in this context because we haven't 244 * assigned a sk_socket yet so have no link to the ops. The work-around 245 * is to check the sk_receive_queue and in these cases read skbs off 246 * queue again. The read_skb hook is not running at this point because 247 * of lock_sock so we avoid having multiple runners in read_skb. 248 */ 249 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) { 250 tcp_data_ready(sk); 251 /* This handles the ENOMEM errors if we both receive data 252 * pre accept and are already under memory pressure. At least 253 * let user know to retry. 254 */ 255 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) { 256 copied = -EAGAIN; 257 goto out; 258 } 259 } 260 261 msg_bytes_ready: 262 copied = sk_msg_recvmsg(sk, psock, msg, len, flags); 263 /* The typical case for EFAULT is the socket was gracefully 264 * shutdown with a FIN pkt. So check here the other case is 265 * some error on copy_page_to_iter which would be unexpected. 266 * On fin return correct return code to zero. 267 */ 268 if (copied == -EFAULT) { 269 bool is_fin = is_next_msg_fin(psock); 270 271 if (is_fin) { 272 copied = 0; 273 seq++; 274 goto out; 275 } 276 } 277 seq += copied; 278 if (!copied) { 279 long timeo; 280 int data; 281 282 if (sock_flag(sk, SOCK_DONE)) 283 goto out; 284 285 if (sk->sk_err) { 286 copied = sock_error(sk); 287 goto out; 288 } 289 290 if (sk->sk_shutdown & RCV_SHUTDOWN) 291 goto out; 292 293 if (sk->sk_state == TCP_CLOSE) { 294 copied = -ENOTCONN; 295 goto out; 296 } 297 298 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 299 if (!timeo) { 300 copied = -EAGAIN; 301 goto out; 302 } 303 304 if (signal_pending(current)) { 305 copied = sock_intr_errno(timeo); 306 goto out; 307 } 308 309 data = tcp_msg_wait_data(sk, psock, timeo); 310 if (data && !sk_psock_queue_empty(psock)) 311 goto msg_bytes_ready; 312 copied = -EAGAIN; 313 } 314 out: 315 if (!peek) 316 WRITE_ONCE(tcp->copied_seq, seq); 317 tcp_rcv_space_adjust(sk); 318 if (copied > 0) 319 __tcp_cleanup_rbuf(sk, copied); 320 release_sock(sk); 321 sk_psock_put(sk, psock); 322 return copied; 323 } 324 325 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 326 int flags, int *addr_len) 327 { 328 struct sk_psock *psock; 329 int copied, ret; 330 331 if (unlikely(flags & MSG_ERRQUEUE)) 332 return inet_recv_error(sk, msg, len, addr_len); 333 334 if (!len) 335 return 0; 336 337 psock = sk_psock_get(sk); 338 if (unlikely(!psock)) 339 return tcp_recvmsg(sk, msg, len, flags, addr_len); 340 if (!skb_queue_empty(&sk->sk_receive_queue) && 341 sk_psock_queue_empty(psock)) { 342 sk_psock_put(sk, psock); 343 return tcp_recvmsg(sk, msg, len, flags, addr_len); 344 } 345 lock_sock(sk); 346 msg_bytes_ready: 347 copied = sk_msg_recvmsg(sk, psock, msg, len, flags); 348 if (!copied) { 349 long timeo; 350 int data; 351 352 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 353 data = tcp_msg_wait_data(sk, psock, timeo); 354 if (data) { 355 if (!sk_psock_queue_empty(psock)) 356 goto msg_bytes_ready; 357 release_sock(sk); 358 sk_psock_put(sk, psock); 359 return tcp_recvmsg(sk, msg, len, flags, addr_len); 360 } 361 copied = -EAGAIN; 362 } 363 ret = copied; 364 release_sock(sk); 365 sk_psock_put(sk, psock); 366 return ret; 367 } 368 369 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock, 370 struct sk_msg *msg, int *copied, int flags) 371 { 372 bool cork = false, enospc = sk_msg_full(msg), redir_ingress; 373 struct sock *sk_redir; 374 u32 tosend, origsize, sent, delta = 0; 375 u32 eval; 376 int ret; 377 378 more_data: 379 if (psock->eval == __SK_NONE) { 380 /* Track delta in msg size to add/subtract it on SK_DROP from 381 * returned to user copied size. This ensures user doesn't 382 * get a positive return code with msg_cut_data and SK_DROP 383 * verdict. 384 */ 385 delta = msg->sg.size; 386 psock->eval = sk_psock_msg_verdict(sk, psock, msg); 387 delta -= msg->sg.size; 388 } 389 390 if (msg->cork_bytes && 391 msg->cork_bytes > msg->sg.size && !enospc) { 392 psock->cork_bytes = msg->cork_bytes - msg->sg.size; 393 if (!psock->cork) { 394 psock->cork = kzalloc(sizeof(*psock->cork), 395 GFP_ATOMIC | __GFP_NOWARN); 396 if (!psock->cork) 397 return -ENOMEM; 398 } 399 memcpy(psock->cork, msg, sizeof(*msg)); 400 return 0; 401 } 402 403 tosend = msg->sg.size; 404 if (psock->apply_bytes && psock->apply_bytes < tosend) 405 tosend = psock->apply_bytes; 406 eval = __SK_NONE; 407 408 switch (psock->eval) { 409 case __SK_PASS: 410 ret = tcp_bpf_push(sk, msg, tosend, flags, true); 411 if (unlikely(ret)) { 412 *copied -= sk_msg_free(sk, msg); 413 break; 414 } 415 sk_msg_apply_bytes(psock, tosend); 416 break; 417 case __SK_REDIRECT: 418 redir_ingress = psock->redir_ingress; 419 sk_redir = psock->sk_redir; 420 sk_msg_apply_bytes(psock, tosend); 421 if (!psock->apply_bytes) { 422 /* Clean up before releasing the sock lock. */ 423 eval = psock->eval; 424 psock->eval = __SK_NONE; 425 psock->sk_redir = NULL; 426 } 427 if (psock->cork) { 428 cork = true; 429 psock->cork = NULL; 430 } 431 sk_msg_return(sk, msg, tosend); 432 release_sock(sk); 433 434 origsize = msg->sg.size; 435 ret = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress, 436 msg, tosend, flags); 437 sent = origsize - msg->sg.size; 438 439 if (eval == __SK_REDIRECT) 440 sock_put(sk_redir); 441 442 lock_sock(sk); 443 if (unlikely(ret < 0)) { 444 int free = sk_msg_free_nocharge(sk, msg); 445 446 if (!cork) 447 *copied -= free; 448 } 449 if (cork) { 450 sk_msg_free(sk, msg); 451 kfree(msg); 452 msg = NULL; 453 ret = 0; 454 } 455 break; 456 case __SK_DROP: 457 default: 458 sk_msg_free_partial(sk, msg, tosend); 459 sk_msg_apply_bytes(psock, tosend); 460 *copied -= (tosend + delta); 461 return -EACCES; 462 } 463 464 if (likely(!ret)) { 465 if (!psock->apply_bytes) { 466 psock->eval = __SK_NONE; 467 if (psock->sk_redir) { 468 sock_put(psock->sk_redir); 469 psock->sk_redir = NULL; 470 } 471 } 472 if (msg && 473 msg->sg.data[msg->sg.start].page_link && 474 msg->sg.data[msg->sg.start].length) { 475 if (eval == __SK_REDIRECT) 476 sk_mem_charge(sk, tosend - sent); 477 goto more_data; 478 } 479 } 480 return ret; 481 } 482 483 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 484 { 485 struct sk_msg tmp, *msg_tx = NULL; 486 int copied = 0, err = 0; 487 struct sk_psock *psock; 488 long timeo; 489 int flags; 490 491 /* Don't let internal flags through */ 492 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED); 493 flags |= MSG_NO_SHARED_FRAGS; 494 495 psock = sk_psock_get(sk); 496 if (unlikely(!psock)) 497 return tcp_sendmsg(sk, msg, size); 498 499 lock_sock(sk); 500 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 501 while (msg_data_left(msg)) { 502 bool enospc = false; 503 u32 copy, osize; 504 505 if (sk->sk_err) { 506 err = -sk->sk_err; 507 goto out_err; 508 } 509 510 copy = msg_data_left(msg); 511 if (!sk_stream_memory_free(sk)) 512 goto wait_for_sndbuf; 513 if (psock->cork) { 514 msg_tx = psock->cork; 515 } else { 516 msg_tx = &tmp; 517 sk_msg_init(msg_tx); 518 } 519 520 osize = msg_tx->sg.size; 521 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1); 522 if (err) { 523 if (err != -ENOSPC) 524 goto wait_for_memory; 525 enospc = true; 526 copy = msg_tx->sg.size - osize; 527 } 528 529 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx, 530 copy); 531 if (err < 0) { 532 sk_msg_trim(sk, msg_tx, osize); 533 goto out_err; 534 } 535 536 copied += copy; 537 if (psock->cork_bytes) { 538 if (size > psock->cork_bytes) 539 psock->cork_bytes = 0; 540 else 541 psock->cork_bytes -= size; 542 if (psock->cork_bytes && !enospc) 543 goto out_err; 544 /* All cork bytes are accounted, rerun the prog. */ 545 psock->eval = __SK_NONE; 546 psock->cork_bytes = 0; 547 } 548 549 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags); 550 if (unlikely(err < 0)) 551 goto out_err; 552 continue; 553 wait_for_sndbuf: 554 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 555 wait_for_memory: 556 err = sk_stream_wait_memory(sk, &timeo); 557 if (err) { 558 if (msg_tx && msg_tx != psock->cork) 559 sk_msg_free(sk, msg_tx); 560 goto out_err; 561 } 562 } 563 out_err: 564 if (err < 0) 565 err = sk_stream_error(sk, msg->msg_flags, err); 566 release_sock(sk); 567 sk_psock_put(sk, psock); 568 return copied ? copied : err; 569 } 570 571 enum { 572 TCP_BPF_IPV4, 573 TCP_BPF_IPV6, 574 TCP_BPF_NUM_PROTS, 575 }; 576 577 enum { 578 TCP_BPF_BASE, 579 TCP_BPF_TX, 580 TCP_BPF_RX, 581 TCP_BPF_TXRX, 582 TCP_BPF_NUM_CFGS, 583 }; 584 585 static struct proto *tcpv6_prot_saved __read_mostly; 586 static DEFINE_SPINLOCK(tcpv6_prot_lock); 587 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS]; 588 589 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS], 590 struct proto *base) 591 { 592 prot[TCP_BPF_BASE] = *base; 593 prot[TCP_BPF_BASE].destroy = sock_map_destroy; 594 prot[TCP_BPF_BASE].close = sock_map_close; 595 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg; 596 prot[TCP_BPF_BASE].sock_is_readable = sk_msg_is_readable; 597 598 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE]; 599 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg; 600 601 prot[TCP_BPF_RX] = prot[TCP_BPF_BASE]; 602 prot[TCP_BPF_RX].recvmsg = tcp_bpf_recvmsg_parser; 603 604 prot[TCP_BPF_TXRX] = prot[TCP_BPF_TX]; 605 prot[TCP_BPF_TXRX].recvmsg = tcp_bpf_recvmsg_parser; 606 } 607 608 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops) 609 { 610 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) { 611 spin_lock_bh(&tcpv6_prot_lock); 612 if (likely(ops != tcpv6_prot_saved)) { 613 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops); 614 smp_store_release(&tcpv6_prot_saved, ops); 615 } 616 spin_unlock_bh(&tcpv6_prot_lock); 617 } 618 } 619 620 static int __init tcp_bpf_v4_build_proto(void) 621 { 622 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot); 623 return 0; 624 } 625 late_initcall(tcp_bpf_v4_build_proto); 626 627 static int tcp_bpf_assert_proto_ops(struct proto *ops) 628 { 629 /* In order to avoid retpoline, we make assumptions when we call 630 * into ops if e.g. a psock is not present. Make sure they are 631 * indeed valid assumptions. 632 */ 633 return ops->recvmsg == tcp_recvmsg && 634 ops->sendmsg == tcp_sendmsg ? 0 : -ENOTSUPP; 635 } 636 637 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore) 638 { 639 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 640 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 641 642 if (psock->progs.stream_verdict || psock->progs.skb_verdict) { 643 config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX; 644 } 645 646 if (restore) { 647 if (inet_csk_has_ulp(sk)) { 648 /* TLS does not have an unhash proto in SW cases, 649 * but we need to ensure we stop using the sock_map 650 * unhash routine because the associated psock is being 651 * removed. So use the original unhash handler. 652 */ 653 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash); 654 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space); 655 } else { 656 sk->sk_write_space = psock->saved_write_space; 657 /* Pairs with lockless read in sk_clone_lock() */ 658 sock_replace_proto(sk, psock->sk_proto); 659 } 660 return 0; 661 } 662 663 if (sk->sk_family == AF_INET6) { 664 if (tcp_bpf_assert_proto_ops(psock->sk_proto)) 665 return -EINVAL; 666 667 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto); 668 } 669 670 /* Pairs with lockless read in sk_clone_lock() */ 671 sock_replace_proto(sk, &tcp_bpf_prots[family][config]); 672 return 0; 673 } 674 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto); 675 676 /* If a child got cloned from a listening socket that had tcp_bpf 677 * protocol callbacks installed, we need to restore the callbacks to 678 * the default ones because the child does not inherit the psock state 679 * that tcp_bpf callbacks expect. 680 */ 681 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk) 682 { 683 struct proto *prot = newsk->sk_prot; 684 685 if (is_insidevar(prot, tcp_bpf_prots)) 686 newsk->sk_prot = sk->sk_prot_creator; 687 } 688 #endif /* CONFIG_BPF_SYSCALL */ 689