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 int peek = flags & MSG_PEEK; 225 struct sk_psock *psock; 226 struct tcp_sock *tcp; 227 int copied = 0; 228 u32 seq; 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 tcp = tcp_sk(sk); 242 seq = tcp->copied_seq; 243 /* We may have received data on the sk_receive_queue pre-accept and 244 * then we can not use read_skb in this context because we haven't 245 * assigned a sk_socket yet so have no link to the ops. The work-around 246 * is to check the sk_receive_queue and in these cases read skbs off 247 * queue again. The read_skb hook is not running at this point because 248 * of lock_sock so we avoid having multiple runners in read_skb. 249 */ 250 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) { 251 tcp_data_ready(sk); 252 /* This handles the ENOMEM errors if we both receive data 253 * pre accept and are already under memory pressure. At least 254 * let user know to retry. 255 */ 256 if (unlikely(!skb_queue_empty(&sk->sk_receive_queue))) { 257 copied = -EAGAIN; 258 goto out; 259 } 260 } 261 262 msg_bytes_ready: 263 copied = sk_msg_recvmsg(sk, psock, msg, len, flags); 264 /* The typical case for EFAULT is the socket was gracefully 265 * shutdown with a FIN pkt. So check here the other case is 266 * some error on copy_page_to_iter which would be unexpected. 267 * On fin return correct return code to zero. 268 */ 269 if (copied == -EFAULT) { 270 bool is_fin = is_next_msg_fin(psock); 271 272 if (is_fin) { 273 copied = 0; 274 seq++; 275 goto out; 276 } 277 } 278 seq += copied; 279 if (!copied) { 280 long timeo; 281 int data; 282 283 if (sock_flag(sk, SOCK_DONE)) 284 goto out; 285 286 if (sk->sk_err) { 287 copied = sock_error(sk); 288 goto out; 289 } 290 291 if (sk->sk_shutdown & RCV_SHUTDOWN) 292 goto out; 293 294 if (sk->sk_state == TCP_CLOSE) { 295 copied = -ENOTCONN; 296 goto out; 297 } 298 299 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 300 if (!timeo) { 301 copied = -EAGAIN; 302 goto out; 303 } 304 305 if (signal_pending(current)) { 306 copied = sock_intr_errno(timeo); 307 goto out; 308 } 309 310 data = tcp_msg_wait_data(sk, psock, timeo); 311 if (data < 0) { 312 copied = data; 313 goto unlock; 314 } 315 if (data && !sk_psock_queue_empty(psock)) 316 goto msg_bytes_ready; 317 copied = -EAGAIN; 318 } 319 out: 320 if (!peek) 321 WRITE_ONCE(tcp->copied_seq, seq); 322 tcp_rcv_space_adjust(sk); 323 if (copied > 0) 324 __tcp_cleanup_rbuf(sk, copied); 325 326 unlock: 327 release_sock(sk); 328 sk_psock_put(sk, psock); 329 return copied; 330 } 331 332 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 333 int flags, int *addr_len) 334 { 335 struct sk_psock *psock; 336 int copied, ret; 337 338 if (unlikely(flags & MSG_ERRQUEUE)) 339 return inet_recv_error(sk, msg, len, addr_len); 340 341 if (!len) 342 return 0; 343 344 psock = sk_psock_get(sk); 345 if (unlikely(!psock)) 346 return tcp_recvmsg(sk, msg, len, flags, addr_len); 347 if (!skb_queue_empty(&sk->sk_receive_queue) && 348 sk_psock_queue_empty(psock)) { 349 sk_psock_put(sk, psock); 350 return tcp_recvmsg(sk, msg, len, flags, addr_len); 351 } 352 lock_sock(sk); 353 msg_bytes_ready: 354 copied = sk_msg_recvmsg(sk, psock, msg, len, flags); 355 if (!copied) { 356 long timeo; 357 int data; 358 359 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 360 data = tcp_msg_wait_data(sk, psock, timeo); 361 if (data < 0) { 362 ret = data; 363 goto unlock; 364 } 365 if (data) { 366 if (!sk_psock_queue_empty(psock)) 367 goto msg_bytes_ready; 368 release_sock(sk); 369 sk_psock_put(sk, psock); 370 return tcp_recvmsg(sk, msg, len, flags, addr_len); 371 } 372 copied = -EAGAIN; 373 } 374 ret = copied; 375 376 unlock: 377 release_sock(sk); 378 sk_psock_put(sk, psock); 379 return ret; 380 } 381 382 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock, 383 struct sk_msg *msg, int *copied, int flags) 384 { 385 bool cork = false, enospc = sk_msg_full(msg), redir_ingress; 386 struct sock *sk_redir; 387 u32 tosend, origsize, sent, delta = 0; 388 u32 eval; 389 int ret; 390 391 more_data: 392 if (psock->eval == __SK_NONE) { 393 /* Track delta in msg size to add/subtract it on SK_DROP from 394 * returned to user copied size. This ensures user doesn't 395 * get a positive return code with msg_cut_data and SK_DROP 396 * verdict. 397 */ 398 delta = msg->sg.size; 399 psock->eval = sk_psock_msg_verdict(sk, psock, msg); 400 delta -= msg->sg.size; 401 } 402 403 if (msg->cork_bytes && 404 msg->cork_bytes > msg->sg.size && !enospc) { 405 psock->cork_bytes = msg->cork_bytes - msg->sg.size; 406 if (!psock->cork) { 407 psock->cork = kzalloc(sizeof(*psock->cork), 408 GFP_ATOMIC | __GFP_NOWARN); 409 if (!psock->cork) 410 return -ENOMEM; 411 } 412 memcpy(psock->cork, msg, sizeof(*msg)); 413 return 0; 414 } 415 416 tosend = msg->sg.size; 417 if (psock->apply_bytes && psock->apply_bytes < tosend) 418 tosend = psock->apply_bytes; 419 eval = __SK_NONE; 420 421 switch (psock->eval) { 422 case __SK_PASS: 423 ret = tcp_bpf_push(sk, msg, tosend, flags, true); 424 if (unlikely(ret)) { 425 *copied -= sk_msg_free(sk, msg); 426 break; 427 } 428 sk_msg_apply_bytes(psock, tosend); 429 break; 430 case __SK_REDIRECT: 431 redir_ingress = psock->redir_ingress; 432 sk_redir = psock->sk_redir; 433 sk_msg_apply_bytes(psock, tosend); 434 if (!psock->apply_bytes) { 435 /* Clean up before releasing the sock lock. */ 436 eval = psock->eval; 437 psock->eval = __SK_NONE; 438 psock->sk_redir = NULL; 439 } 440 if (psock->cork) { 441 cork = true; 442 psock->cork = NULL; 443 } 444 release_sock(sk); 445 446 origsize = msg->sg.size; 447 ret = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress, 448 msg, tosend, flags); 449 sent = origsize - msg->sg.size; 450 451 if (eval == __SK_REDIRECT) 452 sock_put(sk_redir); 453 454 lock_sock(sk); 455 sk_mem_uncharge(sk, sent); 456 if (unlikely(ret < 0)) { 457 int free = sk_msg_free(sk, msg); 458 459 if (!cork) 460 *copied -= free; 461 } 462 if (cork) { 463 sk_msg_free(sk, msg); 464 kfree(msg); 465 msg = NULL; 466 ret = 0; 467 } 468 break; 469 case __SK_DROP: 470 default: 471 sk_msg_free(sk, msg); 472 sk_msg_apply_bytes(psock, tosend); 473 *copied -= (tosend + delta); 474 return -EACCES; 475 } 476 477 if (likely(!ret)) { 478 if (!psock->apply_bytes) { 479 psock->eval = __SK_NONE; 480 if (psock->sk_redir) { 481 sock_put(psock->sk_redir); 482 psock->sk_redir = NULL; 483 } 484 } 485 if (msg && 486 msg->sg.data[msg->sg.start].page_link && 487 msg->sg.data[msg->sg.start].length) 488 goto more_data; 489 } 490 return ret; 491 } 492 493 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 494 { 495 struct sk_msg tmp, *msg_tx = NULL; 496 int copied = 0, err = 0; 497 struct sk_psock *psock; 498 long timeo; 499 int flags; 500 501 /* Don't let internal flags through */ 502 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED); 503 flags |= MSG_NO_SHARED_FRAGS; 504 505 psock = sk_psock_get(sk); 506 if (unlikely(!psock)) 507 return tcp_sendmsg(sk, msg, size); 508 509 lock_sock(sk); 510 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 511 while (msg_data_left(msg)) { 512 bool enospc = false; 513 u32 copy, osize; 514 515 if (sk->sk_err) { 516 err = -sk->sk_err; 517 goto out_err; 518 } 519 520 copy = msg_data_left(msg); 521 if (!sk_stream_memory_free(sk)) 522 goto wait_for_sndbuf; 523 if (psock->cork) { 524 msg_tx = psock->cork; 525 } else { 526 msg_tx = &tmp; 527 sk_msg_init(msg_tx); 528 } 529 530 osize = msg_tx->sg.size; 531 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1); 532 if (err) { 533 if (err != -ENOSPC) 534 goto wait_for_memory; 535 enospc = true; 536 copy = msg_tx->sg.size - osize; 537 } 538 539 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx, 540 copy); 541 if (err < 0) { 542 sk_msg_trim(sk, msg_tx, osize); 543 goto out_err; 544 } 545 546 copied += copy; 547 if (psock->cork_bytes) { 548 if (size > psock->cork_bytes) 549 psock->cork_bytes = 0; 550 else 551 psock->cork_bytes -= size; 552 if (psock->cork_bytes && !enospc) 553 goto out_err; 554 /* All cork bytes are accounted, rerun the prog. */ 555 psock->eval = __SK_NONE; 556 psock->cork_bytes = 0; 557 } 558 559 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags); 560 if (unlikely(err < 0)) 561 goto out_err; 562 continue; 563 wait_for_sndbuf: 564 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 565 wait_for_memory: 566 err = sk_stream_wait_memory(sk, &timeo); 567 if (err) { 568 if (msg_tx && msg_tx != psock->cork) 569 sk_msg_free(sk, msg_tx); 570 goto out_err; 571 } 572 } 573 out_err: 574 if (err < 0) 575 err = sk_stream_error(sk, msg->msg_flags, err); 576 release_sock(sk); 577 sk_psock_put(sk, psock); 578 return copied > 0 ? copied : err; 579 } 580 581 enum { 582 TCP_BPF_IPV4, 583 TCP_BPF_IPV6, 584 TCP_BPF_NUM_PROTS, 585 }; 586 587 enum { 588 TCP_BPF_BASE, 589 TCP_BPF_TX, 590 TCP_BPF_RX, 591 TCP_BPF_TXRX, 592 TCP_BPF_NUM_CFGS, 593 }; 594 595 static struct proto *tcpv6_prot_saved __read_mostly; 596 static DEFINE_SPINLOCK(tcpv6_prot_lock); 597 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS]; 598 599 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS], 600 struct proto *base) 601 { 602 prot[TCP_BPF_BASE] = *base; 603 prot[TCP_BPF_BASE].destroy = sock_map_destroy; 604 prot[TCP_BPF_BASE].close = sock_map_close; 605 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg; 606 prot[TCP_BPF_BASE].sock_is_readable = sk_msg_is_readable; 607 608 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE]; 609 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg; 610 611 prot[TCP_BPF_RX] = prot[TCP_BPF_BASE]; 612 prot[TCP_BPF_RX].recvmsg = tcp_bpf_recvmsg_parser; 613 614 prot[TCP_BPF_TXRX] = prot[TCP_BPF_TX]; 615 prot[TCP_BPF_TXRX].recvmsg = tcp_bpf_recvmsg_parser; 616 } 617 618 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops) 619 { 620 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) { 621 spin_lock_bh(&tcpv6_prot_lock); 622 if (likely(ops != tcpv6_prot_saved)) { 623 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops); 624 smp_store_release(&tcpv6_prot_saved, ops); 625 } 626 spin_unlock_bh(&tcpv6_prot_lock); 627 } 628 } 629 630 static int __init tcp_bpf_v4_build_proto(void) 631 { 632 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot); 633 return 0; 634 } 635 late_initcall(tcp_bpf_v4_build_proto); 636 637 static int tcp_bpf_assert_proto_ops(struct proto *ops) 638 { 639 /* In order to avoid retpoline, we make assumptions when we call 640 * into ops if e.g. a psock is not present. Make sure they are 641 * indeed valid assumptions. 642 */ 643 return ops->recvmsg == tcp_recvmsg && 644 ops->sendmsg == tcp_sendmsg ? 0 : -ENOTSUPP; 645 } 646 647 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore) 648 { 649 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 650 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 651 652 if (psock->progs.stream_verdict || psock->progs.skb_verdict) { 653 config = (config == TCP_BPF_TX) ? TCP_BPF_TXRX : TCP_BPF_RX; 654 } 655 656 if (restore) { 657 if (inet_csk_has_ulp(sk)) { 658 /* TLS does not have an unhash proto in SW cases, 659 * but we need to ensure we stop using the sock_map 660 * unhash routine because the associated psock is being 661 * removed. So use the original unhash handler. 662 */ 663 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash); 664 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space); 665 } else { 666 sk->sk_write_space = psock->saved_write_space; 667 /* Pairs with lockless read in sk_clone_lock() */ 668 sock_replace_proto(sk, psock->sk_proto); 669 } 670 return 0; 671 } 672 673 if (sk->sk_family == AF_INET6) { 674 if (tcp_bpf_assert_proto_ops(psock->sk_proto)) 675 return -EINVAL; 676 677 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto); 678 } 679 680 /* Pairs with lockless read in sk_clone_lock() */ 681 sock_replace_proto(sk, &tcp_bpf_prots[family][config]); 682 return 0; 683 } 684 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto); 685 686 /* If a child got cloned from a listening socket that had tcp_bpf 687 * protocol callbacks installed, we need to restore the callbacks to 688 * the default ones because the child does not inherit the psock state 689 * that tcp_bpf callbacks expect. 690 */ 691 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk) 692 { 693 struct proto *prot = newsk->sk_prot; 694 695 if (is_insidevar(prot, tcp_bpf_prots)) 696 newsk->sk_prot = sk->sk_prot_creator; 697 } 698 #endif /* CONFIG_BPF_SYSCALL */ 699