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 10 #include <net/inet_common.h> 11 #include <net/tls.h> 12 13 static bool tcp_bpf_stream_read(const struct sock *sk) 14 { 15 struct sk_psock *psock; 16 bool empty = true; 17 18 rcu_read_lock(); 19 psock = sk_psock(sk); 20 if (likely(psock)) 21 empty = list_empty(&psock->ingress_msg); 22 rcu_read_unlock(); 23 return !empty; 24 } 25 26 static int tcp_bpf_wait_data(struct sock *sk, struct sk_psock *psock, 27 int flags, long timeo, int *err) 28 { 29 DEFINE_WAIT_FUNC(wait, woken_wake_function); 30 int ret = 0; 31 32 if (!timeo) 33 return ret; 34 35 add_wait_queue(sk_sleep(sk), &wait); 36 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk); 37 ret = sk_wait_event(sk, &timeo, 38 !list_empty(&psock->ingress_msg) || 39 !skb_queue_empty(&sk->sk_receive_queue), &wait); 40 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk); 41 remove_wait_queue(sk_sleep(sk), &wait); 42 return ret; 43 } 44 45 int __tcp_bpf_recvmsg(struct sock *sk, struct sk_psock *psock, 46 struct msghdr *msg, int len, int flags) 47 { 48 struct iov_iter *iter = &msg->msg_iter; 49 int peek = flags & MSG_PEEK; 50 int i, ret, copied = 0; 51 struct sk_msg *msg_rx; 52 53 msg_rx = list_first_entry_or_null(&psock->ingress_msg, 54 struct sk_msg, list); 55 56 while (copied != len) { 57 struct scatterlist *sge; 58 59 if (unlikely(!msg_rx)) 60 break; 61 62 i = msg_rx->sg.start; 63 do { 64 struct page *page; 65 int copy; 66 67 sge = sk_msg_elem(msg_rx, i); 68 copy = sge->length; 69 page = sg_page(sge); 70 if (copied + copy > len) 71 copy = len - copied; 72 ret = copy_page_to_iter(page, sge->offset, copy, iter); 73 if (ret != copy) { 74 msg_rx->sg.start = i; 75 return -EFAULT; 76 } 77 78 copied += copy; 79 if (likely(!peek)) { 80 sge->offset += copy; 81 sge->length -= copy; 82 sk_mem_uncharge(sk, copy); 83 msg_rx->sg.size -= copy; 84 85 if (!sge->length) { 86 sk_msg_iter_var_next(i); 87 if (!msg_rx->skb) 88 put_page(page); 89 } 90 } else { 91 sk_msg_iter_var_next(i); 92 } 93 94 if (copied == len) 95 break; 96 } while (i != msg_rx->sg.end); 97 98 if (unlikely(peek)) { 99 msg_rx = list_next_entry(msg_rx, list); 100 continue; 101 } 102 103 msg_rx->sg.start = i; 104 if (!sge->length && msg_rx->sg.start == msg_rx->sg.end) { 105 list_del(&msg_rx->list); 106 if (msg_rx->skb) 107 consume_skb(msg_rx->skb); 108 kfree(msg_rx); 109 } 110 msg_rx = list_first_entry_or_null(&psock->ingress_msg, 111 struct sk_msg, list); 112 } 113 114 return copied; 115 } 116 EXPORT_SYMBOL_GPL(__tcp_bpf_recvmsg); 117 118 int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 119 int nonblock, int flags, int *addr_len) 120 { 121 struct sk_psock *psock; 122 int copied, ret; 123 124 if (unlikely(flags & MSG_ERRQUEUE)) 125 return inet_recv_error(sk, msg, len, addr_len); 126 if (!skb_queue_empty(&sk->sk_receive_queue)) 127 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 128 129 psock = sk_psock_get(sk); 130 if (unlikely(!psock)) 131 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 132 lock_sock(sk); 133 msg_bytes_ready: 134 copied = __tcp_bpf_recvmsg(sk, psock, msg, len, flags); 135 if (!copied) { 136 int data, err = 0; 137 long timeo; 138 139 timeo = sock_rcvtimeo(sk, nonblock); 140 data = tcp_bpf_wait_data(sk, psock, flags, timeo, &err); 141 if (data) { 142 if (skb_queue_empty(&sk->sk_receive_queue)) 143 goto msg_bytes_ready; 144 release_sock(sk); 145 sk_psock_put(sk, psock); 146 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 147 } 148 if (err) { 149 ret = err; 150 goto out; 151 } 152 copied = -EAGAIN; 153 } 154 ret = copied; 155 out: 156 release_sock(sk); 157 sk_psock_put(sk, psock); 158 return ret; 159 } 160 161 static int bpf_tcp_ingress(struct sock *sk, struct sk_psock *psock, 162 struct sk_msg *msg, u32 apply_bytes, int flags) 163 { 164 bool apply = apply_bytes; 165 struct scatterlist *sge; 166 u32 size, copied = 0; 167 struct sk_msg *tmp; 168 int i, ret = 0; 169 170 tmp = kzalloc(sizeof(*tmp), __GFP_NOWARN | GFP_KERNEL); 171 if (unlikely(!tmp)) 172 return -ENOMEM; 173 174 lock_sock(sk); 175 tmp->sg.start = msg->sg.start; 176 i = msg->sg.start; 177 do { 178 sge = sk_msg_elem(msg, i); 179 size = (apply && apply_bytes < sge->length) ? 180 apply_bytes : sge->length; 181 if (!sk_wmem_schedule(sk, size)) { 182 if (!copied) 183 ret = -ENOMEM; 184 break; 185 } 186 187 sk_mem_charge(sk, size); 188 sk_msg_xfer(tmp, msg, i, size); 189 copied += size; 190 if (sge->length) 191 get_page(sk_msg_page(tmp, i)); 192 sk_msg_iter_var_next(i); 193 tmp->sg.end = i; 194 if (apply) { 195 apply_bytes -= size; 196 if (!apply_bytes) 197 break; 198 } 199 } while (i != msg->sg.end); 200 201 if (!ret) { 202 msg->sg.start = i; 203 msg->sg.size -= apply_bytes; 204 sk_psock_queue_msg(psock, tmp); 205 sk_psock_data_ready(sk, psock); 206 } else { 207 sk_msg_free(sk, tmp); 208 kfree(tmp); 209 } 210 211 release_sock(sk); 212 return ret; 213 } 214 215 static int tcp_bpf_push(struct sock *sk, struct sk_msg *msg, u32 apply_bytes, 216 int flags, bool uncharge) 217 { 218 bool apply = apply_bytes; 219 struct scatterlist *sge; 220 struct page *page; 221 int size, ret = 0; 222 u32 off; 223 224 while (1) { 225 bool has_tx_ulp; 226 227 sge = sk_msg_elem(msg, msg->sg.start); 228 size = (apply && apply_bytes < sge->length) ? 229 apply_bytes : sge->length; 230 off = sge->offset; 231 page = sg_page(sge); 232 233 tcp_rate_check_app_limited(sk); 234 retry: 235 has_tx_ulp = tls_sw_has_ctx_tx(sk); 236 if (has_tx_ulp) { 237 flags |= MSG_SENDPAGE_NOPOLICY; 238 ret = kernel_sendpage_locked(sk, 239 page, off, size, flags); 240 } else { 241 ret = do_tcp_sendpages(sk, page, off, size, flags); 242 } 243 244 if (ret <= 0) 245 return ret; 246 if (apply) 247 apply_bytes -= ret; 248 msg->sg.size -= ret; 249 sge->offset += ret; 250 sge->length -= ret; 251 if (uncharge) 252 sk_mem_uncharge(sk, ret); 253 if (ret != size) { 254 size -= ret; 255 off += ret; 256 goto retry; 257 } 258 if (!sge->length) { 259 put_page(page); 260 sk_msg_iter_next(msg, start); 261 sg_init_table(sge, 1); 262 if (msg->sg.start == msg->sg.end) 263 break; 264 } 265 if (apply && !apply_bytes) 266 break; 267 } 268 269 return 0; 270 } 271 272 static int tcp_bpf_push_locked(struct sock *sk, struct sk_msg *msg, 273 u32 apply_bytes, int flags, bool uncharge) 274 { 275 int ret; 276 277 lock_sock(sk); 278 ret = tcp_bpf_push(sk, msg, apply_bytes, flags, uncharge); 279 release_sock(sk); 280 return ret; 281 } 282 283 int tcp_bpf_sendmsg_redir(struct sock *sk, struct sk_msg *msg, 284 u32 bytes, int flags) 285 { 286 bool ingress = sk_msg_to_ingress(msg); 287 struct sk_psock *psock = sk_psock_get(sk); 288 int ret; 289 290 if (unlikely(!psock)) { 291 sk_msg_free(sk, msg); 292 return 0; 293 } 294 ret = ingress ? bpf_tcp_ingress(sk, psock, msg, bytes, flags) : 295 tcp_bpf_push_locked(sk, msg, bytes, flags, false); 296 sk_psock_put(sk, psock); 297 return ret; 298 } 299 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir); 300 301 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock, 302 struct sk_msg *msg, int *copied, int flags) 303 { 304 bool cork = false, enospc = msg->sg.start == msg->sg.end; 305 struct sock *sk_redir; 306 u32 tosend, delta = 0; 307 int ret; 308 309 more_data: 310 if (psock->eval == __SK_NONE) { 311 /* Track delta in msg size to add/subtract it on SK_DROP from 312 * returned to user copied size. This ensures user doesn't 313 * get a positive return code with msg_cut_data and SK_DROP 314 * verdict. 315 */ 316 delta = msg->sg.size; 317 psock->eval = sk_psock_msg_verdict(sk, psock, msg); 318 if (msg->sg.size < delta) 319 delta -= msg->sg.size; 320 else 321 delta = 0; 322 } 323 324 if (msg->cork_bytes && 325 msg->cork_bytes > msg->sg.size && !enospc) { 326 psock->cork_bytes = msg->cork_bytes - msg->sg.size; 327 if (!psock->cork) { 328 psock->cork = kzalloc(sizeof(*psock->cork), 329 GFP_ATOMIC | __GFP_NOWARN); 330 if (!psock->cork) 331 return -ENOMEM; 332 } 333 memcpy(psock->cork, msg, sizeof(*msg)); 334 return 0; 335 } 336 337 tosend = msg->sg.size; 338 if (psock->apply_bytes && psock->apply_bytes < tosend) 339 tosend = psock->apply_bytes; 340 341 switch (psock->eval) { 342 case __SK_PASS: 343 ret = tcp_bpf_push(sk, msg, tosend, flags, true); 344 if (unlikely(ret)) { 345 *copied -= sk_msg_free(sk, msg); 346 break; 347 } 348 sk_msg_apply_bytes(psock, tosend); 349 break; 350 case __SK_REDIRECT: 351 sk_redir = psock->sk_redir; 352 sk_msg_apply_bytes(psock, tosend); 353 if (psock->cork) { 354 cork = true; 355 psock->cork = NULL; 356 } 357 sk_msg_return(sk, msg, tosend); 358 release_sock(sk); 359 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags); 360 lock_sock(sk); 361 if (unlikely(ret < 0)) { 362 int free = sk_msg_free_nocharge(sk, msg); 363 364 if (!cork) 365 *copied -= free; 366 } 367 if (cork) { 368 sk_msg_free(sk, msg); 369 kfree(msg); 370 msg = NULL; 371 ret = 0; 372 } 373 break; 374 case __SK_DROP: 375 default: 376 sk_msg_free_partial(sk, msg, tosend); 377 sk_msg_apply_bytes(psock, tosend); 378 *copied -= (tosend + delta); 379 return -EACCES; 380 } 381 382 if (likely(!ret)) { 383 if (!psock->apply_bytes) { 384 psock->eval = __SK_NONE; 385 if (psock->sk_redir) { 386 sock_put(psock->sk_redir); 387 psock->sk_redir = NULL; 388 } 389 } 390 if (msg && 391 msg->sg.data[msg->sg.start].page_link && 392 msg->sg.data[msg->sg.start].length) 393 goto more_data; 394 } 395 return ret; 396 } 397 398 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 399 { 400 struct sk_msg tmp, *msg_tx = NULL; 401 int flags = msg->msg_flags | MSG_NO_SHARED_FRAGS; 402 int copied = 0, err = 0; 403 struct sk_psock *psock; 404 long timeo; 405 406 psock = sk_psock_get(sk); 407 if (unlikely(!psock)) 408 return tcp_sendmsg(sk, msg, size); 409 410 lock_sock(sk); 411 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 412 while (msg_data_left(msg)) { 413 bool enospc = false; 414 u32 copy, osize; 415 416 if (sk->sk_err) { 417 err = -sk->sk_err; 418 goto out_err; 419 } 420 421 copy = msg_data_left(msg); 422 if (!sk_stream_memory_free(sk)) 423 goto wait_for_sndbuf; 424 if (psock->cork) { 425 msg_tx = psock->cork; 426 } else { 427 msg_tx = &tmp; 428 sk_msg_init(msg_tx); 429 } 430 431 osize = msg_tx->sg.size; 432 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1); 433 if (err) { 434 if (err != -ENOSPC) 435 goto wait_for_memory; 436 enospc = true; 437 copy = msg_tx->sg.size - osize; 438 } 439 440 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx, 441 copy); 442 if (err < 0) { 443 sk_msg_trim(sk, msg_tx, osize); 444 goto out_err; 445 } 446 447 copied += copy; 448 if (psock->cork_bytes) { 449 if (size > psock->cork_bytes) 450 psock->cork_bytes = 0; 451 else 452 psock->cork_bytes -= size; 453 if (psock->cork_bytes && !enospc) 454 goto out_err; 455 /* All cork bytes are accounted, rerun the prog. */ 456 psock->eval = __SK_NONE; 457 psock->cork_bytes = 0; 458 } 459 460 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags); 461 if (unlikely(err < 0)) 462 goto out_err; 463 continue; 464 wait_for_sndbuf: 465 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 466 wait_for_memory: 467 err = sk_stream_wait_memory(sk, &timeo); 468 if (err) { 469 if (msg_tx && msg_tx != psock->cork) 470 sk_msg_free(sk, msg_tx); 471 goto out_err; 472 } 473 } 474 out_err: 475 if (err < 0) 476 err = sk_stream_error(sk, msg->msg_flags, err); 477 release_sock(sk); 478 sk_psock_put(sk, psock); 479 return copied ? copied : err; 480 } 481 482 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset, 483 size_t size, int flags) 484 { 485 struct sk_msg tmp, *msg = NULL; 486 int err = 0, copied = 0; 487 struct sk_psock *psock; 488 bool enospc = false; 489 490 psock = sk_psock_get(sk); 491 if (unlikely(!psock)) 492 return tcp_sendpage(sk, page, offset, size, flags); 493 494 lock_sock(sk); 495 if (psock->cork) { 496 msg = psock->cork; 497 } else { 498 msg = &tmp; 499 sk_msg_init(msg); 500 } 501 502 /* Catch case where ring is full and sendpage is stalled. */ 503 if (unlikely(sk_msg_full(msg))) 504 goto out_err; 505 506 sk_msg_page_add(msg, page, size, offset); 507 sk_mem_charge(sk, size); 508 copied = size; 509 if (sk_msg_full(msg)) 510 enospc = true; 511 if (psock->cork_bytes) { 512 if (size > psock->cork_bytes) 513 psock->cork_bytes = 0; 514 else 515 psock->cork_bytes -= size; 516 if (psock->cork_bytes && !enospc) 517 goto out_err; 518 /* All cork bytes are accounted, rerun the prog. */ 519 psock->eval = __SK_NONE; 520 psock->cork_bytes = 0; 521 } 522 523 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags); 524 out_err: 525 release_sock(sk); 526 sk_psock_put(sk, psock); 527 return copied ? copied : err; 528 } 529 530 static void tcp_bpf_remove(struct sock *sk, struct sk_psock *psock) 531 { 532 struct sk_psock_link *link; 533 534 while ((link = sk_psock_link_pop(psock))) { 535 sk_psock_unlink(sk, link); 536 sk_psock_free_link(link); 537 } 538 } 539 540 static void tcp_bpf_unhash(struct sock *sk) 541 { 542 void (*saved_unhash)(struct sock *sk); 543 struct sk_psock *psock; 544 545 rcu_read_lock(); 546 psock = sk_psock(sk); 547 if (unlikely(!psock)) { 548 rcu_read_unlock(); 549 if (sk->sk_prot->unhash) 550 sk->sk_prot->unhash(sk); 551 return; 552 } 553 554 saved_unhash = psock->saved_unhash; 555 tcp_bpf_remove(sk, psock); 556 rcu_read_unlock(); 557 saved_unhash(sk); 558 } 559 560 static void tcp_bpf_close(struct sock *sk, long timeout) 561 { 562 void (*saved_close)(struct sock *sk, long timeout); 563 struct sk_psock *psock; 564 565 lock_sock(sk); 566 rcu_read_lock(); 567 psock = sk_psock(sk); 568 if (unlikely(!psock)) { 569 rcu_read_unlock(); 570 release_sock(sk); 571 return sk->sk_prot->close(sk, timeout); 572 } 573 574 saved_close = psock->saved_close; 575 tcp_bpf_remove(sk, psock); 576 rcu_read_unlock(); 577 release_sock(sk); 578 saved_close(sk, timeout); 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_NUM_CFGS, 591 }; 592 593 static struct proto *tcpv6_prot_saved __read_mostly; 594 static DEFINE_SPINLOCK(tcpv6_prot_lock); 595 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS]; 596 597 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS], 598 struct proto *base) 599 { 600 prot[TCP_BPF_BASE] = *base; 601 prot[TCP_BPF_BASE].unhash = tcp_bpf_unhash; 602 prot[TCP_BPF_BASE].close = tcp_bpf_close; 603 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg; 604 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read; 605 606 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE]; 607 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg; 608 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage; 609 } 610 611 static void tcp_bpf_check_v6_needs_rebuild(struct sock *sk, struct proto *ops) 612 { 613 if (sk->sk_family == AF_INET6 && 614 unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) { 615 spin_lock_bh(&tcpv6_prot_lock); 616 if (likely(ops != tcpv6_prot_saved)) { 617 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops); 618 smp_store_release(&tcpv6_prot_saved, ops); 619 } 620 spin_unlock_bh(&tcpv6_prot_lock); 621 } 622 } 623 624 static int __init tcp_bpf_v4_build_proto(void) 625 { 626 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot); 627 return 0; 628 } 629 core_initcall(tcp_bpf_v4_build_proto); 630 631 static void tcp_bpf_update_sk_prot(struct sock *sk, struct sk_psock *psock) 632 { 633 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 634 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 635 636 sk_psock_update_proto(sk, psock, &tcp_bpf_prots[family][config]); 637 } 638 639 static void tcp_bpf_reinit_sk_prot(struct sock *sk, struct sk_psock *psock) 640 { 641 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 642 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 643 644 /* Reinit occurs when program types change e.g. TCP_BPF_TX is removed 645 * or added requiring sk_prot hook updates. We keep original saved 646 * hooks in this case. 647 */ 648 sk->sk_prot = &tcp_bpf_prots[family][config]; 649 } 650 651 static int tcp_bpf_assert_proto_ops(struct proto *ops) 652 { 653 /* In order to avoid retpoline, we make assumptions when we call 654 * into ops if e.g. a psock is not present. Make sure they are 655 * indeed valid assumptions. 656 */ 657 return ops->recvmsg == tcp_recvmsg && 658 ops->sendmsg == tcp_sendmsg && 659 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP; 660 } 661 662 void tcp_bpf_reinit(struct sock *sk) 663 { 664 struct sk_psock *psock; 665 666 sock_owned_by_me(sk); 667 668 rcu_read_lock(); 669 psock = sk_psock(sk); 670 tcp_bpf_reinit_sk_prot(sk, psock); 671 rcu_read_unlock(); 672 } 673 674 int tcp_bpf_init(struct sock *sk) 675 { 676 struct proto *ops = READ_ONCE(sk->sk_prot); 677 struct sk_psock *psock; 678 679 sock_owned_by_me(sk); 680 681 rcu_read_lock(); 682 psock = sk_psock(sk); 683 if (unlikely(!psock || psock->sk_proto || 684 tcp_bpf_assert_proto_ops(ops))) { 685 rcu_read_unlock(); 686 return -EINVAL; 687 } 688 tcp_bpf_check_v6_needs_rebuild(sk, ops); 689 tcp_bpf_update_sk_prot(sk, psock); 690 rcu_read_unlock(); 691 return 0; 692 } 693