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