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, struct sk_msg *msg, 135 u32 bytes, int flags) 136 { 137 bool ingress = sk_msg_to_ingress(msg); 138 struct sk_psock *psock = sk_psock_get(sk); 139 int ret; 140 141 if (unlikely(!psock)) { 142 sk_msg_free(sk, msg); 143 return 0; 144 } 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); 148 return ret; 149 } 150 EXPORT_SYMBOL_GPL(tcp_bpf_sendmsg_redir); 151 152 #ifdef CONFIG_BPF_SYSCALL 153 static bool tcp_bpf_stream_read(const struct sock *sk) 154 { 155 struct sk_psock *psock; 156 bool empty = true; 157 158 rcu_read_lock(); 159 psock = sk_psock(sk); 160 if (likely(psock)) 161 empty = list_empty(&psock->ingress_msg); 162 rcu_read_unlock(); 163 return !empty; 164 } 165 166 static int tcp_bpf_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 167 int nonblock, int flags, int *addr_len) 168 { 169 struct sk_psock *psock; 170 int copied, ret; 171 172 if (unlikely(flags & MSG_ERRQUEUE)) 173 return inet_recv_error(sk, msg, len, addr_len); 174 175 psock = sk_psock_get(sk); 176 if (unlikely(!psock)) 177 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 178 if (!skb_queue_empty(&sk->sk_receive_queue) && 179 sk_psock_queue_empty(psock)) { 180 sk_psock_put(sk, psock); 181 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 182 } 183 lock_sock(sk); 184 msg_bytes_ready: 185 copied = sk_msg_recvmsg(sk, psock, msg, len, flags); 186 if (!copied) { 187 int data, err = 0; 188 long timeo; 189 190 timeo = sock_rcvtimeo(sk, nonblock); 191 data = sk_msg_wait_data(sk, psock, flags, timeo, &err); 192 if (data) { 193 if (!sk_psock_queue_empty(psock)) 194 goto msg_bytes_ready; 195 release_sock(sk); 196 sk_psock_put(sk, psock); 197 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len); 198 } 199 if (err) { 200 ret = err; 201 goto out; 202 } 203 copied = -EAGAIN; 204 } 205 ret = copied; 206 out: 207 release_sock(sk); 208 sk_psock_put(sk, psock); 209 return ret; 210 } 211 212 static int tcp_bpf_send_verdict(struct sock *sk, struct sk_psock *psock, 213 struct sk_msg *msg, int *copied, int flags) 214 { 215 bool cork = false, enospc = sk_msg_full(msg); 216 struct sock *sk_redir; 217 u32 tosend, delta = 0; 218 int ret; 219 220 more_data: 221 if (psock->eval == __SK_NONE) { 222 /* Track delta in msg size to add/subtract it on SK_DROP from 223 * returned to user copied size. This ensures user doesn't 224 * get a positive return code with msg_cut_data and SK_DROP 225 * verdict. 226 */ 227 delta = msg->sg.size; 228 psock->eval = sk_psock_msg_verdict(sk, psock, msg); 229 delta -= msg->sg.size; 230 } 231 232 if (msg->cork_bytes && 233 msg->cork_bytes > msg->sg.size && !enospc) { 234 psock->cork_bytes = msg->cork_bytes - msg->sg.size; 235 if (!psock->cork) { 236 psock->cork = kzalloc(sizeof(*psock->cork), 237 GFP_ATOMIC | __GFP_NOWARN); 238 if (!psock->cork) 239 return -ENOMEM; 240 } 241 memcpy(psock->cork, msg, sizeof(*msg)); 242 return 0; 243 } 244 245 tosend = msg->sg.size; 246 if (psock->apply_bytes && psock->apply_bytes < tosend) 247 tosend = psock->apply_bytes; 248 249 switch (psock->eval) { 250 case __SK_PASS: 251 ret = tcp_bpf_push(sk, msg, tosend, flags, true); 252 if (unlikely(ret)) { 253 *copied -= sk_msg_free(sk, msg); 254 break; 255 } 256 sk_msg_apply_bytes(psock, tosend); 257 break; 258 case __SK_REDIRECT: 259 sk_redir = psock->sk_redir; 260 sk_msg_apply_bytes(psock, tosend); 261 if (psock->cork) { 262 cork = true; 263 psock->cork = NULL; 264 } 265 sk_msg_return(sk, msg, tosend); 266 release_sock(sk); 267 ret = tcp_bpf_sendmsg_redir(sk_redir, msg, tosend, flags); 268 lock_sock(sk); 269 if (unlikely(ret < 0)) { 270 int free = sk_msg_free_nocharge(sk, msg); 271 272 if (!cork) 273 *copied -= free; 274 } 275 if (cork) { 276 sk_msg_free(sk, msg); 277 kfree(msg); 278 msg = NULL; 279 ret = 0; 280 } 281 break; 282 case __SK_DROP: 283 default: 284 sk_msg_free_partial(sk, msg, tosend); 285 sk_msg_apply_bytes(psock, tosend); 286 *copied -= (tosend + delta); 287 return -EACCES; 288 } 289 290 if (likely(!ret)) { 291 if (!psock->apply_bytes) { 292 psock->eval = __SK_NONE; 293 if (psock->sk_redir) { 294 sock_put(psock->sk_redir); 295 psock->sk_redir = NULL; 296 } 297 } 298 if (msg && 299 msg->sg.data[msg->sg.start].page_link && 300 msg->sg.data[msg->sg.start].length) 301 goto more_data; 302 } 303 return ret; 304 } 305 306 static int tcp_bpf_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 307 { 308 struct sk_msg tmp, *msg_tx = NULL; 309 int copied = 0, err = 0; 310 struct sk_psock *psock; 311 long timeo; 312 int flags; 313 314 /* Don't let internal do_tcp_sendpages() flags through */ 315 flags = (msg->msg_flags & ~MSG_SENDPAGE_DECRYPTED); 316 flags |= MSG_NO_SHARED_FRAGS; 317 318 psock = sk_psock_get(sk); 319 if (unlikely(!psock)) 320 return tcp_sendmsg(sk, msg, size); 321 322 lock_sock(sk); 323 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 324 while (msg_data_left(msg)) { 325 bool enospc = false; 326 u32 copy, osize; 327 328 if (sk->sk_err) { 329 err = -sk->sk_err; 330 goto out_err; 331 } 332 333 copy = msg_data_left(msg); 334 if (!sk_stream_memory_free(sk)) 335 goto wait_for_sndbuf; 336 if (psock->cork) { 337 msg_tx = psock->cork; 338 } else { 339 msg_tx = &tmp; 340 sk_msg_init(msg_tx); 341 } 342 343 osize = msg_tx->sg.size; 344 err = sk_msg_alloc(sk, msg_tx, msg_tx->sg.size + copy, msg_tx->sg.end - 1); 345 if (err) { 346 if (err != -ENOSPC) 347 goto wait_for_memory; 348 enospc = true; 349 copy = msg_tx->sg.size - osize; 350 } 351 352 err = sk_msg_memcopy_from_iter(sk, &msg->msg_iter, msg_tx, 353 copy); 354 if (err < 0) { 355 sk_msg_trim(sk, msg_tx, osize); 356 goto out_err; 357 } 358 359 copied += copy; 360 if (psock->cork_bytes) { 361 if (size > psock->cork_bytes) 362 psock->cork_bytes = 0; 363 else 364 psock->cork_bytes -= size; 365 if (psock->cork_bytes && !enospc) 366 goto out_err; 367 /* All cork bytes are accounted, rerun the prog. */ 368 psock->eval = __SK_NONE; 369 psock->cork_bytes = 0; 370 } 371 372 err = tcp_bpf_send_verdict(sk, psock, msg_tx, &copied, flags); 373 if (unlikely(err < 0)) 374 goto out_err; 375 continue; 376 wait_for_sndbuf: 377 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 378 wait_for_memory: 379 err = sk_stream_wait_memory(sk, &timeo); 380 if (err) { 381 if (msg_tx && msg_tx != psock->cork) 382 sk_msg_free(sk, msg_tx); 383 goto out_err; 384 } 385 } 386 out_err: 387 if (err < 0) 388 err = sk_stream_error(sk, msg->msg_flags, err); 389 release_sock(sk); 390 sk_psock_put(sk, psock); 391 return copied ? copied : err; 392 } 393 394 static int tcp_bpf_sendpage(struct sock *sk, struct page *page, int offset, 395 size_t size, int flags) 396 { 397 struct sk_msg tmp, *msg = NULL; 398 int err = 0, copied = 0; 399 struct sk_psock *psock; 400 bool enospc = false; 401 402 psock = sk_psock_get(sk); 403 if (unlikely(!psock)) 404 return tcp_sendpage(sk, page, offset, size, flags); 405 406 lock_sock(sk); 407 if (psock->cork) { 408 msg = psock->cork; 409 } else { 410 msg = &tmp; 411 sk_msg_init(msg); 412 } 413 414 /* Catch case where ring is full and sendpage is stalled. */ 415 if (unlikely(sk_msg_full(msg))) 416 goto out_err; 417 418 sk_msg_page_add(msg, page, size, offset); 419 sk_mem_charge(sk, size); 420 copied = size; 421 if (sk_msg_full(msg)) 422 enospc = true; 423 if (psock->cork_bytes) { 424 if (size > psock->cork_bytes) 425 psock->cork_bytes = 0; 426 else 427 psock->cork_bytes -= size; 428 if (psock->cork_bytes && !enospc) 429 goto out_err; 430 /* All cork bytes are accounted, rerun the prog. */ 431 psock->eval = __SK_NONE; 432 psock->cork_bytes = 0; 433 } 434 435 err = tcp_bpf_send_verdict(sk, psock, msg, &copied, flags); 436 out_err: 437 release_sock(sk); 438 sk_psock_put(sk, psock); 439 return copied ? copied : err; 440 } 441 442 enum { 443 TCP_BPF_IPV4, 444 TCP_BPF_IPV6, 445 TCP_BPF_NUM_PROTS, 446 }; 447 448 enum { 449 TCP_BPF_BASE, 450 TCP_BPF_TX, 451 TCP_BPF_NUM_CFGS, 452 }; 453 454 static struct proto *tcpv6_prot_saved __read_mostly; 455 static DEFINE_SPINLOCK(tcpv6_prot_lock); 456 static struct proto tcp_bpf_prots[TCP_BPF_NUM_PROTS][TCP_BPF_NUM_CFGS]; 457 458 static void tcp_bpf_rebuild_protos(struct proto prot[TCP_BPF_NUM_CFGS], 459 struct proto *base) 460 { 461 prot[TCP_BPF_BASE] = *base; 462 prot[TCP_BPF_BASE].unhash = sock_map_unhash; 463 prot[TCP_BPF_BASE].close = sock_map_close; 464 prot[TCP_BPF_BASE].recvmsg = tcp_bpf_recvmsg; 465 prot[TCP_BPF_BASE].stream_memory_read = tcp_bpf_stream_read; 466 467 prot[TCP_BPF_TX] = prot[TCP_BPF_BASE]; 468 prot[TCP_BPF_TX].sendmsg = tcp_bpf_sendmsg; 469 prot[TCP_BPF_TX].sendpage = tcp_bpf_sendpage; 470 } 471 472 static void tcp_bpf_check_v6_needs_rebuild(struct proto *ops) 473 { 474 if (unlikely(ops != smp_load_acquire(&tcpv6_prot_saved))) { 475 spin_lock_bh(&tcpv6_prot_lock); 476 if (likely(ops != tcpv6_prot_saved)) { 477 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV6], ops); 478 smp_store_release(&tcpv6_prot_saved, ops); 479 } 480 spin_unlock_bh(&tcpv6_prot_lock); 481 } 482 } 483 484 static int __init tcp_bpf_v4_build_proto(void) 485 { 486 tcp_bpf_rebuild_protos(tcp_bpf_prots[TCP_BPF_IPV4], &tcp_prot); 487 return 0; 488 } 489 core_initcall(tcp_bpf_v4_build_proto); 490 491 static int tcp_bpf_assert_proto_ops(struct proto *ops) 492 { 493 /* In order to avoid retpoline, we make assumptions when we call 494 * into ops if e.g. a psock is not present. Make sure they are 495 * indeed valid assumptions. 496 */ 497 return ops->recvmsg == tcp_recvmsg && 498 ops->sendmsg == tcp_sendmsg && 499 ops->sendpage == tcp_sendpage ? 0 : -ENOTSUPP; 500 } 501 502 int tcp_bpf_update_proto(struct sock *sk, struct sk_psock *psock, bool restore) 503 { 504 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 505 int config = psock->progs.msg_parser ? TCP_BPF_TX : TCP_BPF_BASE; 506 507 if (restore) { 508 if (inet_csk_has_ulp(sk)) { 509 /* TLS does not have an unhash proto in SW cases, 510 * but we need to ensure we stop using the sock_map 511 * unhash routine because the associated psock is being 512 * removed. So use the original unhash handler. 513 */ 514 WRITE_ONCE(sk->sk_prot->unhash, psock->saved_unhash); 515 tcp_update_ulp(sk, psock->sk_proto, psock->saved_write_space); 516 } else { 517 sk->sk_write_space = psock->saved_write_space; 518 /* Pairs with lockless read in sk_clone_lock() */ 519 WRITE_ONCE(sk->sk_prot, psock->sk_proto); 520 } 521 return 0; 522 } 523 524 if (inet_csk_has_ulp(sk)) 525 return -EINVAL; 526 527 if (sk->sk_family == AF_INET6) { 528 if (tcp_bpf_assert_proto_ops(psock->sk_proto)) 529 return -EINVAL; 530 531 tcp_bpf_check_v6_needs_rebuild(psock->sk_proto); 532 } 533 534 /* Pairs with lockless read in sk_clone_lock() */ 535 WRITE_ONCE(sk->sk_prot, &tcp_bpf_prots[family][config]); 536 return 0; 537 } 538 EXPORT_SYMBOL_GPL(tcp_bpf_update_proto); 539 540 /* If a child got cloned from a listening socket that had tcp_bpf 541 * protocol callbacks installed, we need to restore the callbacks to 542 * the default ones because the child does not inherit the psock state 543 * that tcp_bpf callbacks expect. 544 */ 545 void tcp_bpf_clone(const struct sock *sk, struct sock *newsk) 546 { 547 int family = sk->sk_family == AF_INET6 ? TCP_BPF_IPV6 : TCP_BPF_IPV4; 548 struct proto *prot = newsk->sk_prot; 549 550 if (prot == &tcp_bpf_prots[family][TCP_BPF_BASE]) 551 newsk->sk_prot = sk->sk_prot_creator; 552 } 553 #endif /* CONFIG_BPF_SYSCALL */ 554