1 // SPDX-License-Identifier: GPL-2.0-only 2 /* Copyright (c) 2016 Thomas Graf <tgraf@tgraf.ch> 3 */ 4 5 #include <linux/filter.h> 6 #include <linux/kernel.h> 7 #include <linux/module.h> 8 #include <linux/skbuff.h> 9 #include <linux/types.h> 10 #include <linux/bpf.h> 11 #include <net/lwtunnel.h> 12 #include <net/gre.h> 13 #include <net/ip.h> 14 #include <net/ip6_route.h> 15 #include <net/ipv6_stubs.h> 16 #include <net/inet_dscp.h> 17 18 struct bpf_lwt_prog { 19 struct bpf_prog *prog; 20 char *name; 21 }; 22 23 struct bpf_lwt { 24 struct bpf_lwt_prog in; 25 struct bpf_lwt_prog out; 26 struct bpf_lwt_prog xmit; 27 int family; 28 }; 29 30 #define MAX_PROG_NAME 256 31 32 static inline struct bpf_lwt *bpf_lwt_lwtunnel(struct lwtunnel_state *lwt) 33 { 34 return (struct bpf_lwt *)lwt->data; 35 } 36 37 #define NO_REDIRECT false 38 #define CAN_REDIRECT true 39 40 static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt, 41 struct dst_entry *dst, bool can_redirect) 42 { 43 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 44 int ret; 45 46 /* Disabling BH is needed to protect per-CPU bpf_redirect_info between 47 * BPF prog and skb_do_redirect(). 48 */ 49 local_bh_disable(); 50 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 51 bpf_compute_data_pointers(skb); 52 ret = bpf_prog_run_save_cb(lwt->prog, skb); 53 54 switch (ret) { 55 case BPF_OK: 56 case BPF_LWT_REROUTE: 57 break; 58 59 case BPF_REDIRECT: 60 if (unlikely(!can_redirect)) { 61 pr_warn_once("Illegal redirect return code in prog %s\n", 62 lwt->name ? : "<unknown>"); 63 ret = BPF_OK; 64 } else { 65 skb_reset_mac_header(skb); 66 skb_do_redirect(skb); 67 ret = BPF_REDIRECT; 68 } 69 break; 70 71 case BPF_DROP: 72 kfree_skb(skb); 73 ret = -EPERM; 74 break; 75 76 default: 77 pr_warn_once("bpf-lwt: Illegal return value %u, expect packet loss\n", ret); 78 kfree_skb(skb); 79 ret = -EINVAL; 80 break; 81 } 82 83 bpf_net_ctx_clear(bpf_net_ctx); 84 local_bh_enable(); 85 86 return ret; 87 } 88 89 static int bpf_lwt_input_reroute(struct sk_buff *skb) 90 { 91 enum skb_drop_reason reason; 92 int err = -EINVAL; 93 94 if (skb->protocol == htons(ETH_P_IP)) { 95 struct net_device *dev = skb_dst(skb)->dev; 96 const struct iphdr *iph = ip_hdr(skb); 97 98 dev_hold(dev); 99 skb_dst_drop(skb); 100 reason = ip_route_input_noref(skb, iph->daddr, iph->saddr, 101 ip4h_dscp(iph), dev); 102 err = reason ? -EINVAL : 0; 103 dev_put(dev); 104 } else if (skb->protocol == htons(ETH_P_IPV6)) { 105 skb_dst_drop(skb); 106 err = ipv6_stub->ipv6_route_input(skb); 107 } else { 108 err = -EAFNOSUPPORT; 109 } 110 111 if (err) 112 goto err; 113 return dst_input(skb); 114 115 err: 116 kfree_skb(skb); 117 return err; 118 } 119 120 static int bpf_input(struct sk_buff *skb) 121 { 122 struct dst_entry *dst = skb_dst(skb); 123 struct bpf_lwt *bpf; 124 int ret; 125 126 bpf = bpf_lwt_lwtunnel(dst->lwtstate); 127 if (bpf->in.prog) { 128 ret = run_lwt_bpf(skb, &bpf->in, dst, NO_REDIRECT); 129 if (ret < 0) 130 return ret; 131 if (ret == BPF_LWT_REROUTE) 132 return bpf_lwt_input_reroute(skb); 133 } 134 135 if (unlikely(!dst->lwtstate->orig_input)) { 136 kfree_skb(skb); 137 return -EINVAL; 138 } 139 140 return dst->lwtstate->orig_input(skb); 141 } 142 143 static int bpf_output(struct net *net, struct sock *sk, struct sk_buff *skb) 144 { 145 struct dst_entry *dst = skb_dst(skb); 146 struct bpf_lwt *bpf; 147 int ret; 148 149 bpf = bpf_lwt_lwtunnel(dst->lwtstate); 150 if (bpf->out.prog) { 151 ret = run_lwt_bpf(skb, &bpf->out, dst, NO_REDIRECT); 152 if (ret < 0) 153 return ret; 154 } 155 156 if (unlikely(!dst->lwtstate->orig_output)) { 157 pr_warn_once("orig_output not set on dst for prog %s\n", 158 bpf->out.name); 159 kfree_skb(skb); 160 return -EINVAL; 161 } 162 163 return dst->lwtstate->orig_output(net, sk, skb); 164 } 165 166 static int xmit_check_hhlen(struct sk_buff *skb, int hh_len) 167 { 168 if (skb_headroom(skb) < hh_len) { 169 int nhead = HH_DATA_ALIGN(hh_len - skb_headroom(skb)); 170 171 if (pskb_expand_head(skb, nhead, 0, GFP_ATOMIC)) 172 return -ENOMEM; 173 } 174 175 return 0; 176 } 177 178 static int bpf_lwt_xmit_reroute(struct sk_buff *skb) 179 { 180 struct net_device *l3mdev = l3mdev_master_dev_rcu(skb_dst(skb)->dev); 181 int oif = l3mdev ? l3mdev->ifindex : 0; 182 struct dst_entry *dst = NULL; 183 int err = -EAFNOSUPPORT; 184 struct sock *sk; 185 struct net *net; 186 bool ipv4; 187 188 if (skb->protocol == htons(ETH_P_IP)) 189 ipv4 = true; 190 else if (skb->protocol == htons(ETH_P_IPV6)) 191 ipv4 = false; 192 else 193 goto err; 194 195 sk = sk_to_full_sk(skb->sk); 196 if (sk) { 197 if (sk->sk_bound_dev_if) 198 oif = sk->sk_bound_dev_if; 199 net = sock_net(sk); 200 } else { 201 net = dev_net(skb_dst(skb)->dev); 202 } 203 204 if (ipv4) { 205 struct iphdr *iph = ip_hdr(skb); 206 struct flowi4 fl4 = {}; 207 struct rtable *rt; 208 209 fl4.flowi4_oif = oif; 210 fl4.flowi4_mark = skb->mark; 211 fl4.flowi4_uid = sock_net_uid(net, sk); 212 fl4.flowi4_tos = inet_dscp_to_dsfield(ip4h_dscp(iph)); 213 fl4.flowi4_flags = FLOWI_FLAG_ANYSRC; 214 fl4.flowi4_proto = iph->protocol; 215 fl4.daddr = iph->daddr; 216 fl4.saddr = iph->saddr; 217 218 rt = ip_route_output_key(net, &fl4); 219 if (IS_ERR(rt)) { 220 err = PTR_ERR(rt); 221 goto err; 222 } 223 dst = &rt->dst; 224 } else { 225 struct ipv6hdr *iph6 = ipv6_hdr(skb); 226 struct flowi6 fl6 = {}; 227 228 fl6.flowi6_oif = oif; 229 fl6.flowi6_mark = skb->mark; 230 fl6.flowi6_uid = sock_net_uid(net, sk); 231 fl6.flowlabel = ip6_flowinfo(iph6); 232 fl6.flowi6_proto = iph6->nexthdr; 233 fl6.daddr = iph6->daddr; 234 fl6.saddr = iph6->saddr; 235 236 dst = ipv6_stub->ipv6_dst_lookup_flow(net, skb->sk, &fl6, NULL); 237 if (IS_ERR(dst)) { 238 err = PTR_ERR(dst); 239 goto err; 240 } 241 } 242 if (unlikely(dst->error)) { 243 err = dst->error; 244 dst_release(dst); 245 goto err; 246 } 247 248 /* Although skb header was reserved in bpf_lwt_push_ip_encap(), it 249 * was done for the previous dst, so we are doing it here again, in 250 * case the new dst needs much more space. The call below is a noop 251 * if there is enough header space in skb. 252 */ 253 err = skb_cow_head(skb, LL_RESERVED_SPACE(dst->dev)); 254 if (unlikely(err)) 255 goto err; 256 257 skb_dst_drop(skb); 258 skb_dst_set(skb, dst); 259 260 err = dst_output(dev_net(skb_dst(skb)->dev), skb->sk, skb); 261 if (unlikely(err)) 262 return net_xmit_errno(err); 263 264 /* ip[6]_finish_output2 understand LWTUNNEL_XMIT_DONE */ 265 return LWTUNNEL_XMIT_DONE; 266 267 err: 268 kfree_skb(skb); 269 return err; 270 } 271 272 static int bpf_xmit(struct sk_buff *skb) 273 { 274 struct dst_entry *dst = skb_dst(skb); 275 struct bpf_lwt *bpf; 276 277 bpf = bpf_lwt_lwtunnel(dst->lwtstate); 278 if (bpf->xmit.prog) { 279 int hh_len = dst->dev->hard_header_len; 280 __be16 proto = skb->protocol; 281 int ret; 282 283 ret = run_lwt_bpf(skb, &bpf->xmit, dst, CAN_REDIRECT); 284 switch (ret) { 285 case BPF_OK: 286 /* If the header changed, e.g. via bpf_lwt_push_encap, 287 * BPF_LWT_REROUTE below should have been used if the 288 * protocol was also changed. 289 */ 290 if (skb->protocol != proto) { 291 kfree_skb(skb); 292 return -EINVAL; 293 } 294 /* If the header was expanded, headroom might be too 295 * small for L2 header to come, expand as needed. 296 */ 297 ret = xmit_check_hhlen(skb, hh_len); 298 if (unlikely(ret)) 299 return ret; 300 301 return LWTUNNEL_XMIT_CONTINUE; 302 case BPF_REDIRECT: 303 return LWTUNNEL_XMIT_DONE; 304 case BPF_LWT_REROUTE: 305 return bpf_lwt_xmit_reroute(skb); 306 default: 307 return ret; 308 } 309 } 310 311 return LWTUNNEL_XMIT_CONTINUE; 312 } 313 314 static void bpf_lwt_prog_destroy(struct bpf_lwt_prog *prog) 315 { 316 if (prog->prog) 317 bpf_prog_put(prog->prog); 318 319 kfree(prog->name); 320 } 321 322 static void bpf_destroy_state(struct lwtunnel_state *lwt) 323 { 324 struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt); 325 326 bpf_lwt_prog_destroy(&bpf->in); 327 bpf_lwt_prog_destroy(&bpf->out); 328 bpf_lwt_prog_destroy(&bpf->xmit); 329 } 330 331 static const struct nla_policy bpf_prog_policy[LWT_BPF_PROG_MAX + 1] = { 332 [LWT_BPF_PROG_FD] = { .type = NLA_U32, }, 333 [LWT_BPF_PROG_NAME] = { .type = NLA_NUL_STRING, 334 .len = MAX_PROG_NAME }, 335 }; 336 337 static int bpf_parse_prog(struct nlattr *attr, struct bpf_lwt_prog *prog, 338 enum bpf_prog_type type) 339 { 340 struct nlattr *tb[LWT_BPF_PROG_MAX + 1]; 341 struct bpf_prog *p; 342 int ret; 343 u32 fd; 344 345 ret = nla_parse_nested_deprecated(tb, LWT_BPF_PROG_MAX, attr, 346 bpf_prog_policy, NULL); 347 if (ret < 0) 348 return ret; 349 350 if (!tb[LWT_BPF_PROG_FD] || !tb[LWT_BPF_PROG_NAME]) 351 return -EINVAL; 352 353 prog->name = nla_memdup(tb[LWT_BPF_PROG_NAME], GFP_ATOMIC); 354 if (!prog->name) 355 return -ENOMEM; 356 357 fd = nla_get_u32(tb[LWT_BPF_PROG_FD]); 358 p = bpf_prog_get_type(fd, type); 359 if (IS_ERR(p)) 360 return PTR_ERR(p); 361 362 prog->prog = p; 363 364 return 0; 365 } 366 367 static const struct nla_policy bpf_nl_policy[LWT_BPF_MAX + 1] = { 368 [LWT_BPF_IN] = { .type = NLA_NESTED, }, 369 [LWT_BPF_OUT] = { .type = NLA_NESTED, }, 370 [LWT_BPF_XMIT] = { .type = NLA_NESTED, }, 371 [LWT_BPF_XMIT_HEADROOM] = { .type = NLA_U32 }, 372 }; 373 374 static int bpf_build_state(struct net *net, struct nlattr *nla, 375 unsigned int family, const void *cfg, 376 struct lwtunnel_state **ts, 377 struct netlink_ext_ack *extack) 378 { 379 struct nlattr *tb[LWT_BPF_MAX + 1]; 380 struct lwtunnel_state *newts; 381 struct bpf_lwt *bpf; 382 int ret; 383 384 if (family != AF_INET && family != AF_INET6) 385 return -EAFNOSUPPORT; 386 387 ret = nla_parse_nested_deprecated(tb, LWT_BPF_MAX, nla, bpf_nl_policy, 388 extack); 389 if (ret < 0) 390 return ret; 391 392 if (!tb[LWT_BPF_IN] && !tb[LWT_BPF_OUT] && !tb[LWT_BPF_XMIT]) 393 return -EINVAL; 394 395 newts = lwtunnel_state_alloc(sizeof(*bpf)); 396 if (!newts) 397 return -ENOMEM; 398 399 newts->type = LWTUNNEL_ENCAP_BPF; 400 bpf = bpf_lwt_lwtunnel(newts); 401 402 if (tb[LWT_BPF_IN]) { 403 newts->flags |= LWTUNNEL_STATE_INPUT_REDIRECT; 404 ret = bpf_parse_prog(tb[LWT_BPF_IN], &bpf->in, 405 BPF_PROG_TYPE_LWT_IN); 406 if (ret < 0) 407 goto errout; 408 } 409 410 if (tb[LWT_BPF_OUT]) { 411 newts->flags |= LWTUNNEL_STATE_OUTPUT_REDIRECT; 412 ret = bpf_parse_prog(tb[LWT_BPF_OUT], &bpf->out, 413 BPF_PROG_TYPE_LWT_OUT); 414 if (ret < 0) 415 goto errout; 416 } 417 418 if (tb[LWT_BPF_XMIT]) { 419 newts->flags |= LWTUNNEL_STATE_XMIT_REDIRECT; 420 ret = bpf_parse_prog(tb[LWT_BPF_XMIT], &bpf->xmit, 421 BPF_PROG_TYPE_LWT_XMIT); 422 if (ret < 0) 423 goto errout; 424 } 425 426 if (tb[LWT_BPF_XMIT_HEADROOM]) { 427 u32 headroom = nla_get_u32(tb[LWT_BPF_XMIT_HEADROOM]); 428 429 if (headroom > LWT_BPF_MAX_HEADROOM) { 430 ret = -ERANGE; 431 goto errout; 432 } 433 434 newts->headroom = headroom; 435 } 436 437 bpf->family = family; 438 *ts = newts; 439 440 return 0; 441 442 errout: 443 bpf_destroy_state(newts); 444 kfree(newts); 445 return ret; 446 } 447 448 static int bpf_fill_lwt_prog(struct sk_buff *skb, int attr, 449 struct bpf_lwt_prog *prog) 450 { 451 struct nlattr *nest; 452 453 if (!prog->prog) 454 return 0; 455 456 nest = nla_nest_start_noflag(skb, attr); 457 if (!nest) 458 return -EMSGSIZE; 459 460 if (prog->name && 461 nla_put_string(skb, LWT_BPF_PROG_NAME, prog->name)) 462 return -EMSGSIZE; 463 464 return nla_nest_end(skb, nest); 465 } 466 467 static int bpf_fill_encap_info(struct sk_buff *skb, struct lwtunnel_state *lwt) 468 { 469 struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt); 470 471 if (bpf_fill_lwt_prog(skb, LWT_BPF_IN, &bpf->in) < 0 || 472 bpf_fill_lwt_prog(skb, LWT_BPF_OUT, &bpf->out) < 0 || 473 bpf_fill_lwt_prog(skb, LWT_BPF_XMIT, &bpf->xmit) < 0) 474 return -EMSGSIZE; 475 476 return 0; 477 } 478 479 static int bpf_encap_nlsize(struct lwtunnel_state *lwtstate) 480 { 481 int nest_len = nla_total_size(sizeof(struct nlattr)) + 482 nla_total_size(MAX_PROG_NAME) + /* LWT_BPF_PROG_NAME */ 483 0; 484 485 return nest_len + /* LWT_BPF_IN */ 486 nest_len + /* LWT_BPF_OUT */ 487 nest_len + /* LWT_BPF_XMIT */ 488 0; 489 } 490 491 static int bpf_lwt_prog_cmp(struct bpf_lwt_prog *a, struct bpf_lwt_prog *b) 492 { 493 /* FIXME: 494 * The LWT state is currently rebuilt for delete requests which 495 * results in a new bpf_prog instance. Comparing names for now. 496 */ 497 if (!a->name && !b->name) 498 return 0; 499 500 if (!a->name || !b->name) 501 return 1; 502 503 return strcmp(a->name, b->name); 504 } 505 506 static int bpf_encap_cmp(struct lwtunnel_state *a, struct lwtunnel_state *b) 507 { 508 struct bpf_lwt *a_bpf = bpf_lwt_lwtunnel(a); 509 struct bpf_lwt *b_bpf = bpf_lwt_lwtunnel(b); 510 511 return bpf_lwt_prog_cmp(&a_bpf->in, &b_bpf->in) || 512 bpf_lwt_prog_cmp(&a_bpf->out, &b_bpf->out) || 513 bpf_lwt_prog_cmp(&a_bpf->xmit, &b_bpf->xmit); 514 } 515 516 static const struct lwtunnel_encap_ops bpf_encap_ops = { 517 .build_state = bpf_build_state, 518 .destroy_state = bpf_destroy_state, 519 .input = bpf_input, 520 .output = bpf_output, 521 .xmit = bpf_xmit, 522 .fill_encap = bpf_fill_encap_info, 523 .get_encap_size = bpf_encap_nlsize, 524 .cmp_encap = bpf_encap_cmp, 525 .owner = THIS_MODULE, 526 }; 527 528 static int handle_gso_type(struct sk_buff *skb, unsigned int gso_type, 529 int encap_len) 530 { 531 struct skb_shared_info *shinfo = skb_shinfo(skb); 532 533 gso_type |= SKB_GSO_DODGY; 534 shinfo->gso_type |= gso_type; 535 skb_decrease_gso_size(shinfo, encap_len); 536 shinfo->gso_segs = 0; 537 return 0; 538 } 539 540 static int handle_gso_encap(struct sk_buff *skb, bool ipv4, int encap_len) 541 { 542 int next_hdr_offset; 543 void *next_hdr; 544 __u8 protocol; 545 546 /* SCTP and UDP_L4 gso need more nuanced handling than what 547 * handle_gso_type() does above: skb_decrease_gso_size() is not enough. 548 * So at the moment only TCP GSO packets are let through. 549 */ 550 if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) 551 return -ENOTSUPP; 552 553 if (ipv4) { 554 protocol = ip_hdr(skb)->protocol; 555 next_hdr_offset = sizeof(struct iphdr); 556 next_hdr = skb_network_header(skb) + next_hdr_offset; 557 } else { 558 protocol = ipv6_hdr(skb)->nexthdr; 559 next_hdr_offset = sizeof(struct ipv6hdr); 560 next_hdr = skb_network_header(skb) + next_hdr_offset; 561 } 562 563 switch (protocol) { 564 case IPPROTO_GRE: 565 next_hdr_offset += sizeof(struct gre_base_hdr); 566 if (next_hdr_offset > encap_len) 567 return -EINVAL; 568 569 if (((struct gre_base_hdr *)next_hdr)->flags & GRE_CSUM) 570 return handle_gso_type(skb, SKB_GSO_GRE_CSUM, 571 encap_len); 572 return handle_gso_type(skb, SKB_GSO_GRE, encap_len); 573 574 case IPPROTO_UDP: 575 next_hdr_offset += sizeof(struct udphdr); 576 if (next_hdr_offset > encap_len) 577 return -EINVAL; 578 579 if (((struct udphdr *)next_hdr)->check) 580 return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL_CSUM, 581 encap_len); 582 return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL, encap_len); 583 584 case IPPROTO_IP: 585 case IPPROTO_IPV6: 586 if (ipv4) 587 return handle_gso_type(skb, SKB_GSO_IPXIP4, encap_len); 588 else 589 return handle_gso_type(skb, SKB_GSO_IPXIP6, encap_len); 590 591 default: 592 return -EPROTONOSUPPORT; 593 } 594 } 595 596 int bpf_lwt_push_ip_encap(struct sk_buff *skb, void *hdr, u32 len, bool ingress) 597 { 598 struct iphdr *iph; 599 bool ipv4; 600 int err; 601 602 if (unlikely(len < sizeof(struct iphdr) || len > LWT_BPF_MAX_HEADROOM)) 603 return -EINVAL; 604 605 /* validate protocol and length */ 606 iph = (struct iphdr *)hdr; 607 if (iph->version == 4) { 608 ipv4 = true; 609 if (unlikely(len < iph->ihl * 4)) 610 return -EINVAL; 611 } else if (iph->version == 6) { 612 ipv4 = false; 613 if (unlikely(len < sizeof(struct ipv6hdr))) 614 return -EINVAL; 615 } else { 616 return -EINVAL; 617 } 618 619 if (ingress) 620 err = skb_cow_head(skb, len + skb->mac_len); 621 else 622 err = skb_cow_head(skb, 623 len + LL_RESERVED_SPACE(skb_dst(skb)->dev)); 624 if (unlikely(err)) 625 return err; 626 627 /* push the encap headers and fix pointers */ 628 skb_reset_inner_headers(skb); 629 skb_reset_inner_mac_header(skb); /* mac header is not yet set */ 630 skb_set_inner_protocol(skb, skb->protocol); 631 skb->encapsulation = 1; 632 skb_push(skb, len); 633 if (ingress) 634 skb_postpush_rcsum(skb, iph, len); 635 skb_reset_network_header(skb); 636 memcpy(skb_network_header(skb), hdr, len); 637 bpf_compute_data_pointers(skb); 638 skb_clear_hash(skb); 639 640 if (ipv4) { 641 skb->protocol = htons(ETH_P_IP); 642 iph = ip_hdr(skb); 643 644 if (!iph->check) 645 iph->check = ip_fast_csum((unsigned char *)iph, 646 iph->ihl); 647 } else { 648 skb->protocol = htons(ETH_P_IPV6); 649 } 650 651 if (skb_is_gso(skb)) 652 return handle_gso_encap(skb, ipv4, len); 653 654 return 0; 655 } 656 657 static int __init bpf_lwt_init(void) 658 { 659 return lwtunnel_encap_add_ops(&bpf_encap_ops, LWTUNNEL_ENCAP_BPF); 660 } 661 662 subsys_initcall(bpf_lwt_init) 663