| | 1 | #include <linux/module.h> |
| | 2 | #include <linux/errno.h> |
| | 3 | #include <linux/socket.h> |
| | 4 | #include <linux/skbuff.h> |
| | 5 | #include <linux/ip.h> |
| | 6 | #include <linux/udp.h> |
| | 7 | #include <linux/types.h> |
| | 8 | #include <linux/kernel.h> |
| | 9 | #include <net/genetlink.h> |
| | 10 | #include <net/gue.h> |
| | 11 | #include <net/ip.h> |
| | 12 | #include <net/protocol.h> |
| | 13 | #include <net/udp.h> |
| | 14 | #include <net/udp_tunnel.h> |
| | 15 | #include <net/xfrm.h> |
| | 16 | #include <uapi/linux/fou.h> |
| | 17 | #include <uapi/linux/genetlink.h> |
| | 18 | |
| | 19 | struct fou { |
| | 20 | struct socket *sock; |
| | 21 | u8 protocol; |
| | 22 | u8 flags; |
| | 23 | __be16 port; |
| | 24 | u16 type; |
| | 25 | struct udp_offload udp_offloads; |
| | 26 | struct list_head list; |
| | 27 | struct rcu_head rcu; |
| | 28 | }; |
| | 29 | |
| | 30 | #define FOU_F_REMCSUM_NOPARTIAL BIT(0) |
| | 31 | |
| | 32 | struct fou_cfg { |
| | 33 | u16 type; |
| | 34 | u8 protocol; |
| | 35 | u8 flags; |
| | 36 | struct udp_port_cfg udp_config; |
| | 37 | }; |
| | 38 | |
| | 39 | static unsigned int fou_net_id; |
| | 40 | |
| | 41 | struct fou_net { |
| | 42 | struct list_head fou_list; |
| | 43 | struct mutex fou_lock; |
| | 44 | }; |
| | 45 | |
| | 46 | static inline struct fou *fou_from_sock(struct sock *sk) |
| | 47 | { |
| | 48 | return sk->sk_user_data; |
| | 49 | } |
| | 50 | |
| | 51 | static void fou_recv_pull(struct sk_buff *skb, size_t len) |
| | 52 | { |
| | 53 | struct iphdr *iph = ip_hdr(skb); |
| | 54 | |
| | 55 | /* Remove 'len' bytes from the packet (UDP header and |
| | 56 | * FOU header if present). |
| | 57 | */ |
| | 58 | iph->tot_len = htons(ntohs(iph->tot_len) - len); |
| | 59 | __skb_pull(skb, len); |
| | 60 | skb_postpull_rcsum(skb, udp_hdr(skb), len); |
| | 61 | skb_reset_transport_header(skb); |
| | 62 | } |
| | 63 | |
| | 64 | static int fou_udp_recv(struct sock *sk, struct sk_buff *skb) |
| | 65 | { |
| | 66 | struct fou *fou = fou_from_sock(sk); |
| | 67 | |
| | 68 | if (!fou) |
| | 69 | return 1; |
| | 70 | |
| | 71 | fou_recv_pull(skb, sizeof(struct udphdr)); |
| | 72 | |
| | 73 | return -fou->protocol; |
| | 74 | } |
| | 75 | |
| | 76 | static struct guehdr *gue_remcsum(struct sk_buff *skb, struct guehdr *guehdr, |
| | 77 | void *data, size_t hdrlen, u8 ipproto, |
| | 78 | bool nopartial) |
| | 79 | { |
| | 80 | __be16 *pd = data; |
| | 81 | size_t start = ntohs(pd[0]); |
| | 82 | size_t offset = ntohs(pd[1]); |
| | 83 | size_t plen = sizeof(struct udphdr) + hdrlen + |
| | 84 | max_t(size_t, offset + sizeof(u16), start); |
| | 85 | |
| | 86 | if (skb->remcsum_offload) |
| | 87 | return guehdr; |
| | 88 | |
| | 89 | if (!pskb_may_pull(skb, plen)) |
| | 90 | return NULL; |
| | 91 | guehdr = (struct guehdr *)&udp_hdr(skb)[1]; |
| | 92 | |
| | 93 | skb_remcsum_process(skb, (void *)guehdr + hdrlen, |
| | 94 | start, offset, nopartial); |
| | 95 | |
| | 96 | return guehdr; |
| | 97 | } |
| | 98 | |
| | 99 | static int gue_control_message(struct sk_buff *skb, struct guehdr *guehdr) |
| | 100 | { |
| | 101 | /* No support yet */ |
| | 102 | kfree_skb(skb); |
| | 103 | return 0; |
| | 104 | } |
| | 105 | |
| | 106 | static int gue_udp_recv(struct sock *sk, struct sk_buff *skb) |
| | 107 | { |
| | 108 | struct fou *fou = fou_from_sock(sk); |
| | 109 | size_t len, optlen, hdrlen; |
| | 110 | struct guehdr *guehdr; |
| | 111 | void *data; |
| | 112 | u16 doffset = 0; |
| | 113 | |
| | 114 | if (!fou) |
| | 115 | return 1; |
| | 116 | |
| | 117 | len = sizeof(struct udphdr) + sizeof(struct guehdr); |
| | 118 | if (!pskb_may_pull(skb, len)) |
| | 119 | goto drop; |
| | 120 | |
| | 121 | guehdr = (struct guehdr *)&udp_hdr(skb)[1]; |
| | 122 | |
| | 123 | optlen = guehdr->hlen << 2; |
| | 124 | len += optlen; |
| | 125 | |
| | 126 | if (!pskb_may_pull(skb, len)) |
| | 127 | goto drop; |
| | 128 | |
| | 129 | /* guehdr may change after pull */ |
| | 130 | guehdr = (struct guehdr *)&udp_hdr(skb)[1]; |
| | 131 | |
| | 132 | hdrlen = sizeof(struct guehdr) + optlen; |
| | 133 | |
| | 134 | if (guehdr->version != 0 || validate_gue_flags(guehdr, optlen)) |
| | 135 | goto drop; |
| | 136 | |
| | 137 | hdrlen = sizeof(struct guehdr) + optlen; |
| | 138 | |
| | 139 | ip_hdr(skb)->tot_len = htons(ntohs(ip_hdr(skb)->tot_len) - len); |
| | 140 | |
| | 141 | /* Pull csum through the guehdr now . This can be used if |
| | 142 | * there is a remote checksum offload. |
| | 143 | */ |
| | 144 | skb_postpull_rcsum(skb, udp_hdr(skb), len); |
| | 145 | |
| | 146 | data = &guehdr[1]; |
| | 147 | |
| | 148 | if (guehdr->flags & GUE_FLAG_PRIV) { |
| | 149 | __be32 flags = *(__be32 *)(data + doffset); |
| | 150 | |
| | 151 | doffset += GUE_LEN_PRIV; |
| | 152 | |
| | 153 | if (flags & GUE_PFLAG_REMCSUM) { |
| | 154 | guehdr = gue_remcsum(skb, guehdr, data + doffset, |
| | 155 | hdrlen, guehdr->proto_ctype, |
| | 156 | !!(fou->flags & |
| | 157 | FOU_F_REMCSUM_NOPARTIAL)); |
| | 158 | if (!guehdr) |
| | 159 | goto drop; |
| | 160 | |
| | 161 | data = &guehdr[1]; |
| | 162 | |
| | 163 | doffset += GUE_PLEN_REMCSUM; |
| | 164 | } |
| | 165 | } |
| | 166 | |
| | 167 | if (unlikely(guehdr->control)) |
| | 168 | return gue_control_message(skb, guehdr); |
| | 169 | |
| | 170 | __skb_pull(skb, sizeof(struct udphdr) + hdrlen); |
| | 171 | skb_reset_transport_header(skb); |
| | 172 | |
| | 173 | return -guehdr->proto_ctype; |
| | 174 | |
| | 175 | drop: |
| | 176 | kfree_skb(skb); |
| | 177 | return 0; |
| | 178 | } |
| | 179 | |
| | 180 | static struct sk_buff **fou_gro_receive(struct sk_buff **head, |
| | 181 | struct sk_buff *skb, |
| | 182 | struct udp_offload *uoff) |
| | 183 | { |
| | 184 | const struct net_offload *ops; |
| | 185 | struct sk_buff **pp = NULL; |
| | 186 | u8 proto = NAPI_GRO_CB(skb)->proto; |
| | 187 | const struct net_offload **offloads; |
| | 188 | |
| | 189 | rcu_read_lock(); |
| | 190 | offloads = NAPI_GRO_CB(skb)->is_ipv6 ? inet6_offloads : inet_offloads; |
| | 191 | ops = rcu_dereference(offloads[proto]); |
| | 192 | if (!ops || !ops->callbacks.gro_receive) |
| | 193 | goto out_unlock; |
| | 194 | |
| | 195 | pp = ops->callbacks.gro_receive(head, skb); |
| | 196 | |
| | 197 | out_unlock: |
| | 198 | rcu_read_unlock(); |
| | 199 | |
| | 200 | return pp; |
| | 201 | } |
| | 202 | |
| | 203 | static int fou_gro_complete(struct sk_buff *skb, int nhoff, |
| | 204 | struct udp_offload *uoff) |
| | 205 | { |
| | 206 | const struct net_offload *ops; |
| | 207 | u8 proto = NAPI_GRO_CB(skb)->proto; |
| | 208 | int err = -ENOSYS; |
| | 209 | const struct net_offload **offloads; |
| | 210 | |
| | 211 | udp_tunnel_gro_complete(skb, nhoff); |
| | 212 | |
| | 213 | rcu_read_lock(); |
| | 214 | offloads = NAPI_GRO_CB(skb)->is_ipv6 ? inet6_offloads : inet_offloads; |
| | 215 | ops = rcu_dereference(offloads[proto]); |
| | 216 | if (WARN_ON(!ops || !ops->callbacks.gro_complete)) |
| | 217 | goto out_unlock; |
| | 218 | |
| | 219 | err = ops->callbacks.gro_complete(skb, nhoff); |
| | 220 | |
| | 221 | out_unlock: |
| | 222 | rcu_read_unlock(); |
| | 223 | |
| | 224 | return err; |
| | 225 | } |
| | 226 | |
| | 227 | static struct guehdr *gue_gro_remcsum(struct sk_buff *skb, unsigned int off, |
| | 228 | struct guehdr *guehdr, void *data, |
| | 229 | size_t hdrlen, struct gro_remcsum *grc, |
| | 230 | bool nopartial) |
| | 231 | { |
| | 232 | __be16 *pd = data; |
| | 233 | size_t start = ntohs(pd[0]); |
| | 234 | size_t offset = ntohs(pd[1]); |
| | 235 | |
| | 236 | if (skb->remcsum_offload) |
| | 237 | return guehdr; |
| | 238 | |
| | 239 | if (!NAPI_GRO_CB(skb)->csum_valid) |
| | 240 | return NULL; |
| | 241 | |
| | 242 | guehdr = skb_gro_remcsum_process(skb, (void *)guehdr, off, hdrlen, |
| | 243 | start, offset, grc, nopartial); |
| | 244 | |
| | 245 | skb->remcsum_offload = 1; |
| | 246 | |
| | 247 | return guehdr; |
| | 248 | } |
| | 249 | |
| | 250 | static struct sk_buff **gue_gro_receive(struct sk_buff **head, |
| | 251 | struct sk_buff *skb, |
| | 252 | struct udp_offload *uoff) |
| | 253 | { |
| | 254 | const struct net_offload **offloads; |
| | 255 | const struct net_offload *ops; |
| | 256 | struct sk_buff **pp = NULL; |
| | 257 | struct sk_buff *p; |
| | 258 | struct guehdr *guehdr; |
| | 259 | size_t len, optlen, hdrlen, off; |
| | 260 | void *data; |
| | 261 | u16 doffset = 0; |
| | 262 | int flush = 1; |
| | 263 | struct fou *fou = container_of(uoff, struct fou, udp_offloads); |
| | 264 | struct gro_remcsum grc; |
| | 265 | |
| | 266 | skb_gro_remcsum_init(&grc); |
| | 267 | |
| | 268 | off = skb_gro_offset(skb); |
| | 269 | len = off + sizeof(*guehdr); |
| | 270 | |
| | 271 | guehdr = skb_gro_header_fast(skb, off); |
| | 272 | if (skb_gro_header_hard(skb, len)) { |
| | 273 | guehdr = skb_gro_header_slow(skb, len, off); |
| | 274 | if (unlikely(!guehdr)) |
| | 275 | goto out; |
| | 276 | } |
| | 277 | |
| | 278 | optlen = guehdr->hlen << 2; |
| | 279 | len += optlen; |
| | 280 | |
| | 281 | if (skb_gro_header_hard(skb, len)) { |
| | 282 | guehdr = skb_gro_header_slow(skb, len, off); |
| | 283 | if (unlikely(!guehdr)) |
| | 284 | goto out; |
| | 285 | } |
| | 286 | |
| | 287 | if (unlikely(guehdr->control) || guehdr->version != 0 || |
| | 288 | validate_gue_flags(guehdr, optlen)) |
| | 289 | goto out; |
| | 290 | |
| | 291 | hdrlen = sizeof(*guehdr) + optlen; |
| | 292 | |
| | 293 | /* Adjust NAPI_GRO_CB(skb)->csum to account for guehdr, |
| | 294 | * this is needed if there is a remote checkcsum offload. |
| | 295 | */ |
| | 296 | skb_gro_postpull_rcsum(skb, guehdr, hdrlen); |
| | 297 | |
| | 298 | data = &guehdr[1]; |
| | 299 | |
| | 300 | if (guehdr->flags & GUE_FLAG_PRIV) { |
| | 301 | __be32 flags = *(__be32 *)(data + doffset); |
| | 302 | |
| | 303 | doffset += GUE_LEN_PRIV; |
| | 304 | |
| | 305 | if (flags & GUE_PFLAG_REMCSUM) { |
| | 306 | guehdr = gue_gro_remcsum(skb, off, guehdr, |
| | 307 | data + doffset, hdrlen, &grc, |
| | 308 | !!(fou->flags & |
| | 309 | FOU_F_REMCSUM_NOPARTIAL)); |
| | 310 | |
| | 311 | if (!guehdr) |
| | 312 | goto out; |
| | 313 | |
| | 314 | data = &guehdr[1]; |
| | 315 | |
| | 316 | doffset += GUE_PLEN_REMCSUM; |
| | 317 | } |
| | 318 | } |
| | 319 | |
| | 320 | skb_gro_pull(skb, hdrlen); |
| | 321 | |
| | 322 | for (p = *head; p; p = p->next) { |
| | 323 | const struct guehdr *guehdr2; |
| | 324 | |
| | 325 | if (!NAPI_GRO_CB(p)->same_flow) |
| | 326 | continue; |
| | 327 | |
| | 328 | guehdr2 = (struct guehdr *)(p->data + off); |
| | 329 | |
| | 330 | /* Compare base GUE header to be equal (covers |
| | 331 | * hlen, version, proto_ctype, and flags. |
| | 332 | */ |
| | 333 | if (guehdr->word != guehdr2->word) { |
| | 334 | NAPI_GRO_CB(p)->same_flow = 0; |
| | 335 | continue; |
| | 336 | } |
| | 337 | |
| | 338 | /* Compare optional fields are the same. */ |
| | 339 | if (guehdr->hlen && memcmp(&guehdr[1], &guehdr2[1], |
| | 340 | guehdr->hlen << 2)) { |
| | 341 | NAPI_GRO_CB(p)->same_flow = 0; |
| | 342 | continue; |
| | 343 | } |
| | 344 | } |
| | 345 | |
| | 346 | rcu_read_lock(); |
| | 347 | offloads = NAPI_GRO_CB(skb)->is_ipv6 ? inet6_offloads : inet_offloads; |
| | 348 | ops = rcu_dereference(offloads[guehdr->proto_ctype]); |
| | 349 | if (WARN_ON_ONCE(!ops || !ops->callbacks.gro_receive)) |
| | 350 | goto out_unlock; |
| | 351 | |
| | 352 | pp = ops->callbacks.gro_receive(head, skb); |
| | 353 | flush = 0; |
| | 354 | |
| | 355 | out_unlock: |
| | 356 | rcu_read_unlock(); |
| | 357 | out: |
| | 358 | NAPI_GRO_CB(skb)->flush |= flush; |
| | 359 | skb_gro_remcsum_cleanup(skb, &grc); |
| | 360 | |
| | 361 | return pp; |
| | 362 | } |
| | 363 | |
| | 364 | static int gue_gro_complete(struct sk_buff *skb, int nhoff, |
| | 365 | struct udp_offload *uoff) |
| | 366 | { |
| | 367 | const struct net_offload **offloads; |
| | 368 | struct guehdr *guehdr = (struct guehdr *)(skb->data + nhoff); |
| | 369 | const struct net_offload *ops; |
| | 370 | unsigned int guehlen; |
| | 371 | u8 proto; |
| | 372 | int err = -ENOENT; |
| | 373 | |
| | 374 | proto = guehdr->proto_ctype; |
| | 375 | |
| | 376 | guehlen = sizeof(*guehdr) + (guehdr->hlen << 2); |
| | 377 | |
| | 378 | rcu_read_lock(); |
| | 379 | offloads = NAPI_GRO_CB(skb)->is_ipv6 ? inet6_offloads : inet_offloads; |
| | 380 | ops = rcu_dereference(offloads[proto]); |
| | 381 | if (WARN_ON(!ops || !ops->callbacks.gro_complete)) |
| | 382 | goto out_unlock; |
| | 383 | |
| | 384 | err = ops->callbacks.gro_complete(skb, nhoff + guehlen); |
| | 385 | |
| | 386 | out_unlock: |
| | 387 | rcu_read_unlock(); |
| | 388 | return err; |
| | 389 | } |
| | 390 | |
| | 391 | static int fou_add_to_port_list(struct net *net, struct fou *fou) |
| | 392 | { |
| | 393 | struct fou_net *fn = net_generic(net, fou_net_id); |
| | 394 | struct fou *fout; |
| | 395 | |
| | 396 | mutex_lock(&fn->fou_lock); |
| | 397 | list_for_each_entry(fout, &fn->fou_list, list) { |
| | 398 | if (fou->port == fout->port) { |
| | 399 | mutex_unlock(&fn->fou_lock); |
| | 400 | return -EALREADY; |
| | 401 | } |
| | 402 | } |
| | 403 | |
| | 404 | list_add(&fou->list, &fn->fou_list); |
| | 405 | mutex_unlock(&fn->fou_lock); |
| | 406 | |
| | 407 | return 0; |
| | 408 | } |
| | 409 | |
| | 410 | static void fou_release(struct fou *fou) |
| | 411 | { |
| | 412 | struct socket *sock = fou->sock; |
| | 413 | struct sock *sk = sock->sk; |
| | 414 | |
| | 415 | if (sk->sk_family == AF_INET) |
| | 416 | udp_del_offload(&fou->udp_offloads); |
| | 417 | list_del(&fou->list); |
| | 418 | udp_tunnel_sock_release(sock); |
| | 419 | |
| | 420 | kfree_rcu(fou, rcu); |
| | 421 | } |
| | 422 | |
| | 423 | static int fou_encap_init(struct sock *sk, struct fou *fou, struct fou_cfg *cfg) |
| | 424 | { |
| | 425 | udp_sk(sk)->encap_rcv = fou_udp_recv; |
| | 426 | fou->protocol = cfg->protocol; |
| | 427 | fou->udp_offloads.callbacks.gro_receive = fou_gro_receive; |
| | 428 | fou->udp_offloads.callbacks.gro_complete = fou_gro_complete; |
| | 429 | fou->udp_offloads.port = cfg->udp_config.local_udp_port; |
| | 430 | fou->udp_offloads.ipproto = cfg->protocol; |
| | 431 | |
| | 432 | return 0; |
| | 433 | } |
| | 434 | |
| | 435 | static int gue_encap_init(struct sock *sk, struct fou *fou, struct fou_cfg *cfg) |
| | 436 | { |
| | 437 | udp_sk(sk)->encap_rcv = gue_udp_recv; |
| | 438 | fou->udp_offloads.callbacks.gro_receive = gue_gro_receive; |
| | 439 | fou->udp_offloads.callbacks.gro_complete = gue_gro_complete; |
| | 440 | fou->udp_offloads.port = cfg->udp_config.local_udp_port; |
| | 441 | |
| | 442 | return 0; |
| | 443 | } |
| | 444 | |
| | 445 | static int fou_create(struct net *net, struct fou_cfg *cfg, |
| | 446 | struct socket **sockp) |
| | 447 | { |
| | 448 | struct socket *sock = NULL; |
| | 449 | struct fou *fou = NULL; |
| | 450 | struct sock *sk; |
| | 451 | int err; |
| | 452 | |
| | 453 | /* Open UDP socket */ |
| | 454 | err = udp_sock_create(net, &cfg->udp_config, &sock); |
| | 455 | if (err < 0) |
| | 456 | goto error; |
| | 457 | |
| | 458 | /* Allocate FOU port structure */ |
| | 459 | fou = kzalloc(sizeof(*fou), GFP_KERNEL); |
| | 460 | if (!fou) { |
| | 461 | err = -ENOMEM; |
| | 462 | goto error; |
| | 463 | } |
| | 464 | |
| | 465 | sk = sock->sk; |
| | 466 | |
| | 467 | fou->flags = cfg->flags; |
| | 468 | fou->port = cfg->udp_config.local_udp_port; |
| | 469 | |
| | 470 | /* Initial for fou type */ |
| | 471 | switch (cfg->type) { |
| | 472 | case FOU_ENCAP_DIRECT: |
| | 473 | err = fou_encap_init(sk, fou, cfg); |
| | 474 | if (err) |
| | 475 | goto error; |
| | 476 | break; |
| | 477 | case FOU_ENCAP_GUE: |
| | 478 | err = gue_encap_init(sk, fou, cfg); |
| | 479 | if (err) |
| | 480 | goto error; |
| | 481 | break; |
| | 482 | default: |
| | 483 | err = -EINVAL; |
| | 484 | goto error; |
| | 485 | } |
| | 486 | |
| | 487 | fou->type = cfg->type; |
| | 488 | |
| | 489 | udp_sk(sk)->encap_type = 1; |
| | 490 | udp_encap_enable(); |
| | 491 | |
| | 492 | sk->sk_user_data = fou; |
| | 493 | fou->sock = sock; |
| | 494 | |
| | 495 | inet_inc_convert_csum(sk); |
| | 496 | |
| | 497 | sk->sk_allocation = GFP_ATOMIC; |
| | 498 | |
| | 499 | if (cfg->udp_config.family == AF_INET) { |
| | 500 | err = udp_add_offload(net, &fou->udp_offloads); |
| | 501 | if (err) |
| | 502 | goto error; |
| | 503 | } |
| | 504 | |
| | 505 | err = fou_add_to_port_list(net, fou); |
| | 506 | if (err) |
| | 507 | goto error; |
| | 508 | |
| | 509 | if (sockp) |
| | 510 | *sockp = sock; |
| | 511 | |
| | 512 | return 0; |
| | 513 | |
| | 514 | error: |
| | 515 | kfree(fou); |
| | 516 | if (sock) |
| | 517 | udp_tunnel_sock_release(sock); |
| | 518 | |
| | 519 | return err; |
| | 520 | } |
| | 521 | |
| | 522 | static int fou_destroy(struct net *net, struct fou_cfg *cfg) |
| | 523 | { |
| | 524 | struct fou_net *fn = net_generic(net, fou_net_id); |
| | 525 | __be16 port = cfg->udp_config.local_udp_port; |
| | 526 | int err = -EINVAL; |
| | 527 | struct fou *fou; |
| | 528 | |
| | 529 | mutex_lock(&fn->fou_lock); |
| | 530 | list_for_each_entry(fou, &fn->fou_list, list) { |
| | 531 | if (fou->port == port) { |
| | 532 | fou_release(fou); |
| | 533 | err = 0; |
| | 534 | break; |
| | 535 | } |
| | 536 | } |
| | 537 | mutex_unlock(&fn->fou_lock); |
| | 538 | |
| | 539 | return err; |
| | 540 | } |
| | 541 | |
| | 542 | static struct genl_family fou_nl_family = { |
| | 543 | .id = GENL_ID_GENERATE, |
| | 544 | .hdrsize = 0, |
| | 545 | .name = FOU_GENL_NAME, |
| | 546 | .version = FOU_GENL_VERSION, |
| | 547 | .maxattr = FOU_ATTR_MAX, |
| | 548 | .netnsok = true, |
| | 549 | }; |
| | 550 | |
| | 551 | static struct nla_policy fou_nl_policy[FOU_ATTR_MAX + 1] = { |
| | 552 | [FOU_ATTR_PORT] = { .type = NLA_U16, }, |
| | 553 | [FOU_ATTR_AF] = { .type = NLA_U8, }, |
| | 554 | [FOU_ATTR_IPPROTO] = { .type = NLA_U8, }, |
| | 555 | [FOU_ATTR_TYPE] = { .type = NLA_U8, }, |
| | 556 | [FOU_ATTR_REMCSUM_NOPARTIAL] = { .type = NLA_FLAG, }, |
| | 557 | }; |
| | 558 | |
| | 559 | static int parse_nl_config(struct genl_info *info, |
| | 560 | struct fou_cfg *cfg) |
| | 561 | { |
| | 562 | memset(cfg, 0, sizeof(*cfg)); |
| | 563 | |
| | 564 | cfg->udp_config.family = AF_INET; |
| | 565 | |
| | 566 | if (info->attrs[FOU_ATTR_AF]) { |
| | 567 | u8 family = nla_get_u8(info->attrs[FOU_ATTR_AF]); |
| | 568 | |
| | 569 | if (family != AF_INET) |
| | 570 | return -EINVAL; |
| | 571 | |
| | 572 | cfg->udp_config.family = family; |
| | 573 | } |
| | 574 | |
| | 575 | if (info->attrs[FOU_ATTR_PORT]) { |
| | 576 | __be16 port = nla_get_be16(info->attrs[FOU_ATTR_PORT]); |
| | 577 | |
| | 578 | cfg->udp_config.local_udp_port = port; |
| | 579 | } |
| | 580 | |
| | 581 | if (info->attrs[FOU_ATTR_IPPROTO]) |
| | 582 | cfg->protocol = nla_get_u8(info->attrs[FOU_ATTR_IPPROTO]); |
| | 583 | |
| | 584 | if (info->attrs[FOU_ATTR_TYPE]) |
| | 585 | cfg->type = nla_get_u8(info->attrs[FOU_ATTR_TYPE]); |
| | 586 | |
| | 587 | if (info->attrs[FOU_ATTR_REMCSUM_NOPARTIAL]) |
| | 588 | cfg->flags |= FOU_F_REMCSUM_NOPARTIAL; |
| | 589 | |
| | 590 | return 0; |
| | 591 | } |
| | 592 | |
| | 593 | static int fou_nl_cmd_add_port(struct sk_buff *skb, struct genl_info *info) |
| | 594 | { |
| | 595 | struct net *net = genl_info_net(info); |
| | 596 | struct fou_cfg cfg; |
| | 597 | int err; |
| | 598 | |
| | 599 | err = parse_nl_config(info, &cfg); |
| | 600 | if (err) |
| | 601 | return err; |
| | 602 | |
| | 603 | return fou_create(net, &cfg, NULL); |
| | 604 | } |
| | 605 | |
| | 606 | static int fou_nl_cmd_rm_port(struct sk_buff *skb, struct genl_info *info) |
| | 607 | { |
| | 608 | struct net *net = genl_info_net(info); |
| | 609 | struct fou_cfg cfg; |
| | 610 | int err; |
| | 611 | |
| | 612 | err = parse_nl_config(info, &cfg); |
| | 613 | if (err) |
| | 614 | return err; |
| | 615 | |
| | 616 | return fou_destroy(net, &cfg); |
| | 617 | } |
| | 618 | |
| | 619 | static int fou_fill_info(struct fou *fou, struct sk_buff *msg) |
| | 620 | { |
| | 621 | if (nla_put_u8(msg, FOU_ATTR_AF, fou->sock->sk->sk_family) || |
| | 622 | nla_put_be16(msg, FOU_ATTR_PORT, fou->port) || |
| | 623 | nla_put_u8(msg, FOU_ATTR_IPPROTO, fou->protocol) || |
| | 624 | nla_put_u8(msg, FOU_ATTR_TYPE, fou->type)) |
| | 625 | return -1; |
| | 626 | |
| | 627 | if (fou->flags & FOU_F_REMCSUM_NOPARTIAL) |
| | 628 | if (nla_put_flag(msg, FOU_ATTR_REMCSUM_NOPARTIAL)) |
| | 629 | return -1; |
| | 630 | return 0; |
| | 631 | } |
| | 632 | |
| | 633 | static int fou_dump_info(struct fou *fou, u32 portid, u32 seq, |
| | 634 | u32 flags, struct sk_buff *skb, u8 cmd) |
| | 635 | { |
| | 636 | void *hdr; |
| | 637 | |
| | 638 | hdr = genlmsg_put(skb, portid, seq, &fou_nl_family, flags, cmd); |
| | 639 | if (!hdr) |
| | 640 | return -ENOMEM; |
| | 641 | |
| | 642 | if (fou_fill_info(fou, skb) < 0) |
| | 643 | goto nla_put_failure; |
| | 644 | |
| | 645 | genlmsg_end(skb, hdr); |
| | 646 | return 0; |
| | 647 | |
| | 648 | nla_put_failure: |
| | 649 | genlmsg_cancel(skb, hdr); |
| | 650 | return -EMSGSIZE; |
| | 651 | } |
| | 652 | |
| | 653 | static int fou_nl_cmd_get_port(struct sk_buff *skb, struct genl_info *info) |
| | 654 | { |
| | 655 | struct net *net = genl_info_net(info); |
| | 656 | struct fou_net *fn = net_generic(net, fou_net_id); |
| | 657 | struct sk_buff *msg; |
| | 658 | struct fou_cfg cfg; |
| | 659 | struct fou *fout; |
| | 660 | __be16 port; |
| | 661 | int ret; |
| | 662 | |
| | 663 | ret = parse_nl_config(info, &cfg); |
| | 664 | if (ret) |
| | 665 | return ret; |
| | 666 | port = cfg.udp_config.local_udp_port; |
| | 667 | if (port == 0) |
| | 668 | return -EINVAL; |
| | 669 | |
| | 670 | msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL); |
| | 671 | if (!msg) |
| | 672 | return -ENOMEM; |
| | 673 | |
| | 674 | ret = -ESRCH; |
| | 675 | mutex_lock(&fn->fou_lock); |
| | 676 | list_for_each_entry(fout, &fn->fou_list, list) { |
| | 677 | if (port == fout->port) { |
| | 678 | ret = fou_dump_info(fout, info->snd_portid, |
| | 679 | info->snd_seq, 0, msg, |
| | 680 | info->genlhdr->cmd); |
| | 681 | break; |
| | 682 | } |
| | 683 | } |
| | 684 | mutex_unlock(&fn->fou_lock); |
| | 685 | if (ret < 0) |
| | 686 | goto out_free; |
| | 687 | |
| | 688 | return genlmsg_reply(msg, info); |
| | 689 | |
| | 690 | out_free: |
| | 691 | nlmsg_free(msg); |
| | 692 | return ret; |
| | 693 | } |
| | 694 | |
| | 695 | static int fou_nl_dump(struct sk_buff *skb, struct netlink_callback *cb) |
| | 696 | { |
| | 697 | struct net *net = sock_net(skb->sk); |
| | 698 | struct fou_net *fn = net_generic(net, fou_net_id); |
| | 699 | struct fou *fout; |
| | 700 | int idx = 0, ret; |
| | 701 | |
| | 702 | mutex_lock(&fn->fou_lock); |
| | 703 | list_for_each_entry(fout, &fn->fou_list, list) { |
| | 704 | if (idx++ < cb->args[0]) |
| | 705 | continue; |
| | 706 | ret = fou_dump_info(fout, NETLINK_CB(cb->skb).portid, |
| | 707 | cb->nlh->nlmsg_seq, NLM_F_MULTI, |
| | 708 | skb, FOU_CMD_GET); |
| | 709 | if (ret) |
| | 710 | break; |
| | 711 | } |
| | 712 | mutex_unlock(&fn->fou_lock); |
| | 713 | |
| | 714 | cb->args[0] = idx; |
| | 715 | return skb->len; |
| | 716 | } |
| | 717 | |
| | 718 | static const struct genl_ops fou_nl_ops[] = { |
| | 719 | { |
| | 720 | .cmd = FOU_CMD_ADD, |
| | 721 | .doit = fou_nl_cmd_add_port, |
| | 722 | .policy = fou_nl_policy, |
| | 723 | .flags = GENL_ADMIN_PERM, |
| | 724 | }, |
| | 725 | { |
| | 726 | .cmd = FOU_CMD_DEL, |
| | 727 | .doit = fou_nl_cmd_rm_port, |
| | 728 | .policy = fou_nl_policy, |
| | 729 | .flags = GENL_ADMIN_PERM, |
| | 730 | }, |
| | 731 | { |
| | 732 | .cmd = FOU_CMD_GET, |
| | 733 | .doit = fou_nl_cmd_get_port, |
| | 734 | .dumpit = fou_nl_dump, |
| | 735 | .policy = fou_nl_policy, |
| | 736 | }, |
| | 737 | }; |
| | 738 | |
| | 739 | size_t fou_encap_hlen(struct ip_tunnel_encap *e) |
| | 740 | { |
| | 741 | return sizeof(struct udphdr); |
| | 742 | } |
| | 743 | EXPORT_SYMBOL(fou_encap_hlen); |
| | 744 | |
| | 745 | size_t gue_encap_hlen(struct ip_tunnel_encap *e) |
| | 746 | { |
| | 747 | size_t len; |
| | 748 | bool need_priv = false; |
| | 749 | |
| | 750 | len = sizeof(struct udphdr) + sizeof(struct guehdr); |
| | 751 | |
| | 752 | if (e->flags & TUNNEL_ENCAP_FLAG_REMCSUM) { |
| | 753 | len += GUE_PLEN_REMCSUM; |
| | 754 | need_priv = true; |
| | 755 | } |
| | 756 | |
| | 757 | len += need_priv ? GUE_LEN_PRIV : 0; |
| | 758 | |
| | 759 | return len; |
| | 760 | } |
| | 761 | EXPORT_SYMBOL(gue_encap_hlen); |
| | 762 | |
| | 763 | static void fou_build_udp(struct sk_buff *skb, struct ip_tunnel_encap *e, |
| | 764 | struct flowi4 *fl4, u8 *protocol, __be16 sport) |
| | 765 | { |
| | 766 | struct udphdr *uh; |
| | 767 | |
| | 768 | skb_push(skb, sizeof(struct udphdr)); |
| | 769 | skb_reset_transport_header(skb); |
| | 770 | |
| | 771 | uh = udp_hdr(skb); |
| | 772 | |
| | 773 | uh->dest = e->dport; |
| | 774 | uh->source = sport; |
| | 775 | uh->len = htons(skb->len); |
| | 776 | udp_set_csum(!(e->flags & TUNNEL_ENCAP_FLAG_CSUM), skb, |
| | 777 | fl4->saddr, fl4->daddr, skb->len); |
| | 778 | |
| | 779 | *protocol = IPPROTO_UDP; |
| | 780 | } |
| | 781 | |
| | 782 | int fou_build_header(struct sk_buff *skb, struct ip_tunnel_encap *e, |
| | 783 | u8 *protocol, struct flowi4 *fl4) |
| | 784 | { |
| | 785 | int type = e->flags & TUNNEL_ENCAP_FLAG_CSUM ? SKB_GSO_UDP_TUNNEL_CSUM : |
| | 786 | SKB_GSO_UDP_TUNNEL; |
| | 787 | __be16 sport; |
| | 788 | |
| | 789 | skb = iptunnel_handle_offloads(skb, type); |
| | 790 | |
| | 791 | if (IS_ERR(skb)) |
| | 792 | return PTR_ERR(skb); |
| | 793 | |
| | 794 | sport = e->sport ? : udp_flow_src_port(dev_net(skb->dev), |
| | 795 | skb, 0, 0, false); |
| | 796 | fou_build_udp(skb, e, fl4, protocol, sport); |
| | 797 | |
| | 798 | return 0; |
| | 799 | } |
| | 800 | EXPORT_SYMBOL(fou_build_header); |
| | 801 | |
| | 802 | int gue_build_header(struct sk_buff *skb, struct ip_tunnel_encap *e, |
| | 803 | u8 *protocol, struct flowi4 *fl4) |
| | 804 | { |
| | 805 | int type = e->flags & TUNNEL_ENCAP_FLAG_CSUM ? SKB_GSO_UDP_TUNNEL_CSUM : |
| | 806 | SKB_GSO_UDP_TUNNEL; |
| | 807 | struct guehdr *guehdr; |
| | 808 | size_t hdrlen, optlen = 0; |
| | 809 | __be16 sport; |
| | 810 | void *data; |
| | 811 | bool need_priv = false; |
| | 812 | |
| | 813 | if ((e->flags & TUNNEL_ENCAP_FLAG_REMCSUM) && |
| | 814 | skb->ip_summed == CHECKSUM_PARTIAL) { |
| | 815 | optlen += GUE_PLEN_REMCSUM; |
| | 816 | type |= SKB_GSO_TUNNEL_REMCSUM; |
| | 817 | need_priv = true; |
| | 818 | } |
| | 819 | |
| | 820 | optlen += need_priv ? GUE_LEN_PRIV : 0; |
| | 821 | |
| | 822 | skb = iptunnel_handle_offloads(skb, type); |
| | 823 | |
| | 824 | if (IS_ERR(skb)) |
| | 825 | return PTR_ERR(skb); |
| | 826 | |
| | 827 | /* Get source port (based on flow hash) before skb_push */ |
| | 828 | sport = e->sport ? : udp_flow_src_port(dev_net(skb->dev), |
| | 829 | skb, 0, 0, false); |
| | 830 | |
| | 831 | hdrlen = sizeof(struct guehdr) + optlen; |
| | 832 | |
| | 833 | skb_push(skb, hdrlen); |
| | 834 | |
| | 835 | guehdr = (struct guehdr *)skb->data; |
| | 836 | |
| | 837 | guehdr->control = 0; |
| | 838 | guehdr->version = 0; |
| | 839 | guehdr->hlen = optlen >> 2; |
| | 840 | guehdr->flags = 0; |
| | 841 | guehdr->proto_ctype = *protocol; |
| | 842 | |
| | 843 | data = &guehdr[1]; |
| | 844 | |
| | 845 | if (need_priv) { |
| | 846 | __be32 *flags = data; |
| | 847 | |
| | 848 | guehdr->flags |= GUE_FLAG_PRIV; |
| | 849 | *flags = 0; |
| | 850 | data += GUE_LEN_PRIV; |
| | 851 | |
| | 852 | if (type & SKB_GSO_TUNNEL_REMCSUM) { |
| | 853 | u16 csum_start = skb_checksum_start_offset(skb); |
| | 854 | __be16 *pd = data; |
| | 855 | |
| | 856 | if (csum_start < hdrlen) |
| | 857 | return -EINVAL; |
| | 858 | |
| | 859 | csum_start -= hdrlen; |
| | 860 | pd[0] = htons(csum_start); |
| | 861 | pd[1] = htons(csum_start + skb->csum_offset); |
| | 862 | |
| | 863 | if (!skb_is_gso(skb)) { |
| | 864 | skb->ip_summed = CHECKSUM_NONE; |
| | 865 | skb->encapsulation = 0; |
| | 866 | } |
| | 867 | |
| | 868 | *flags |= GUE_PFLAG_REMCSUM; |
| | 869 | data += GUE_PLEN_REMCSUM; |
| | 870 | } |
| | 871 | |
| | 872 | } |
| | 873 | |
| | 874 | fou_build_udp(skb, e, fl4, protocol, sport); |
| | 875 | |
| | 876 | return 0; |
| | 877 | } |
| | 878 | EXPORT_SYMBOL(gue_build_header); |
| | 879 | |
| | 880 | #ifdef CONFIG_NET_FOU_IP_TUNNELS |
| | 881 | |
| | 882 | static const struct ip_tunnel_encap_ops fou_iptun_ops = { |
| | 883 | .encap_hlen = fou_encap_hlen, |
| | 884 | .build_header = fou_build_header, |
| | 885 | }; |
| | 886 | |
| | 887 | static const struct ip_tunnel_encap_ops gue_iptun_ops = { |
| | 888 | .encap_hlen = gue_encap_hlen, |
| | 889 | .build_header = gue_build_header, |
| | 890 | }; |
| | 891 | |
| | 892 | static int ip_tunnel_encap_add_fou_ops(void) |
| | 893 | { |
| | 894 | int ret; |
| | 895 | |
| | 896 | ret = ip_tunnel_encap_add_ops(&fou_iptun_ops, TUNNEL_ENCAP_FOU); |
| | 897 | if (ret < 0) { |
| | 898 | pr_err("can't add fou ops\n"); |
| | 899 | return ret; |
| | 900 | } |
| | 901 | |
| | 902 | ret = ip_tunnel_encap_add_ops(&gue_iptun_ops, TUNNEL_ENCAP_GUE); |
| | 903 | if (ret < 0) { |
| | 904 | pr_err("can't add gue ops\n"); |
| | 905 | ip_tunnel_encap_del_ops(&fou_iptun_ops, TUNNEL_ENCAP_FOU); |
| | 906 | return ret; |
| | 907 | } |
| | 908 | |
| | 909 | return 0; |
| | 910 | } |
| | 911 | |
| | 912 | static void ip_tunnel_encap_del_fou_ops(void) |
| | 913 | { |
| | 914 | ip_tunnel_encap_del_ops(&fou_iptun_ops, TUNNEL_ENCAP_FOU); |
| | 915 | ip_tunnel_encap_del_ops(&gue_iptun_ops, TUNNEL_ENCAP_GUE); |
| | 916 | } |
| | 917 | |
| | 918 | #else |
| | 919 | |
| | 920 | static int ip_tunnel_encap_add_fou_ops(void) |
| | 921 | { |
| | 922 | return 0; |
| | 923 | } |
| | 924 | |
| | 925 | static void ip_tunnel_encap_del_fou_ops(void) |
| | 926 | { |
| | 927 | } |
| | 928 | |
| | 929 | #endif |
| | 930 | |
| | 931 | static __net_init int fou_init_net(struct net *net) |
| | 932 | { |
| | 933 | struct fou_net *fn = net_generic(net, fou_net_id); |
| | 934 | |
| | 935 | INIT_LIST_HEAD(&fn->fou_list); |
| | 936 | mutex_init(&fn->fou_lock); |
| | 937 | return 0; |
| | 938 | } |
| | 939 | |
| | 940 | static __net_exit void fou_exit_net(struct net *net) |
| | 941 | { |
| | 942 | struct fou_net *fn = net_generic(net, fou_net_id); |
| | 943 | struct fou *fou, *next; |
| | 944 | |
| | 945 | /* Close all the FOU sockets */ |
| | 946 | mutex_lock(&fn->fou_lock); |
| | 947 | list_for_each_entry_safe(fou, next, &fn->fou_list, list) |
| | 948 | fou_release(fou); |
| | 949 | mutex_unlock(&fn->fou_lock); |
| | 950 | } |
| | 951 | |
| | 952 | static struct pernet_operations fou_net_ops = { |
| | 953 | .init = fou_init_net, |
| | 954 | .exit = fou_exit_net, |
| | 955 | .id = &fou_net_id, |
| | 956 | .size = sizeof(struct fou_net), |
| | 957 | }; |
| | 958 | |
| | 959 | static int __init fou_init(void) |
| | 960 | { |
| | 961 | int ret; |
| | 962 | |
| | 963 | ret = register_pernet_device(&fou_net_ops); |
| | 964 | if (ret) |
| | 965 | goto exit; |
| | 966 | |
| | 967 | ret = genl_register_family_with_ops(&fou_nl_family, |
| | 968 | fou_nl_ops); |
| | 969 | if (ret < 0) |
| | 970 | goto unregister; |
| | 971 | |
| | 972 | ret = ip_tunnel_encap_add_fou_ops(); |
| | 973 | if (ret == 0) |
| | 974 | return 0; |
| | 975 | |
| | 976 | genl_unregister_family(&fou_nl_family); |
| | 977 | unregister: |
| | 978 | unregister_pernet_device(&fou_net_ops); |
| | 979 | exit: |
| | 980 | return ret; |
| | 981 | } |
| | 982 | |
| | 983 | static void __exit fou_fini(void) |
| | 984 | { |
| | 985 | ip_tunnel_encap_del_fou_ops(); |
| | 986 | genl_unregister_family(&fou_nl_family); |
| | 987 | unregister_pernet_device(&fou_net_ops); |
| | 988 | } |
| | 989 | |
| | 990 | module_init(fou_init); |
| | 991 | module_exit(fou_fini); |
| | 992 | MODULE_AUTHOR("Tom Herbert <therbert@google.com>"); |
| | 993 | MODULE_LICENSE("GPL"); |
projects / deliverable / linux.git / blame_incremental