1 /* SPDX-License-Identifier: GPL-2.0-or-later */ 2 /* 3 * INET An implementation of the TCP/IP protocol suite for the LINUX 4 * operating system. INET is implemented using the BSD Socket 5 * interface as the means of communication with the user level. 6 * 7 * Definitions for inet_sock 8 * 9 * Authors: Many, reorganised here by 10 * Arnaldo Carvalho de Melo <acme@mandriva.com> 11 */ 12 #ifndef _INET_SOCK_H 13 #define _INET_SOCK_H 14 15 #include <linux/bitops.h> 16 #include <linux/string.h> 17 #include <linux/types.h> 18 #include <linux/jhash.h> 19 #include <linux/netdevice.h> 20 21 #include <net/flow.h> 22 #include <net/inet_dscp.h> 23 #include <net/sock.h> 24 #include <net/request_sock.h> 25 #include <net/netns/hash.h> 26 #include <net/tcp_states.h> 27 #include <net/l3mdev.h> 28 29 #define IP_OPTIONS_DATA_FIXED_SIZE 40 30 31 /** struct ip_options - IP Options 32 * 33 * @faddr - Saved first hop address 34 * @nexthop - Saved nexthop address in LSRR and SSRR 35 * @is_strictroute - Strict source route 36 * @srr_is_hit - Packet destination addr was our one 37 * @is_changed - IP checksum more not valid 38 * @rr_needaddr - Need to record addr of outgoing dev 39 * @ts_needtime - Need to record timestamp 40 * @ts_needaddr - Need to record addr of outgoing dev 41 */ 42 struct ip_options { 43 __be32 faddr; 44 __be32 nexthop; 45 unsigned char optlen; 46 unsigned char srr; 47 unsigned char rr; 48 unsigned char ts; 49 unsigned char is_strictroute:1, 50 srr_is_hit:1, 51 is_changed:1, 52 rr_needaddr:1, 53 ts_needtime:1, 54 ts_needaddr:1; 55 unsigned char router_alert; 56 unsigned char cipso; 57 unsigned char __pad2; 58 unsigned char __data[]; 59 }; 60 61 struct ip_options_rcu { 62 struct rcu_head rcu; 63 64 /* Must be last as it ends in a flexible-array member. */ 65 struct ip_options opt; 66 }; 67 68 struct inet_request_sock { 69 struct request_sock req; 70 #define ir_loc_addr req.__req_common.skc_rcv_saddr 71 #define ir_rmt_addr req.__req_common.skc_daddr 72 #define ir_num req.__req_common.skc_num 73 #define ir_rmt_port req.__req_common.skc_dport 74 #define ir_v6_rmt_addr req.__req_common.skc_v6_daddr 75 #define ir_v6_loc_addr req.__req_common.skc_v6_rcv_saddr 76 #define ir_iif req.__req_common.skc_bound_dev_if 77 #define ir_cookie req.__req_common.skc_cookie 78 #define ireq_net req.__req_common.skc_net 79 #define ireq_state req.__req_common.skc_state 80 #define ireq_family req.__req_common.skc_family 81 82 u16 snd_wscale : 4, 83 rcv_wscale : 4, 84 tstamp_ok : 1, 85 sack_ok : 1, 86 wscale_ok : 1, 87 ecn_ok : 1, 88 acked : 1, 89 no_srccheck: 1, 90 smc_ok : 1; 91 u32 ir_mark; 92 union { 93 struct ip_options_rcu __rcu *ireq_opt; 94 #if IS_ENABLED(CONFIG_IPV6) 95 struct { 96 struct ipv6_txoptions *ipv6_opt; 97 struct sk_buff *pktopts; 98 }; 99 #endif 100 }; 101 }; 102 103 static inline struct inet_request_sock *inet_rsk(const struct request_sock *sk) 104 { 105 return (struct inet_request_sock *)sk; 106 } 107 108 static inline u32 inet_request_mark(const struct sock *sk, struct sk_buff *skb) 109 { 110 u32 mark = READ_ONCE(sk->sk_mark); 111 112 if (!mark && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fwmark_accept)) 113 return skb->mark; 114 115 return mark; 116 } 117 118 static inline int inet_request_bound_dev_if(const struct sock *sk, 119 struct sk_buff *skb) 120 { 121 int bound_dev_if = READ_ONCE(sk->sk_bound_dev_if); 122 #ifdef CONFIG_NET_L3_MASTER_DEV 123 struct net *net = sock_net(sk); 124 125 if (!bound_dev_if && READ_ONCE(net->ipv4.sysctl_tcp_l3mdev_accept)) 126 return l3mdev_master_ifindex_by_index(net, skb->skb_iif); 127 #endif 128 129 return bound_dev_if; 130 } 131 132 static inline int inet_sk_bound_l3mdev(const struct sock *sk) 133 { 134 #ifdef CONFIG_NET_L3_MASTER_DEV 135 struct net *net = sock_net(sk); 136 137 if (!READ_ONCE(net->ipv4.sysctl_tcp_l3mdev_accept)) 138 return l3mdev_master_ifindex_by_index(net, 139 sk->sk_bound_dev_if); 140 #endif 141 142 return 0; 143 } 144 145 static inline bool inet_bound_dev_eq(bool l3mdev_accept, int bound_dev_if, 146 int dif, int sdif) 147 { 148 if (!bound_dev_if) 149 return !sdif || l3mdev_accept; 150 return bound_dev_if == dif || bound_dev_if == sdif; 151 } 152 153 static inline bool inet_sk_bound_dev_eq(const struct net *net, 154 int bound_dev_if, 155 int dif, int sdif) 156 { 157 #if IS_ENABLED(CONFIG_NET_L3_MASTER_DEV) 158 return inet_bound_dev_eq(!!READ_ONCE(net->ipv4.sysctl_tcp_l3mdev_accept), 159 bound_dev_if, dif, sdif); 160 #else 161 return inet_bound_dev_eq(true, bound_dev_if, dif, sdif); 162 #endif 163 } 164 165 struct inet_cork { 166 unsigned int flags; 167 __be32 addr; 168 struct ip_options *opt; 169 unsigned int fragsize; 170 int length; /* Total length of all frames */ 171 struct dst_entry *dst; 172 u8 tx_flags; 173 __u8 ttl; 174 __s16 tos; 175 u32 priority; 176 __u16 gso_size; 177 u32 ts_opt_id; 178 u64 transmit_time; 179 u32 mark; 180 }; 181 182 struct inet_cork_full { 183 struct inet_cork base; 184 struct flowi fl; 185 }; 186 187 struct ip_mc_socklist; 188 struct ipv6_pinfo; 189 struct rtable; 190 191 /** struct inet_sock - representation of INET sockets 192 * 193 * @sk - ancestor class 194 * @pinet6 - pointer to IPv6 control block 195 * @inet_daddr - Foreign IPv4 addr 196 * @inet_rcv_saddr - Bound local IPv4 addr 197 * @inet_dport - Destination port 198 * @inet_num - Local port 199 * @inet_flags - various atomic flags 200 * @inet_saddr - Sending source 201 * @uc_ttl - Unicast TTL 202 * @inet_sport - Source port 203 * @inet_id - ID counter for DF pkts 204 * @tos - TOS 205 * @mc_ttl - Multicasting TTL 206 * @uc_index - Unicast outgoing device index 207 * @mc_index - Multicast device index 208 * @mc_list - Group array 209 * @cork - info to build ip hdr on each ip frag while socket is corked 210 */ 211 struct inet_sock { 212 /* sk and pinet6 has to be the first two members of inet_sock */ 213 struct sock sk; 214 #if IS_ENABLED(CONFIG_IPV6) 215 struct ipv6_pinfo *pinet6; 216 struct ipv6_fl_socklist __rcu *ipv6_fl_list; 217 #endif 218 /* Socket demultiplex comparisons on incoming packets. */ 219 #define inet_daddr sk.__sk_common.skc_daddr 220 #define inet_rcv_saddr sk.__sk_common.skc_rcv_saddr 221 #define inet_dport sk.__sk_common.skc_dport 222 #define inet_num sk.__sk_common.skc_num 223 224 unsigned long inet_flags; 225 __be32 inet_saddr; 226 __s16 uc_ttl; 227 __be16 inet_sport; 228 struct ip_options_rcu __rcu *inet_opt; 229 atomic_t inet_id; 230 231 __u8 tos; 232 __u8 min_ttl; 233 __u8 mc_ttl; 234 __u8 pmtudisc; 235 __u8 rcv_tos; 236 __u8 convert_csum; 237 int uc_index; 238 int mc_index; 239 __be32 mc_addr; 240 u32 local_port_range; /* high << 16 | low */ 241 242 struct ip_mc_socklist __rcu *mc_list; 243 struct inet_cork_full cork; 244 }; 245 246 #define IPCORK_OPT 1 /* ip-options has been held in ipcork.opt */ 247 #define IPCORK_TS_OPT_ID 2 /* ts_opt_id field is valid, overriding sk_tskey */ 248 249 enum { 250 INET_FLAGS_PKTINFO = 0, 251 INET_FLAGS_TTL = 1, 252 INET_FLAGS_TOS = 2, 253 INET_FLAGS_RECVOPTS = 3, 254 INET_FLAGS_RETOPTS = 4, 255 INET_FLAGS_PASSSEC = 5, 256 INET_FLAGS_ORIGDSTADDR = 6, 257 INET_FLAGS_CHECKSUM = 7, 258 INET_FLAGS_RECVFRAGSIZE = 8, 259 260 INET_FLAGS_RECVERR = 9, 261 INET_FLAGS_RECVERR_RFC4884 = 10, 262 INET_FLAGS_FREEBIND = 11, 263 INET_FLAGS_HDRINCL = 12, 264 INET_FLAGS_MC_LOOP = 13, 265 INET_FLAGS_MC_ALL = 14, 266 INET_FLAGS_TRANSPARENT = 15, 267 INET_FLAGS_IS_ICSK = 16, 268 INET_FLAGS_NODEFRAG = 17, 269 INET_FLAGS_BIND_ADDRESS_NO_PORT = 18, 270 INET_FLAGS_DEFER_CONNECT = 19, 271 INET_FLAGS_MC6_LOOP = 20, 272 INET_FLAGS_RECVERR6_RFC4884 = 21, 273 INET_FLAGS_MC6_ALL = 22, 274 INET_FLAGS_AUTOFLOWLABEL_SET = 23, 275 INET_FLAGS_AUTOFLOWLABEL = 24, 276 INET_FLAGS_DONTFRAG = 25, 277 INET_FLAGS_RECVERR6 = 26, 278 INET_FLAGS_REPFLOW = 27, 279 INET_FLAGS_RTALERT_ISOLATE = 28, 280 INET_FLAGS_SNDFLOW = 29, 281 INET_FLAGS_RTALERT = 30, 282 }; 283 284 /* cmsg flags for inet */ 285 #define IP_CMSG_PKTINFO BIT(INET_FLAGS_PKTINFO) 286 #define IP_CMSG_TTL BIT(INET_FLAGS_TTL) 287 #define IP_CMSG_TOS BIT(INET_FLAGS_TOS) 288 #define IP_CMSG_RECVOPTS BIT(INET_FLAGS_RECVOPTS) 289 #define IP_CMSG_RETOPTS BIT(INET_FLAGS_RETOPTS) 290 #define IP_CMSG_PASSSEC BIT(INET_FLAGS_PASSSEC) 291 #define IP_CMSG_ORIGDSTADDR BIT(INET_FLAGS_ORIGDSTADDR) 292 #define IP_CMSG_CHECKSUM BIT(INET_FLAGS_CHECKSUM) 293 #define IP_CMSG_RECVFRAGSIZE BIT(INET_FLAGS_RECVFRAGSIZE) 294 295 #define IP_CMSG_ALL (IP_CMSG_PKTINFO | IP_CMSG_TTL | \ 296 IP_CMSG_TOS | IP_CMSG_RECVOPTS | \ 297 IP_CMSG_RETOPTS | IP_CMSG_PASSSEC | \ 298 IP_CMSG_ORIGDSTADDR | IP_CMSG_CHECKSUM | \ 299 IP_CMSG_RECVFRAGSIZE) 300 301 static inline unsigned long inet_cmsg_flags(const struct inet_sock *inet) 302 { 303 return READ_ONCE(inet->inet_flags) & IP_CMSG_ALL; 304 } 305 306 static inline dscp_t inet_sk_dscp(const struct inet_sock *inet) 307 { 308 return inet_dsfield_to_dscp(READ_ONCE(inet->tos)); 309 } 310 311 #define inet_test_bit(nr, sk) \ 312 test_bit(INET_FLAGS_##nr, &inet_sk(sk)->inet_flags) 313 #define inet_set_bit(nr, sk) \ 314 set_bit(INET_FLAGS_##nr, &inet_sk(sk)->inet_flags) 315 #define inet_clear_bit(nr, sk) \ 316 clear_bit(INET_FLAGS_##nr, &inet_sk(sk)->inet_flags) 317 #define inet_assign_bit(nr, sk, val) \ 318 assign_bit(INET_FLAGS_##nr, &inet_sk(sk)->inet_flags, val) 319 320 /** 321 * sk_to_full_sk - Access to a full socket 322 * @sk: pointer to a socket 323 * 324 * SYNACK messages might be attached to request sockets. 325 * Some places want to reach the listener in this case. 326 */ 327 static inline struct sock *sk_to_full_sk(struct sock *sk) 328 { 329 #ifdef CONFIG_INET 330 if (sk && READ_ONCE(sk->sk_state) == TCP_NEW_SYN_RECV) 331 sk = inet_reqsk(sk)->rsk_listener; 332 if (sk && READ_ONCE(sk->sk_state) == TCP_TIME_WAIT) 333 sk = NULL; 334 #endif 335 return sk; 336 } 337 338 /* sk_to_full_sk() variant with a const argument */ 339 static inline const struct sock *sk_const_to_full_sk(const struct sock *sk) 340 { 341 #ifdef CONFIG_INET 342 if (sk && READ_ONCE(sk->sk_state) == TCP_NEW_SYN_RECV) 343 sk = ((const struct request_sock *)sk)->rsk_listener; 344 if (sk && READ_ONCE(sk->sk_state) == TCP_TIME_WAIT) 345 sk = NULL; 346 #endif 347 return sk; 348 } 349 350 static inline struct sock *skb_to_full_sk(const struct sk_buff *skb) 351 { 352 return sk_to_full_sk(skb->sk); 353 } 354 355 #define inet_sk(ptr) container_of_const(ptr, struct inet_sock, sk) 356 357 int inet_sk_rebuild_header(struct sock *sk); 358 359 /** 360 * inet_sk_state_load - read sk->sk_state for lockless contexts 361 * @sk: socket pointer 362 * 363 * Paired with inet_sk_state_store(). Used in places we don't hold socket lock: 364 * tcp_diag_get_info(), tcp_get_info(), tcp_poll(), get_tcp4_sock() ... 365 */ 366 static inline int inet_sk_state_load(const struct sock *sk) 367 { 368 /* state change might impact lockless readers. */ 369 return smp_load_acquire(&sk->sk_state); 370 } 371 372 /** 373 * inet_sk_state_store - update sk->sk_state 374 * @sk: socket pointer 375 * @newstate: new state 376 * 377 * Paired with inet_sk_state_load(). Should be used in contexts where 378 * state change might impact lockless readers. 379 */ 380 void inet_sk_state_store(struct sock *sk, int newstate); 381 382 void inet_sk_set_state(struct sock *sk, int state); 383 384 static inline unsigned int __inet_ehashfn(const __be32 laddr, 385 const __u16 lport, 386 const __be32 faddr, 387 const __be16 fport, 388 u32 initval) 389 { 390 return jhash_3words((__force __u32) laddr, 391 (__force __u32) faddr, 392 ((__u32) lport) << 16 | (__force __u32)fport, 393 initval); 394 } 395 396 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops, 397 struct sock *sk_listener, 398 bool attach_listener); 399 400 static inline __u8 inet_sk_flowi_flags(const struct sock *sk) 401 { 402 __u8 flags = 0; 403 404 if (inet_test_bit(TRANSPARENT, sk) || inet_test_bit(HDRINCL, sk)) 405 flags |= FLOWI_FLAG_ANYSRC; 406 return flags; 407 } 408 409 static inline void inet_inc_convert_csum(struct sock *sk) 410 { 411 inet_sk(sk)->convert_csum++; 412 } 413 414 static inline void inet_dec_convert_csum(struct sock *sk) 415 { 416 if (inet_sk(sk)->convert_csum > 0) 417 inet_sk(sk)->convert_csum--; 418 } 419 420 static inline bool inet_get_convert_csum(struct sock *sk) 421 { 422 return !!inet_sk(sk)->convert_csum; 423 } 424 425 426 static inline bool inet_can_nonlocal_bind(struct net *net, 427 struct inet_sock *inet) 428 { 429 return READ_ONCE(net->ipv4.sysctl_ip_nonlocal_bind) || 430 test_bit(INET_FLAGS_FREEBIND, &inet->inet_flags) || 431 test_bit(INET_FLAGS_TRANSPARENT, &inet->inet_flags); 432 } 433 434 static inline bool inet_addr_valid_or_nonlocal(struct net *net, 435 struct inet_sock *inet, 436 __be32 addr, 437 int addr_type) 438 { 439 return inet_can_nonlocal_bind(net, inet) || 440 addr == htonl(INADDR_ANY) || 441 addr_type == RTN_LOCAL || 442 addr_type == RTN_MULTICAST || 443 addr_type == RTN_BROADCAST; 444 } 445 446 #endif /* _INET_SOCK_H */ 447