net/ipv6/inet6_hashtables.c

   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  *              Generic INET6 transport hashtables
   8  *
   9  * Authors:     Lotsa people, from code originally in tcp, generalised here
  10  *              by Arnaldo Carvalho de Melo <acme@mandriva.com>
  11  */
  12
  13 #include <linux/module.h>
  14 #include <linux/random.h>
  15
  16 #include <net/addrconf.h>
  17 #include <net/inet_connection_sock.h>
  18 #include <net/inet_hashtables.h>
  19 #include <net/inet6_hashtables.h>
  20 #include <net/secure_seq.h>
  21 #include <net/ip.h>
  22 #include <net/sock_reuseport.h>
  23
  24 u32 inet6_ehashfn(const struct net *net,
  25                   const struct in6_addr *laddr, const u16 lport,
  26                   const struct in6_addr *faddr, const __be16 fport)
  27 {
  28         static u32 inet6_ehash_secret __read_mostly;
  29         static u32 ipv6_hash_secret __read_mostly;
  30
  31         u32 lhash, fhash;
  32
  33         net_get_random_once(&inet6_ehash_secret, sizeof(inet6_ehash_secret));
  34         net_get_random_once(&ipv6_hash_secret, sizeof(ipv6_hash_secret));
  35
  36         lhash = (__force u32)laddr->s6_addr32[3];
  37         fhash = __ipv6_addr_jhash(faddr, ipv6_hash_secret);
  38
  39         return __inet6_ehashfn(lhash, lport, fhash, fport,
  40                                inet6_ehash_secret + net_hash_mix(net));
  41 }
  42 EXPORT_SYMBOL_GPL(inet6_ehashfn);
  43
  44 /*
  45  * Sockets in TCP_CLOSE state are _always_ taken out of the hash, so
  46  * we need not check it for TCP lookups anymore, thanks Alexey. -DaveM
  47  *
  48  * The sockhash lock must be held as a reader here.
  49  */
  50 struct sock *__inet6_lookup_established(struct net *net,
  51                                         struct inet_hashinfo *hashinfo,
  52                                            const struct in6_addr *saddr,
  53                                            const __be16 sport,
  54                                            const struct in6_addr *daddr,
  55                                            const u16 hnum,
  56                                            const int dif, const int sdif)
  57 {
  58         struct sock *sk;
  59         const struct hlist_nulls_node *node;
  60         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
  61         /* Optimize here for direct hit, only listening connections can
  62          * have wildcards anyways.
  63          */
  64         unsigned int hash = inet6_ehashfn(net, daddr, hnum, saddr, sport);
  65         unsigned int slot = hash & hashinfo->ehash_mask;
  66         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
  67
  68
  69 begin:
  70         sk_nulls_for_each_rcu(sk, node, &head->chain) {
  71                 if (sk->sk_hash != hash)
  72                         continue;
  73                 if (!inet6_match(net, sk, saddr, daddr, ports, dif, sdif))
  74                         continue;
  75                 if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
  76                         goto out;
  77
  78                 if (unlikely(!inet6_match(net, sk, saddr, daddr, ports, dif, sdif))) {
  79                         sock_gen_put(sk);
  80                         goto begin;
  81                 }
  82                 goto found;
  83         }
  84         if (get_nulls_value(node) != slot)
  85                 goto begin;
  86 out:
  87         sk = NULL;
  88 found:
  89         return sk;
  90 }
  91 EXPORT_SYMBOL(__inet6_lookup_established);
  92
  93 static inline int compute_score(struct sock *sk, struct net *net,
  94                                 const unsigned short hnum,
  95                                 const struct in6_addr *daddr,
  96                                 const int dif, const int sdif)
  97 {
  98         int score = -1;
  99
 100         if (net_eq(sock_net(sk), net) && inet_sk(sk)->inet_num == hnum &&
 101             sk->sk_family == PF_INET6) {
 102                 if (!ipv6_addr_equal(&sk->sk_v6_rcv_saddr, daddr))
 103                         return -1;
 104
 105                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 106                         return -1;
 107
 108                 score =  sk->sk_bound_dev_if ? 2 : 1;
 109                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 110                         score++;
 111         }
 112         return score;
 113 }
 114
 115 INDIRECT_CALLABLE_DECLARE(inet6_ehashfn_t udp6_ehashfn);
 116
 117 /**
 118  * inet6_lookup_reuseport() - execute reuseport logic on AF_INET6 socket if necessary.
 119  * @net: network namespace.
 120  * @sk: AF_INET6 socket, must be in TCP_LISTEN state for TCP or TCP_CLOSE for UDP.
 121  * @skb: context for a potential SK_REUSEPORT program.
 122  * @doff: header offset.
 123  * @saddr: source address.
 124  * @sport: source port.
 125  * @daddr: destination address.
 126  * @hnum: destination port in host byte order.
 127  * @ehashfn: hash function used to generate the fallback hash.
 128  *
 129  * Return: NULL if sk doesn't have SO_REUSEPORT set, otherwise a pointer to
 130  *         the selected sock or an error.
 131  */
 132 struct sock *inet6_lookup_reuseport(struct net *net, struct sock *sk,
 133                                     struct sk_buff *skb, int doff,
 134                                     const struct in6_addr *saddr,
 135                                     __be16 sport,
 136                                     const struct in6_addr *daddr,
 137                                     unsigned short hnum,
 138                                     inet6_ehashfn_t *ehashfn)
 139 {
 140         struct sock *reuse_sk = NULL;
 141         u32 phash;
 142
 143         if (sk->sk_reuseport) {
 144                 phash = INDIRECT_CALL_INET(ehashfn, udp6_ehashfn, inet6_ehashfn,
 145                                            net, daddr, hnum, saddr, sport);
 146                 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
 147         }
 148         return reuse_sk;
 149 }
 150 EXPORT_SYMBOL_GPL(inet6_lookup_reuseport);
 151
 152 /* called with rcu_read_lock() */
 153 static struct sock *inet6_lhash2_lookup(struct net *net,
 154                 struct inet_listen_hashbucket *ilb2,
 155                 struct sk_buff *skb, int doff,
 156                 const struct in6_addr *saddr,
 157                 const __be16 sport, const struct in6_addr *daddr,
 158                 const unsigned short hnum, const int dif, const int sdif)
 159 {
 160         struct sock *sk, *result = NULL;
 161         struct hlist_nulls_node *node;
 162         int score, hiscore = 0;
 163
 164         sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
 165                 score = compute_score(sk, net, hnum, daddr, dif, sdif);
 166                 if (score > hiscore) {
 167                         result = inet6_lookup_reuseport(net, sk, skb, doff,
 168                                                         saddr, sport, daddr, hnum, inet6_ehashfn);
 169                         if (result)
 170                                 return result;
 171
 172                         result = sk;
 173                         hiscore = score;
 174                 }
 175         }
 176
 177         return result;
 178 }
 179
 180 static inline struct sock *inet6_lookup_run_bpf(struct net *net,
 181                                                 struct inet_hashinfo *hashinfo,
 182                                                 struct sk_buff *skb, int doff,
 183                                                 const struct in6_addr *saddr,
 184                                                 const __be16 sport,
 185                                                 const struct in6_addr *daddr,
 186                                                 const u16 hnum, const int dif)
 187 {
 188         struct sock *sk, *reuse_sk;
 189         bool no_reuseport;
 190
 191         if (hashinfo != net->ipv4.tcp_death_row.hashinfo)
 192                 return NULL; /* only TCP is supported */
 193
 194         no_reuseport = bpf_sk_lookup_run_v6(net, IPPROTO_TCP, saddr, sport,
 195                                             daddr, hnum, dif, &sk);
 196         if (no_reuseport || IS_ERR_OR_NULL(sk))
 197                 return sk;
 198
 199         reuse_sk = inet6_lookup_reuseport(net, sk, skb, doff,
 200                                           saddr, sport, daddr, hnum, inet6_ehashfn);
 201         if (reuse_sk)
 202                 sk = reuse_sk;
 203         return sk;
 204 }
 205
 206 struct sock *inet6_lookup_listener(struct net *net,
 207                 struct inet_hashinfo *hashinfo,
 208                 struct sk_buff *skb, int doff,
 209                 const struct in6_addr *saddr,
 210                 const __be16 sport, const struct in6_addr *daddr,
 211                 const unsigned short hnum, const int dif, const int sdif)
 212 {
 213         struct inet_listen_hashbucket *ilb2;
 214         struct sock *result = NULL;
 215         unsigned int hash2;
 216
 217         /* Lookup redirect from BPF */
 218         if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
 219                 result = inet6_lookup_run_bpf(net, hashinfo, skb, doff,
 220                                               saddr, sport, daddr, hnum, dif);
 221                 if (result)
 222                         goto done;
 223         }
 224
 225         hash2 = ipv6_portaddr_hash(net, daddr, hnum);
 226         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 227
 228         result = inet6_lhash2_lookup(net, ilb2, skb, doff,
 229                                      saddr, sport, daddr, hnum,
 230                                      dif, sdif);
 231         if (result)
 232                 goto done;
 233
 234         /* Lookup lhash2 with in6addr_any */
 235         hash2 = ipv6_portaddr_hash(net, &in6addr_any, hnum);
 236         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 237
 238         result = inet6_lhash2_lookup(net, ilb2, skb, doff,
 239                                      saddr, sport, &in6addr_any, hnum,
 240                                      dif, sdif);
 241 done:
 242         if (IS_ERR(result))
 243                 return NULL;
 244         return result;
 245 }
 246 EXPORT_SYMBOL_GPL(inet6_lookup_listener);
 247
 248 struct sock *inet6_lookup(struct net *net, struct inet_hashinfo *hashinfo,
 249                           struct sk_buff *skb, int doff,
 250                           const struct in6_addr *saddr, const __be16 sport,
 251                           const struct in6_addr *daddr, const __be16 dport,
 252                           const int dif)
 253 {
 254         struct sock *sk;
 255         bool refcounted;
 256
 257         sk = __inet6_lookup(net, hashinfo, skb, doff, saddr, sport, daddr,
 258                             ntohs(dport), dif, 0, &refcounted);
 259         if (sk && !refcounted && !refcount_inc_not_zero(&sk->sk_refcnt))
 260                 sk = NULL;
 261         return sk;
 262 }
 263 EXPORT_SYMBOL_GPL(inet6_lookup);
 264
 265 static int __inet6_check_established(struct inet_timewait_death_row *death_row,
 266                                      struct sock *sk, const __u16 lport,
 267                                      struct inet_timewait_sock **twp)
 268 {
 269         struct inet_hashinfo *hinfo = death_row->hashinfo;
 270         struct inet_sock *inet = inet_sk(sk);
 271         const struct in6_addr *daddr = &sk->sk_v6_rcv_saddr;
 272         const struct in6_addr *saddr = &sk->sk_v6_daddr;
 273         const int dif = sk->sk_bound_dev_if;
 274         struct net *net = sock_net(sk);
 275         const int sdif = l3mdev_master_ifindex_by_index(net, dif);
 276         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 277         const unsigned int hash = inet6_ehashfn(net, daddr, lport, saddr,
 278                                                 inet->inet_dport);
 279         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 280         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 281         struct sock *sk2;
 282         const struct hlist_nulls_node *node;
 283         struct inet_timewait_sock *tw = NULL;
 284
 285         spin_lock(lock);
 286
 287         sk_nulls_for_each(sk2, node, &head->chain) {
 288                 if (sk2->sk_hash != hash)
 289                         continue;
 290
 291                 if (likely(inet6_match(net, sk2, saddr, daddr, ports,
 292                                        dif, sdif))) {
 293                         if (sk2->sk_state == TCP_TIME_WAIT) {
 294                                 tw = inet_twsk(sk2);
 295                                 if (twsk_unique(sk, sk2, twp))
 296                                         break;
 297                         }
 298                         goto not_unique;
 299                 }
 300         }
 301
 302         /* Must record num and sport now. Otherwise we will see
 303          * in hash table socket with a funny identity.
 304          */
 305         inet->inet_num = lport;
 306         inet->inet_sport = htons(lport);
 307         sk->sk_hash = hash;
 308         WARN_ON(!sk_unhashed(sk));
 309         __sk_nulls_add_node_rcu(sk, &head->chain);
 310         if (tw) {
 311                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 312                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 313         }
 314         spin_unlock(lock);
 315         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 316
 317         if (twp) {
 318                 *twp = tw;
 319         } else if (tw) {
 320                 /* Silly. Should hash-dance instead... */
 321                 inet_twsk_deschedule_put(tw);
 322         }
 323         return 0;
 324
 325 not_unique:
 326         spin_unlock(lock);
 327         return -EADDRNOTAVAIL;
 328 }
 329
 330 static u64 inet6_sk_port_offset(const struct sock *sk)
 331 {
 332         const struct inet_sock *inet = inet_sk(sk);
 333
 334         return secure_ipv6_port_ephemeral(sk->sk_v6_rcv_saddr.s6_addr32,
 335                                           sk->sk_v6_daddr.s6_addr32,
 336                                           inet->inet_dport);
 337 }
 338
 339 int inet6_hash_connect(struct inet_timewait_death_row *death_row,
 340                        struct sock *sk)
 341 {
 342         u64 port_offset = 0;
 343
 344         if (!inet_sk(sk)->inet_num)
 345                 port_offset = inet6_sk_port_offset(sk);
 346         return __inet_hash_connect(death_row, sk, port_offset,
 347                                    __inet6_check_established);
 348 }
 349 EXPORT_SYMBOL_GPL(inet6_hash_connect);
 350
 351 int inet6_hash(struct sock *sk)
 352 {
 353         int err = 0;
 354
 355         if (sk->sk_state != TCP_CLOSE)
 356                 err = __inet_hash(sk, NULL);
 357
 358         return err;
 359 }
 360 EXPORT_SYMBOL_GPL(inet6_hash);