net/core/sock_map.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
   3
   4 #include <linux/bpf.h>
   5 #include <linux/filter.h>
   6 #include <linux/errno.h>
   7 #include <linux/file.h>
   8 #include <linux/net.h>
   9 #include <linux/workqueue.h>
  10 #include <linux/skmsg.h>
  11 #include <linux/list.h>
  12 #include <linux/jhash.h>
  13 #include <linux/sock_diag.h>
  14 #include <net/udp.h>
  15
  16 struct bpf_stab {
  17         struct bpf_map map;
  18         struct sock **sks;
  19         struct sk_psock_progs progs;
  20         raw_spinlock_t lock;
  21 };
  22
  23 #define SOCK_CREATE_FLAG_MASK                           \
  24         (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
  25
  26 static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
  27 {
  28         struct bpf_stab *stab;
  29         u64 cost;
  30         int err;
  31
  32         if (!capable(CAP_NET_ADMIN))
  33                 return ERR_PTR(-EPERM);
  34         if (attr->max_entries == 0 ||
  35             attr->key_size    != 4 ||
  36             (attr->value_size != sizeof(u32) &&
  37              attr->value_size != sizeof(u64)) ||
  38             attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
  39                 return ERR_PTR(-EINVAL);
  40
  41         stab = kzalloc(sizeof(*stab), GFP_USER);
  42         if (!stab)
  43                 return ERR_PTR(-ENOMEM);
  44
  45         bpf_map_init_from_attr(&stab->map, attr);
  46         raw_spin_lock_init(&stab->lock);
  47
  48         /* Make sure page count doesn't overflow. */
  49         cost = (u64) stab->map.max_entries * sizeof(struct sock *);
  50         err = bpf_map_charge_init(&stab->map.memory, cost);
  51         if (err)
  52                 goto free_stab;
  53
  54         stab->sks = bpf_map_area_alloc(stab->map.max_entries *
  55                                        sizeof(struct sock *),
  56                                        stab->map.numa_node);
  57         if (stab->sks)
  58                 return &stab->map;
  59         err = -ENOMEM;
  60         bpf_map_charge_finish(&stab->map.memory);
  61 free_stab:
  62         kfree(stab);
  63         return ERR_PTR(err);
  64 }
  65
  66 int sock_map_get_from_fd(const union bpf_attr *attr, struct bpf_prog *prog)
  67 {
  68         u32 ufd = attr->target_fd;
  69         struct bpf_map *map;
  70         struct fd f;
  71         int ret;
  72
  73         f = fdget(ufd);
  74         map = __bpf_map_get(f);
  75         if (IS_ERR(map))
  76                 return PTR_ERR(map);
  77         ret = sock_map_prog_update(map, prog, attr->attach_type);
  78         fdput(f);
  79         return ret;
  80 }
  81
  82 static void sock_map_sk_acquire(struct sock *sk)
  83         __acquires(&sk->sk_lock.slock)
  84 {
  85         lock_sock(sk);
  86         preempt_disable();
  87         rcu_read_lock();
  88 }
  89
  90 static void sock_map_sk_release(struct sock *sk)
  91         __releases(&sk->sk_lock.slock)
  92 {
  93         rcu_read_unlock();
  94         preempt_enable();
  95         release_sock(sk);
  96 }
  97
  98 static void sock_map_add_link(struct sk_psock *psock,
  99                               struct sk_psock_link *link,
 100                               struct bpf_map *map, void *link_raw)
 101 {
 102         link->link_raw = link_raw;
 103         link->map = map;
 104         spin_lock_bh(&psock->link_lock);
 105         list_add_tail(&link->list, &psock->link);
 106         spin_unlock_bh(&psock->link_lock);
 107 }
 108
 109 static void sock_map_del_link(struct sock *sk,
 110                               struct sk_psock *psock, void *link_raw)
 111 {
 112         struct sk_psock_link *link, *tmp;
 113         bool strp_stop = false;
 114
 115         spin_lock_bh(&psock->link_lock);
 116         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 117                 if (link->link_raw == link_raw) {
 118                         struct bpf_map *map = link->map;
 119                         struct bpf_stab *stab = container_of(map, struct bpf_stab,
 120                                                              map);
 121                         if (psock->parser.enabled && stab->progs.skb_parser)
 122                                 strp_stop = true;
 123                         list_del(&link->list);
 124                         sk_psock_free_link(link);
 125                 }
 126         }
 127         spin_unlock_bh(&psock->link_lock);
 128         if (strp_stop) {
 129                 write_lock_bh(&sk->sk_callback_lock);
 130                 sk_psock_stop_strp(sk, psock);
 131                 write_unlock_bh(&sk->sk_callback_lock);
 132         }
 133 }
 134
 135 static void sock_map_unref(struct sock *sk, void *link_raw)
 136 {
 137         struct sk_psock *psock = sk_psock(sk);
 138
 139         if (likely(psock)) {
 140                 sock_map_del_link(sk, psock, link_raw);
 141                 sk_psock_put(sk, psock);
 142         }
 143 }
 144
 145 static int sock_map_init_proto(struct sock *sk, struct sk_psock *psock)
 146 {
 147         struct proto *prot;
 148
 149         sock_owned_by_me(sk);
 150
 151         switch (sk->sk_type) {
 152         case SOCK_STREAM:
 153                 prot = tcp_bpf_get_proto(sk, psock);
 154                 break;
 155
 156         case SOCK_DGRAM:
 157                 prot = udp_bpf_get_proto(sk, psock);
 158                 break;
 159
 160         default:
 161                 return -EINVAL;
 162         }
 163
 164         if (IS_ERR(prot))
 165                 return PTR_ERR(prot);
 166
 167         sk_psock_update_proto(sk, psock, prot);
 168         return 0;
 169 }
 170
 171 static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
 172 {
 173         struct sk_psock *psock;
 174
 175         rcu_read_lock();
 176         psock = sk_psock(sk);
 177         if (psock) {
 178                 if (sk->sk_prot->close != sock_map_close) {
 179                         psock = ERR_PTR(-EBUSY);
 180                         goto out;
 181                 }
 182
 183                 if (!refcount_inc_not_zero(&psock->refcnt))
 184                         psock = ERR_PTR(-EBUSY);
 185         }
 186 out:
 187         rcu_read_unlock();
 188         return psock;
 189 }
 190
 191 static int sock_map_link(struct bpf_map *map, struct sk_psock_progs *progs,
 192                          struct sock *sk)
 193 {
 194         struct bpf_prog *msg_parser, *skb_parser, *skb_verdict;
 195         struct sk_psock *psock;
 196         bool skb_progs;
 197         int ret;
 198
 199         skb_verdict = READ_ONCE(progs->skb_verdict);
 200         skb_parser = READ_ONCE(progs->skb_parser);
 201         skb_progs = skb_parser && skb_verdict;
 202         if (skb_progs) {
 203                 skb_verdict = bpf_prog_inc_not_zero(skb_verdict);
 204                 if (IS_ERR(skb_verdict))
 205                         return PTR_ERR(skb_verdict);
 206                 skb_parser = bpf_prog_inc_not_zero(skb_parser);
 207                 if (IS_ERR(skb_parser)) {
 208                         bpf_prog_put(skb_verdict);
 209                         return PTR_ERR(skb_parser);
 210                 }
 211         }
 212
 213         msg_parser = READ_ONCE(progs->msg_parser);
 214         if (msg_parser) {
 215                 msg_parser = bpf_prog_inc_not_zero(msg_parser);
 216                 if (IS_ERR(msg_parser)) {
 217                         ret = PTR_ERR(msg_parser);
 218                         goto out;
 219                 }
 220         }
 221
 222         psock = sock_map_psock_get_checked(sk);
 223         if (IS_ERR(psock)) {
 224                 ret = PTR_ERR(psock);
 225                 goto out_progs;
 226         }
 227
 228         if (psock) {
 229                 if ((msg_parser && READ_ONCE(psock->progs.msg_parser)) ||
 230                     (skb_progs  && READ_ONCE(psock->progs.skb_parser))) {
 231                         sk_psock_put(sk, psock);
 232                         ret = -EBUSY;
 233                         goto out_progs;
 234                 }
 235         } else {
 236                 psock = sk_psock_init(sk, map->numa_node);
 237                 if (!psock) {
 238                         ret = -ENOMEM;
 239                         goto out_progs;
 240                 }
 241         }
 242
 243         if (msg_parser)
 244                 psock_set_prog(&psock->progs.msg_parser, msg_parser);
 245
 246         ret = sock_map_init_proto(sk, psock);
 247         if (ret < 0)
 248                 goto out_drop;
 249
 250         write_lock_bh(&sk->sk_callback_lock);
 251         if (skb_progs && !psock->parser.enabled) {
 252                 ret = sk_psock_init_strp(sk, psock);
 253                 if (ret) {
 254                         write_unlock_bh(&sk->sk_callback_lock);
 255                         goto out_drop;
 256                 }
 257                 psock_set_prog(&psock->progs.skb_verdict, skb_verdict);
 258                 psock_set_prog(&psock->progs.skb_parser, skb_parser);
 259                 sk_psock_start_strp(sk, psock);
 260         }
 261         write_unlock_bh(&sk->sk_callback_lock);
 262         return 0;
 263 out_drop:
 264         sk_psock_put(sk, psock);
 265 out_progs:
 266         if (msg_parser)
 267                 bpf_prog_put(msg_parser);
 268 out:
 269         if (skb_progs) {
 270                 bpf_prog_put(skb_verdict);
 271                 bpf_prog_put(skb_parser);
 272         }
 273         return ret;
 274 }
 275
 276 static int sock_map_link_no_progs(struct bpf_map *map, struct sock *sk)
 277 {
 278         struct sk_psock *psock;
 279         int ret;
 280
 281         psock = sock_map_psock_get_checked(sk);
 282         if (IS_ERR(psock))
 283                 return PTR_ERR(psock);
 284
 285         if (!psock) {
 286                 psock = sk_psock_init(sk, map->numa_node);
 287                 if (!psock)
 288                         return -ENOMEM;
 289         }
 290
 291         ret = sock_map_init_proto(sk, psock);
 292         if (ret < 0)
 293                 sk_psock_put(sk, psock);
 294         return ret;
 295 }
 296
 297 static void sock_map_free(struct bpf_map *map)
 298 {
 299         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 300         int i;
 301
 302         /* After the sync no updates or deletes will be in-flight so it
 303          * is safe to walk map and remove entries without risking a race
 304          * in EEXIST update case.
 305          */
 306         synchronize_rcu();
 307         for (i = 0; i < stab->map.max_entries; i++) {
 308                 struct sock **psk = &stab->sks[i];
 309                 struct sock *sk;
 310
 311                 sk = xchg(psk, NULL);
 312                 if (sk) {
 313                         lock_sock(sk);
 314                         rcu_read_lock();
 315                         sock_map_unref(sk, psk);
 316                         rcu_read_unlock();
 317                         release_sock(sk);
 318                 }
 319         }
 320
 321         /* wait for psock readers accessing its map link */
 322         synchronize_rcu();
 323
 324         bpf_map_area_free(stab->sks);
 325         kfree(stab);
 326 }
 327
 328 static void sock_map_release_progs(struct bpf_map *map)
 329 {
 330         psock_progs_drop(&container_of(map, struct bpf_stab, map)->progs);
 331 }
 332
 333 static struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key)
 334 {
 335         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 336
 337         WARN_ON_ONCE(!rcu_read_lock_held());
 338
 339         if (unlikely(key >= map->max_entries))
 340                 return NULL;
 341         return READ_ONCE(stab->sks[key]);
 342 }
 343
 344 static void *sock_map_lookup(struct bpf_map *map, void *key)
 345 {
 346         struct sock *sk;
 347
 348         sk = __sock_map_lookup_elem(map, *(u32 *)key);
 349         if (!sk || !sk_fullsock(sk))
 350                 return NULL;
 351         if (sk_is_refcounted(sk) && !refcount_inc_not_zero(&sk->sk_refcnt))
 352                 return NULL;
 353         return sk;
 354 }
 355
 356 static void *sock_map_lookup_sys(struct bpf_map *map, void *key)
 357 {
 358         struct sock *sk;
 359
 360         if (map->value_size != sizeof(u64))
 361                 return ERR_PTR(-ENOSPC);
 362
 363         sk = __sock_map_lookup_elem(map, *(u32 *)key);
 364         if (!sk)
 365                 return ERR_PTR(-ENOENT);
 366
 367         sock_gen_cookie(sk);
 368         return &sk->sk_cookie;
 369 }
 370
 371 static int __sock_map_delete(struct bpf_stab *stab, struct sock *sk_test,
 372                              struct sock **psk)
 373 {
 374         struct sock *sk;
 375         int err = 0;
 376
 377         raw_spin_lock_bh(&stab->lock);
 378         sk = *psk;
 379         if (!sk_test || sk_test == sk)
 380                 sk = xchg(psk, NULL);
 381
 382         if (likely(sk))
 383                 sock_map_unref(sk, psk);
 384         else
 385                 err = -EINVAL;
 386
 387         raw_spin_unlock_bh(&stab->lock);
 388         return err;
 389 }
 390
 391 static void sock_map_delete_from_link(struct bpf_map *map, struct sock *sk,
 392                                       void *link_raw)
 393 {
 394         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 395
 396         __sock_map_delete(stab, sk, link_raw);
 397 }
 398
 399 static int sock_map_delete_elem(struct bpf_map *map, void *key)
 400 {
 401         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 402         u32 i = *(u32 *)key;
 403         struct sock **psk;
 404
 405         if (unlikely(i >= map->max_entries))
 406                 return -EINVAL;
 407
 408         psk = &stab->sks[i];
 409         return __sock_map_delete(stab, NULL, psk);
 410 }
 411
 412 static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next)
 413 {
 414         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 415         u32 i = key ? *(u32 *)key : U32_MAX;
 416         u32 *key_next = next;
 417
 418         if (i == stab->map.max_entries - 1)
 419                 return -ENOENT;
 420         if (i >= stab->map.max_entries)
 421                 *key_next = 0;
 422         else
 423                 *key_next = i + 1;
 424         return 0;
 425 }
 426
 427 static bool sock_map_redirect_allowed(const struct sock *sk);
 428
 429 static int sock_map_update_common(struct bpf_map *map, u32 idx,
 430                                   struct sock *sk, u64 flags)
 431 {
 432         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 433         struct sk_psock_link *link;
 434         struct sk_psock *psock;
 435         struct sock *osk;
 436         int ret;
 437
 438         WARN_ON_ONCE(!rcu_read_lock_held());
 439         if (unlikely(flags > BPF_EXIST))
 440                 return -EINVAL;
 441         if (unlikely(idx >= map->max_entries))
 442                 return -E2BIG;
 443         if (inet_csk_has_ulp(sk))
 444                 return -EINVAL;
 445
 446         link = sk_psock_init_link();
 447         if (!link)
 448                 return -ENOMEM;
 449
 450         /* Only sockets we can redirect into/from in BPF need to hold
 451          * refs to parser/verdict progs and have their sk_data_ready
 452          * and sk_write_space callbacks overridden.
 453          */
 454         if (sock_map_redirect_allowed(sk))
 455                 ret = sock_map_link(map, &stab->progs, sk);
 456         else
 457                 ret = sock_map_link_no_progs(map, sk);
 458         if (ret < 0)
 459                 goto out_free;
 460
 461         psock = sk_psock(sk);
 462         WARN_ON_ONCE(!psock);
 463
 464         raw_spin_lock_bh(&stab->lock);
 465         osk = stab->sks[idx];
 466         if (osk && flags == BPF_NOEXIST) {
 467                 ret = -EEXIST;
 468                 goto out_unlock;
 469         } else if (!osk && flags == BPF_EXIST) {
 470                 ret = -ENOENT;
 471                 goto out_unlock;
 472         }
 473
 474         sock_map_add_link(psock, link, map, &stab->sks[idx]);
 475         stab->sks[idx] = sk;
 476         if (osk)
 477                 sock_map_unref(osk, &stab->sks[idx]);
 478         raw_spin_unlock_bh(&stab->lock);
 479         return 0;
 480 out_unlock:
 481         raw_spin_unlock_bh(&stab->lock);
 482         if (psock)
 483                 sk_psock_put(sk, psock);
 484 out_free:
 485         sk_psock_free_link(link);
 486         return ret;
 487 }
 488
 489 static bool sock_map_op_okay(const struct bpf_sock_ops_kern *ops)
 490 {
 491         return ops->op == BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB ||
 492                ops->op == BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB ||
 493                ops->op == BPF_SOCK_OPS_TCP_LISTEN_CB;
 494 }
 495
 496 static bool sk_is_tcp(const struct sock *sk)
 497 {
 498         return sk->sk_type == SOCK_STREAM &&
 499                sk->sk_protocol == IPPROTO_TCP;
 500 }
 501
 502 static bool sk_is_udp(const struct sock *sk)
 503 {
 504         return sk->sk_type == SOCK_DGRAM &&
 505                sk->sk_protocol == IPPROTO_UDP;
 506 }
 507
 508 static bool sock_map_redirect_allowed(const struct sock *sk)
 509 {
 510         return sk_is_tcp(sk) && sk->sk_state != TCP_LISTEN;
 511 }
 512
 513 static bool sock_map_sk_is_suitable(const struct sock *sk)
 514 {
 515         return sk_is_tcp(sk) || sk_is_udp(sk);
 516 }
 517
 518 static bool sock_map_sk_state_allowed(const struct sock *sk)
 519 {
 520         if (sk_is_tcp(sk))
 521                 return (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_LISTEN);
 522         else if (sk_is_udp(sk))
 523                 return sk_hashed(sk);
 524
 525         return false;
 526 }
 527
 528 static int sock_map_update_elem(struct bpf_map *map, void *key,
 529                                 void *value, u64 flags)
 530 {
 531         u32 idx = *(u32 *)key;
 532         struct socket *sock;
 533         struct sock *sk;
 534         int ret;
 535         u64 ufd;
 536
 537         if (map->value_size == sizeof(u64))
 538                 ufd = *(u64 *)value;
 539         else
 540                 ufd = *(u32 *)value;
 541         if (ufd > S32_MAX)
 542                 return -EINVAL;
 543
 544         sock = sockfd_lookup(ufd, &ret);
 545         if (!sock)
 546                 return ret;
 547         sk = sock->sk;
 548         if (!sk) {
 549                 ret = -EINVAL;
 550                 goto out;
 551         }
 552         if (!sock_map_sk_is_suitable(sk)) {
 553                 ret = -EOPNOTSUPP;
 554                 goto out;
 555         }
 556
 557         sock_map_sk_acquire(sk);
 558         if (!sock_map_sk_state_allowed(sk))
 559                 ret = -EOPNOTSUPP;
 560         else
 561                 ret = sock_map_update_common(map, idx, sk, flags);
 562         sock_map_sk_release(sk);
 563 out:
 564         fput(sock->file);
 565         return ret;
 566 }
 567
 568 BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, sops,
 569            struct bpf_map *, map, void *, key, u64, flags)
 570 {
 571         WARN_ON_ONCE(!rcu_read_lock_held());
 572
 573         if (likely(sock_map_sk_is_suitable(sops->sk) &&
 574                    sock_map_op_okay(sops)))
 575                 return sock_map_update_common(map, *(u32 *)key, sops->sk,
 576                                               flags);
 577         return -EOPNOTSUPP;
 578 }
 579
 580 const struct bpf_func_proto bpf_sock_map_update_proto = {
 581         .func           = bpf_sock_map_update,
 582         .gpl_only       = false,
 583         .pkt_access     = true,
 584         .ret_type       = RET_INTEGER,
 585         .arg1_type      = ARG_PTR_TO_CTX,
 586         .arg2_type      = ARG_CONST_MAP_PTR,
 587         .arg3_type      = ARG_PTR_TO_MAP_KEY,
 588         .arg4_type      = ARG_ANYTHING,
 589 };
 590
 591 BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
 592            struct bpf_map *, map, u32, key, u64, flags)
 593 {
 594         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 595         struct sock *sk;
 596
 597         if (unlikely(flags & ~(BPF_F_INGRESS)))
 598                 return SK_DROP;
 599
 600         sk = __sock_map_lookup_elem(map, key);
 601         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
 602                 return SK_DROP;
 603
 604         tcb->bpf.flags = flags;
 605         tcb->bpf.sk_redir = sk;
 606         return SK_PASS;
 607 }
 608
 609 const struct bpf_func_proto bpf_sk_redirect_map_proto = {
 610         .func           = bpf_sk_redirect_map,
 611         .gpl_only       = false,
 612         .ret_type       = RET_INTEGER,
 613         .arg1_type      = ARG_PTR_TO_CTX,
 614         .arg2_type      = ARG_CONST_MAP_PTR,
 615         .arg3_type      = ARG_ANYTHING,
 616         .arg4_type      = ARG_ANYTHING,
 617 };
 618
 619 BPF_CALL_4(bpf_msg_redirect_map, struct sk_msg *, msg,
 620            struct bpf_map *, map, u32, key, u64, flags)
 621 {
 622         struct sock *sk;
 623
 624         if (unlikely(flags & ~(BPF_F_INGRESS)))
 625                 return SK_DROP;
 626
 627         sk = __sock_map_lookup_elem(map, key);
 628         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
 629                 return SK_DROP;
 630
 631         msg->flags = flags;
 632         msg->sk_redir = sk;
 633         return SK_PASS;
 634 }
 635
 636 const struct bpf_func_proto bpf_msg_redirect_map_proto = {
 637         .func           = bpf_msg_redirect_map,
 638         .gpl_only       = false,
 639         .ret_type       = RET_INTEGER,
 640         .arg1_type      = ARG_PTR_TO_CTX,
 641         .arg2_type      = ARG_CONST_MAP_PTR,
 642         .arg3_type      = ARG_ANYTHING,
 643         .arg4_type      = ARG_ANYTHING,
 644 };
 645
 646 static int sock_map_btf_id;
 647 const struct bpf_map_ops sock_map_ops = {
 648         .map_alloc              = sock_map_alloc,
 649         .map_free               = sock_map_free,
 650         .map_get_next_key       = sock_map_get_next_key,
 651         .map_lookup_elem_sys_only = sock_map_lookup_sys,
 652         .map_update_elem        = sock_map_update_elem,
 653         .map_delete_elem        = sock_map_delete_elem,
 654         .map_lookup_elem        = sock_map_lookup,
 655         .map_release_uref       = sock_map_release_progs,
 656         .map_check_btf          = map_check_no_btf,
 657         .map_btf_name           = "bpf_stab",
 658         .map_btf_id             = &sock_map_btf_id,
 659 };
 660
 661 struct bpf_shtab_elem {
 662         struct rcu_head rcu;
 663         u32 hash;
 664         struct sock *sk;
 665         struct hlist_node node;
 666         u8 key[];
 667 };
 668
 669 struct bpf_shtab_bucket {
 670         struct hlist_head head;
 671         raw_spinlock_t lock;
 672 };
 673
 674 struct bpf_shtab {
 675         struct bpf_map map;
 676         struct bpf_shtab_bucket *buckets;
 677         u32 buckets_num;
 678         u32 elem_size;
 679         struct sk_psock_progs progs;
 680         atomic_t count;
 681 };
 682
 683 static inline u32 sock_hash_bucket_hash(const void *key, u32 len)
 684 {
 685         return jhash(key, len, 0);
 686 }
 687
 688 static struct bpf_shtab_bucket *sock_hash_select_bucket(struct bpf_shtab *htab,
 689                                                         u32 hash)
 690 {
 691         return &htab->buckets[hash & (htab->buckets_num - 1)];
 692 }
 693
 694 static struct bpf_shtab_elem *
 695 sock_hash_lookup_elem_raw(struct hlist_head *head, u32 hash, void *key,
 696                           u32 key_size)
 697 {
 698         struct bpf_shtab_elem *elem;
 699
 700         hlist_for_each_entry_rcu(elem, head, node) {
 701                 if (elem->hash == hash &&
 702                     !memcmp(&elem->key, key, key_size))
 703                         return elem;
 704         }
 705
 706         return NULL;
 707 }
 708
 709 static struct sock *__sock_hash_lookup_elem(struct bpf_map *map, void *key)
 710 {
 711         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 712         u32 key_size = map->key_size, hash;
 713         struct bpf_shtab_bucket *bucket;
 714         struct bpf_shtab_elem *elem;
 715
 716         WARN_ON_ONCE(!rcu_read_lock_held());
 717
 718         hash = sock_hash_bucket_hash(key, key_size);
 719         bucket = sock_hash_select_bucket(htab, hash);
 720         elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 721
 722         return elem ? elem->sk : NULL;
 723 }
 724
 725 static void sock_hash_free_elem(struct bpf_shtab *htab,
 726                                 struct bpf_shtab_elem *elem)
 727 {
 728         atomic_dec(&htab->count);
 729         kfree_rcu(elem, rcu);
 730 }
 731
 732 static void sock_hash_delete_from_link(struct bpf_map *map, struct sock *sk,
 733                                        void *link_raw)
 734 {
 735         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 736         struct bpf_shtab_elem *elem_probe, *elem = link_raw;
 737         struct bpf_shtab_bucket *bucket;
 738
 739         WARN_ON_ONCE(!rcu_read_lock_held());
 740         bucket = sock_hash_select_bucket(htab, elem->hash);
 741
 742         /* elem may be deleted in parallel from the map, but access here
 743          * is okay since it's going away only after RCU grace period.
 744          * However, we need to check whether it's still present.
 745          */
 746         raw_spin_lock_bh(&bucket->lock);
 747         elem_probe = sock_hash_lookup_elem_raw(&bucket->head, elem->hash,
 748                                                elem->key, map->key_size);
 749         if (elem_probe && elem_probe == elem) {
 750                 hlist_del_rcu(&elem->node);
 751                 sock_map_unref(elem->sk, elem);
 752                 sock_hash_free_elem(htab, elem);
 753         }
 754         raw_spin_unlock_bh(&bucket->lock);
 755 }
 756
 757 static int sock_hash_delete_elem(struct bpf_map *map, void *key)
 758 {
 759         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 760         u32 hash, key_size = map->key_size;
 761         struct bpf_shtab_bucket *bucket;
 762         struct bpf_shtab_elem *elem;
 763         int ret = -ENOENT;
 764
 765         hash = sock_hash_bucket_hash(key, key_size);
 766         bucket = sock_hash_select_bucket(htab, hash);
 767
 768         raw_spin_lock_bh(&bucket->lock);
 769         elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 770         if (elem) {
 771                 hlist_del_rcu(&elem->node);
 772                 sock_map_unref(elem->sk, elem);
 773                 sock_hash_free_elem(htab, elem);
 774                 ret = 0;
 775         }
 776         raw_spin_unlock_bh(&bucket->lock);
 777         return ret;
 778 }
 779
 780 static struct bpf_shtab_elem *sock_hash_alloc_elem(struct bpf_shtab *htab,
 781                                                    void *key, u32 key_size,
 782                                                    u32 hash, struct sock *sk,
 783                                                    struct bpf_shtab_elem *old)
 784 {
 785         struct bpf_shtab_elem *new;
 786
 787         if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
 788                 if (!old) {
 789                         atomic_dec(&htab->count);
 790                         return ERR_PTR(-E2BIG);
 791                 }
 792         }
 793
 794         new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 795                            htab->map.numa_node);
 796         if (!new) {
 797                 atomic_dec(&htab->count);
 798                 return ERR_PTR(-ENOMEM);
 799         }
 800         memcpy(new->key, key, key_size);
 801         new->sk = sk;
 802         new->hash = hash;
 803         return new;
 804 }
 805
 806 static int sock_hash_update_common(struct bpf_map *map, void *key,
 807                                    struct sock *sk, u64 flags)
 808 {
 809         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 810         u32 key_size = map->key_size, hash;
 811         struct bpf_shtab_elem *elem, *elem_new;
 812         struct bpf_shtab_bucket *bucket;
 813         struct sk_psock_link *link;
 814         struct sk_psock *psock;
 815         int ret;
 816
 817         WARN_ON_ONCE(!rcu_read_lock_held());
 818         if (unlikely(flags > BPF_EXIST))
 819                 return -EINVAL;
 820         if (inet_csk_has_ulp(sk))
 821                 return -EINVAL;
 822
 823         link = sk_psock_init_link();
 824         if (!link)
 825                 return -ENOMEM;
 826
 827         /* Only sockets we can redirect into/from in BPF need to hold
 828          * refs to parser/verdict progs and have their sk_data_ready
 829          * and sk_write_space callbacks overridden.
 830          */
 831         if (sock_map_redirect_allowed(sk))
 832                 ret = sock_map_link(map, &htab->progs, sk);
 833         else
 834                 ret = sock_map_link_no_progs(map, sk);
 835         if (ret < 0)
 836                 goto out_free;
 837
 838         psock = sk_psock(sk);
 839         WARN_ON_ONCE(!psock);
 840
 841         hash = sock_hash_bucket_hash(key, key_size);
 842         bucket = sock_hash_select_bucket(htab, hash);
 843
 844         raw_spin_lock_bh(&bucket->lock);
 845         elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 846         if (elem && flags == BPF_NOEXIST) {
 847                 ret = -EEXIST;
 848                 goto out_unlock;
 849         } else if (!elem && flags == BPF_EXIST) {
 850                 ret = -ENOENT;
 851                 goto out_unlock;
 852         }
 853
 854         elem_new = sock_hash_alloc_elem(htab, key, key_size, hash, sk, elem);
 855         if (IS_ERR(elem_new)) {
 856                 ret = PTR_ERR(elem_new);
 857                 goto out_unlock;
 858         }
 859
 860         sock_map_add_link(psock, link, map, elem_new);
 861         /* Add new element to the head of the list, so that
 862          * concurrent search will find it before old elem.
 863          */
 864         hlist_add_head_rcu(&elem_new->node, &bucket->head);
 865         if (elem) {
 866                 hlist_del_rcu(&elem->node);
 867                 sock_map_unref(elem->sk, elem);
 868                 sock_hash_free_elem(htab, elem);
 869         }
 870         raw_spin_unlock_bh(&bucket->lock);
 871         return 0;
 872 out_unlock:
 873         raw_spin_unlock_bh(&bucket->lock);
 874         sk_psock_put(sk, psock);
 875 out_free:
 876         sk_psock_free_link(link);
 877         return ret;
 878 }
 879
 880 static int sock_hash_update_elem(struct bpf_map *map, void *key,
 881                                  void *value, u64 flags)
 882 {
 883         struct socket *sock;
 884         struct sock *sk;
 885         int ret;
 886         u64 ufd;
 887
 888         if (map->value_size == sizeof(u64))
 889                 ufd = *(u64 *)value;
 890         else
 891                 ufd = *(u32 *)value;
 892         if (ufd > S32_MAX)
 893                 return -EINVAL;
 894
 895         sock = sockfd_lookup(ufd, &ret);
 896         if (!sock)
 897                 return ret;
 898         sk = sock->sk;
 899         if (!sk) {
 900                 ret = -EINVAL;
 901                 goto out;
 902         }
 903         if (!sock_map_sk_is_suitable(sk)) {
 904                 ret = -EOPNOTSUPP;
 905                 goto out;
 906         }
 907
 908         sock_map_sk_acquire(sk);
 909         if (!sock_map_sk_state_allowed(sk))
 910                 ret = -EOPNOTSUPP;
 911         else
 912                 ret = sock_hash_update_common(map, key, sk, flags);
 913         sock_map_sk_release(sk);
 914 out:
 915         fput(sock->file);
 916         return ret;
 917 }
 918
 919 static int sock_hash_get_next_key(struct bpf_map *map, void *key,
 920                                   void *key_next)
 921 {
 922         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 923         struct bpf_shtab_elem *elem, *elem_next;
 924         u32 hash, key_size = map->key_size;
 925         struct hlist_head *head;
 926         int i = 0;
 927
 928         if (!key)
 929                 goto find_first_elem;
 930         hash = sock_hash_bucket_hash(key, key_size);
 931         head = &sock_hash_select_bucket(htab, hash)->head;
 932         elem = sock_hash_lookup_elem_raw(head, hash, key, key_size);
 933         if (!elem)
 934                 goto find_first_elem;
 935
 936         elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&elem->node)),
 937                                      struct bpf_shtab_elem, node);
 938         if (elem_next) {
 939                 memcpy(key_next, elem_next->key, key_size);
 940                 return 0;
 941         }
 942
 943         i = hash & (htab->buckets_num - 1);
 944         i++;
 945 find_first_elem:
 946         for (; i < htab->buckets_num; i++) {
 947                 head = &sock_hash_select_bucket(htab, i)->head;
 948                 elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
 949                                              struct bpf_shtab_elem, node);
 950                 if (elem_next) {
 951                         memcpy(key_next, elem_next->key, key_size);
 952                         return 0;
 953                 }
 954         }
 955
 956         return -ENOENT;
 957 }
 958
 959 static struct bpf_map *sock_hash_alloc(union bpf_attr *attr)
 960 {
 961         struct bpf_shtab *htab;
 962         int i, err;
 963         u64 cost;
 964
 965         if (!capable(CAP_NET_ADMIN))
 966                 return ERR_PTR(-EPERM);
 967         if (attr->max_entries == 0 ||
 968             attr->key_size    == 0 ||
 969             (attr->value_size != sizeof(u32) &&
 970              attr->value_size != sizeof(u64)) ||
 971             attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
 972                 return ERR_PTR(-EINVAL);
 973         if (attr->key_size > MAX_BPF_STACK)
 974                 return ERR_PTR(-E2BIG);
 975
 976         htab = kzalloc(sizeof(*htab), GFP_USER);
 977         if (!htab)
 978                 return ERR_PTR(-ENOMEM);
 979
 980         bpf_map_init_from_attr(&htab->map, attr);
 981
 982         htab->buckets_num = roundup_pow_of_two(htab->map.max_entries);
 983         htab->elem_size = sizeof(struct bpf_shtab_elem) +
 984                           round_up(htab->map.key_size, 8);
 985         if (htab->buckets_num == 0 ||
 986             htab->buckets_num > U32_MAX / sizeof(struct bpf_shtab_bucket)) {
 987                 err = -EINVAL;
 988                 goto free_htab;
 989         }
 990
 991         cost = (u64) htab->buckets_num * sizeof(struct bpf_shtab_bucket) +
 992                (u64) htab->elem_size * htab->map.max_entries;
 993         if (cost >= U32_MAX - PAGE_SIZE) {
 994                 err = -EINVAL;
 995                 goto free_htab;
 996         }
 997         err = bpf_map_charge_init(&htab->map.memory, cost);
 998         if (err)
 999                 goto free_htab;
1000
1001         htab->buckets = bpf_map_area_alloc(htab->buckets_num *
1002                                            sizeof(struct bpf_shtab_bucket),
1003                                            htab->map.numa_node);
1004         if (!htab->buckets) {
1005                 bpf_map_charge_finish(&htab->map.memory);
1006                 err = -ENOMEM;
1007                 goto free_htab;
1008         }
1009
1010         for (i = 0; i < htab->buckets_num; i++) {
1011                 INIT_HLIST_HEAD(&htab->buckets[i].head);
1012                 raw_spin_lock_init(&htab->buckets[i].lock);
1013         }
1014
1015         return &htab->map;
1016 free_htab:
1017         kfree(htab);
1018         return ERR_PTR(err);
1019 }
1020
1021 static void sock_hash_free(struct bpf_map *map)
1022 {
1023         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
1024         struct bpf_shtab_bucket *bucket;
1025         struct hlist_head unlink_list;
1026         struct bpf_shtab_elem *elem;
1027         struct hlist_node *node;
1028         int i;
1029
1030         /* After the sync no updates or deletes will be in-flight so it
1031          * is safe to walk map and remove entries without risking a race
1032          * in EEXIST update case.
1033          */
1034         synchronize_rcu();
1035         for (i = 0; i < htab->buckets_num; i++) {
1036                 bucket = sock_hash_select_bucket(htab, i);
1037
1038                 /* We are racing with sock_hash_delete_from_link to
1039                  * enter the spin-lock critical section. Every socket on
1040                  * the list is still linked to sockhash. Since link
1041                  * exists, psock exists and holds a ref to socket. That
1042                  * lets us to grab a socket ref too.
1043                  */
1044                 raw_spin_lock_bh(&bucket->lock);
1045                 hlist_for_each_entry(elem, &bucket->head, node)
1046                         sock_hold(elem->sk);
1047                 hlist_move_list(&bucket->head, &unlink_list);
1048                 raw_spin_unlock_bh(&bucket->lock);
1049
1050                 /* Process removed entries out of atomic context to
1051                  * block for socket lock before deleting the psock's
1052                  * link to sockhash.
1053                  */
1054                 hlist_for_each_entry_safe(elem, node, &unlink_list, node) {
1055                         hlist_del(&elem->node);
1056                         lock_sock(elem->sk);
1057                         rcu_read_lock();
1058                         sock_map_unref(elem->sk, elem);
1059                         rcu_read_unlock();
1060                         release_sock(elem->sk);
1061                         sock_put(elem->sk);
1062                         sock_hash_free_elem(htab, elem);
1063                 }
1064         }
1065
1066         /* wait for psock readers accessing its map link */
1067         synchronize_rcu();
1068
1069         bpf_map_area_free(htab->buckets);
1070         kfree(htab);
1071 }
1072
1073 static void *sock_hash_lookup_sys(struct bpf_map *map, void *key)
1074 {
1075         struct sock *sk;
1076
1077         if (map->value_size != sizeof(u64))
1078                 return ERR_PTR(-ENOSPC);
1079
1080         sk = __sock_hash_lookup_elem(map, key);
1081         if (!sk)
1082                 return ERR_PTR(-ENOENT);
1083
1084         sock_gen_cookie(sk);
1085         return &sk->sk_cookie;
1086 }
1087
1088 static void *sock_hash_lookup(struct bpf_map *map, void *key)
1089 {
1090         struct sock *sk;
1091
1092         sk = __sock_hash_lookup_elem(map, key);
1093         if (!sk || !sk_fullsock(sk))
1094                 return NULL;
1095         if (sk_is_refcounted(sk) && !refcount_inc_not_zero(&sk->sk_refcnt))
1096                 return NULL;
1097         return sk;
1098 }
1099
1100 static void sock_hash_release_progs(struct bpf_map *map)
1101 {
1102         psock_progs_drop(&container_of(map, struct bpf_shtab, map)->progs);
1103 }
1104
1105 BPF_CALL_4(bpf_sock_hash_update, struct bpf_sock_ops_kern *, sops,
1106            struct bpf_map *, map, void *, key, u64, flags)
1107 {
1108         WARN_ON_ONCE(!rcu_read_lock_held());
1109
1110         if (likely(sock_map_sk_is_suitable(sops->sk) &&
1111                    sock_map_op_okay(sops)))
1112                 return sock_hash_update_common(map, key, sops->sk, flags);
1113         return -EOPNOTSUPP;
1114 }
1115
1116 const struct bpf_func_proto bpf_sock_hash_update_proto = {
1117         .func           = bpf_sock_hash_update,
1118         .gpl_only       = false,
1119         .pkt_access     = true,
1120         .ret_type       = RET_INTEGER,
1121         .arg1_type      = ARG_PTR_TO_CTX,
1122         .arg2_type      = ARG_CONST_MAP_PTR,
1123         .arg3_type      = ARG_PTR_TO_MAP_KEY,
1124         .arg4_type      = ARG_ANYTHING,
1125 };
1126
1127 BPF_CALL_4(bpf_sk_redirect_hash, struct sk_buff *, skb,
1128            struct bpf_map *, map, void *, key, u64, flags)
1129 {
1130         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1131         struct sock *sk;
1132
1133         if (unlikely(flags & ~(BPF_F_INGRESS)))
1134                 return SK_DROP;
1135
1136         sk = __sock_hash_lookup_elem(map, key);
1137         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1138                 return SK_DROP;
1139
1140         tcb->bpf.flags = flags;
1141         tcb->bpf.sk_redir = sk;
1142         return SK_PASS;
1143 }
1144
1145 const struct bpf_func_proto bpf_sk_redirect_hash_proto = {
1146         .func           = bpf_sk_redirect_hash,
1147         .gpl_only       = false,
1148         .ret_type       = RET_INTEGER,
1149         .arg1_type      = ARG_PTR_TO_CTX,
1150         .arg2_type      = ARG_CONST_MAP_PTR,
1151         .arg3_type      = ARG_PTR_TO_MAP_KEY,
1152         .arg4_type      = ARG_ANYTHING,
1153 };
1154
1155 BPF_CALL_4(bpf_msg_redirect_hash, struct sk_msg *, msg,
1156            struct bpf_map *, map, void *, key, u64, flags)
1157 {
1158         struct sock *sk;
1159
1160         if (unlikely(flags & ~(BPF_F_INGRESS)))
1161                 return SK_DROP;
1162
1163         sk = __sock_hash_lookup_elem(map, key);
1164         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1165                 return SK_DROP;
1166
1167         msg->flags = flags;
1168         msg->sk_redir = sk;
1169         return SK_PASS;
1170 }
1171
1172 const struct bpf_func_proto bpf_msg_redirect_hash_proto = {
1173         .func           = bpf_msg_redirect_hash,
1174         .gpl_only       = false,
1175         .ret_type       = RET_INTEGER,
1176         .arg1_type      = ARG_PTR_TO_CTX,
1177         .arg2_type      = ARG_CONST_MAP_PTR,
1178         .arg3_type      = ARG_PTR_TO_MAP_KEY,
1179         .arg4_type      = ARG_ANYTHING,
1180 };
1181
1182 static int sock_hash_map_btf_id;
1183 const struct bpf_map_ops sock_hash_ops = {
1184         .map_alloc              = sock_hash_alloc,
1185         .map_free               = sock_hash_free,
1186         .map_get_next_key       = sock_hash_get_next_key,
1187         .map_update_elem        = sock_hash_update_elem,
1188         .map_delete_elem        = sock_hash_delete_elem,
1189         .map_lookup_elem        = sock_hash_lookup,
1190         .map_lookup_elem_sys_only = sock_hash_lookup_sys,
1191         .map_release_uref       = sock_hash_release_progs,
1192         .map_check_btf          = map_check_no_btf,
1193         .map_btf_name           = "bpf_shtab",
1194         .map_btf_id             = &sock_hash_map_btf_id,
1195 };
1196
1197 static struct sk_psock_progs *sock_map_progs(struct bpf_map *map)
1198 {
1199         switch (map->map_type) {
1200         case BPF_MAP_TYPE_SOCKMAP:
1201                 return &container_of(map, struct bpf_stab, map)->progs;
1202         case BPF_MAP_TYPE_SOCKHASH:
1203                 return &container_of(map, struct bpf_shtab, map)->progs;
1204         default:
1205                 break;
1206         }
1207
1208         return NULL;
1209 }
1210
1211 int sock_map_prog_update(struct bpf_map *map, struct bpf_prog *prog,
1212                          u32 which)
1213 {
1214         struct sk_psock_progs *progs = sock_map_progs(map);
1215
1216         if (!progs)
1217                 return -EOPNOTSUPP;
1218
1219         switch (which) {
1220         case BPF_SK_MSG_VERDICT:
1221                 psock_set_prog(&progs->msg_parser, prog);
1222                 break;
1223         case BPF_SK_SKB_STREAM_PARSER:
1224                 psock_set_prog(&progs->skb_parser, prog);
1225                 break;
1226         case BPF_SK_SKB_STREAM_VERDICT:
1227                 psock_set_prog(&progs->skb_verdict, prog);
1228                 break;
1229         default:
1230                 return -EOPNOTSUPP;
1231         }
1232
1233         return 0;
1234 }
1235
1236 static void sock_map_unlink(struct sock *sk, struct sk_psock_link *link)
1237 {
1238         switch (link->map->map_type) {
1239         case BPF_MAP_TYPE_SOCKMAP:
1240                 return sock_map_delete_from_link(link->map, sk,
1241                                                  link->link_raw);
1242         case BPF_MAP_TYPE_SOCKHASH:
1243                 return sock_hash_delete_from_link(link->map, sk,
1244                                                   link->link_raw);
1245         default:
1246                 break;
1247         }
1248 }
1249
1250 static void sock_map_remove_links(struct sock *sk, struct sk_psock *psock)
1251 {
1252         struct sk_psock_link *link;
1253
1254         while ((link = sk_psock_link_pop(psock))) {
1255                 sock_map_unlink(sk, link);
1256                 sk_psock_free_link(link);
1257         }
1258 }
1259
1260 void sock_map_unhash(struct sock *sk)
1261 {
1262         void (*saved_unhash)(struct sock *sk);
1263         struct sk_psock *psock;
1264
1265         rcu_read_lock();
1266         psock = sk_psock(sk);
1267         if (unlikely(!psock)) {
1268                 rcu_read_unlock();
1269                 if (sk->sk_prot->unhash)
1270                         sk->sk_prot->unhash(sk);
1271                 return;
1272         }
1273
1274         saved_unhash = psock->saved_unhash;
1275         sock_map_remove_links(sk, psock);
1276         rcu_read_unlock();
1277         saved_unhash(sk);
1278 }
1279
1280 void sock_map_close(struct sock *sk, long timeout)
1281 {
1282         void (*saved_close)(struct sock *sk, long timeout);
1283         struct sk_psock *psock;
1284
1285         lock_sock(sk);
1286         rcu_read_lock();
1287         psock = sk_psock(sk);
1288         if (unlikely(!psock)) {
1289                 rcu_read_unlock();
1290                 release_sock(sk);
1291                 return sk->sk_prot->close(sk, timeout);
1292         }
1293
1294         saved_close = psock->saved_close;
1295         sock_map_remove_links(sk, psock);
1296         rcu_read_unlock();
1297         release_sock(sk);
1298         saved_close(sk, timeout);
1299 }