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 const struct bpf_map_ops sock_map_ops = {
 647         .map_alloc              = sock_map_alloc,
 648         .map_free               = sock_map_free,
 649         .map_get_next_key       = sock_map_get_next_key,
 650         .map_lookup_elem_sys_only = sock_map_lookup_sys,
 651         .map_update_elem        = sock_map_update_elem,
 652         .map_delete_elem        = sock_map_delete_elem,
 653         .map_lookup_elem        = sock_map_lookup,
 654         .map_release_uref       = sock_map_release_progs,
 655         .map_check_btf          = map_check_no_btf,
 656 };
 657
 658 struct bpf_htab_elem {
 659         struct rcu_head rcu;
 660         u32 hash;
 661         struct sock *sk;
 662         struct hlist_node node;
 663         u8 key[];
 664 };
 665
 666 struct bpf_htab_bucket {
 667         struct hlist_head head;
 668         raw_spinlock_t lock;
 669 };
 670
 671 struct bpf_htab {
 672         struct bpf_map map;
 673         struct bpf_htab_bucket *buckets;
 674         u32 buckets_num;
 675         u32 elem_size;
 676         struct sk_psock_progs progs;
 677         atomic_t count;
 678 };
 679
 680 static inline u32 sock_hash_bucket_hash(const void *key, u32 len)
 681 {
 682         return jhash(key, len, 0);
 683 }
 684
 685 static struct bpf_htab_bucket *sock_hash_select_bucket(struct bpf_htab *htab,
 686                                                        u32 hash)
 687 {
 688         return &htab->buckets[hash & (htab->buckets_num - 1)];
 689 }
 690
 691 static struct bpf_htab_elem *
 692 sock_hash_lookup_elem_raw(struct hlist_head *head, u32 hash, void *key,
 693                           u32 key_size)
 694 {
 695         struct bpf_htab_elem *elem;
 696
 697         hlist_for_each_entry_rcu(elem, head, node) {
 698                 if (elem->hash == hash &&
 699                     !memcmp(&elem->key, key, key_size))
 700                         return elem;
 701         }
 702
 703         return NULL;
 704 }
 705
 706 static struct sock *__sock_hash_lookup_elem(struct bpf_map *map, void *key)
 707 {
 708         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 709         u32 key_size = map->key_size, hash;
 710         struct bpf_htab_bucket *bucket;
 711         struct bpf_htab_elem *elem;
 712
 713         WARN_ON_ONCE(!rcu_read_lock_held());
 714
 715         hash = sock_hash_bucket_hash(key, key_size);
 716         bucket = sock_hash_select_bucket(htab, hash);
 717         elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 718
 719         return elem ? elem->sk : NULL;
 720 }
 721
 722 static void sock_hash_free_elem(struct bpf_htab *htab,
 723                                 struct bpf_htab_elem *elem)
 724 {
 725         atomic_dec(&htab->count);
 726         kfree_rcu(elem, rcu);
 727 }
 728
 729 static void sock_hash_delete_from_link(struct bpf_map *map, struct sock *sk,
 730                                        void *link_raw)
 731 {
 732         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 733         struct bpf_htab_elem *elem_probe, *elem = link_raw;
 734         struct bpf_htab_bucket *bucket;
 735
 736         WARN_ON_ONCE(!rcu_read_lock_held());
 737         bucket = sock_hash_select_bucket(htab, elem->hash);
 738
 739         /* elem may be deleted in parallel from the map, but access here
 740          * is okay since it's going away only after RCU grace period.
 741          * However, we need to check whether it's still present.
 742          */
 743         raw_spin_lock_bh(&bucket->lock);
 744         elem_probe = sock_hash_lookup_elem_raw(&bucket->head, elem->hash,
 745                                                elem->key, map->key_size);
 746         if (elem_probe && elem_probe == elem) {
 747                 hlist_del_rcu(&elem->node);
 748                 sock_map_unref(elem->sk, elem);
 749                 sock_hash_free_elem(htab, elem);
 750         }
 751         raw_spin_unlock_bh(&bucket->lock);
 752 }
 753
 754 static int sock_hash_delete_elem(struct bpf_map *map, void *key)
 755 {
 756         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 757         u32 hash, key_size = map->key_size;
 758         struct bpf_htab_bucket *bucket;
 759         struct bpf_htab_elem *elem;
 760         int ret = -ENOENT;
 761
 762         hash = sock_hash_bucket_hash(key, key_size);
 763         bucket = sock_hash_select_bucket(htab, hash);
 764
 765         raw_spin_lock_bh(&bucket->lock);
 766         elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 767         if (elem) {
 768                 hlist_del_rcu(&elem->node);
 769                 sock_map_unref(elem->sk, elem);
 770                 sock_hash_free_elem(htab, elem);
 771                 ret = 0;
 772         }
 773         raw_spin_unlock_bh(&bucket->lock);
 774         return ret;
 775 }
 776
 777 static struct bpf_htab_elem *sock_hash_alloc_elem(struct bpf_htab *htab,
 778                                                   void *key, u32 key_size,
 779                                                   u32 hash, struct sock *sk,
 780                                                   struct bpf_htab_elem *old)
 781 {
 782         struct bpf_htab_elem *new;
 783
 784         if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
 785                 if (!old) {
 786                         atomic_dec(&htab->count);
 787                         return ERR_PTR(-E2BIG);
 788                 }
 789         }
 790
 791         new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 792                            htab->map.numa_node);
 793         if (!new) {
 794                 atomic_dec(&htab->count);
 795                 return ERR_PTR(-ENOMEM);
 796         }
 797         memcpy(new->key, key, key_size);
 798         new->sk = sk;
 799         new->hash = hash;
 800         return new;
 801 }
 802
 803 static int sock_hash_update_common(struct bpf_map *map, void *key,
 804                                    struct sock *sk, u64 flags)
 805 {
 806         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 807         u32 key_size = map->key_size, hash;
 808         struct bpf_htab_elem *elem, *elem_new;
 809         struct bpf_htab_bucket *bucket;
 810         struct sk_psock_link *link;
 811         struct sk_psock *psock;
 812         int ret;
 813
 814         WARN_ON_ONCE(!rcu_read_lock_held());
 815         if (unlikely(flags > BPF_EXIST))
 816                 return -EINVAL;
 817         if (inet_csk_has_ulp(sk))
 818                 return -EINVAL;
 819
 820         link = sk_psock_init_link();
 821         if (!link)
 822                 return -ENOMEM;
 823
 824         /* Only sockets we can redirect into/from in BPF need to hold
 825          * refs to parser/verdict progs and have their sk_data_ready
 826          * and sk_write_space callbacks overridden.
 827          */
 828         if (sock_map_redirect_allowed(sk))
 829                 ret = sock_map_link(map, &htab->progs, sk);
 830         else
 831                 ret = sock_map_link_no_progs(map, sk);
 832         if (ret < 0)
 833                 goto out_free;
 834
 835         psock = sk_psock(sk);
 836         WARN_ON_ONCE(!psock);
 837
 838         hash = sock_hash_bucket_hash(key, key_size);
 839         bucket = sock_hash_select_bucket(htab, hash);
 840
 841         raw_spin_lock_bh(&bucket->lock);
 842         elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 843         if (elem && flags == BPF_NOEXIST) {
 844                 ret = -EEXIST;
 845                 goto out_unlock;
 846         } else if (!elem && flags == BPF_EXIST) {
 847                 ret = -ENOENT;
 848                 goto out_unlock;
 849         }
 850
 851         elem_new = sock_hash_alloc_elem(htab, key, key_size, hash, sk, elem);
 852         if (IS_ERR(elem_new)) {
 853                 ret = PTR_ERR(elem_new);
 854                 goto out_unlock;
 855         }
 856
 857         sock_map_add_link(psock, link, map, elem_new);
 858         /* Add new element to the head of the list, so that
 859          * concurrent search will find it before old elem.
 860          */
 861         hlist_add_head_rcu(&elem_new->node, &bucket->head);
 862         if (elem) {
 863                 hlist_del_rcu(&elem->node);
 864                 sock_map_unref(elem->sk, elem);
 865                 sock_hash_free_elem(htab, elem);
 866         }
 867         raw_spin_unlock_bh(&bucket->lock);
 868         return 0;
 869 out_unlock:
 870         raw_spin_unlock_bh(&bucket->lock);
 871         sk_psock_put(sk, psock);
 872 out_free:
 873         sk_psock_free_link(link);
 874         return ret;
 875 }
 876
 877 static int sock_hash_update_elem(struct bpf_map *map, void *key,
 878                                  void *value, u64 flags)
 879 {
 880         struct socket *sock;
 881         struct sock *sk;
 882         int ret;
 883         u64 ufd;
 884
 885         if (map->value_size == sizeof(u64))
 886                 ufd = *(u64 *)value;
 887         else
 888                 ufd = *(u32 *)value;
 889         if (ufd > S32_MAX)
 890                 return -EINVAL;
 891
 892         sock = sockfd_lookup(ufd, &ret);
 893         if (!sock)
 894                 return ret;
 895         sk = sock->sk;
 896         if (!sk) {
 897                 ret = -EINVAL;
 898                 goto out;
 899         }
 900         if (!sock_map_sk_is_suitable(sk)) {
 901                 ret = -EOPNOTSUPP;
 902                 goto out;
 903         }
 904
 905         sock_map_sk_acquire(sk);
 906         if (!sock_map_sk_state_allowed(sk))
 907                 ret = -EOPNOTSUPP;
 908         else
 909                 ret = sock_hash_update_common(map, key, sk, flags);
 910         sock_map_sk_release(sk);
 911 out:
 912         fput(sock->file);
 913         return ret;
 914 }
 915
 916 static int sock_hash_get_next_key(struct bpf_map *map, void *key,
 917                                   void *key_next)
 918 {
 919         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 920         struct bpf_htab_elem *elem, *elem_next;
 921         u32 hash, key_size = map->key_size;
 922         struct hlist_head *head;
 923         int i = 0;
 924
 925         if (!key)
 926                 goto find_first_elem;
 927         hash = sock_hash_bucket_hash(key, key_size);
 928         head = &sock_hash_select_bucket(htab, hash)->head;
 929         elem = sock_hash_lookup_elem_raw(head, hash, key, key_size);
 930         if (!elem)
 931                 goto find_first_elem;
 932
 933         elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&elem->node)),
 934                                      struct bpf_htab_elem, node);
 935         if (elem_next) {
 936                 memcpy(key_next, elem_next->key, key_size);
 937                 return 0;
 938         }
 939
 940         i = hash & (htab->buckets_num - 1);
 941         i++;
 942 find_first_elem:
 943         for (; i < htab->buckets_num; i++) {
 944                 head = &sock_hash_select_bucket(htab, i)->head;
 945                 elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
 946                                              struct bpf_htab_elem, node);
 947                 if (elem_next) {
 948                         memcpy(key_next, elem_next->key, key_size);
 949                         return 0;
 950                 }
 951         }
 952
 953         return -ENOENT;
 954 }
 955
 956 static struct bpf_map *sock_hash_alloc(union bpf_attr *attr)
 957 {
 958         struct bpf_htab *htab;
 959         int i, err;
 960         u64 cost;
 961
 962         if (!capable(CAP_NET_ADMIN))
 963                 return ERR_PTR(-EPERM);
 964         if (attr->max_entries == 0 ||
 965             attr->key_size    == 0 ||
 966             (attr->value_size != sizeof(u32) &&
 967              attr->value_size != sizeof(u64)) ||
 968             attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
 969                 return ERR_PTR(-EINVAL);
 970         if (attr->key_size > MAX_BPF_STACK)
 971                 return ERR_PTR(-E2BIG);
 972
 973         htab = kzalloc(sizeof(*htab), GFP_USER);
 974         if (!htab)
 975                 return ERR_PTR(-ENOMEM);
 976
 977         bpf_map_init_from_attr(&htab->map, attr);
 978
 979         htab->buckets_num = roundup_pow_of_two(htab->map.max_entries);
 980         htab->elem_size = sizeof(struct bpf_htab_elem) +
 981                           round_up(htab->map.key_size, 8);
 982         if (htab->buckets_num == 0 ||
 983             htab->buckets_num > U32_MAX / sizeof(struct bpf_htab_bucket)) {
 984                 err = -EINVAL;
 985                 goto free_htab;
 986         }
 987
 988         cost = (u64) htab->buckets_num * sizeof(struct bpf_htab_bucket) +
 989                (u64) htab->elem_size * htab->map.max_entries;
 990         if (cost >= U32_MAX - PAGE_SIZE) {
 991                 err = -EINVAL;
 992                 goto free_htab;
 993         }
 994
 995         htab->buckets = bpf_map_area_alloc(htab->buckets_num *
 996                                            sizeof(struct bpf_htab_bucket),
 997                                            htab->map.numa_node);
 998         if (!htab->buckets) {
 999                 err = -ENOMEM;
1000                 goto free_htab;
1001         }
1002
1003         for (i = 0; i < htab->buckets_num; i++) {
1004                 INIT_HLIST_HEAD(&htab->buckets[i].head);
1005                 raw_spin_lock_init(&htab->buckets[i].lock);
1006         }
1007
1008         return &htab->map;
1009 free_htab:
1010         kfree(htab);
1011         return ERR_PTR(err);
1012 }
1013
1014 static void sock_hash_free(struct bpf_map *map)
1015 {
1016         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1017         struct bpf_htab_bucket *bucket;
1018         struct hlist_head unlink_list;
1019         struct bpf_htab_elem *elem;
1020         struct hlist_node *node;
1021         int i;
1022
1023         /* After the sync no updates or deletes will be in-flight so it
1024          * is safe to walk map and remove entries without risking a race
1025          * in EEXIST update case.
1026          */
1027         synchronize_rcu();
1028         for (i = 0; i < htab->buckets_num; i++) {
1029                 bucket = sock_hash_select_bucket(htab, i);
1030
1031                 /* We are racing with sock_hash_delete_from_link to
1032                  * enter the spin-lock critical section. Every socket on
1033                  * the list is still linked to sockhash. Since link
1034                  * exists, psock exists and holds a ref to socket. That
1035                  * lets us to grab a socket ref too.
1036                  */
1037                 raw_spin_lock_bh(&bucket->lock);
1038                 hlist_for_each_entry(elem, &bucket->head, node)
1039                         sock_hold(elem->sk);
1040                 hlist_move_list(&bucket->head, &unlink_list);
1041                 raw_spin_unlock_bh(&bucket->lock);
1042
1043                 /* Process removed entries out of atomic context to
1044                  * block for socket lock before deleting the psock's
1045                  * link to sockhash.
1046                  */
1047                 hlist_for_each_entry_safe(elem, node, &unlink_list, node) {
1048                         hlist_del(&elem->node);
1049                         lock_sock(elem->sk);
1050                         rcu_read_lock();
1051                         sock_map_unref(elem->sk, elem);
1052                         rcu_read_unlock();
1053                         release_sock(elem->sk);
1054                         sock_put(elem->sk);
1055                         sock_hash_free_elem(htab, elem);
1056                 }
1057         }
1058
1059         /* wait for psock readers accessing its map link */
1060         synchronize_rcu();
1061
1062         bpf_map_area_free(htab->buckets);
1063         kfree(htab);
1064 }
1065
1066 static void *sock_hash_lookup_sys(struct bpf_map *map, void *key)
1067 {
1068         struct sock *sk;
1069
1070         if (map->value_size != sizeof(u64))
1071                 return ERR_PTR(-ENOSPC);
1072
1073         sk = __sock_hash_lookup_elem(map, key);
1074         if (!sk)
1075                 return ERR_PTR(-ENOENT);
1076
1077         sock_gen_cookie(sk);
1078         return &sk->sk_cookie;
1079 }
1080
1081 static void *sock_hash_lookup(struct bpf_map *map, void *key)
1082 {
1083         struct sock *sk;
1084
1085         sk = __sock_hash_lookup_elem(map, key);
1086         if (!sk || !sk_fullsock(sk))
1087                 return NULL;
1088         if (sk_is_refcounted(sk) && !refcount_inc_not_zero(&sk->sk_refcnt))
1089                 return NULL;
1090         return sk;
1091 }
1092
1093 static void sock_hash_release_progs(struct bpf_map *map)
1094 {
1095         psock_progs_drop(&container_of(map, struct bpf_htab, map)->progs);
1096 }
1097
1098 BPF_CALL_4(bpf_sock_hash_update, struct bpf_sock_ops_kern *, sops,
1099            struct bpf_map *, map, void *, key, u64, flags)
1100 {
1101         WARN_ON_ONCE(!rcu_read_lock_held());
1102
1103         if (likely(sock_map_sk_is_suitable(sops->sk) &&
1104                    sock_map_op_okay(sops)))
1105                 return sock_hash_update_common(map, key, sops->sk, flags);
1106         return -EOPNOTSUPP;
1107 }
1108
1109 const struct bpf_func_proto bpf_sock_hash_update_proto = {
1110         .func           = bpf_sock_hash_update,
1111         .gpl_only       = false,
1112         .pkt_access     = true,
1113         .ret_type       = RET_INTEGER,
1114         .arg1_type      = ARG_PTR_TO_CTX,
1115         .arg2_type      = ARG_CONST_MAP_PTR,
1116         .arg3_type      = ARG_PTR_TO_MAP_KEY,
1117         .arg4_type      = ARG_ANYTHING,
1118 };
1119
1120 BPF_CALL_4(bpf_sk_redirect_hash, struct sk_buff *, skb,
1121            struct bpf_map *, map, void *, key, u64, flags)
1122 {
1123         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1124         struct sock *sk;
1125
1126         if (unlikely(flags & ~(BPF_F_INGRESS)))
1127                 return SK_DROP;
1128
1129         sk = __sock_hash_lookup_elem(map, key);
1130         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1131                 return SK_DROP;
1132
1133         tcb->bpf.flags = flags;
1134         tcb->bpf.sk_redir = sk;
1135         return SK_PASS;
1136 }
1137
1138 const struct bpf_func_proto bpf_sk_redirect_hash_proto = {
1139         .func           = bpf_sk_redirect_hash,
1140         .gpl_only       = false,
1141         .ret_type       = RET_INTEGER,
1142         .arg1_type      = ARG_PTR_TO_CTX,
1143         .arg2_type      = ARG_CONST_MAP_PTR,
1144         .arg3_type      = ARG_PTR_TO_MAP_KEY,
1145         .arg4_type      = ARG_ANYTHING,
1146 };
1147
1148 BPF_CALL_4(bpf_msg_redirect_hash, struct sk_msg *, msg,
1149            struct bpf_map *, map, void *, key, u64, flags)
1150 {
1151         struct sock *sk;
1152
1153         if (unlikely(flags & ~(BPF_F_INGRESS)))
1154                 return SK_DROP;
1155
1156         sk = __sock_hash_lookup_elem(map, key);
1157         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1158                 return SK_DROP;
1159
1160         msg->flags = flags;
1161         msg->sk_redir = sk;
1162         return SK_PASS;
1163 }
1164
1165 const struct bpf_func_proto bpf_msg_redirect_hash_proto = {
1166         .func           = bpf_msg_redirect_hash,
1167         .gpl_only       = false,
1168         .ret_type       = RET_INTEGER,
1169         .arg1_type      = ARG_PTR_TO_CTX,
1170         .arg2_type      = ARG_CONST_MAP_PTR,
1171         .arg3_type      = ARG_PTR_TO_MAP_KEY,
1172         .arg4_type      = ARG_ANYTHING,
1173 };
1174
1175 const struct bpf_map_ops sock_hash_ops = {
1176         .map_alloc              = sock_hash_alloc,
1177         .map_free               = sock_hash_free,
1178         .map_get_next_key       = sock_hash_get_next_key,
1179         .map_update_elem        = sock_hash_update_elem,
1180         .map_delete_elem        = sock_hash_delete_elem,
1181         .map_lookup_elem        = sock_hash_lookup,
1182         .map_lookup_elem_sys_only = sock_hash_lookup_sys,
1183         .map_release_uref       = sock_hash_release_progs,
1184         .map_check_btf          = map_check_no_btf,
1185 };
1186
1187 static struct sk_psock_progs *sock_map_progs(struct bpf_map *map)
1188 {
1189         switch (map->map_type) {
1190         case BPF_MAP_TYPE_SOCKMAP:
1191                 return &container_of(map, struct bpf_stab, map)->progs;
1192         case BPF_MAP_TYPE_SOCKHASH:
1193                 return &container_of(map, struct bpf_htab, map)->progs;
1194         default:
1195                 break;
1196         }
1197
1198         return NULL;
1199 }
1200
1201 int sock_map_prog_update(struct bpf_map *map, struct bpf_prog *prog,
1202                          u32 which)
1203 {
1204         struct sk_psock_progs *progs = sock_map_progs(map);
1205
1206         if (!progs)
1207                 return -EOPNOTSUPP;
1208
1209         switch (which) {
1210         case BPF_SK_MSG_VERDICT:
1211                 psock_set_prog(&progs->msg_parser, prog);
1212                 break;
1213         case BPF_SK_SKB_STREAM_PARSER:
1214                 psock_set_prog(&progs->skb_parser, prog);
1215                 break;
1216         case BPF_SK_SKB_STREAM_VERDICT:
1217                 psock_set_prog(&progs->skb_verdict, prog);
1218                 break;
1219         default:
1220                 return -EOPNOTSUPP;
1221         }
1222
1223         return 0;
1224 }
1225
1226 static void sock_map_unlink(struct sock *sk, struct sk_psock_link *link)
1227 {
1228         switch (link->map->map_type) {
1229         case BPF_MAP_TYPE_SOCKMAP:
1230                 return sock_map_delete_from_link(link->map, sk,
1231                                                  link->link_raw);
1232         case BPF_MAP_TYPE_SOCKHASH:
1233                 return sock_hash_delete_from_link(link->map, sk,
1234                                                   link->link_raw);
1235         default:
1236                 break;
1237         }
1238 }
1239
1240 static void sock_map_remove_links(struct sock *sk, struct sk_psock *psock)
1241 {
1242         struct sk_psock_link *link;
1243
1244         while ((link = sk_psock_link_pop(psock))) {
1245                 sock_map_unlink(sk, link);
1246                 sk_psock_free_link(link);
1247         }
1248 }
1249
1250 void sock_map_unhash(struct sock *sk)
1251 {
1252         void (*saved_unhash)(struct sock *sk);
1253         struct sk_psock *psock;
1254
1255         rcu_read_lock();
1256         psock = sk_psock(sk);
1257         if (unlikely(!psock)) {
1258                 rcu_read_unlock();
1259                 if (sk->sk_prot->unhash)
1260                         sk->sk_prot->unhash(sk);
1261                 return;
1262         }
1263
1264         saved_unhash = psock->saved_unhash;
1265         sock_map_remove_links(sk, psock);
1266         rcu_read_unlock();
1267         saved_unhash(sk);
1268 }
1269
1270 void sock_map_close(struct sock *sk, long timeout)
1271 {
1272         void (*saved_close)(struct sock *sk, long timeout);
1273         struct sk_psock *psock;
1274
1275         lock_sock(sk);
1276         rcu_read_lock();
1277         psock = sk_psock(sk);
1278         if (unlikely(!psock)) {
1279                 rcu_read_unlock();
1280                 release_sock(sk);
1281                 return sk->sk_prot->close(sk, timeout);
1282         }
1283
1284         saved_close = psock->saved_close;
1285         sock_map_remove_links(sk, psock);
1286         rcu_read_unlock();
1287         release_sock(sk);
1288         saved_close(sk, timeout);
1289 }