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         if (attr->attach_flags || attr->replace_bpf_fd)
  74                 return -EINVAL;
  75
  76         f = fdget(ufd);
  77         map = __bpf_map_get(f);
  78         if (IS_ERR(map))
  79                 return PTR_ERR(map);
  80         ret = sock_map_prog_update(map, prog, NULL, attr->attach_type);
  81         fdput(f);
  82         return ret;
  83 }
  84
  85 int sock_map_prog_detach(const union bpf_attr *attr, enum bpf_prog_type ptype)
  86 {
  87         u32 ufd = attr->target_fd;
  88         struct bpf_prog *prog;
  89         struct bpf_map *map;
  90         struct fd f;
  91         int ret;
  92
  93         if (attr->attach_flags || attr->replace_bpf_fd)
  94                 return -EINVAL;
  95
  96         f = fdget(ufd);
  97         map = __bpf_map_get(f);
  98         if (IS_ERR(map))
  99                 return PTR_ERR(map);
 100
 101         prog = bpf_prog_get(attr->attach_bpf_fd);
 102         if (IS_ERR(prog)) {
 103                 ret = PTR_ERR(prog);
 104                 goto put_map;
 105         }
 106
 107         if (prog->type != ptype) {
 108                 ret = -EINVAL;
 109                 goto put_prog;
 110         }
 111
 112         ret = sock_map_prog_update(map, NULL, prog, attr->attach_type);
 113 put_prog:
 114         bpf_prog_put(prog);
 115 put_map:
 116         fdput(f);
 117         return ret;
 118 }
 119
 120 static void sock_map_sk_acquire(struct sock *sk)
 121         __acquires(&sk->sk_lock.slock)
 122 {
 123         lock_sock(sk);
 124         preempt_disable();
 125         rcu_read_lock();
 126 }
 127
 128 static void sock_map_sk_release(struct sock *sk)
 129         __releases(&sk->sk_lock.slock)
 130 {
 131         rcu_read_unlock();
 132         preempt_enable();
 133         release_sock(sk);
 134 }
 135
 136 static void sock_map_add_link(struct sk_psock *psock,
 137                               struct sk_psock_link *link,
 138                               struct bpf_map *map, void *link_raw)
 139 {
 140         link->link_raw = link_raw;
 141         link->map = map;
 142         spin_lock_bh(&psock->link_lock);
 143         list_add_tail(&link->list, &psock->link);
 144         spin_unlock_bh(&psock->link_lock);
 145 }
 146
 147 static void sock_map_del_link(struct sock *sk,
 148                               struct sk_psock *psock, void *link_raw)
 149 {
 150         struct sk_psock_link *link, *tmp;
 151         bool strp_stop = false;
 152
 153         spin_lock_bh(&psock->link_lock);
 154         list_for_each_entry_safe(link, tmp, &psock->link, list) {
 155                 if (link->link_raw == link_raw) {
 156                         struct bpf_map *map = link->map;
 157                         struct bpf_stab *stab = container_of(map, struct bpf_stab,
 158                                                              map);
 159                         if (psock->parser.enabled && stab->progs.skb_parser)
 160                                 strp_stop = true;
 161                         list_del(&link->list);
 162                         sk_psock_free_link(link);
 163                 }
 164         }
 165         spin_unlock_bh(&psock->link_lock);
 166         if (strp_stop) {
 167                 write_lock_bh(&sk->sk_callback_lock);
 168                 sk_psock_stop_strp(sk, psock);
 169                 write_unlock_bh(&sk->sk_callback_lock);
 170         }
 171 }
 172
 173 static void sock_map_unref(struct sock *sk, void *link_raw)
 174 {
 175         struct sk_psock *psock = sk_psock(sk);
 176
 177         if (likely(psock)) {
 178                 sock_map_del_link(sk, psock, link_raw);
 179                 sk_psock_put(sk, psock);
 180         }
 181 }
 182
 183 static int sock_map_init_proto(struct sock *sk, struct sk_psock *psock)
 184 {
 185         struct proto *prot;
 186
 187         switch (sk->sk_type) {
 188         case SOCK_STREAM:
 189                 prot = tcp_bpf_get_proto(sk, psock);
 190                 break;
 191
 192         case SOCK_DGRAM:
 193                 prot = udp_bpf_get_proto(sk, psock);
 194                 break;
 195
 196         default:
 197                 return -EINVAL;
 198         }
 199
 200         if (IS_ERR(prot))
 201                 return PTR_ERR(prot);
 202
 203         sk_psock_update_proto(sk, psock, prot);
 204         return 0;
 205 }
 206
 207 static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
 208 {
 209         struct sk_psock *psock;
 210
 211         rcu_read_lock();
 212         psock = sk_psock(sk);
 213         if (psock) {
 214                 if (sk->sk_prot->close != sock_map_close) {
 215                         psock = ERR_PTR(-EBUSY);
 216                         goto out;
 217                 }
 218
 219                 if (!refcount_inc_not_zero(&psock->refcnt))
 220                         psock = ERR_PTR(-EBUSY);
 221         }
 222 out:
 223         rcu_read_unlock();
 224         return psock;
 225 }
 226
 227 static int sock_map_link(struct bpf_map *map, struct sk_psock_progs *progs,
 228                          struct sock *sk)
 229 {
 230         struct bpf_prog *msg_parser, *skb_parser, *skb_verdict;
 231         struct sk_psock *psock;
 232         bool skb_progs;
 233         int ret;
 234
 235         skb_verdict = READ_ONCE(progs->skb_verdict);
 236         skb_parser = READ_ONCE(progs->skb_parser);
 237         skb_progs = skb_parser && skb_verdict;
 238         if (skb_progs) {
 239                 skb_verdict = bpf_prog_inc_not_zero(skb_verdict);
 240                 if (IS_ERR(skb_verdict))
 241                         return PTR_ERR(skb_verdict);
 242                 skb_parser = bpf_prog_inc_not_zero(skb_parser);
 243                 if (IS_ERR(skb_parser)) {
 244                         bpf_prog_put(skb_verdict);
 245                         return PTR_ERR(skb_parser);
 246                 }
 247         }
 248
 249         msg_parser = READ_ONCE(progs->msg_parser);
 250         if (msg_parser) {
 251                 msg_parser = bpf_prog_inc_not_zero(msg_parser);
 252                 if (IS_ERR(msg_parser)) {
 253                         ret = PTR_ERR(msg_parser);
 254                         goto out;
 255                 }
 256         }
 257
 258         psock = sock_map_psock_get_checked(sk);
 259         if (IS_ERR(psock)) {
 260                 ret = PTR_ERR(psock);
 261                 goto out_progs;
 262         }
 263
 264         if (psock) {
 265                 if ((msg_parser && READ_ONCE(psock->progs.msg_parser)) ||
 266                     (skb_progs  && READ_ONCE(psock->progs.skb_parser))) {
 267                         sk_psock_put(sk, psock);
 268                         ret = -EBUSY;
 269                         goto out_progs;
 270                 }
 271         } else {
 272                 psock = sk_psock_init(sk, map->numa_node);
 273                 if (IS_ERR(psock)) {
 274                         ret = PTR_ERR(psock);
 275                         goto out_progs;
 276                 }
 277         }
 278
 279         if (msg_parser)
 280                 psock_set_prog(&psock->progs.msg_parser, msg_parser);
 281
 282         ret = sock_map_init_proto(sk, psock);
 283         if (ret < 0)
 284                 goto out_drop;
 285
 286         write_lock_bh(&sk->sk_callback_lock);
 287         if (skb_progs && !psock->parser.enabled) {
 288                 ret = sk_psock_init_strp(sk, psock);
 289                 if (ret) {
 290                         write_unlock_bh(&sk->sk_callback_lock);
 291                         goto out_drop;
 292                 }
 293                 psock_set_prog(&psock->progs.skb_verdict, skb_verdict);
 294                 psock_set_prog(&psock->progs.skb_parser, skb_parser);
 295                 sk_psock_start_strp(sk, psock);
 296         }
 297         write_unlock_bh(&sk->sk_callback_lock);
 298         return 0;
 299 out_drop:
 300         sk_psock_put(sk, psock);
 301 out_progs:
 302         if (msg_parser)
 303                 bpf_prog_put(msg_parser);
 304 out:
 305         if (skb_progs) {
 306                 bpf_prog_put(skb_verdict);
 307                 bpf_prog_put(skb_parser);
 308         }
 309         return ret;
 310 }
 311
 312 static int sock_map_link_no_progs(struct bpf_map *map, struct sock *sk)
 313 {
 314         struct sk_psock *psock;
 315         int ret;
 316
 317         psock = sock_map_psock_get_checked(sk);
 318         if (IS_ERR(psock))
 319                 return PTR_ERR(psock);
 320
 321         if (!psock) {
 322                 psock = sk_psock_init(sk, map->numa_node);
 323                 if (IS_ERR(psock))
 324                         return PTR_ERR(psock);
 325         }
 326
 327         ret = sock_map_init_proto(sk, psock);
 328         if (ret < 0)
 329                 sk_psock_put(sk, psock);
 330         return ret;
 331 }
 332
 333 static void sock_map_free(struct bpf_map *map)
 334 {
 335         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 336         int i;
 337
 338         /* After the sync no updates or deletes will be in-flight so it
 339          * is safe to walk map and remove entries without risking a race
 340          * in EEXIST update case.
 341          */
 342         synchronize_rcu();
 343         for (i = 0; i < stab->map.max_entries; i++) {
 344                 struct sock **psk = &stab->sks[i];
 345                 struct sock *sk;
 346
 347                 sk = xchg(psk, NULL);
 348                 if (sk) {
 349                         lock_sock(sk);
 350                         rcu_read_lock();
 351                         sock_map_unref(sk, psk);
 352                         rcu_read_unlock();
 353                         release_sock(sk);
 354                 }
 355         }
 356
 357         /* wait for psock readers accessing its map link */
 358         synchronize_rcu();
 359
 360         bpf_map_area_free(stab->sks);
 361         kfree(stab);
 362 }
 363
 364 static void sock_map_release_progs(struct bpf_map *map)
 365 {
 366         psock_progs_drop(&container_of(map, struct bpf_stab, map)->progs);
 367 }
 368
 369 static struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key)
 370 {
 371         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 372
 373         WARN_ON_ONCE(!rcu_read_lock_held());
 374
 375         if (unlikely(key >= map->max_entries))
 376                 return NULL;
 377         return READ_ONCE(stab->sks[key]);
 378 }
 379
 380 static void *sock_map_lookup(struct bpf_map *map, void *key)
 381 {
 382         struct sock *sk;
 383
 384         sk = __sock_map_lookup_elem(map, *(u32 *)key);
 385         if (!sk || !sk_fullsock(sk))
 386                 return NULL;
 387         if (sk_is_refcounted(sk) && !refcount_inc_not_zero(&sk->sk_refcnt))
 388                 return NULL;
 389         return sk;
 390 }
 391
 392 static void *sock_map_lookup_sys(struct bpf_map *map, void *key)
 393 {
 394         struct sock *sk;
 395
 396         if (map->value_size != sizeof(u64))
 397                 return ERR_PTR(-ENOSPC);
 398
 399         sk = __sock_map_lookup_elem(map, *(u32 *)key);
 400         if (!sk)
 401                 return ERR_PTR(-ENOENT);
 402
 403         sock_gen_cookie(sk);
 404         return &sk->sk_cookie;
 405 }
 406
 407 static int __sock_map_delete(struct bpf_stab *stab, struct sock *sk_test,
 408                              struct sock **psk)
 409 {
 410         struct sock *sk;
 411         int err = 0;
 412
 413         raw_spin_lock_bh(&stab->lock);
 414         sk = *psk;
 415         if (!sk_test || sk_test == sk)
 416                 sk = xchg(psk, NULL);
 417
 418         if (likely(sk))
 419                 sock_map_unref(sk, psk);
 420         else
 421                 err = -EINVAL;
 422
 423         raw_spin_unlock_bh(&stab->lock);
 424         return err;
 425 }
 426
 427 static void sock_map_delete_from_link(struct bpf_map *map, struct sock *sk,
 428                                       void *link_raw)
 429 {
 430         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 431
 432         __sock_map_delete(stab, sk, link_raw);
 433 }
 434
 435 static int sock_map_delete_elem(struct bpf_map *map, void *key)
 436 {
 437         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 438         u32 i = *(u32 *)key;
 439         struct sock **psk;
 440
 441         if (unlikely(i >= map->max_entries))
 442                 return -EINVAL;
 443
 444         psk = &stab->sks[i];
 445         return __sock_map_delete(stab, NULL, psk);
 446 }
 447
 448 static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next)
 449 {
 450         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 451         u32 i = key ? *(u32 *)key : U32_MAX;
 452         u32 *key_next = next;
 453
 454         if (i == stab->map.max_entries - 1)
 455                 return -ENOENT;
 456         if (i >= stab->map.max_entries)
 457                 *key_next = 0;
 458         else
 459                 *key_next = i + 1;
 460         return 0;
 461 }
 462
 463 static bool sock_map_redirect_allowed(const struct sock *sk);
 464
 465 static int sock_map_update_common(struct bpf_map *map, u32 idx,
 466                                   struct sock *sk, u64 flags)
 467 {
 468         struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 469         struct sk_psock_link *link;
 470         struct sk_psock *psock;
 471         struct sock *osk;
 472         int ret;
 473
 474         WARN_ON_ONCE(!rcu_read_lock_held());
 475         if (unlikely(flags > BPF_EXIST))
 476                 return -EINVAL;
 477         if (unlikely(idx >= map->max_entries))
 478                 return -E2BIG;
 479
 480         link = sk_psock_init_link();
 481         if (!link)
 482                 return -ENOMEM;
 483
 484         /* Only sockets we can redirect into/from in BPF need to hold
 485          * refs to parser/verdict progs and have their sk_data_ready
 486          * and sk_write_space callbacks overridden.
 487          */
 488         if (sock_map_redirect_allowed(sk))
 489                 ret = sock_map_link(map, &stab->progs, sk);
 490         else
 491                 ret = sock_map_link_no_progs(map, sk);
 492         if (ret < 0)
 493                 goto out_free;
 494
 495         psock = sk_psock(sk);
 496         WARN_ON_ONCE(!psock);
 497
 498         raw_spin_lock_bh(&stab->lock);
 499         osk = stab->sks[idx];
 500         if (osk && flags == BPF_NOEXIST) {
 501                 ret = -EEXIST;
 502                 goto out_unlock;
 503         } else if (!osk && flags == BPF_EXIST) {
 504                 ret = -ENOENT;
 505                 goto out_unlock;
 506         }
 507
 508         sock_map_add_link(psock, link, map, &stab->sks[idx]);
 509         stab->sks[idx] = sk;
 510         if (osk)
 511                 sock_map_unref(osk, &stab->sks[idx]);
 512         raw_spin_unlock_bh(&stab->lock);
 513         return 0;
 514 out_unlock:
 515         raw_spin_unlock_bh(&stab->lock);
 516         if (psock)
 517                 sk_psock_put(sk, psock);
 518 out_free:
 519         sk_psock_free_link(link);
 520         return ret;
 521 }
 522
 523 static bool sock_map_op_okay(const struct bpf_sock_ops_kern *ops)
 524 {
 525         return ops->op == BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB ||
 526                ops->op == BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB ||
 527                ops->op == BPF_SOCK_OPS_TCP_LISTEN_CB;
 528 }
 529
 530 static bool sk_is_tcp(const struct sock *sk)
 531 {
 532         return sk->sk_type == SOCK_STREAM &&
 533                sk->sk_protocol == IPPROTO_TCP;
 534 }
 535
 536 static bool sk_is_udp(const struct sock *sk)
 537 {
 538         return sk->sk_type == SOCK_DGRAM &&
 539                sk->sk_protocol == IPPROTO_UDP;
 540 }
 541
 542 static bool sock_map_redirect_allowed(const struct sock *sk)
 543 {
 544         return sk_is_tcp(sk) && sk->sk_state != TCP_LISTEN;
 545 }
 546
 547 static bool sock_map_sk_is_suitable(const struct sock *sk)
 548 {
 549         return sk_is_tcp(sk) || sk_is_udp(sk);
 550 }
 551
 552 static bool sock_map_sk_state_allowed(const struct sock *sk)
 553 {
 554         if (sk_is_tcp(sk))
 555                 return (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_LISTEN);
 556         else if (sk_is_udp(sk))
 557                 return sk_hashed(sk);
 558
 559         return false;
 560 }
 561
 562 static int sock_hash_update_common(struct bpf_map *map, void *key,
 563                                    struct sock *sk, u64 flags);
 564
 565 static int sock_map_update_elem(struct bpf_map *map, void *key,
 566                                 void *value, u64 flags)
 567 {
 568         struct socket *sock;
 569         struct sock *sk;
 570         int ret;
 571         u64 ufd;
 572
 573         if (map->value_size == sizeof(u64))
 574                 ufd = *(u64 *)value;
 575         else
 576                 ufd = *(u32 *)value;
 577         if (ufd > S32_MAX)
 578                 return -EINVAL;
 579
 580         sock = sockfd_lookup(ufd, &ret);
 581         if (!sock)
 582                 return ret;
 583         sk = sock->sk;
 584         if (!sk) {
 585                 ret = -EINVAL;
 586                 goto out;
 587         }
 588         if (!sock_map_sk_is_suitable(sk)) {
 589                 ret = -EOPNOTSUPP;
 590                 goto out;
 591         }
 592
 593         sock_map_sk_acquire(sk);
 594         if (!sock_map_sk_state_allowed(sk))
 595                 ret = -EOPNOTSUPP;
 596         else if (map->map_type == BPF_MAP_TYPE_SOCKMAP)
 597                 ret = sock_map_update_common(map, *(u32 *)key, sk, flags);
 598         else
 599                 ret = sock_hash_update_common(map, key, sk, flags);
 600         sock_map_sk_release(sk);
 601 out:
 602         fput(sock->file);
 603         return ret;
 604 }
 605
 606 BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, sops,
 607            struct bpf_map *, map, void *, key, u64, flags)
 608 {
 609         WARN_ON_ONCE(!rcu_read_lock_held());
 610
 611         if (likely(sock_map_sk_is_suitable(sops->sk) &&
 612                    sock_map_op_okay(sops)))
 613                 return sock_map_update_common(map, *(u32 *)key, sops->sk,
 614                                               flags);
 615         return -EOPNOTSUPP;
 616 }
 617
 618 const struct bpf_func_proto bpf_sock_map_update_proto = {
 619         .func           = bpf_sock_map_update,
 620         .gpl_only       = false,
 621         .pkt_access     = true,
 622         .ret_type       = RET_INTEGER,
 623         .arg1_type      = ARG_PTR_TO_CTX,
 624         .arg2_type      = ARG_CONST_MAP_PTR,
 625         .arg3_type      = ARG_PTR_TO_MAP_KEY,
 626         .arg4_type      = ARG_ANYTHING,
 627 };
 628
 629 BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
 630            struct bpf_map *, map, u32, key, u64, flags)
 631 {
 632         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 633         struct sock *sk;
 634
 635         if (unlikely(flags & ~(BPF_F_INGRESS)))
 636                 return SK_DROP;
 637
 638         sk = __sock_map_lookup_elem(map, key);
 639         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
 640                 return SK_DROP;
 641
 642         tcb->bpf.flags = flags;
 643         tcb->bpf.sk_redir = sk;
 644         return SK_PASS;
 645 }
 646
 647 const struct bpf_func_proto bpf_sk_redirect_map_proto = {
 648         .func           = bpf_sk_redirect_map,
 649         .gpl_only       = false,
 650         .ret_type       = RET_INTEGER,
 651         .arg1_type      = ARG_PTR_TO_CTX,
 652         .arg2_type      = ARG_CONST_MAP_PTR,
 653         .arg3_type      = ARG_ANYTHING,
 654         .arg4_type      = ARG_ANYTHING,
 655 };
 656
 657 BPF_CALL_4(bpf_msg_redirect_map, struct sk_msg *, msg,
 658            struct bpf_map *, map, u32, key, u64, flags)
 659 {
 660         struct sock *sk;
 661
 662         if (unlikely(flags & ~(BPF_F_INGRESS)))
 663                 return SK_DROP;
 664
 665         sk = __sock_map_lookup_elem(map, key);
 666         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
 667                 return SK_DROP;
 668
 669         msg->flags = flags;
 670         msg->sk_redir = sk;
 671         return SK_PASS;
 672 }
 673
 674 const struct bpf_func_proto bpf_msg_redirect_map_proto = {
 675         .func           = bpf_msg_redirect_map,
 676         .gpl_only       = false,
 677         .ret_type       = RET_INTEGER,
 678         .arg1_type      = ARG_PTR_TO_CTX,
 679         .arg2_type      = ARG_CONST_MAP_PTR,
 680         .arg3_type      = ARG_ANYTHING,
 681         .arg4_type      = ARG_ANYTHING,
 682 };
 683
 684 static int sock_map_btf_id;
 685 const struct bpf_map_ops sock_map_ops = {
 686         .map_alloc              = sock_map_alloc,
 687         .map_free               = sock_map_free,
 688         .map_get_next_key       = sock_map_get_next_key,
 689         .map_lookup_elem_sys_only = sock_map_lookup_sys,
 690         .map_update_elem        = sock_map_update_elem,
 691         .map_delete_elem        = sock_map_delete_elem,
 692         .map_lookup_elem        = sock_map_lookup,
 693         .map_release_uref       = sock_map_release_progs,
 694         .map_check_btf          = map_check_no_btf,
 695         .map_btf_name           = "bpf_stab",
 696         .map_btf_id             = &sock_map_btf_id,
 697 };
 698
 699 struct bpf_shtab_elem {
 700         struct rcu_head rcu;
 701         u32 hash;
 702         struct sock *sk;
 703         struct hlist_node node;
 704         u8 key[];
 705 };
 706
 707 struct bpf_shtab_bucket {
 708         struct hlist_head head;
 709         raw_spinlock_t lock;
 710 };
 711
 712 struct bpf_shtab {
 713         struct bpf_map map;
 714         struct bpf_shtab_bucket *buckets;
 715         u32 buckets_num;
 716         u32 elem_size;
 717         struct sk_psock_progs progs;
 718         atomic_t count;
 719 };
 720
 721 static inline u32 sock_hash_bucket_hash(const void *key, u32 len)
 722 {
 723         return jhash(key, len, 0);
 724 }
 725
 726 static struct bpf_shtab_bucket *sock_hash_select_bucket(struct bpf_shtab *htab,
 727                                                         u32 hash)
 728 {
 729         return &htab->buckets[hash & (htab->buckets_num - 1)];
 730 }
 731
 732 static struct bpf_shtab_elem *
 733 sock_hash_lookup_elem_raw(struct hlist_head *head, u32 hash, void *key,
 734                           u32 key_size)
 735 {
 736         struct bpf_shtab_elem *elem;
 737
 738         hlist_for_each_entry_rcu(elem, head, node) {
 739                 if (elem->hash == hash &&
 740                     !memcmp(&elem->key, key, key_size))
 741                         return elem;
 742         }
 743
 744         return NULL;
 745 }
 746
 747 static struct sock *__sock_hash_lookup_elem(struct bpf_map *map, void *key)
 748 {
 749         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 750         u32 key_size = map->key_size, hash;
 751         struct bpf_shtab_bucket *bucket;
 752         struct bpf_shtab_elem *elem;
 753
 754         WARN_ON_ONCE(!rcu_read_lock_held());
 755
 756         hash = sock_hash_bucket_hash(key, key_size);
 757         bucket = sock_hash_select_bucket(htab, hash);
 758         elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 759
 760         return elem ? elem->sk : NULL;
 761 }
 762
 763 static void sock_hash_free_elem(struct bpf_shtab *htab,
 764                                 struct bpf_shtab_elem *elem)
 765 {
 766         atomic_dec(&htab->count);
 767         kfree_rcu(elem, rcu);
 768 }
 769
 770 static void sock_hash_delete_from_link(struct bpf_map *map, struct sock *sk,
 771                                        void *link_raw)
 772 {
 773         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 774         struct bpf_shtab_elem *elem_probe, *elem = link_raw;
 775         struct bpf_shtab_bucket *bucket;
 776
 777         WARN_ON_ONCE(!rcu_read_lock_held());
 778         bucket = sock_hash_select_bucket(htab, elem->hash);
 779
 780         /* elem may be deleted in parallel from the map, but access here
 781          * is okay since it's going away only after RCU grace period.
 782          * However, we need to check whether it's still present.
 783          */
 784         raw_spin_lock_bh(&bucket->lock);
 785         elem_probe = sock_hash_lookup_elem_raw(&bucket->head, elem->hash,
 786                                                elem->key, map->key_size);
 787         if (elem_probe && elem_probe == elem) {
 788                 hlist_del_rcu(&elem->node);
 789                 sock_map_unref(elem->sk, elem);
 790                 sock_hash_free_elem(htab, elem);
 791         }
 792         raw_spin_unlock_bh(&bucket->lock);
 793 }
 794
 795 static int sock_hash_delete_elem(struct bpf_map *map, void *key)
 796 {
 797         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 798         u32 hash, key_size = map->key_size;
 799         struct bpf_shtab_bucket *bucket;
 800         struct bpf_shtab_elem *elem;
 801         int ret = -ENOENT;
 802
 803         hash = sock_hash_bucket_hash(key, key_size);
 804         bucket = sock_hash_select_bucket(htab, hash);
 805
 806         raw_spin_lock_bh(&bucket->lock);
 807         elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 808         if (elem) {
 809                 hlist_del_rcu(&elem->node);
 810                 sock_map_unref(elem->sk, elem);
 811                 sock_hash_free_elem(htab, elem);
 812                 ret = 0;
 813         }
 814         raw_spin_unlock_bh(&bucket->lock);
 815         return ret;
 816 }
 817
 818 static struct bpf_shtab_elem *sock_hash_alloc_elem(struct bpf_shtab *htab,
 819                                                    void *key, u32 key_size,
 820                                                    u32 hash, struct sock *sk,
 821                                                    struct bpf_shtab_elem *old)
 822 {
 823         struct bpf_shtab_elem *new;
 824
 825         if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
 826                 if (!old) {
 827                         atomic_dec(&htab->count);
 828                         return ERR_PTR(-E2BIG);
 829                 }
 830         }
 831
 832         new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 833                            htab->map.numa_node);
 834         if (!new) {
 835                 atomic_dec(&htab->count);
 836                 return ERR_PTR(-ENOMEM);
 837         }
 838         memcpy(new->key, key, key_size);
 839         new->sk = sk;
 840         new->hash = hash;
 841         return new;
 842 }
 843
 844 static int sock_hash_update_common(struct bpf_map *map, void *key,
 845                                    struct sock *sk, u64 flags)
 846 {
 847         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 848         u32 key_size = map->key_size, hash;
 849         struct bpf_shtab_elem *elem, *elem_new;
 850         struct bpf_shtab_bucket *bucket;
 851         struct sk_psock_link *link;
 852         struct sk_psock *psock;
 853         int ret;
 854
 855         WARN_ON_ONCE(!rcu_read_lock_held());
 856         if (unlikely(flags > BPF_EXIST))
 857                 return -EINVAL;
 858
 859         link = sk_psock_init_link();
 860         if (!link)
 861                 return -ENOMEM;
 862
 863         /* Only sockets we can redirect into/from in BPF need to hold
 864          * refs to parser/verdict progs and have their sk_data_ready
 865          * and sk_write_space callbacks overridden.
 866          */
 867         if (sock_map_redirect_allowed(sk))
 868                 ret = sock_map_link(map, &htab->progs, sk);
 869         else
 870                 ret = sock_map_link_no_progs(map, sk);
 871         if (ret < 0)
 872                 goto out_free;
 873
 874         psock = sk_psock(sk);
 875         WARN_ON_ONCE(!psock);
 876
 877         hash = sock_hash_bucket_hash(key, key_size);
 878         bucket = sock_hash_select_bucket(htab, hash);
 879
 880         raw_spin_lock_bh(&bucket->lock);
 881         elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 882         if (elem && flags == BPF_NOEXIST) {
 883                 ret = -EEXIST;
 884                 goto out_unlock;
 885         } else if (!elem && flags == BPF_EXIST) {
 886                 ret = -ENOENT;
 887                 goto out_unlock;
 888         }
 889
 890         elem_new = sock_hash_alloc_elem(htab, key, key_size, hash, sk, elem);
 891         if (IS_ERR(elem_new)) {
 892                 ret = PTR_ERR(elem_new);
 893                 goto out_unlock;
 894         }
 895
 896         sock_map_add_link(psock, link, map, elem_new);
 897         /* Add new element to the head of the list, so that
 898          * concurrent search will find it before old elem.
 899          */
 900         hlist_add_head_rcu(&elem_new->node, &bucket->head);
 901         if (elem) {
 902                 hlist_del_rcu(&elem->node);
 903                 sock_map_unref(elem->sk, elem);
 904                 sock_hash_free_elem(htab, elem);
 905         }
 906         raw_spin_unlock_bh(&bucket->lock);
 907         return 0;
 908 out_unlock:
 909         raw_spin_unlock_bh(&bucket->lock);
 910         sk_psock_put(sk, psock);
 911 out_free:
 912         sk_psock_free_link(link);
 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_shtab *htab = container_of(map, struct bpf_shtab, map);
 920         struct bpf_shtab_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_shtab_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_shtab_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_shtab *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_shtab_elem) +
 981                           round_up(htab->map.key_size, 8);
 982         if (htab->buckets_num == 0 ||
 983             htab->buckets_num > U32_MAX / sizeof(struct bpf_shtab_bucket)) {
 984                 err = -EINVAL;
 985                 goto free_htab;
 986         }
 987
 988         cost = (u64) htab->buckets_num * sizeof(struct bpf_shtab_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         err = bpf_map_charge_init(&htab->map.memory, cost);
 995         if (err)
 996                 goto free_htab;
 997
 998         htab->buckets = bpf_map_area_alloc(htab->buckets_num *
 999                                            sizeof(struct bpf_shtab_bucket),
1000                                            htab->map.numa_node);
1001         if (!htab->buckets) {
1002                 bpf_map_charge_finish(&htab->map.memory);
1003                 err = -ENOMEM;
1004                 goto free_htab;
1005         }
1006
1007         for (i = 0; i < htab->buckets_num; i++) {
1008                 INIT_HLIST_HEAD(&htab->buckets[i].head);
1009                 raw_spin_lock_init(&htab->buckets[i].lock);
1010         }
1011
1012         return &htab->map;
1013 free_htab:
1014         kfree(htab);
1015         return ERR_PTR(err);
1016 }
1017
1018 static void sock_hash_free(struct bpf_map *map)
1019 {
1020         struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
1021         struct bpf_shtab_bucket *bucket;
1022         struct hlist_head unlink_list;
1023         struct bpf_shtab_elem *elem;
1024         struct hlist_node *node;
1025         int i;
1026
1027         /* After the sync no updates or deletes will be in-flight so it
1028          * is safe to walk map and remove entries without risking a race
1029          * in EEXIST update case.
1030          */
1031         synchronize_rcu();
1032         for (i = 0; i < htab->buckets_num; i++) {
1033                 bucket = sock_hash_select_bucket(htab, i);
1034
1035                 /* We are racing with sock_hash_delete_from_link to
1036                  * enter the spin-lock critical section. Every socket on
1037                  * the list is still linked to sockhash. Since link
1038                  * exists, psock exists and holds a ref to socket. That
1039                  * lets us to grab a socket ref too.
1040                  */
1041                 raw_spin_lock_bh(&bucket->lock);
1042                 hlist_for_each_entry(elem, &bucket->head, node)
1043                         sock_hold(elem->sk);
1044                 hlist_move_list(&bucket->head, &unlink_list);
1045                 raw_spin_unlock_bh(&bucket->lock);
1046
1047                 /* Process removed entries out of atomic context to
1048                  * block for socket lock before deleting the psock's
1049                  * link to sockhash.
1050                  */
1051                 hlist_for_each_entry_safe(elem, node, &unlink_list, node) {
1052                         hlist_del(&elem->node);
1053                         lock_sock(elem->sk);
1054                         rcu_read_lock();
1055                         sock_map_unref(elem->sk, elem);
1056                         rcu_read_unlock();
1057                         release_sock(elem->sk);
1058                         sock_put(elem->sk);
1059                         sock_hash_free_elem(htab, elem);
1060                 }
1061         }
1062
1063         /* wait for psock readers accessing its map link */
1064         synchronize_rcu();
1065
1066         bpf_map_area_free(htab->buckets);
1067         kfree(htab);
1068 }
1069
1070 static void *sock_hash_lookup_sys(struct bpf_map *map, void *key)
1071 {
1072         struct sock *sk;
1073
1074         if (map->value_size != sizeof(u64))
1075                 return ERR_PTR(-ENOSPC);
1076
1077         sk = __sock_hash_lookup_elem(map, key);
1078         if (!sk)
1079                 return ERR_PTR(-ENOENT);
1080
1081         sock_gen_cookie(sk);
1082         return &sk->sk_cookie;
1083 }
1084
1085 static void *sock_hash_lookup(struct bpf_map *map, void *key)
1086 {
1087         struct sock *sk;
1088
1089         sk = __sock_hash_lookup_elem(map, key);
1090         if (!sk || !sk_fullsock(sk))
1091                 return NULL;
1092         if (sk_is_refcounted(sk) && !refcount_inc_not_zero(&sk->sk_refcnt))
1093                 return NULL;
1094         return sk;
1095 }
1096
1097 static void sock_hash_release_progs(struct bpf_map *map)
1098 {
1099         psock_progs_drop(&container_of(map, struct bpf_shtab, map)->progs);
1100 }
1101
1102 BPF_CALL_4(bpf_sock_hash_update, struct bpf_sock_ops_kern *, sops,
1103            struct bpf_map *, map, void *, key, u64, flags)
1104 {
1105         WARN_ON_ONCE(!rcu_read_lock_held());
1106
1107         if (likely(sock_map_sk_is_suitable(sops->sk) &&
1108                    sock_map_op_okay(sops)))
1109                 return sock_hash_update_common(map, key, sops->sk, flags);
1110         return -EOPNOTSUPP;
1111 }
1112
1113 const struct bpf_func_proto bpf_sock_hash_update_proto = {
1114         .func           = bpf_sock_hash_update,
1115         .gpl_only       = false,
1116         .pkt_access     = true,
1117         .ret_type       = RET_INTEGER,
1118         .arg1_type      = ARG_PTR_TO_CTX,
1119         .arg2_type      = ARG_CONST_MAP_PTR,
1120         .arg3_type      = ARG_PTR_TO_MAP_KEY,
1121         .arg4_type      = ARG_ANYTHING,
1122 };
1123
1124 BPF_CALL_4(bpf_sk_redirect_hash, struct sk_buff *, skb,
1125            struct bpf_map *, map, void *, key, u64, flags)
1126 {
1127         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1128         struct sock *sk;
1129
1130         if (unlikely(flags & ~(BPF_F_INGRESS)))
1131                 return SK_DROP;
1132
1133         sk = __sock_hash_lookup_elem(map, key);
1134         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1135                 return SK_DROP;
1136
1137         tcb->bpf.flags = flags;
1138         tcb->bpf.sk_redir = sk;
1139         return SK_PASS;
1140 }
1141
1142 const struct bpf_func_proto bpf_sk_redirect_hash_proto = {
1143         .func           = bpf_sk_redirect_hash,
1144         .gpl_only       = false,
1145         .ret_type       = RET_INTEGER,
1146         .arg1_type      = ARG_PTR_TO_CTX,
1147         .arg2_type      = ARG_CONST_MAP_PTR,
1148         .arg3_type      = ARG_PTR_TO_MAP_KEY,
1149         .arg4_type      = ARG_ANYTHING,
1150 };
1151
1152 BPF_CALL_4(bpf_msg_redirect_hash, struct sk_msg *, msg,
1153            struct bpf_map *, map, void *, key, u64, flags)
1154 {
1155         struct sock *sk;
1156
1157         if (unlikely(flags & ~(BPF_F_INGRESS)))
1158                 return SK_DROP;
1159
1160         sk = __sock_hash_lookup_elem(map, key);
1161         if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1162                 return SK_DROP;
1163
1164         msg->flags = flags;
1165         msg->sk_redir = sk;
1166         return SK_PASS;
1167 }
1168
1169 const struct bpf_func_proto bpf_msg_redirect_hash_proto = {
1170         .func           = bpf_msg_redirect_hash,
1171         .gpl_only       = false,
1172         .ret_type       = RET_INTEGER,
1173         .arg1_type      = ARG_PTR_TO_CTX,
1174         .arg2_type      = ARG_CONST_MAP_PTR,
1175         .arg3_type      = ARG_PTR_TO_MAP_KEY,
1176         .arg4_type      = ARG_ANYTHING,
1177 };
1178
1179 static int sock_hash_map_btf_id;
1180 const struct bpf_map_ops sock_hash_ops = {
1181         .map_alloc              = sock_hash_alloc,
1182         .map_free               = sock_hash_free,
1183         .map_get_next_key       = sock_hash_get_next_key,
1184         .map_update_elem        = sock_map_update_elem,
1185         .map_delete_elem        = sock_hash_delete_elem,
1186         .map_lookup_elem        = sock_hash_lookup,
1187         .map_lookup_elem_sys_only = sock_hash_lookup_sys,
1188         .map_release_uref       = sock_hash_release_progs,
1189         .map_check_btf          = map_check_no_btf,
1190         .map_btf_name           = "bpf_shtab",
1191         .map_btf_id             = &sock_hash_map_btf_id,
1192 };
1193
1194 static struct sk_psock_progs *sock_map_progs(struct bpf_map *map)
1195 {
1196         switch (map->map_type) {
1197         case BPF_MAP_TYPE_SOCKMAP:
1198                 return &container_of(map, struct bpf_stab, map)->progs;
1199         case BPF_MAP_TYPE_SOCKHASH:
1200                 return &container_of(map, struct bpf_shtab, map)->progs;
1201         default:
1202                 break;
1203         }
1204
1205         return NULL;
1206 }
1207
1208 int sock_map_prog_update(struct bpf_map *map, struct bpf_prog *prog,
1209                          struct bpf_prog *old, u32 which)
1210 {
1211         struct sk_psock_progs *progs = sock_map_progs(map);
1212         struct bpf_prog **pprog;
1213
1214         if (!progs)
1215                 return -EOPNOTSUPP;
1216
1217         switch (which) {
1218         case BPF_SK_MSG_VERDICT:
1219                 pprog = &progs->msg_parser;
1220                 break;
1221         case BPF_SK_SKB_STREAM_PARSER:
1222                 pprog = &progs->skb_parser;
1223                 break;
1224         case BPF_SK_SKB_STREAM_VERDICT:
1225                 pprog = &progs->skb_verdict;
1226                 break;
1227         default:
1228                 return -EOPNOTSUPP;
1229         }
1230
1231         if (old)
1232                 return psock_replace_prog(pprog, prog, old);
1233
1234         psock_set_prog(pprog, prog);
1235         return 0;
1236 }
1237
1238 static void sock_map_unlink(struct sock *sk, struct sk_psock_link *link)
1239 {
1240         switch (link->map->map_type) {
1241         case BPF_MAP_TYPE_SOCKMAP:
1242                 return sock_map_delete_from_link(link->map, sk,
1243                                                  link->link_raw);
1244         case BPF_MAP_TYPE_SOCKHASH:
1245                 return sock_hash_delete_from_link(link->map, sk,
1246                                                   link->link_raw);
1247         default:
1248                 break;
1249         }
1250 }
1251
1252 static void sock_map_remove_links(struct sock *sk, struct sk_psock *psock)
1253 {
1254         struct sk_psock_link *link;
1255
1256         while ((link = sk_psock_link_pop(psock))) {
1257                 sock_map_unlink(sk, link);
1258                 sk_psock_free_link(link);
1259         }
1260 }
1261
1262 void sock_map_unhash(struct sock *sk)
1263 {
1264         void (*saved_unhash)(struct sock *sk);
1265         struct sk_psock *psock;
1266
1267         rcu_read_lock();
1268         psock = sk_psock(sk);
1269         if (unlikely(!psock)) {
1270                 rcu_read_unlock();
1271                 if (sk->sk_prot->unhash)
1272                         sk->sk_prot->unhash(sk);
1273                 return;
1274         }
1275
1276         saved_unhash = psock->saved_unhash;
1277         sock_map_remove_links(sk, psock);
1278         rcu_read_unlock();
1279         saved_unhash(sk);
1280 }
1281
1282 void sock_map_close(struct sock *sk, long timeout)
1283 {
1284         void (*saved_close)(struct sock *sk, long timeout);
1285         struct sk_psock *psock;
1286
1287         lock_sock(sk);
1288         rcu_read_lock();
1289         psock = sk_psock(sk);
1290         if (unlikely(!psock)) {
1291                 rcu_read_unlock();
1292                 release_sock(sk);
1293                 return sk->sk_prot->close(sk, timeout);
1294         }
1295
1296         saved_close = psock->saved_close;
1297         sock_map_remove_links(sk, psock);
1298         rcu_read_unlock();
1299         release_sock(sk);
1300         saved_close(sk, timeout);
1301 }