linux/net/netfilter/nf_tables_api.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (c) 2007-2009 Patrick McHardy <kaber@trash.net>
   4 *
   5 * Development of this code funded by Astaro AG (http://www.astaro.com/)
   6 */
   7
   8#include <linux/module.h>
   9#include <linux/init.h>
  10#include <linux/list.h>
  11#include <linux/skbuff.h>
  12#include <linux/netlink.h>
  13#include <linux/vmalloc.h>
  14#include <linux/rhashtable.h>
  15#include <linux/audit.h>
  16#include <linux/netfilter.h>
  17#include <linux/netfilter/nfnetlink.h>
  18#include <linux/netfilter/nf_tables.h>
  19#include <net/netfilter/nf_flow_table.h>
  20#include <net/netfilter/nf_tables_core.h>
  21#include <net/netfilter/nf_tables.h>
  22#include <net/netfilter/nf_tables_offload.h>
  23#include <net/net_namespace.h>
  24#include <net/sock.h>
  25
  26#define NFT_MODULE_AUTOLOAD_LIMIT (MODULE_NAME_LEN - sizeof("nft-expr-255-"))
  27
  28unsigned int nf_tables_net_id __read_mostly;
  29
  30static LIST_HEAD(nf_tables_expressions);
  31static LIST_HEAD(nf_tables_objects);
  32static LIST_HEAD(nf_tables_flowtables);
  33static LIST_HEAD(nf_tables_destroy_list);
  34static DEFINE_SPINLOCK(nf_tables_destroy_list_lock);
  35static u64 table_handle;
  36
  37enum {
  38        NFT_VALIDATE_SKIP       = 0,
  39        NFT_VALIDATE_NEED,
  40        NFT_VALIDATE_DO,
  41};
  42
  43static struct rhltable nft_objname_ht;
  44
  45static u32 nft_chain_hash(const void *data, u32 len, u32 seed);
  46static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed);
  47static int nft_chain_hash_cmp(struct rhashtable_compare_arg *, const void *);
  48
  49static u32 nft_objname_hash(const void *data, u32 len, u32 seed);
  50static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed);
  51static int nft_objname_hash_cmp(struct rhashtable_compare_arg *, const void *);
  52
  53static const struct rhashtable_params nft_chain_ht_params = {
  54        .head_offset            = offsetof(struct nft_chain, rhlhead),
  55        .key_offset             = offsetof(struct nft_chain, name),
  56        .hashfn                 = nft_chain_hash,
  57        .obj_hashfn             = nft_chain_hash_obj,
  58        .obj_cmpfn              = nft_chain_hash_cmp,
  59        .automatic_shrinking    = true,
  60};
  61
  62static const struct rhashtable_params nft_objname_ht_params = {
  63        .head_offset            = offsetof(struct nft_object, rhlhead),
  64        .key_offset             = offsetof(struct nft_object, key),
  65        .hashfn                 = nft_objname_hash,
  66        .obj_hashfn             = nft_objname_hash_obj,
  67        .obj_cmpfn              = nft_objname_hash_cmp,
  68        .automatic_shrinking    = true,
  69};
  70
  71struct nft_audit_data {
  72        struct nft_table *table;
  73        int entries;
  74        int op;
  75        struct list_head list;
  76};
  77
  78static const u8 nft2audit_op[NFT_MSG_MAX] = { // enum nf_tables_msg_types
  79        [NFT_MSG_NEWTABLE]      = AUDIT_NFT_OP_TABLE_REGISTER,
  80        [NFT_MSG_GETTABLE]      = AUDIT_NFT_OP_INVALID,
  81        [NFT_MSG_DELTABLE]      = AUDIT_NFT_OP_TABLE_UNREGISTER,
  82        [NFT_MSG_NEWCHAIN]      = AUDIT_NFT_OP_CHAIN_REGISTER,
  83        [NFT_MSG_GETCHAIN]      = AUDIT_NFT_OP_INVALID,
  84        [NFT_MSG_DELCHAIN]      = AUDIT_NFT_OP_CHAIN_UNREGISTER,
  85        [NFT_MSG_NEWRULE]       = AUDIT_NFT_OP_RULE_REGISTER,
  86        [NFT_MSG_GETRULE]       = AUDIT_NFT_OP_INVALID,
  87        [NFT_MSG_DELRULE]       = AUDIT_NFT_OP_RULE_UNREGISTER,
  88        [NFT_MSG_NEWSET]        = AUDIT_NFT_OP_SET_REGISTER,
  89        [NFT_MSG_GETSET]        = AUDIT_NFT_OP_INVALID,
  90        [NFT_MSG_DELSET]        = AUDIT_NFT_OP_SET_UNREGISTER,
  91        [NFT_MSG_NEWSETELEM]    = AUDIT_NFT_OP_SETELEM_REGISTER,
  92        [NFT_MSG_GETSETELEM]    = AUDIT_NFT_OP_INVALID,
  93        [NFT_MSG_DELSETELEM]    = AUDIT_NFT_OP_SETELEM_UNREGISTER,
  94        [NFT_MSG_NEWGEN]        = AUDIT_NFT_OP_GEN_REGISTER,
  95        [NFT_MSG_GETGEN]        = AUDIT_NFT_OP_INVALID,
  96        [NFT_MSG_TRACE]         = AUDIT_NFT_OP_INVALID,
  97        [NFT_MSG_NEWOBJ]        = AUDIT_NFT_OP_OBJ_REGISTER,
  98        [NFT_MSG_GETOBJ]        = AUDIT_NFT_OP_INVALID,
  99        [NFT_MSG_DELOBJ]        = AUDIT_NFT_OP_OBJ_UNREGISTER,
 100        [NFT_MSG_GETOBJ_RESET]  = AUDIT_NFT_OP_OBJ_RESET,
 101        [NFT_MSG_NEWFLOWTABLE]  = AUDIT_NFT_OP_FLOWTABLE_REGISTER,
 102        [NFT_MSG_GETFLOWTABLE]  = AUDIT_NFT_OP_INVALID,
 103        [NFT_MSG_DELFLOWTABLE]  = AUDIT_NFT_OP_FLOWTABLE_UNREGISTER,
 104};
 105
 106static void nft_validate_state_update(struct net *net, u8 new_validate_state)
 107{
 108        struct nftables_pernet *nft_net = nft_pernet(net);
 109
 110        switch (nft_net->validate_state) {
 111        case NFT_VALIDATE_SKIP:
 112                WARN_ON_ONCE(new_validate_state == NFT_VALIDATE_DO);
 113                break;
 114        case NFT_VALIDATE_NEED:
 115                break;
 116        case NFT_VALIDATE_DO:
 117                if (new_validate_state == NFT_VALIDATE_NEED)
 118                        return;
 119        }
 120
 121        nft_net->validate_state = new_validate_state;
 122}
 123static void nf_tables_trans_destroy_work(struct work_struct *w);
 124static DECLARE_WORK(trans_destroy_work, nf_tables_trans_destroy_work);
 125
 126static void nft_ctx_init(struct nft_ctx *ctx,
 127                         struct net *net,
 128                         const struct sk_buff *skb,
 129                         const struct nlmsghdr *nlh,
 130                         u8 family,
 131                         struct nft_table *table,
 132                         struct nft_chain *chain,
 133                         const struct nlattr * const *nla)
 134{
 135        ctx->net        = net;
 136        ctx->family     = family;
 137        ctx->level      = 0;
 138        ctx->table      = table;
 139        ctx->chain      = chain;
 140        ctx->nla        = nla;
 141        ctx->portid     = NETLINK_CB(skb).portid;
 142        ctx->report     = nlmsg_report(nlh);
 143        ctx->flags      = nlh->nlmsg_flags;
 144        ctx->seq        = nlh->nlmsg_seq;
 145}
 146
 147static struct nft_trans *nft_trans_alloc_gfp(const struct nft_ctx *ctx,
 148                                             int msg_type, u32 size, gfp_t gfp)
 149{
 150        struct nft_trans *trans;
 151
 152        trans = kzalloc(sizeof(struct nft_trans) + size, gfp);
 153        if (trans == NULL)
 154                return NULL;
 155
 156        trans->msg_type = msg_type;
 157        trans->ctx      = *ctx;
 158
 159        return trans;
 160}
 161
 162static struct nft_trans *nft_trans_alloc(const struct nft_ctx *ctx,
 163                                         int msg_type, u32 size)
 164{
 165        return nft_trans_alloc_gfp(ctx, msg_type, size, GFP_KERNEL);
 166}
 167
 168static void nft_trans_destroy(struct nft_trans *trans)
 169{
 170        list_del(&trans->list);
 171        kfree(trans);
 172}
 173
 174static void nft_set_trans_bind(const struct nft_ctx *ctx, struct nft_set *set)
 175{
 176        struct nftables_pernet *nft_net;
 177        struct net *net = ctx->net;
 178        struct nft_trans *trans;
 179
 180        if (!nft_set_is_anonymous(set))
 181                return;
 182
 183        nft_net = nft_pernet(net);
 184        list_for_each_entry_reverse(trans, &nft_net->commit_list, list) {
 185                switch (trans->msg_type) {
 186                case NFT_MSG_NEWSET:
 187                        if (nft_trans_set(trans) == set)
 188                                nft_trans_set_bound(trans) = true;
 189                        break;
 190                case NFT_MSG_NEWSETELEM:
 191                        if (nft_trans_elem_set(trans) == set)
 192                                nft_trans_elem_set_bound(trans) = true;
 193                        break;
 194                }
 195        }
 196}
 197
 198static int nft_netdev_register_hooks(struct net *net,
 199                                     struct list_head *hook_list)
 200{
 201        struct nft_hook *hook;
 202        int err, j;
 203
 204        j = 0;
 205        list_for_each_entry(hook, hook_list, list) {
 206                err = nf_register_net_hook(net, &hook->ops);
 207                if (err < 0)
 208                        goto err_register;
 209
 210                j++;
 211        }
 212        return 0;
 213
 214err_register:
 215        list_for_each_entry(hook, hook_list, list) {
 216                if (j-- <= 0)
 217                        break;
 218
 219                nf_unregister_net_hook(net, &hook->ops);
 220        }
 221        return err;
 222}
 223
 224static void nft_netdev_unregister_hooks(struct net *net,
 225                                        struct list_head *hook_list)
 226{
 227        struct nft_hook *hook;
 228
 229        list_for_each_entry(hook, hook_list, list)
 230                nf_unregister_net_hook(net, &hook->ops);
 231}
 232
 233static int nf_tables_register_hook(struct net *net,
 234                                   const struct nft_table *table,
 235                                   struct nft_chain *chain)
 236{
 237        struct nft_base_chain *basechain;
 238        const struct nf_hook_ops *ops;
 239
 240        if (table->flags & NFT_TABLE_F_DORMANT ||
 241            !nft_is_base_chain(chain))
 242                return 0;
 243
 244        basechain = nft_base_chain(chain);
 245        ops = &basechain->ops;
 246
 247        if (basechain->type->ops_register)
 248                return basechain->type->ops_register(net, ops);
 249
 250        if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
 251                return nft_netdev_register_hooks(net, &basechain->hook_list);
 252
 253        return nf_register_net_hook(net, &basechain->ops);
 254}
 255
 256static void nf_tables_unregister_hook(struct net *net,
 257                                      const struct nft_table *table,
 258                                      struct nft_chain *chain)
 259{
 260        struct nft_base_chain *basechain;
 261        const struct nf_hook_ops *ops;
 262
 263        if (table->flags & NFT_TABLE_F_DORMANT ||
 264            !nft_is_base_chain(chain))
 265                return;
 266        basechain = nft_base_chain(chain);
 267        ops = &basechain->ops;
 268
 269        if (basechain->type->ops_unregister)
 270                return basechain->type->ops_unregister(net, ops);
 271
 272        if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
 273                nft_netdev_unregister_hooks(net, &basechain->hook_list);
 274        else
 275                nf_unregister_net_hook(net, &basechain->ops);
 276}
 277
 278static void nft_trans_commit_list_add_tail(struct net *net, struct nft_trans *trans)
 279{
 280        struct nftables_pernet *nft_net = nft_pernet(net);
 281
 282        list_add_tail(&trans->list, &nft_net->commit_list);
 283}
 284
 285static int nft_trans_table_add(struct nft_ctx *ctx, int msg_type)
 286{
 287        struct nft_trans *trans;
 288
 289        trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_table));
 290        if (trans == NULL)
 291                return -ENOMEM;
 292
 293        if (msg_type == NFT_MSG_NEWTABLE)
 294                nft_activate_next(ctx->net, ctx->table);
 295
 296        nft_trans_commit_list_add_tail(ctx->net, trans);
 297        return 0;
 298}
 299
 300static int nft_deltable(struct nft_ctx *ctx)
 301{
 302        int err;
 303
 304        err = nft_trans_table_add(ctx, NFT_MSG_DELTABLE);
 305        if (err < 0)
 306                return err;
 307
 308        nft_deactivate_next(ctx->net, ctx->table);
 309        return err;
 310}
 311
 312static struct nft_trans *nft_trans_chain_add(struct nft_ctx *ctx, int msg_type)
 313{
 314        struct nft_trans *trans;
 315
 316        trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_chain));
 317        if (trans == NULL)
 318                return ERR_PTR(-ENOMEM);
 319
 320        if (msg_type == NFT_MSG_NEWCHAIN) {
 321                nft_activate_next(ctx->net, ctx->chain);
 322
 323                if (ctx->nla[NFTA_CHAIN_ID]) {
 324                        nft_trans_chain_id(trans) =
 325                                ntohl(nla_get_be32(ctx->nla[NFTA_CHAIN_ID]));
 326                }
 327        }
 328
 329        nft_trans_commit_list_add_tail(ctx->net, trans);
 330        return trans;
 331}
 332
 333static int nft_delchain(struct nft_ctx *ctx)
 334{
 335        struct nft_trans *trans;
 336
 337        trans = nft_trans_chain_add(ctx, NFT_MSG_DELCHAIN);
 338        if (IS_ERR(trans))
 339                return PTR_ERR(trans);
 340
 341        ctx->table->use--;
 342        nft_deactivate_next(ctx->net, ctx->chain);
 343
 344        return 0;
 345}
 346
 347static void nft_rule_expr_activate(const struct nft_ctx *ctx,
 348                                   struct nft_rule *rule)
 349{
 350        struct nft_expr *expr;
 351
 352        expr = nft_expr_first(rule);
 353        while (nft_expr_more(rule, expr)) {
 354                if (expr->ops->activate)
 355                        expr->ops->activate(ctx, expr);
 356
 357                expr = nft_expr_next(expr);
 358        }
 359}
 360
 361static void nft_rule_expr_deactivate(const struct nft_ctx *ctx,
 362                                     struct nft_rule *rule,
 363                                     enum nft_trans_phase phase)
 364{
 365        struct nft_expr *expr;
 366
 367        expr = nft_expr_first(rule);
 368        while (nft_expr_more(rule, expr)) {
 369                if (expr->ops->deactivate)
 370                        expr->ops->deactivate(ctx, expr, phase);
 371
 372                expr = nft_expr_next(expr);
 373        }
 374}
 375
 376static int
 377nf_tables_delrule_deactivate(struct nft_ctx *ctx, struct nft_rule *rule)
 378{
 379        /* You cannot delete the same rule twice */
 380        if (nft_is_active_next(ctx->net, rule)) {
 381                nft_deactivate_next(ctx->net, rule);
 382                ctx->chain->use--;
 383                return 0;
 384        }
 385        return -ENOENT;
 386}
 387
 388static struct nft_trans *nft_trans_rule_add(struct nft_ctx *ctx, int msg_type,
 389                                            struct nft_rule *rule)
 390{
 391        struct nft_trans *trans;
 392
 393        trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_rule));
 394        if (trans == NULL)
 395                return NULL;
 396
 397        if (msg_type == NFT_MSG_NEWRULE && ctx->nla[NFTA_RULE_ID] != NULL) {
 398                nft_trans_rule_id(trans) =
 399                        ntohl(nla_get_be32(ctx->nla[NFTA_RULE_ID]));
 400        }
 401        nft_trans_rule(trans) = rule;
 402        nft_trans_commit_list_add_tail(ctx->net, trans);
 403
 404        return trans;
 405}
 406
 407static int nft_delrule(struct nft_ctx *ctx, struct nft_rule *rule)
 408{
 409        struct nft_flow_rule *flow;
 410        struct nft_trans *trans;
 411        int err;
 412
 413        trans = nft_trans_rule_add(ctx, NFT_MSG_DELRULE, rule);
 414        if (trans == NULL)
 415                return -ENOMEM;
 416
 417        if (ctx->chain->flags & NFT_CHAIN_HW_OFFLOAD) {
 418                flow = nft_flow_rule_create(ctx->net, rule);
 419                if (IS_ERR(flow)) {
 420                        nft_trans_destroy(trans);
 421                        return PTR_ERR(flow);
 422                }
 423
 424                nft_trans_flow_rule(trans) = flow;
 425        }
 426
 427        err = nf_tables_delrule_deactivate(ctx, rule);
 428        if (err < 0) {
 429                nft_trans_destroy(trans);
 430                return err;
 431        }
 432        nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_PREPARE);
 433
 434        return 0;
 435}
 436
 437static int nft_delrule_by_chain(struct nft_ctx *ctx)
 438{
 439        struct nft_rule *rule;
 440        int err;
 441
 442        list_for_each_entry(rule, &ctx->chain->rules, list) {
 443                if (!nft_is_active_next(ctx->net, rule))
 444                        continue;
 445
 446                err = nft_delrule(ctx, rule);
 447                if (err < 0)
 448                        return err;
 449        }
 450        return 0;
 451}
 452
 453static int nft_trans_set_add(const struct nft_ctx *ctx, int msg_type,
 454                             struct nft_set *set)
 455{
 456        struct nft_trans *trans;
 457
 458        trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_set));
 459        if (trans == NULL)
 460                return -ENOMEM;
 461
 462        if (msg_type == NFT_MSG_NEWSET && ctx->nla[NFTA_SET_ID] != NULL) {
 463                nft_trans_set_id(trans) =
 464                        ntohl(nla_get_be32(ctx->nla[NFTA_SET_ID]));
 465                nft_activate_next(ctx->net, set);
 466        }
 467        nft_trans_set(trans) = set;
 468        nft_trans_commit_list_add_tail(ctx->net, trans);
 469
 470        return 0;
 471}
 472
 473static int nft_delset(const struct nft_ctx *ctx, struct nft_set *set)
 474{
 475        int err;
 476
 477        err = nft_trans_set_add(ctx, NFT_MSG_DELSET, set);
 478        if (err < 0)
 479                return err;
 480
 481        nft_deactivate_next(ctx->net, set);
 482        ctx->table->use--;
 483
 484        return err;
 485}
 486
 487static int nft_trans_obj_add(struct nft_ctx *ctx, int msg_type,
 488                             struct nft_object *obj)
 489{
 490        struct nft_trans *trans;
 491
 492        trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_obj));
 493        if (trans == NULL)
 494                return -ENOMEM;
 495
 496        if (msg_type == NFT_MSG_NEWOBJ)
 497                nft_activate_next(ctx->net, obj);
 498
 499        nft_trans_obj(trans) = obj;
 500        nft_trans_commit_list_add_tail(ctx->net, trans);
 501
 502        return 0;
 503}
 504
 505static int nft_delobj(struct nft_ctx *ctx, struct nft_object *obj)
 506{
 507        int err;
 508
 509        err = nft_trans_obj_add(ctx, NFT_MSG_DELOBJ, obj);
 510        if (err < 0)
 511                return err;
 512
 513        nft_deactivate_next(ctx->net, obj);
 514        ctx->table->use--;
 515
 516        return err;
 517}
 518
 519static int nft_trans_flowtable_add(struct nft_ctx *ctx, int msg_type,
 520                                   struct nft_flowtable *flowtable)
 521{
 522        struct nft_trans *trans;
 523
 524        trans = nft_trans_alloc(ctx, msg_type,
 525                                sizeof(struct nft_trans_flowtable));
 526        if (trans == NULL)
 527                return -ENOMEM;
 528
 529        if (msg_type == NFT_MSG_NEWFLOWTABLE)
 530                nft_activate_next(ctx->net, flowtable);
 531
 532        nft_trans_flowtable(trans) = flowtable;
 533        nft_trans_commit_list_add_tail(ctx->net, trans);
 534
 535        return 0;
 536}
 537
 538static int nft_delflowtable(struct nft_ctx *ctx,
 539                            struct nft_flowtable *flowtable)
 540{
 541        int err;
 542
 543        err = nft_trans_flowtable_add(ctx, NFT_MSG_DELFLOWTABLE, flowtable);
 544        if (err < 0)
 545                return err;
 546
 547        nft_deactivate_next(ctx->net, flowtable);
 548        ctx->table->use--;
 549
 550        return err;
 551}
 552
 553/*
 554 * Tables
 555 */
 556
 557static struct nft_table *nft_table_lookup(const struct net *net,
 558                                          const struct nlattr *nla,
 559                                          u8 family, u8 genmask, u32 nlpid)
 560{
 561        struct nftables_pernet *nft_net;
 562        struct nft_table *table;
 563
 564        if (nla == NULL)
 565                return ERR_PTR(-EINVAL);
 566
 567        nft_net = nft_pernet(net);
 568        list_for_each_entry_rcu(table, &nft_net->tables, list,
 569                                lockdep_is_held(&nft_net->commit_mutex)) {
 570                if (!nla_strcmp(nla, table->name) &&
 571                    table->family == family &&
 572                    nft_active_genmask(table, genmask)) {
 573                        if (nft_table_has_owner(table) &&
 574                            nlpid && table->nlpid != nlpid)
 575                                return ERR_PTR(-EPERM);
 576
 577                        return table;
 578                }
 579        }
 580
 581        return ERR_PTR(-ENOENT);
 582}
 583
 584static struct nft_table *nft_table_lookup_byhandle(const struct net *net,
 585                                                   const struct nlattr *nla,
 586                                                   u8 genmask, u32 nlpid)
 587{
 588        struct nftables_pernet *nft_net;
 589        struct nft_table *table;
 590
 591        nft_net = nft_pernet(net);
 592        list_for_each_entry(table, &nft_net->tables, list) {
 593                if (be64_to_cpu(nla_get_be64(nla)) == table->handle &&
 594                    nft_active_genmask(table, genmask)) {
 595                        if (nft_table_has_owner(table) &&
 596                            nlpid && table->nlpid != nlpid)
 597                                return ERR_PTR(-EPERM);
 598
 599                        return table;
 600                }
 601        }
 602
 603        return ERR_PTR(-ENOENT);
 604}
 605
 606static inline u64 nf_tables_alloc_handle(struct nft_table *table)
 607{
 608        return ++table->hgenerator;
 609}
 610
 611static const struct nft_chain_type *chain_type[NFPROTO_NUMPROTO][NFT_CHAIN_T_MAX];
 612
 613static const struct nft_chain_type *
 614__nft_chain_type_get(u8 family, enum nft_chain_types type)
 615{
 616        if (family >= NFPROTO_NUMPROTO ||
 617            type >= NFT_CHAIN_T_MAX)
 618                return NULL;
 619
 620        return chain_type[family][type];
 621}
 622
 623static const struct nft_chain_type *
 624__nf_tables_chain_type_lookup(const struct nlattr *nla, u8 family)
 625{
 626        const struct nft_chain_type *type;
 627        int i;
 628
 629        for (i = 0; i < NFT_CHAIN_T_MAX; i++) {
 630                type = __nft_chain_type_get(family, i);
 631                if (!type)
 632                        continue;
 633                if (!nla_strcmp(nla, type->name))
 634                        return type;
 635        }
 636        return NULL;
 637}
 638
 639struct nft_module_request {
 640        struct list_head        list;
 641        char                    module[MODULE_NAME_LEN];
 642        bool                    done;
 643};
 644
 645#ifdef CONFIG_MODULES
 646__printf(2, 3) int nft_request_module(struct net *net, const char *fmt,
 647                                      ...)
 648{
 649        char module_name[MODULE_NAME_LEN];
 650        struct nftables_pernet *nft_net;
 651        struct nft_module_request *req;
 652        va_list args;
 653        int ret;
 654
 655        va_start(args, fmt);
 656        ret = vsnprintf(module_name, MODULE_NAME_LEN, fmt, args);
 657        va_end(args);
 658        if (ret >= MODULE_NAME_LEN)
 659                return 0;
 660
 661        nft_net = nft_pernet(net);
 662        list_for_each_entry(req, &nft_net->module_list, list) {
 663                if (!strcmp(req->module, module_name)) {
 664                        if (req->done)
 665                                return 0;
 666
 667                        /* A request to load this module already exists. */
 668                        return -EAGAIN;
 669                }
 670        }
 671
 672        req = kmalloc(sizeof(*req), GFP_KERNEL);
 673        if (!req)
 674                return -ENOMEM;
 675
 676        req->done = false;
 677        strlcpy(req->module, module_name, MODULE_NAME_LEN);
 678        list_add_tail(&req->list, &nft_net->module_list);
 679
 680        return -EAGAIN;
 681}
 682EXPORT_SYMBOL_GPL(nft_request_module);
 683#endif
 684
 685static void lockdep_nfnl_nft_mutex_not_held(void)
 686{
 687#ifdef CONFIG_PROVE_LOCKING
 688        if (debug_locks)
 689                WARN_ON_ONCE(lockdep_nfnl_is_held(NFNL_SUBSYS_NFTABLES));
 690#endif
 691}
 692
 693static const struct nft_chain_type *
 694nf_tables_chain_type_lookup(struct net *net, const struct nlattr *nla,
 695                            u8 family, bool autoload)
 696{
 697        const struct nft_chain_type *type;
 698
 699        type = __nf_tables_chain_type_lookup(nla, family);
 700        if (type != NULL)
 701                return type;
 702
 703        lockdep_nfnl_nft_mutex_not_held();
 704#ifdef CONFIG_MODULES
 705        if (autoload) {
 706                if (nft_request_module(net, "nft-chain-%u-%.*s", family,
 707                                       nla_len(nla),
 708                                       (const char *)nla_data(nla)) == -EAGAIN)
 709                        return ERR_PTR(-EAGAIN);
 710        }
 711#endif
 712        return ERR_PTR(-ENOENT);
 713}
 714
 715static __be16 nft_base_seq(const struct net *net)
 716{
 717        struct nftables_pernet *nft_net = nft_pernet(net);
 718
 719        return htons(nft_net->base_seq & 0xffff);
 720}
 721
 722static const struct nla_policy nft_table_policy[NFTA_TABLE_MAX + 1] = {
 723        [NFTA_TABLE_NAME]       = { .type = NLA_STRING,
 724                                    .len = NFT_TABLE_MAXNAMELEN - 1 },
 725        [NFTA_TABLE_FLAGS]      = { .type = NLA_U32 },
 726        [NFTA_TABLE_HANDLE]     = { .type = NLA_U64 },
 727        [NFTA_TABLE_USERDATA]   = { .type = NLA_BINARY,
 728                                    .len = NFT_USERDATA_MAXLEN }
 729};
 730
 731static int nf_tables_fill_table_info(struct sk_buff *skb, struct net *net,
 732                                     u32 portid, u32 seq, int event, u32 flags,
 733                                     int family, const struct nft_table *table)
 734{
 735        struct nlmsghdr *nlh;
 736
 737        event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
 738        nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
 739                           NFNETLINK_V0, nft_base_seq(net));
 740        if (!nlh)
 741                goto nla_put_failure;
 742
 743        if (nla_put_string(skb, NFTA_TABLE_NAME, table->name) ||
 744            nla_put_be32(skb, NFTA_TABLE_FLAGS,
 745                         htonl(table->flags & NFT_TABLE_F_MASK)) ||
 746            nla_put_be32(skb, NFTA_TABLE_USE, htonl(table->use)) ||
 747            nla_put_be64(skb, NFTA_TABLE_HANDLE, cpu_to_be64(table->handle),
 748                         NFTA_TABLE_PAD))
 749                goto nla_put_failure;
 750        if (nft_table_has_owner(table) &&
 751            nla_put_be32(skb, NFTA_TABLE_OWNER, htonl(table->nlpid)))
 752                goto nla_put_failure;
 753
 754        if (table->udata) {
 755                if (nla_put(skb, NFTA_TABLE_USERDATA, table->udlen, table->udata))
 756                        goto nla_put_failure;
 757        }
 758
 759        nlmsg_end(skb, nlh);
 760        return 0;
 761
 762nla_put_failure:
 763        nlmsg_trim(skb, nlh);
 764        return -1;
 765}
 766
 767struct nftnl_skb_parms {
 768        bool report;
 769};
 770#define NFT_CB(skb)     (*(struct nftnl_skb_parms*)&((skb)->cb))
 771
 772static void nft_notify_enqueue(struct sk_buff *skb, bool report,
 773                               struct list_head *notify_list)
 774{
 775        NFT_CB(skb).report = report;
 776        list_add_tail(&skb->list, notify_list);
 777}
 778
 779static void nf_tables_table_notify(const struct nft_ctx *ctx, int event)
 780{
 781        struct nftables_pernet *nft_net;
 782        struct sk_buff *skb;
 783        int err;
 784
 785        if (!ctx->report &&
 786            !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
 787                return;
 788
 789        skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
 790        if (skb == NULL)
 791                goto err;
 792
 793        err = nf_tables_fill_table_info(skb, ctx->net, ctx->portid, ctx->seq,
 794                                        event, 0, ctx->family, ctx->table);
 795        if (err < 0) {
 796                kfree_skb(skb);
 797                goto err;
 798        }
 799
 800        nft_net = nft_pernet(ctx->net);
 801        nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
 802        return;
 803err:
 804        nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
 805}
 806
 807static int nf_tables_dump_tables(struct sk_buff *skb,
 808                                 struct netlink_callback *cb)
 809{
 810        const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 811        struct nftables_pernet *nft_net;
 812        const struct nft_table *table;
 813        unsigned int idx = 0, s_idx = cb->args[0];
 814        struct net *net = sock_net(skb->sk);
 815        int family = nfmsg->nfgen_family;
 816
 817        rcu_read_lock();
 818        nft_net = nft_pernet(net);
 819        cb->seq = nft_net->base_seq;
 820
 821        list_for_each_entry_rcu(table, &nft_net->tables, list) {
 822                if (family != NFPROTO_UNSPEC && family != table->family)
 823                        continue;
 824
 825                if (idx < s_idx)
 826                        goto cont;
 827                if (idx > s_idx)
 828                        memset(&cb->args[1], 0,
 829                               sizeof(cb->args) - sizeof(cb->args[0]));
 830                if (!nft_is_active(net, table))
 831                        continue;
 832                if (nf_tables_fill_table_info(skb, net,
 833                                              NETLINK_CB(cb->skb).portid,
 834                                              cb->nlh->nlmsg_seq,
 835                                              NFT_MSG_NEWTABLE, NLM_F_MULTI,
 836                                              table->family, table) < 0)
 837                        goto done;
 838
 839                nl_dump_check_consistent(cb, nlmsg_hdr(skb));
 840cont:
 841                idx++;
 842        }
 843done:
 844        rcu_read_unlock();
 845        cb->args[0] = idx;
 846        return skb->len;
 847}
 848
 849static int nft_netlink_dump_start_rcu(struct sock *nlsk, struct sk_buff *skb,
 850                                      const struct nlmsghdr *nlh,
 851                                      struct netlink_dump_control *c)
 852{
 853        int err;
 854
 855        if (!try_module_get(THIS_MODULE))
 856                return -EINVAL;
 857
 858        rcu_read_unlock();
 859        err = netlink_dump_start(nlsk, skb, nlh, c);
 860        rcu_read_lock();
 861        module_put(THIS_MODULE);
 862
 863        return err;
 864}
 865
 866/* called with rcu_read_lock held */
 867static int nf_tables_gettable(struct sk_buff *skb, const struct nfnl_info *info,
 868                              const struct nlattr * const nla[])
 869{
 870        struct netlink_ext_ack *extack = info->extack;
 871        u8 genmask = nft_genmask_cur(info->net);
 872        u8 family = info->nfmsg->nfgen_family;
 873        const struct nft_table *table;
 874        struct net *net = info->net;
 875        struct sk_buff *skb2;
 876        int err;
 877
 878        if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
 879                struct netlink_dump_control c = {
 880                        .dump = nf_tables_dump_tables,
 881                        .module = THIS_MODULE,
 882                };
 883
 884                return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
 885        }
 886
 887        table = nft_table_lookup(net, nla[NFTA_TABLE_NAME], family, genmask, 0);
 888        if (IS_ERR(table)) {
 889                NL_SET_BAD_ATTR(extack, nla[NFTA_TABLE_NAME]);
 890                return PTR_ERR(table);
 891        }
 892
 893        skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
 894        if (!skb2)
 895                return -ENOMEM;
 896
 897        err = nf_tables_fill_table_info(skb2, net, NETLINK_CB(skb).portid,
 898                                        info->nlh->nlmsg_seq, NFT_MSG_NEWTABLE,
 899                                        0, family, table);
 900        if (err < 0)
 901                goto err_fill_table_info;
 902
 903        return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
 904
 905err_fill_table_info:
 906        kfree_skb(skb2);
 907        return err;
 908}
 909
 910static void nft_table_disable(struct net *net, struct nft_table *table, u32 cnt)
 911{
 912        struct nft_chain *chain;
 913        u32 i = 0;
 914
 915        list_for_each_entry(chain, &table->chains, list) {
 916                if (!nft_is_active_next(net, chain))
 917                        continue;
 918                if (!nft_is_base_chain(chain))
 919                        continue;
 920
 921                if (cnt && i++ == cnt)
 922                        break;
 923
 924                nf_tables_unregister_hook(net, table, chain);
 925        }
 926}
 927
 928static int nf_tables_table_enable(struct net *net, struct nft_table *table)
 929{
 930        struct nft_chain *chain;
 931        int err, i = 0;
 932
 933        list_for_each_entry(chain, &table->chains, list) {
 934                if (!nft_is_active_next(net, chain))
 935                        continue;
 936                if (!nft_is_base_chain(chain))
 937                        continue;
 938
 939                err = nf_tables_register_hook(net, table, chain);
 940                if (err < 0)
 941                        goto err_register_hooks;
 942
 943                i++;
 944        }
 945        return 0;
 946
 947err_register_hooks:
 948        if (i)
 949                nft_table_disable(net, table, i);
 950        return err;
 951}
 952
 953static void nf_tables_table_disable(struct net *net, struct nft_table *table)
 954{
 955        table->flags &= ~NFT_TABLE_F_DORMANT;
 956        nft_table_disable(net, table, 0);
 957        table->flags |= NFT_TABLE_F_DORMANT;
 958}
 959
 960#define __NFT_TABLE_F_INTERNAL          (NFT_TABLE_F_MASK + 1)
 961#define __NFT_TABLE_F_WAS_DORMANT       (__NFT_TABLE_F_INTERNAL << 0)
 962#define __NFT_TABLE_F_WAS_AWAKEN        (__NFT_TABLE_F_INTERNAL << 1)
 963#define __NFT_TABLE_F_UPDATE            (__NFT_TABLE_F_WAS_DORMANT | \
 964                                         __NFT_TABLE_F_WAS_AWAKEN)
 965
 966static int nf_tables_updtable(struct nft_ctx *ctx)
 967{
 968        struct nft_trans *trans;
 969        u32 flags;
 970        int ret;
 971
 972        if (!ctx->nla[NFTA_TABLE_FLAGS])
 973                return 0;
 974
 975        flags = ntohl(nla_get_be32(ctx->nla[NFTA_TABLE_FLAGS]));
 976        if (flags & ~NFT_TABLE_F_MASK)
 977                return -EOPNOTSUPP;
 978
 979        if (flags == ctx->table->flags)
 980                return 0;
 981
 982        if ((nft_table_has_owner(ctx->table) &&
 983             !(flags & NFT_TABLE_F_OWNER)) ||
 984            (!nft_table_has_owner(ctx->table) &&
 985             flags & NFT_TABLE_F_OWNER))
 986                return -EOPNOTSUPP;
 987
 988        trans = nft_trans_alloc(ctx, NFT_MSG_NEWTABLE,
 989                                sizeof(struct nft_trans_table));
 990        if (trans == NULL)
 991                return -ENOMEM;
 992
 993        if ((flags & NFT_TABLE_F_DORMANT) &&
 994            !(ctx->table->flags & NFT_TABLE_F_DORMANT)) {
 995                ctx->table->flags |= NFT_TABLE_F_DORMANT;
 996                if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE))
 997                        ctx->table->flags |= __NFT_TABLE_F_WAS_AWAKEN;
 998        } else if (!(flags & NFT_TABLE_F_DORMANT) &&
 999                   ctx->table->flags & NFT_TABLE_F_DORMANT) {
1000                ctx->table->flags &= ~NFT_TABLE_F_DORMANT;
1001                if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE)) {
1002                        ret = nf_tables_table_enable(ctx->net, ctx->table);
1003                        if (ret < 0)
1004                                goto err_register_hooks;
1005
1006                        ctx->table->flags |= __NFT_TABLE_F_WAS_DORMANT;
1007                }
1008        }
1009
1010        nft_trans_table_update(trans) = true;
1011        nft_trans_commit_list_add_tail(ctx->net, trans);
1012
1013        return 0;
1014
1015err_register_hooks:
1016        nft_trans_destroy(trans);
1017        return ret;
1018}
1019
1020static u32 nft_chain_hash(const void *data, u32 len, u32 seed)
1021{
1022        const char *name = data;
1023
1024        return jhash(name, strlen(name), seed);
1025}
1026
1027static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed)
1028{
1029        const struct nft_chain *chain = data;
1030
1031        return nft_chain_hash(chain->name, 0, seed);
1032}
1033
1034static int nft_chain_hash_cmp(struct rhashtable_compare_arg *arg,
1035                              const void *ptr)
1036{
1037        const struct nft_chain *chain = ptr;
1038        const char *name = arg->key;
1039
1040        return strcmp(chain->name, name);
1041}
1042
1043static u32 nft_objname_hash(const void *data, u32 len, u32 seed)
1044{
1045        const struct nft_object_hash_key *k = data;
1046
1047        seed ^= hash_ptr(k->table, 32);
1048
1049        return jhash(k->name, strlen(k->name), seed);
1050}
1051
1052static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed)
1053{
1054        const struct nft_object *obj = data;
1055
1056        return nft_objname_hash(&obj->key, 0, seed);
1057}
1058
1059static int nft_objname_hash_cmp(struct rhashtable_compare_arg *arg,
1060                                const void *ptr)
1061{
1062        const struct nft_object_hash_key *k = arg->key;
1063        const struct nft_object *obj = ptr;
1064
1065        if (obj->key.table != k->table)
1066                return -1;
1067
1068        return strcmp(obj->key.name, k->name);
1069}
1070
1071static int nf_tables_newtable(struct sk_buff *skb, const struct nfnl_info *info,
1072                              const struct nlattr * const nla[])
1073{
1074        struct nftables_pernet *nft_net = nft_pernet(info->net);
1075        struct netlink_ext_ack *extack = info->extack;
1076        u8 genmask = nft_genmask_next(info->net);
1077        u8 family = info->nfmsg->nfgen_family;
1078        struct net *net = info->net;
1079        const struct nlattr *attr;
1080        struct nft_table *table;
1081        struct nft_ctx ctx;
1082        u32 flags = 0;
1083        int err;
1084
1085        lockdep_assert_held(&nft_net->commit_mutex);
1086        attr = nla[NFTA_TABLE_NAME];
1087        table = nft_table_lookup(net, attr, family, genmask,
1088                                 NETLINK_CB(skb).portid);
1089        if (IS_ERR(table)) {
1090                if (PTR_ERR(table) != -ENOENT)
1091                        return PTR_ERR(table);
1092        } else {
1093                if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
1094                        NL_SET_BAD_ATTR(extack, attr);
1095                        return -EEXIST;
1096                }
1097                if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
1098                        return -EOPNOTSUPP;
1099
1100                nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1101
1102                return nf_tables_updtable(&ctx);
1103        }
1104
1105        if (nla[NFTA_TABLE_FLAGS]) {
1106                flags = ntohl(nla_get_be32(nla[NFTA_TABLE_FLAGS]));
1107                if (flags & ~NFT_TABLE_F_MASK)
1108                        return -EOPNOTSUPP;
1109        }
1110
1111        err = -ENOMEM;
1112        table = kzalloc(sizeof(*table), GFP_KERNEL);
1113        if (table == NULL)
1114                goto err_kzalloc;
1115
1116        table->name = nla_strdup(attr, GFP_KERNEL);
1117        if (table->name == NULL)
1118                goto err_strdup;
1119
1120        if (nla[NFTA_TABLE_USERDATA]) {
1121                table->udata = nla_memdup(nla[NFTA_TABLE_USERDATA], GFP_KERNEL);
1122                if (table->udata == NULL)
1123                        goto err_table_udata;
1124
1125                table->udlen = nla_len(nla[NFTA_TABLE_USERDATA]);
1126        }
1127
1128        err = rhltable_init(&table->chains_ht, &nft_chain_ht_params);
1129        if (err)
1130                goto err_chain_ht;
1131
1132        INIT_LIST_HEAD(&table->chains);
1133        INIT_LIST_HEAD(&table->sets);
1134        INIT_LIST_HEAD(&table->objects);
1135        INIT_LIST_HEAD(&table->flowtables);
1136        table->family = family;
1137        table->flags = flags;
1138        table->handle = ++table_handle;
1139        if (table->flags & NFT_TABLE_F_OWNER)
1140                table->nlpid = NETLINK_CB(skb).portid;
1141
1142        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1143        err = nft_trans_table_add(&ctx, NFT_MSG_NEWTABLE);
1144        if (err < 0)
1145                goto err_trans;
1146
1147        list_add_tail_rcu(&table->list, &nft_net->tables);
1148        return 0;
1149err_trans:
1150        rhltable_destroy(&table->chains_ht);
1151err_chain_ht:
1152        kfree(table->udata);
1153err_table_udata:
1154        kfree(table->name);
1155err_strdup:
1156        kfree(table);
1157err_kzalloc:
1158        return err;
1159}
1160
1161static int nft_flush_table(struct nft_ctx *ctx)
1162{
1163        struct nft_flowtable *flowtable, *nft;
1164        struct nft_chain *chain, *nc;
1165        struct nft_object *obj, *ne;
1166        struct nft_set *set, *ns;
1167        int err;
1168
1169        list_for_each_entry(chain, &ctx->table->chains, list) {
1170                if (!nft_is_active_next(ctx->net, chain))
1171                        continue;
1172
1173                if (nft_chain_is_bound(chain))
1174                        continue;
1175
1176                ctx->chain = chain;
1177
1178                err = nft_delrule_by_chain(ctx);
1179                if (err < 0)
1180                        goto out;
1181        }
1182
1183        list_for_each_entry_safe(set, ns, &ctx->table->sets, list) {
1184                if (!nft_is_active_next(ctx->net, set))
1185                        continue;
1186
1187                if (nft_set_is_anonymous(set) &&
1188                    !list_empty(&set->bindings))
1189                        continue;
1190
1191                err = nft_delset(ctx, set);
1192                if (err < 0)
1193                        goto out;
1194        }
1195
1196        list_for_each_entry_safe(flowtable, nft, &ctx->table->flowtables, list) {
1197                if (!nft_is_active_next(ctx->net, flowtable))
1198                        continue;
1199
1200                err = nft_delflowtable(ctx, flowtable);
1201                if (err < 0)
1202                        goto out;
1203        }
1204
1205        list_for_each_entry_safe(obj, ne, &ctx->table->objects, list) {
1206                if (!nft_is_active_next(ctx->net, obj))
1207                        continue;
1208
1209                err = nft_delobj(ctx, obj);
1210                if (err < 0)
1211                        goto out;
1212        }
1213
1214        list_for_each_entry_safe(chain, nc, &ctx->table->chains, list) {
1215                if (!nft_is_active_next(ctx->net, chain))
1216                        continue;
1217
1218                if (nft_chain_is_bound(chain))
1219                        continue;
1220
1221                ctx->chain = chain;
1222
1223                err = nft_delchain(ctx);
1224                if (err < 0)
1225                        goto out;
1226        }
1227
1228        err = nft_deltable(ctx);
1229out:
1230        return err;
1231}
1232
1233static int nft_flush(struct nft_ctx *ctx, int family)
1234{
1235        struct nftables_pernet *nft_net = nft_pernet(ctx->net);
1236        const struct nlattr * const *nla = ctx->nla;
1237        struct nft_table *table, *nt;
1238        int err = 0;
1239
1240        list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
1241                if (family != AF_UNSPEC && table->family != family)
1242                        continue;
1243
1244                ctx->family = table->family;
1245
1246                if (!nft_is_active_next(ctx->net, table))
1247                        continue;
1248
1249                if (nft_table_has_owner(table) && table->nlpid != ctx->portid)
1250                        continue;
1251
1252                if (nla[NFTA_TABLE_NAME] &&
1253                    nla_strcmp(nla[NFTA_TABLE_NAME], table->name) != 0)
1254                        continue;
1255
1256                ctx->table = table;
1257
1258                err = nft_flush_table(ctx);
1259                if (err < 0)
1260                        goto out;
1261        }
1262out:
1263        return err;
1264}
1265
1266static int nf_tables_deltable(struct sk_buff *skb, const struct nfnl_info *info,
1267                              const struct nlattr * const nla[])
1268{
1269        struct netlink_ext_ack *extack = info->extack;
1270        u8 genmask = nft_genmask_next(info->net);
1271        u8 family = info->nfmsg->nfgen_family;
1272        struct net *net = info->net;
1273        const struct nlattr *attr;
1274        struct nft_table *table;
1275        struct nft_ctx ctx;
1276
1277        nft_ctx_init(&ctx, net, skb, info->nlh, 0, NULL, NULL, nla);
1278        if (family == AF_UNSPEC ||
1279            (!nla[NFTA_TABLE_NAME] && !nla[NFTA_TABLE_HANDLE]))
1280                return nft_flush(&ctx, family);
1281
1282        if (nla[NFTA_TABLE_HANDLE]) {
1283                attr = nla[NFTA_TABLE_HANDLE];
1284                table = nft_table_lookup_byhandle(net, attr, genmask,
1285                                                  NETLINK_CB(skb).portid);
1286        } else {
1287                attr = nla[NFTA_TABLE_NAME];
1288                table = nft_table_lookup(net, attr, family, genmask,
1289                                         NETLINK_CB(skb).portid);
1290        }
1291
1292        if (IS_ERR(table)) {
1293                NL_SET_BAD_ATTR(extack, attr);
1294                return PTR_ERR(table);
1295        }
1296
1297        if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
1298            table->use > 0)
1299                return -EBUSY;
1300
1301        ctx.family = family;
1302        ctx.table = table;
1303
1304        return nft_flush_table(&ctx);
1305}
1306
1307static void nf_tables_table_destroy(struct nft_ctx *ctx)
1308{
1309        if (WARN_ON(ctx->table->use > 0))
1310                return;
1311
1312        rhltable_destroy(&ctx->table->chains_ht);
1313        kfree(ctx->table->name);
1314        kfree(ctx->table->udata);
1315        kfree(ctx->table);
1316}
1317
1318void nft_register_chain_type(const struct nft_chain_type *ctype)
1319{
1320        nfnl_lock(NFNL_SUBSYS_NFTABLES);
1321        if (WARN_ON(__nft_chain_type_get(ctype->family, ctype->type))) {
1322                nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1323                return;
1324        }
1325        chain_type[ctype->family][ctype->type] = ctype;
1326        nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1327}
1328EXPORT_SYMBOL_GPL(nft_register_chain_type);
1329
1330void nft_unregister_chain_type(const struct nft_chain_type *ctype)
1331{
1332        nfnl_lock(NFNL_SUBSYS_NFTABLES);
1333        chain_type[ctype->family][ctype->type] = NULL;
1334        nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1335}
1336EXPORT_SYMBOL_GPL(nft_unregister_chain_type);
1337
1338/*
1339 * Chains
1340 */
1341
1342static struct nft_chain *
1343nft_chain_lookup_byhandle(const struct nft_table *table, u64 handle, u8 genmask)
1344{
1345        struct nft_chain *chain;
1346
1347        list_for_each_entry(chain, &table->chains, list) {
1348                if (chain->handle == handle &&
1349                    nft_active_genmask(chain, genmask))
1350                        return chain;
1351        }
1352
1353        return ERR_PTR(-ENOENT);
1354}
1355
1356static bool lockdep_commit_lock_is_held(const struct net *net)
1357{
1358#ifdef CONFIG_PROVE_LOCKING
1359        struct nftables_pernet *nft_net = nft_pernet(net);
1360
1361        return lockdep_is_held(&nft_net->commit_mutex);
1362#else
1363        return true;
1364#endif
1365}
1366
1367static struct nft_chain *nft_chain_lookup(struct net *net,
1368                                          struct nft_table *table,
1369                                          const struct nlattr *nla, u8 genmask)
1370{
1371        char search[NFT_CHAIN_MAXNAMELEN + 1];
1372        struct rhlist_head *tmp, *list;
1373        struct nft_chain *chain;
1374
1375        if (nla == NULL)
1376                return ERR_PTR(-EINVAL);
1377
1378        nla_strscpy(search, nla, sizeof(search));
1379
1380        WARN_ON(!rcu_read_lock_held() &&
1381                !lockdep_commit_lock_is_held(net));
1382
1383        chain = ERR_PTR(-ENOENT);
1384        rcu_read_lock();
1385        list = rhltable_lookup(&table->chains_ht, search, nft_chain_ht_params);
1386        if (!list)
1387                goto out_unlock;
1388
1389        rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
1390                if (nft_active_genmask(chain, genmask))
1391                        goto out_unlock;
1392        }
1393        chain = ERR_PTR(-ENOENT);
1394out_unlock:
1395        rcu_read_unlock();
1396        return chain;
1397}
1398
1399static const struct nla_policy nft_chain_policy[NFTA_CHAIN_MAX + 1] = {
1400        [NFTA_CHAIN_TABLE]      = { .type = NLA_STRING,
1401                                    .len = NFT_TABLE_MAXNAMELEN - 1 },
1402        [NFTA_CHAIN_HANDLE]     = { .type = NLA_U64 },
1403        [NFTA_CHAIN_NAME]       = { .type = NLA_STRING,
1404                                    .len = NFT_CHAIN_MAXNAMELEN - 1 },
1405        [NFTA_CHAIN_HOOK]       = { .type = NLA_NESTED },
1406        [NFTA_CHAIN_POLICY]     = { .type = NLA_U32 },
1407        [NFTA_CHAIN_TYPE]       = { .type = NLA_STRING,
1408                                    .len = NFT_MODULE_AUTOLOAD_LIMIT },
1409        [NFTA_CHAIN_COUNTERS]   = { .type = NLA_NESTED },
1410        [NFTA_CHAIN_FLAGS]      = { .type = NLA_U32 },
1411        [NFTA_CHAIN_ID]         = { .type = NLA_U32 },
1412        [NFTA_CHAIN_USERDATA]   = { .type = NLA_BINARY,
1413                                    .len = NFT_USERDATA_MAXLEN },
1414};
1415
1416static const struct nla_policy nft_hook_policy[NFTA_HOOK_MAX + 1] = {
1417        [NFTA_HOOK_HOOKNUM]     = { .type = NLA_U32 },
1418        [NFTA_HOOK_PRIORITY]    = { .type = NLA_U32 },
1419        [NFTA_HOOK_DEV]         = { .type = NLA_STRING,
1420                                    .len = IFNAMSIZ - 1 },
1421};
1422
1423static int nft_dump_stats(struct sk_buff *skb, struct nft_stats __percpu *stats)
1424{
1425        struct nft_stats *cpu_stats, total;
1426        struct nlattr *nest;
1427        unsigned int seq;
1428        u64 pkts, bytes;
1429        int cpu;
1430
1431        if (!stats)
1432                return 0;
1433
1434        memset(&total, 0, sizeof(total));
1435        for_each_possible_cpu(cpu) {
1436                cpu_stats = per_cpu_ptr(stats, cpu);
1437                do {
1438                        seq = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1439                        pkts = cpu_stats->pkts;
1440                        bytes = cpu_stats->bytes;
1441                } while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, seq));
1442                total.pkts += pkts;
1443                total.bytes += bytes;
1444        }
1445        nest = nla_nest_start_noflag(skb, NFTA_CHAIN_COUNTERS);
1446        if (nest == NULL)
1447                goto nla_put_failure;
1448
1449        if (nla_put_be64(skb, NFTA_COUNTER_PACKETS, cpu_to_be64(total.pkts),
1450                         NFTA_COUNTER_PAD) ||
1451            nla_put_be64(skb, NFTA_COUNTER_BYTES, cpu_to_be64(total.bytes),
1452                         NFTA_COUNTER_PAD))
1453                goto nla_put_failure;
1454
1455        nla_nest_end(skb, nest);
1456        return 0;
1457
1458nla_put_failure:
1459        return -ENOSPC;
1460}
1461
1462static int nft_dump_basechain_hook(struct sk_buff *skb, int family,
1463                                   const struct nft_base_chain *basechain)
1464{
1465        const struct nf_hook_ops *ops = &basechain->ops;
1466        struct nft_hook *hook, *first = NULL;
1467        struct nlattr *nest, *nest_devs;
1468        int n = 0;
1469
1470        nest = nla_nest_start_noflag(skb, NFTA_CHAIN_HOOK);
1471        if (nest == NULL)
1472                goto nla_put_failure;
1473        if (nla_put_be32(skb, NFTA_HOOK_HOOKNUM, htonl(ops->hooknum)))
1474                goto nla_put_failure;
1475        if (nla_put_be32(skb, NFTA_HOOK_PRIORITY, htonl(ops->priority)))
1476                goto nla_put_failure;
1477
1478        if (nft_base_chain_netdev(family, ops->hooknum)) {
1479                nest_devs = nla_nest_start_noflag(skb, NFTA_HOOK_DEVS);
1480                list_for_each_entry(hook, &basechain->hook_list, list) {
1481                        if (!first)
1482                                first = hook;
1483
1484                        if (nla_put_string(skb, NFTA_DEVICE_NAME,
1485                                           hook->ops.dev->name))
1486                                goto nla_put_failure;
1487                        n++;
1488                }
1489                nla_nest_end(skb, nest_devs);
1490
1491                if (n == 1 &&
1492                    nla_put_string(skb, NFTA_HOOK_DEV, first->ops.dev->name))
1493                        goto nla_put_failure;
1494        }
1495        nla_nest_end(skb, nest);
1496
1497        return 0;
1498nla_put_failure:
1499        return -1;
1500}
1501
1502static int nf_tables_fill_chain_info(struct sk_buff *skb, struct net *net,
1503                                     u32 portid, u32 seq, int event, u32 flags,
1504                                     int family, const struct nft_table *table,
1505                                     const struct nft_chain *chain)
1506{
1507        struct nlmsghdr *nlh;
1508
1509        event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
1510        nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
1511                           NFNETLINK_V0, nft_base_seq(net));
1512        if (!nlh)
1513                goto nla_put_failure;
1514
1515        if (nla_put_string(skb, NFTA_CHAIN_TABLE, table->name))
1516                goto nla_put_failure;
1517        if (nla_put_be64(skb, NFTA_CHAIN_HANDLE, cpu_to_be64(chain->handle),
1518                         NFTA_CHAIN_PAD))
1519                goto nla_put_failure;
1520        if (nla_put_string(skb, NFTA_CHAIN_NAME, chain->name))
1521                goto nla_put_failure;
1522
1523        if (nft_is_base_chain(chain)) {
1524                const struct nft_base_chain *basechain = nft_base_chain(chain);
1525                struct nft_stats __percpu *stats;
1526
1527                if (nft_dump_basechain_hook(skb, family, basechain))
1528                        goto nla_put_failure;
1529
1530                if (nla_put_be32(skb, NFTA_CHAIN_POLICY,
1531                                 htonl(basechain->policy)))
1532                        goto nla_put_failure;
1533
1534                if (nla_put_string(skb, NFTA_CHAIN_TYPE, basechain->type->name))
1535                        goto nla_put_failure;
1536
1537                stats = rcu_dereference_check(basechain->stats,
1538                                              lockdep_commit_lock_is_held(net));
1539                if (nft_dump_stats(skb, stats))
1540                        goto nla_put_failure;
1541        }
1542
1543        if (chain->flags &&
1544            nla_put_be32(skb, NFTA_CHAIN_FLAGS, htonl(chain->flags)))
1545                goto nla_put_failure;
1546
1547        if (nla_put_be32(skb, NFTA_CHAIN_USE, htonl(chain->use)))
1548                goto nla_put_failure;
1549
1550        if (chain->udata &&
1551            nla_put(skb, NFTA_CHAIN_USERDATA, chain->udlen, chain->udata))
1552                goto nla_put_failure;
1553
1554        nlmsg_end(skb, nlh);
1555        return 0;
1556
1557nla_put_failure:
1558        nlmsg_trim(skb, nlh);
1559        return -1;
1560}
1561
1562static void nf_tables_chain_notify(const struct nft_ctx *ctx, int event)
1563{
1564        struct nftables_pernet *nft_net;
1565        struct sk_buff *skb;
1566        int err;
1567
1568        if (!ctx->report &&
1569            !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1570                return;
1571
1572        skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1573        if (skb == NULL)
1574                goto err;
1575
1576        err = nf_tables_fill_chain_info(skb, ctx->net, ctx->portid, ctx->seq,
1577                                        event, 0, ctx->family, ctx->table,
1578                                        ctx->chain);
1579        if (err < 0) {
1580                kfree_skb(skb);
1581                goto err;
1582        }
1583
1584        nft_net = nft_pernet(ctx->net);
1585        nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
1586        return;
1587err:
1588        nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
1589}
1590
1591static int nf_tables_dump_chains(struct sk_buff *skb,
1592                                 struct netlink_callback *cb)
1593{
1594        const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1595        unsigned int idx = 0, s_idx = cb->args[0];
1596        struct net *net = sock_net(skb->sk);
1597        int family = nfmsg->nfgen_family;
1598        struct nftables_pernet *nft_net;
1599        const struct nft_table *table;
1600        const struct nft_chain *chain;
1601
1602        rcu_read_lock();
1603        nft_net = nft_pernet(net);
1604        cb->seq = nft_net->base_seq;
1605
1606        list_for_each_entry_rcu(table, &nft_net->tables, list) {
1607                if (family != NFPROTO_UNSPEC && family != table->family)
1608                        continue;
1609
1610                list_for_each_entry_rcu(chain, &table->chains, list) {
1611                        if (idx < s_idx)
1612                                goto cont;
1613                        if (idx > s_idx)
1614                                memset(&cb->args[1], 0,
1615                                       sizeof(cb->args) - sizeof(cb->args[0]));
1616                        if (!nft_is_active(net, chain))
1617                                continue;
1618                        if (nf_tables_fill_chain_info(skb, net,
1619                                                      NETLINK_CB(cb->skb).portid,
1620                                                      cb->nlh->nlmsg_seq,
1621                                                      NFT_MSG_NEWCHAIN,
1622                                                      NLM_F_MULTI,
1623                                                      table->family, table,
1624                                                      chain) < 0)
1625                                goto done;
1626
1627                        nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1628cont:
1629                        idx++;
1630                }
1631        }
1632done:
1633        rcu_read_unlock();
1634        cb->args[0] = idx;
1635        return skb->len;
1636}
1637
1638/* called with rcu_read_lock held */
1639static int nf_tables_getchain(struct sk_buff *skb, const struct nfnl_info *info,
1640                              const struct nlattr * const nla[])
1641{
1642        struct netlink_ext_ack *extack = info->extack;
1643        u8 genmask = nft_genmask_cur(info->net);
1644        u8 family = info->nfmsg->nfgen_family;
1645        const struct nft_chain *chain;
1646        struct net *net = info->net;
1647        struct nft_table *table;
1648        struct sk_buff *skb2;
1649        int err;
1650
1651        if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1652                struct netlink_dump_control c = {
1653                        .dump = nf_tables_dump_chains,
1654                        .module = THIS_MODULE,
1655                };
1656
1657                return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
1658        }
1659
1660        table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask, 0);
1661        if (IS_ERR(table)) {
1662                NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
1663                return PTR_ERR(table);
1664        }
1665
1666        chain = nft_chain_lookup(net, table, nla[NFTA_CHAIN_NAME], genmask);
1667        if (IS_ERR(chain)) {
1668                NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
1669                return PTR_ERR(chain);
1670        }
1671
1672        skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1673        if (!skb2)
1674                return -ENOMEM;
1675
1676        err = nf_tables_fill_chain_info(skb2, net, NETLINK_CB(skb).portid,
1677                                        info->nlh->nlmsg_seq, NFT_MSG_NEWCHAIN,
1678                                        0, family, table, chain);
1679        if (err < 0)
1680                goto err_fill_chain_info;
1681
1682        return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
1683
1684err_fill_chain_info:
1685        kfree_skb(skb2);
1686        return err;
1687}
1688
1689static const struct nla_policy nft_counter_policy[NFTA_COUNTER_MAX + 1] = {
1690        [NFTA_COUNTER_PACKETS]  = { .type = NLA_U64 },
1691        [NFTA_COUNTER_BYTES]    = { .type = NLA_U64 },
1692};
1693
1694static struct nft_stats __percpu *nft_stats_alloc(const struct nlattr *attr)
1695{
1696        struct nlattr *tb[NFTA_COUNTER_MAX+1];
1697        struct nft_stats __percpu *newstats;
1698        struct nft_stats *stats;
1699        int err;
1700
1701        err = nla_parse_nested_deprecated(tb, NFTA_COUNTER_MAX, attr,
1702                                          nft_counter_policy, NULL);
1703        if (err < 0)
1704                return ERR_PTR(err);
1705
1706        if (!tb[NFTA_COUNTER_BYTES] || !tb[NFTA_COUNTER_PACKETS])
1707                return ERR_PTR(-EINVAL);
1708
1709        newstats = netdev_alloc_pcpu_stats(struct nft_stats);
1710        if (newstats == NULL)
1711                return ERR_PTR(-ENOMEM);
1712
1713        /* Restore old counters on this cpu, no problem. Per-cpu statistics
1714         * are not exposed to userspace.
1715         */
1716        preempt_disable();
1717        stats = this_cpu_ptr(newstats);
1718        stats->bytes = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_BYTES]));
1719        stats->pkts = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_PACKETS]));
1720        preempt_enable();
1721
1722        return newstats;
1723}
1724
1725static void nft_chain_stats_replace(struct nft_trans *trans)
1726{
1727        struct nft_base_chain *chain = nft_base_chain(trans->ctx.chain);
1728
1729        if (!nft_trans_chain_stats(trans))
1730                return;
1731
1732        nft_trans_chain_stats(trans) =
1733                rcu_replace_pointer(chain->stats, nft_trans_chain_stats(trans),
1734                                    lockdep_commit_lock_is_held(trans->ctx.net));
1735
1736        if (!nft_trans_chain_stats(trans))
1737                static_branch_inc(&nft_counters_enabled);
1738}
1739
1740static void nf_tables_chain_free_chain_rules(struct nft_chain *chain)
1741{
1742        struct nft_rule **g0 = rcu_dereference_raw(chain->rules_gen_0);
1743        struct nft_rule **g1 = rcu_dereference_raw(chain->rules_gen_1);
1744
1745        if (g0 != g1)
1746                kvfree(g1);
1747        kvfree(g0);
1748
1749        /* should be NULL either via abort or via successful commit */
1750        WARN_ON_ONCE(chain->rules_next);
1751        kvfree(chain->rules_next);
1752}
1753
1754void nf_tables_chain_destroy(struct nft_ctx *ctx)
1755{
1756        struct nft_chain *chain = ctx->chain;
1757        struct nft_hook *hook, *next;
1758
1759        if (WARN_ON(chain->use > 0))
1760                return;
1761
1762        /* no concurrent access possible anymore */
1763        nf_tables_chain_free_chain_rules(chain);
1764
1765        if (nft_is_base_chain(chain)) {
1766                struct nft_base_chain *basechain = nft_base_chain(chain);
1767
1768                if (nft_base_chain_netdev(ctx->family, basechain->ops.hooknum)) {
1769                        list_for_each_entry_safe(hook, next,
1770                                                 &basechain->hook_list, list) {
1771                                list_del_rcu(&hook->list);
1772                                kfree_rcu(hook, rcu);
1773                        }
1774                }
1775                module_put(basechain->type->owner);
1776                if (rcu_access_pointer(basechain->stats)) {
1777                        static_branch_dec(&nft_counters_enabled);
1778                        free_percpu(rcu_dereference_raw(basechain->stats));
1779                }
1780                kfree(chain->name);
1781                kfree(chain->udata);
1782                kfree(basechain);
1783        } else {
1784                kfree(chain->name);
1785                kfree(chain->udata);
1786                kfree(chain);
1787        }
1788}
1789
1790static struct nft_hook *nft_netdev_hook_alloc(struct net *net,
1791                                              const struct nlattr *attr)
1792{
1793        struct net_device *dev;
1794        char ifname[IFNAMSIZ];
1795        struct nft_hook *hook;
1796        int err;
1797
1798        hook = kmalloc(sizeof(struct nft_hook), GFP_KERNEL);
1799        if (!hook) {
1800                err = -ENOMEM;
1801                goto err_hook_alloc;
1802        }
1803
1804        nla_strscpy(ifname, attr, IFNAMSIZ);
1805        /* nf_tables_netdev_event() is called under rtnl_mutex, this is
1806         * indirectly serializing all the other holders of the commit_mutex with
1807         * the rtnl_mutex.
1808         */
1809        dev = __dev_get_by_name(net, ifname);
1810        if (!dev) {
1811                err = -ENOENT;
1812                goto err_hook_dev;
1813        }
1814        hook->ops.dev = dev;
1815        hook->inactive = false;
1816
1817        return hook;
1818
1819err_hook_dev:
1820        kfree(hook);
1821err_hook_alloc:
1822        return ERR_PTR(err);
1823}
1824
1825static struct nft_hook *nft_hook_list_find(struct list_head *hook_list,
1826                                           const struct nft_hook *this)
1827{
1828        struct nft_hook *hook;
1829
1830        list_for_each_entry(hook, hook_list, list) {
1831                if (this->ops.dev == hook->ops.dev)
1832                        return hook;
1833        }
1834
1835        return NULL;
1836}
1837
1838static int nf_tables_parse_netdev_hooks(struct net *net,
1839                                        const struct nlattr *attr,
1840                                        struct list_head *hook_list)
1841{
1842        struct nft_hook *hook, *next;
1843        const struct nlattr *tmp;
1844        int rem, n = 0, err;
1845
1846        nla_for_each_nested(tmp, attr, rem) {
1847                if (nla_type(tmp) != NFTA_DEVICE_NAME) {
1848                        err = -EINVAL;
1849                        goto err_hook;
1850                }
1851
1852                hook = nft_netdev_hook_alloc(net, tmp);
1853                if (IS_ERR(hook)) {
1854                        err = PTR_ERR(hook);
1855                        goto err_hook;
1856                }
1857                if (nft_hook_list_find(hook_list, hook)) {
1858                        kfree(hook);
1859                        err = -EEXIST;
1860                        goto err_hook;
1861                }
1862                list_add_tail(&hook->list, hook_list);
1863                n++;
1864
1865                if (n == NFT_NETDEVICE_MAX) {
1866                        err = -EFBIG;
1867                        goto err_hook;
1868                }
1869        }
1870
1871        return 0;
1872
1873err_hook:
1874        list_for_each_entry_safe(hook, next, hook_list, list) {
1875                list_del(&hook->list);
1876                kfree(hook);
1877        }
1878        return err;
1879}
1880
1881struct nft_chain_hook {
1882        u32                             num;
1883        s32                             priority;
1884        const struct nft_chain_type     *type;
1885        struct list_head                list;
1886};
1887
1888static int nft_chain_parse_netdev(struct net *net,
1889                                  struct nlattr *tb[],
1890                                  struct list_head *hook_list)
1891{
1892        struct nft_hook *hook;
1893        int err;
1894
1895        if (tb[NFTA_HOOK_DEV]) {
1896                hook = nft_netdev_hook_alloc(net, tb[NFTA_HOOK_DEV]);
1897                if (IS_ERR(hook))
1898                        return PTR_ERR(hook);
1899
1900                list_add_tail(&hook->list, hook_list);
1901        } else if (tb[NFTA_HOOK_DEVS]) {
1902                err = nf_tables_parse_netdev_hooks(net, tb[NFTA_HOOK_DEVS],
1903                                                   hook_list);
1904                if (err < 0)
1905                        return err;
1906
1907                if (list_empty(hook_list))
1908                        return -EINVAL;
1909        } else {
1910                return -EINVAL;
1911        }
1912
1913        return 0;
1914}
1915
1916static int nft_chain_parse_hook(struct net *net,
1917                                const struct nlattr * const nla[],
1918                                struct nft_chain_hook *hook, u8 family,
1919                                struct netlink_ext_ack *extack, bool autoload)
1920{
1921        struct nftables_pernet *nft_net = nft_pernet(net);
1922        struct nlattr *ha[NFTA_HOOK_MAX + 1];
1923        const struct nft_chain_type *type;
1924        int err;
1925
1926        lockdep_assert_held(&nft_net->commit_mutex);
1927        lockdep_nfnl_nft_mutex_not_held();
1928
1929        err = nla_parse_nested_deprecated(ha, NFTA_HOOK_MAX,
1930                                          nla[NFTA_CHAIN_HOOK],
1931                                          nft_hook_policy, NULL);
1932        if (err < 0)
1933                return err;
1934
1935        if (ha[NFTA_HOOK_HOOKNUM] == NULL ||
1936            ha[NFTA_HOOK_PRIORITY] == NULL)
1937                return -EINVAL;
1938
1939        hook->num = ntohl(nla_get_be32(ha[NFTA_HOOK_HOOKNUM]));
1940        hook->priority = ntohl(nla_get_be32(ha[NFTA_HOOK_PRIORITY]));
1941
1942        type = __nft_chain_type_get(family, NFT_CHAIN_T_DEFAULT);
1943        if (!type)
1944                return -EOPNOTSUPP;
1945
1946        if (nla[NFTA_CHAIN_TYPE]) {
1947                type = nf_tables_chain_type_lookup(net, nla[NFTA_CHAIN_TYPE],
1948                                                   family, autoload);
1949                if (IS_ERR(type)) {
1950                        NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
1951                        return PTR_ERR(type);
1952                }
1953        }
1954        if (hook->num >= NFT_MAX_HOOKS || !(type->hook_mask & (1 << hook->num)))
1955                return -EOPNOTSUPP;
1956
1957        if (type->type == NFT_CHAIN_T_NAT &&
1958            hook->priority <= NF_IP_PRI_CONNTRACK)
1959                return -EOPNOTSUPP;
1960
1961        if (!try_module_get(type->owner)) {
1962                if (nla[NFTA_CHAIN_TYPE])
1963                        NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
1964                return -ENOENT;
1965        }
1966
1967        hook->type = type;
1968
1969        INIT_LIST_HEAD(&hook->list);
1970        if (nft_base_chain_netdev(family, hook->num)) {
1971                err = nft_chain_parse_netdev(net, ha, &hook->list);
1972                if (err < 0) {
1973                        module_put(type->owner);
1974                        return err;
1975                }
1976        } else if (ha[NFTA_HOOK_DEV] || ha[NFTA_HOOK_DEVS]) {
1977                module_put(type->owner);
1978                return -EOPNOTSUPP;
1979        }
1980
1981        return 0;
1982}
1983
1984static void nft_chain_release_hook(struct nft_chain_hook *hook)
1985{
1986        struct nft_hook *h, *next;
1987
1988        list_for_each_entry_safe(h, next, &hook->list, list) {
1989                list_del(&h->list);
1990                kfree(h);
1991        }
1992        module_put(hook->type->owner);
1993}
1994
1995struct nft_rules_old {
1996        struct rcu_head h;
1997        struct nft_rule **start;
1998};
1999
2000static struct nft_rule **nf_tables_chain_alloc_rules(const struct nft_chain *chain,
2001                                                     unsigned int alloc)
2002{
2003        if (alloc > INT_MAX)
2004                return NULL;
2005
2006        alloc += 1;     /* NULL, ends rules */
2007        if (sizeof(struct nft_rule *) > INT_MAX / alloc)
2008                return NULL;
2009
2010        alloc *= sizeof(struct nft_rule *);
2011        alloc += sizeof(struct nft_rules_old);
2012
2013        return kvmalloc(alloc, GFP_KERNEL);
2014}
2015
2016static void nft_basechain_hook_init(struct nf_hook_ops *ops, u8 family,
2017                                    const struct nft_chain_hook *hook,
2018                                    struct nft_chain *chain)
2019{
2020        ops->pf                 = family;
2021        ops->hooknum            = hook->num;
2022        ops->priority           = hook->priority;
2023        ops->priv               = chain;
2024        ops->hook               = hook->type->hooks[ops->hooknum];
2025        ops->hook_ops_type      = NF_HOOK_OP_NF_TABLES;
2026}
2027
2028static int nft_basechain_init(struct nft_base_chain *basechain, u8 family,
2029                              struct nft_chain_hook *hook, u32 flags)
2030{
2031        struct nft_chain *chain;
2032        struct nft_hook *h;
2033
2034        basechain->type = hook->type;
2035        INIT_LIST_HEAD(&basechain->hook_list);
2036        chain = &basechain->chain;
2037
2038        if (nft_base_chain_netdev(family, hook->num)) {
2039                list_splice_init(&hook->list, &basechain->hook_list);
2040                list_for_each_entry(h, &basechain->hook_list, list)
2041                        nft_basechain_hook_init(&h->ops, family, hook, chain);
2042
2043                basechain->ops.hooknum  = hook->num;
2044                basechain->ops.priority = hook->priority;
2045        } else {
2046                nft_basechain_hook_init(&basechain->ops, family, hook, chain);
2047        }
2048
2049        chain->flags |= NFT_CHAIN_BASE | flags;
2050        basechain->policy = NF_ACCEPT;
2051        if (chain->flags & NFT_CHAIN_HW_OFFLOAD &&
2052            nft_chain_offload_priority(basechain) < 0)
2053                return -EOPNOTSUPP;
2054
2055        flow_block_init(&basechain->flow_block);
2056
2057        return 0;
2058}
2059
2060static int nft_chain_add(struct nft_table *table, struct nft_chain *chain)
2061{
2062        int err;
2063
2064        err = rhltable_insert_key(&table->chains_ht, chain->name,
2065                                  &chain->rhlhead, nft_chain_ht_params);
2066        if (err)
2067                return err;
2068
2069        list_add_tail_rcu(&chain->list, &table->chains);
2070
2071        return 0;
2072}
2073
2074static u64 chain_id;
2075
2076static int nf_tables_addchain(struct nft_ctx *ctx, u8 family, u8 genmask,
2077                              u8 policy, u32 flags,
2078                              struct netlink_ext_ack *extack)
2079{
2080        const struct nlattr * const *nla = ctx->nla;
2081        struct nft_table *table = ctx->table;
2082        struct nft_base_chain *basechain;
2083        struct nft_stats __percpu *stats;
2084        struct net *net = ctx->net;
2085        char name[NFT_NAME_MAXLEN];
2086        struct nft_trans *trans;
2087        struct nft_chain *chain;
2088        struct nft_rule **rules;
2089        int err;
2090
2091        if (table->use == UINT_MAX)
2092                return -EOVERFLOW;
2093
2094        if (nla[NFTA_CHAIN_HOOK]) {
2095                struct nft_chain_hook hook;
2096
2097                if (flags & NFT_CHAIN_BINDING)
2098                        return -EOPNOTSUPP;
2099
2100                err = nft_chain_parse_hook(net, nla, &hook, family, extack,
2101                                           true);
2102                if (err < 0)
2103                        return err;
2104
2105                basechain = kzalloc(sizeof(*basechain), GFP_KERNEL);
2106                if (basechain == NULL) {
2107                        nft_chain_release_hook(&hook);
2108                        return -ENOMEM;
2109                }
2110                chain = &basechain->chain;
2111
2112                if (nla[NFTA_CHAIN_COUNTERS]) {
2113                        stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2114                        if (IS_ERR(stats)) {
2115                                nft_chain_release_hook(&hook);
2116                                kfree(basechain);
2117                                return PTR_ERR(stats);
2118                        }
2119                        rcu_assign_pointer(basechain->stats, stats);
2120                        static_branch_inc(&nft_counters_enabled);
2121                }
2122
2123                err = nft_basechain_init(basechain, family, &hook, flags);
2124                if (err < 0) {
2125                        nft_chain_release_hook(&hook);
2126                        kfree(basechain);
2127                        return err;
2128                }
2129        } else {
2130                if (flags & NFT_CHAIN_BASE)
2131                        return -EINVAL;
2132                if (flags & NFT_CHAIN_HW_OFFLOAD)
2133                        return -EOPNOTSUPP;
2134
2135                chain = kzalloc(sizeof(*chain), GFP_KERNEL);
2136                if (chain == NULL)
2137                        return -ENOMEM;
2138
2139                chain->flags = flags;
2140        }
2141        ctx->chain = chain;
2142
2143        INIT_LIST_HEAD(&chain->rules);
2144        chain->handle = nf_tables_alloc_handle(table);
2145        chain->table = table;
2146
2147        if (nla[NFTA_CHAIN_NAME]) {
2148                chain->name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL);
2149        } else {
2150                if (!(flags & NFT_CHAIN_BINDING)) {
2151                        err = -EINVAL;
2152                        goto err_destroy_chain;
2153                }
2154
2155                snprintf(name, sizeof(name), "__chain%llu", ++chain_id);
2156                chain->name = kstrdup(name, GFP_KERNEL);
2157        }
2158
2159        if (!chain->name) {
2160                err = -ENOMEM;
2161                goto err_destroy_chain;
2162        }
2163
2164        if (nla[NFTA_CHAIN_USERDATA]) {
2165                chain->udata = nla_memdup(nla[NFTA_CHAIN_USERDATA], GFP_KERNEL);
2166                if (chain->udata == NULL) {
2167                        err = -ENOMEM;
2168                        goto err_destroy_chain;
2169                }
2170                chain->udlen = nla_len(nla[NFTA_CHAIN_USERDATA]);
2171        }
2172
2173        rules = nf_tables_chain_alloc_rules(chain, 0);
2174        if (!rules) {
2175                err = -ENOMEM;
2176                goto err_destroy_chain;
2177        }
2178
2179        *rules = NULL;
2180        rcu_assign_pointer(chain->rules_gen_0, rules);
2181        rcu_assign_pointer(chain->rules_gen_1, rules);
2182
2183        err = nf_tables_register_hook(net, table, chain);
2184        if (err < 0)
2185                goto err_destroy_chain;
2186
2187        trans = nft_trans_chain_add(ctx, NFT_MSG_NEWCHAIN);
2188        if (IS_ERR(trans)) {
2189                err = PTR_ERR(trans);
2190                goto err_unregister_hook;
2191        }
2192
2193        nft_trans_chain_policy(trans) = NFT_CHAIN_POLICY_UNSET;
2194        if (nft_is_base_chain(chain))
2195                nft_trans_chain_policy(trans) = policy;
2196
2197        err = nft_chain_add(table, chain);
2198        if (err < 0) {
2199                nft_trans_destroy(trans);
2200                goto err_unregister_hook;
2201        }
2202
2203        table->use++;
2204
2205        return 0;
2206err_unregister_hook:
2207        nf_tables_unregister_hook(net, table, chain);
2208err_destroy_chain:
2209        nf_tables_chain_destroy(ctx);
2210
2211        return err;
2212}
2213
2214static bool nft_hook_list_equal(struct list_head *hook_list1,
2215                                struct list_head *hook_list2)
2216{
2217        struct nft_hook *hook;
2218        int n = 0, m = 0;
2219
2220        n = 0;
2221        list_for_each_entry(hook, hook_list2, list) {
2222                if (!nft_hook_list_find(hook_list1, hook))
2223                        return false;
2224
2225                n++;
2226        }
2227        list_for_each_entry(hook, hook_list1, list)
2228                m++;
2229
2230        return n == m;
2231}
2232
2233static int nf_tables_updchain(struct nft_ctx *ctx, u8 genmask, u8 policy,
2234                              u32 flags, const struct nlattr *attr,
2235                              struct netlink_ext_ack *extack)
2236{
2237        const struct nlattr * const *nla = ctx->nla;
2238        struct nft_table *table = ctx->table;
2239        struct nft_chain *chain = ctx->chain;
2240        struct nft_base_chain *basechain;
2241        struct nft_stats *stats = NULL;
2242        struct nft_chain_hook hook;
2243        struct nf_hook_ops *ops;
2244        struct nft_trans *trans;
2245        int err;
2246
2247        if (chain->flags ^ flags)
2248                return -EOPNOTSUPP;
2249
2250        if (nla[NFTA_CHAIN_HOOK]) {
2251                if (!nft_is_base_chain(chain)) {
2252                        NL_SET_BAD_ATTR(extack, attr);
2253                        return -EEXIST;
2254                }
2255                err = nft_chain_parse_hook(ctx->net, nla, &hook, ctx->family,
2256                                           extack, false);
2257                if (err < 0)
2258                        return err;
2259
2260                basechain = nft_base_chain(chain);
2261                if (basechain->type != hook.type) {
2262                        nft_chain_release_hook(&hook);
2263                        NL_SET_BAD_ATTR(extack, attr);
2264                        return -EEXIST;
2265                }
2266
2267                if (nft_base_chain_netdev(ctx->family, hook.num)) {
2268                        if (!nft_hook_list_equal(&basechain->hook_list,
2269                                                 &hook.list)) {
2270                                nft_chain_release_hook(&hook);
2271                                NL_SET_BAD_ATTR(extack, attr);
2272                                return -EEXIST;
2273                        }
2274                } else {
2275                        ops = &basechain->ops;
2276                        if (ops->hooknum != hook.num ||
2277                            ops->priority != hook.priority) {
2278                                nft_chain_release_hook(&hook);
2279                                NL_SET_BAD_ATTR(extack, attr);
2280                                return -EEXIST;
2281                        }
2282                }
2283                nft_chain_release_hook(&hook);
2284        }
2285
2286        if (nla[NFTA_CHAIN_HANDLE] &&
2287            nla[NFTA_CHAIN_NAME]) {
2288                struct nft_chain *chain2;
2289
2290                chain2 = nft_chain_lookup(ctx->net, table,
2291                                          nla[NFTA_CHAIN_NAME], genmask);
2292                if (!IS_ERR(chain2)) {
2293                        NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2294                        return -EEXIST;
2295                }
2296        }
2297
2298        if (nla[NFTA_CHAIN_COUNTERS]) {
2299                if (!nft_is_base_chain(chain))
2300                        return -EOPNOTSUPP;
2301
2302                stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2303                if (IS_ERR(stats))
2304                        return PTR_ERR(stats);
2305        }
2306
2307        err = -ENOMEM;
2308        trans = nft_trans_alloc(ctx, NFT_MSG_NEWCHAIN,
2309                                sizeof(struct nft_trans_chain));
2310        if (trans == NULL)
2311                goto err;
2312
2313        nft_trans_chain_stats(trans) = stats;
2314        nft_trans_chain_update(trans) = true;
2315
2316        if (nla[NFTA_CHAIN_POLICY])
2317                nft_trans_chain_policy(trans) = policy;
2318        else
2319                nft_trans_chain_policy(trans) = -1;
2320
2321        if (nla[NFTA_CHAIN_HANDLE] &&
2322            nla[NFTA_CHAIN_NAME]) {
2323                struct nftables_pernet *nft_net = nft_pernet(ctx->net);
2324                struct nft_trans *tmp;
2325                char *name;
2326
2327                err = -ENOMEM;
2328                name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL);
2329                if (!name)
2330                        goto err;
2331
2332                err = -EEXIST;
2333                list_for_each_entry(tmp, &nft_net->commit_list, list) {
2334                        if (tmp->msg_type == NFT_MSG_NEWCHAIN &&
2335                            tmp->ctx.table == table &&
2336                            nft_trans_chain_update(tmp) &&
2337                            nft_trans_chain_name(tmp) &&
2338                            strcmp(name, nft_trans_chain_name(tmp)) == 0) {
2339                                NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2340                                kfree(name);
2341                                goto err;
2342                        }
2343                }
2344
2345                nft_trans_chain_name(trans) = name;
2346        }
2347        nft_trans_commit_list_add_tail(ctx->net, trans);
2348
2349        return 0;
2350err:
2351        free_percpu(stats);
2352        kfree(trans);
2353        return err;
2354}
2355
2356static struct nft_chain *nft_chain_lookup_byid(const struct net *net,
2357                                               const struct nlattr *nla)
2358{
2359        struct nftables_pernet *nft_net = nft_pernet(net);
2360        u32 id = ntohl(nla_get_be32(nla));
2361        struct nft_trans *trans;
2362
2363        list_for_each_entry(trans, &nft_net->commit_list, list) {
2364                struct nft_chain *chain = trans->ctx.chain;
2365
2366                if (trans->msg_type == NFT_MSG_NEWCHAIN &&
2367                    id == nft_trans_chain_id(trans))
2368                        return chain;
2369        }
2370        return ERR_PTR(-ENOENT);
2371}
2372
2373static int nf_tables_newchain(struct sk_buff *skb, const struct nfnl_info *info,
2374                              const struct nlattr * const nla[])
2375{
2376        struct nftables_pernet *nft_net = nft_pernet(info->net);
2377        struct netlink_ext_ack *extack = info->extack;
2378        u8 genmask = nft_genmask_next(info->net);
2379        u8 family = info->nfmsg->nfgen_family;
2380        struct nft_chain *chain = NULL;
2381        struct net *net = info->net;
2382        const struct nlattr *attr;
2383        struct nft_table *table;
2384        u8 policy = NF_ACCEPT;
2385        struct nft_ctx ctx;
2386        u64 handle = 0;
2387        u32 flags = 0;
2388
2389        lockdep_assert_held(&nft_net->commit_mutex);
2390
2391        table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
2392                                 NETLINK_CB(skb).portid);
2393        if (IS_ERR(table)) {
2394                NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2395                return PTR_ERR(table);
2396        }
2397
2398        chain = NULL;
2399        attr = nla[NFTA_CHAIN_NAME];
2400
2401        if (nla[NFTA_CHAIN_HANDLE]) {
2402                handle = be64_to_cpu(nla_get_be64(nla[NFTA_CHAIN_HANDLE]));
2403                chain = nft_chain_lookup_byhandle(table, handle, genmask);
2404                if (IS_ERR(chain)) {
2405                        NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_HANDLE]);
2406                        return PTR_ERR(chain);
2407                }
2408                attr = nla[NFTA_CHAIN_HANDLE];
2409        } else if (nla[NFTA_CHAIN_NAME]) {
2410                chain = nft_chain_lookup(net, table, attr, genmask);
2411                if (IS_ERR(chain)) {
2412                        if (PTR_ERR(chain) != -ENOENT) {
2413                                NL_SET_BAD_ATTR(extack, attr);
2414                                return PTR_ERR(chain);
2415                        }
2416                        chain = NULL;
2417                }
2418        } else if (!nla[NFTA_CHAIN_ID]) {
2419                return -EINVAL;
2420        }
2421
2422        if (nla[NFTA_CHAIN_POLICY]) {
2423                if (chain != NULL &&
2424                    !nft_is_base_chain(chain)) {
2425                        NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2426                        return -EOPNOTSUPP;
2427                }
2428
2429                if (chain == NULL &&
2430                    nla[NFTA_CHAIN_HOOK] == NULL) {
2431                        NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2432                        return -EOPNOTSUPP;
2433                }
2434
2435                policy = ntohl(nla_get_be32(nla[NFTA_CHAIN_POLICY]));
2436                switch (policy) {
2437                case NF_DROP:
2438                case NF_ACCEPT:
2439                        break;
2440                default:
2441                        return -EINVAL;
2442                }
2443        }
2444
2445        if (nla[NFTA_CHAIN_FLAGS])
2446                flags = ntohl(nla_get_be32(nla[NFTA_CHAIN_FLAGS]));
2447        else if (chain)
2448                flags = chain->flags;
2449
2450        if (flags & ~NFT_CHAIN_FLAGS)
2451                return -EOPNOTSUPP;
2452
2453        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
2454
2455        if (chain != NULL) {
2456                if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
2457                        NL_SET_BAD_ATTR(extack, attr);
2458                        return -EEXIST;
2459                }
2460                if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
2461                        return -EOPNOTSUPP;
2462
2463                flags |= chain->flags & NFT_CHAIN_BASE;
2464                return nf_tables_updchain(&ctx, genmask, policy, flags, attr,
2465                                          extack);
2466        }
2467
2468        return nf_tables_addchain(&ctx, family, genmask, policy, flags, extack);
2469}
2470
2471static int nf_tables_delchain(struct sk_buff *skb, const struct nfnl_info *info,
2472                              const struct nlattr * const nla[])
2473{
2474        struct netlink_ext_ack *extack = info->extack;
2475        u8 genmask = nft_genmask_next(info->net);
2476        u8 family = info->nfmsg->nfgen_family;
2477        struct net *net = info->net;
2478        const struct nlattr *attr;
2479        struct nft_table *table;
2480        struct nft_chain *chain;
2481        struct nft_rule *rule;
2482        struct nft_ctx ctx;
2483        u64 handle;
2484        u32 use;
2485        int err;
2486
2487        table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
2488                                 NETLINK_CB(skb).portid);
2489        if (IS_ERR(table)) {
2490                NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2491                return PTR_ERR(table);
2492        }
2493
2494        if (nla[NFTA_CHAIN_HANDLE]) {
2495                attr = nla[NFTA_CHAIN_HANDLE];
2496                handle = be64_to_cpu(nla_get_be64(attr));
2497                chain = nft_chain_lookup_byhandle(table, handle, genmask);
2498        } else {
2499                attr = nla[NFTA_CHAIN_NAME];
2500                chain = nft_chain_lookup(net, table, attr, genmask);
2501        }
2502        if (IS_ERR(chain)) {
2503                NL_SET_BAD_ATTR(extack, attr);
2504                return PTR_ERR(chain);
2505        }
2506
2507        if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
2508            chain->use > 0)
2509                return -EBUSY;
2510
2511        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
2512
2513        use = chain->use;
2514        list_for_each_entry(rule, &chain->rules, list) {
2515                if (!nft_is_active_next(net, rule))
2516                        continue;
2517                use--;
2518
2519                err = nft_delrule(&ctx, rule);
2520                if (err < 0)
2521                        return err;
2522        }
2523
2524        /* There are rules and elements that are still holding references to us,
2525         * we cannot do a recursive removal in this case.
2526         */
2527        if (use > 0) {
2528                NL_SET_BAD_ATTR(extack, attr);
2529                return -EBUSY;
2530        }
2531
2532        return nft_delchain(&ctx);
2533}
2534
2535/*
2536 * Expressions
2537 */
2538
2539/**
2540 *      nft_register_expr - register nf_tables expr type
2541 *      @type: expr type
2542 *
2543 *      Registers the expr type for use with nf_tables. Returns zero on
2544 *      success or a negative errno code otherwise.
2545 */
2546int nft_register_expr(struct nft_expr_type *type)
2547{
2548        nfnl_lock(NFNL_SUBSYS_NFTABLES);
2549        if (type->family == NFPROTO_UNSPEC)
2550                list_add_tail_rcu(&type->list, &nf_tables_expressions);
2551        else
2552                list_add_rcu(&type->list, &nf_tables_expressions);
2553        nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2554        return 0;
2555}
2556EXPORT_SYMBOL_GPL(nft_register_expr);
2557
2558/**
2559 *      nft_unregister_expr - unregister nf_tables expr type
2560 *      @type: expr type
2561 *
2562 *      Unregisters the expr typefor use with nf_tables.
2563 */
2564void nft_unregister_expr(struct nft_expr_type *type)
2565{
2566        nfnl_lock(NFNL_SUBSYS_NFTABLES);
2567        list_del_rcu(&type->list);
2568        nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2569}
2570EXPORT_SYMBOL_GPL(nft_unregister_expr);
2571
2572static const struct nft_expr_type *__nft_expr_type_get(u8 family,
2573                                                       struct nlattr *nla)
2574{
2575        const struct nft_expr_type *type, *candidate = NULL;
2576
2577        list_for_each_entry(type, &nf_tables_expressions, list) {
2578                if (!nla_strcmp(nla, type->name)) {
2579                        if (!type->family && !candidate)
2580                                candidate = type;
2581                        else if (type->family == family)
2582                                candidate = type;
2583                }
2584        }
2585        return candidate;
2586}
2587
2588#ifdef CONFIG_MODULES
2589static int nft_expr_type_request_module(struct net *net, u8 family,
2590                                        struct nlattr *nla)
2591{
2592        if (nft_request_module(net, "nft-expr-%u-%.*s", family,
2593                               nla_len(nla), (char *)nla_data(nla)) == -EAGAIN)
2594                return -EAGAIN;
2595
2596        return 0;
2597}
2598#endif
2599
2600static const struct nft_expr_type *nft_expr_type_get(struct net *net,
2601                                                     u8 family,
2602                                                     struct nlattr *nla)
2603{
2604        const struct nft_expr_type *type;
2605
2606        if (nla == NULL)
2607                return ERR_PTR(-EINVAL);
2608
2609        type = __nft_expr_type_get(family, nla);
2610        if (type != NULL && try_module_get(type->owner))
2611                return type;
2612
2613        lockdep_nfnl_nft_mutex_not_held();
2614#ifdef CONFIG_MODULES
2615        if (type == NULL) {
2616                if (nft_expr_type_request_module(net, family, nla) == -EAGAIN)
2617                        return ERR_PTR(-EAGAIN);
2618
2619                if (nft_request_module(net, "nft-expr-%.*s",
2620                                       nla_len(nla),
2621                                       (char *)nla_data(nla)) == -EAGAIN)
2622                        return ERR_PTR(-EAGAIN);
2623        }
2624#endif
2625        return ERR_PTR(-ENOENT);
2626}
2627
2628static const struct nla_policy nft_expr_policy[NFTA_EXPR_MAX + 1] = {
2629        [NFTA_EXPR_NAME]        = { .type = NLA_STRING,
2630                                    .len = NFT_MODULE_AUTOLOAD_LIMIT },
2631        [NFTA_EXPR_DATA]        = { .type = NLA_NESTED },
2632};
2633
2634static int nf_tables_fill_expr_info(struct sk_buff *skb,
2635                                    const struct nft_expr *expr)
2636{
2637        if (nla_put_string(skb, NFTA_EXPR_NAME, expr->ops->type->name))
2638                goto nla_put_failure;
2639
2640        if (expr->ops->dump) {
2641                struct nlattr *data = nla_nest_start_noflag(skb,
2642                                                            NFTA_EXPR_DATA);
2643                if (data == NULL)
2644                        goto nla_put_failure;
2645                if (expr->ops->dump(skb, expr) < 0)
2646                        goto nla_put_failure;
2647                nla_nest_end(skb, data);
2648        }
2649
2650        return skb->len;
2651
2652nla_put_failure:
2653        return -1;
2654};
2655
2656int nft_expr_dump(struct sk_buff *skb, unsigned int attr,
2657                  const struct nft_expr *expr)
2658{
2659        struct nlattr *nest;
2660
2661        nest = nla_nest_start_noflag(skb, attr);
2662        if (!nest)
2663                goto nla_put_failure;
2664        if (nf_tables_fill_expr_info(skb, expr) < 0)
2665                goto nla_put_failure;
2666        nla_nest_end(skb, nest);
2667        return 0;
2668
2669nla_put_failure:
2670        return -1;
2671}
2672
2673struct nft_expr_info {
2674        const struct nft_expr_ops       *ops;
2675        const struct nlattr             *attr;
2676        struct nlattr                   *tb[NFT_EXPR_MAXATTR + 1];
2677};
2678
2679static int nf_tables_expr_parse(const struct nft_ctx *ctx,
2680                                const struct nlattr *nla,
2681                                struct nft_expr_info *info)
2682{
2683        const struct nft_expr_type *type;
2684        const struct nft_expr_ops *ops;
2685        struct nlattr *tb[NFTA_EXPR_MAX + 1];
2686        int err;
2687
2688        err = nla_parse_nested_deprecated(tb, NFTA_EXPR_MAX, nla,
2689                                          nft_expr_policy, NULL);
2690        if (err < 0)
2691                return err;
2692
2693        type = nft_expr_type_get(ctx->net, ctx->family, tb[NFTA_EXPR_NAME]);
2694        if (IS_ERR(type))
2695                return PTR_ERR(type);
2696
2697        if (tb[NFTA_EXPR_DATA]) {
2698                err = nla_parse_nested_deprecated(info->tb, type->maxattr,
2699                                                  tb[NFTA_EXPR_DATA],
2700                                                  type->policy, NULL);
2701                if (err < 0)
2702                        goto err1;
2703        } else
2704                memset(info->tb, 0, sizeof(info->tb[0]) * (type->maxattr + 1));
2705
2706        if (type->select_ops != NULL) {
2707                ops = type->select_ops(ctx,
2708                                       (const struct nlattr * const *)info->tb);
2709                if (IS_ERR(ops)) {
2710                        err = PTR_ERR(ops);
2711#ifdef CONFIG_MODULES
2712                        if (err == -EAGAIN)
2713                                if (nft_expr_type_request_module(ctx->net,
2714                                                                 ctx->family,
2715                                                                 tb[NFTA_EXPR_NAME]) != -EAGAIN)
2716                                        err = -ENOENT;
2717#endif
2718                        goto err1;
2719                }
2720        } else
2721                ops = type->ops;
2722
2723        info->attr = nla;
2724        info->ops = ops;
2725
2726        return 0;
2727
2728err1:
2729        module_put(type->owner);
2730        return err;
2731}
2732
2733static int nf_tables_newexpr(const struct nft_ctx *ctx,
2734                             const struct nft_expr_info *expr_info,
2735                             struct nft_expr *expr)
2736{
2737        const struct nft_expr_ops *ops = expr_info->ops;
2738        int err;
2739
2740        expr->ops = ops;
2741        if (ops->init) {
2742                err = ops->init(ctx, expr, (const struct nlattr **)expr_info->tb);
2743                if (err < 0)
2744                        goto err1;
2745        }
2746
2747        return 0;
2748err1:
2749        expr->ops = NULL;
2750        return err;
2751}
2752
2753static void nf_tables_expr_destroy(const struct nft_ctx *ctx,
2754                                   struct nft_expr *expr)
2755{
2756        const struct nft_expr_type *type = expr->ops->type;
2757
2758        if (expr->ops->destroy)
2759                expr->ops->destroy(ctx, expr);
2760        module_put(type->owner);
2761}
2762
2763static struct nft_expr *nft_expr_init(const struct nft_ctx *ctx,
2764                                      const struct nlattr *nla)
2765{
2766        struct nft_expr_info expr_info;
2767        struct nft_expr *expr;
2768        struct module *owner;
2769        int err;
2770
2771        err = nf_tables_expr_parse(ctx, nla, &expr_info);
2772        if (err < 0)
2773                goto err1;
2774
2775        err = -ENOMEM;
2776        expr = kzalloc(expr_info.ops->size, GFP_KERNEL);
2777        if (expr == NULL)
2778                goto err2;
2779
2780        err = nf_tables_newexpr(ctx, &expr_info, expr);
2781        if (err < 0)
2782                goto err3;
2783
2784        return expr;
2785err3:
2786        kfree(expr);
2787err2:
2788        owner = expr_info.ops->type->owner;
2789        if (expr_info.ops->type->release_ops)
2790                expr_info.ops->type->release_ops(expr_info.ops);
2791
2792        module_put(owner);
2793err1:
2794        return ERR_PTR(err);
2795}
2796
2797int nft_expr_clone(struct nft_expr *dst, struct nft_expr *src)
2798{
2799        int err;
2800
2801        if (src->ops->clone) {
2802                dst->ops = src->ops;
2803                err = src->ops->clone(dst, src);
2804                if (err < 0)
2805                        return err;
2806        } else {
2807                memcpy(dst, src, src->ops->size);
2808        }
2809
2810        __module_get(src->ops->type->owner);
2811
2812        return 0;
2813}
2814
2815void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
2816{
2817        nf_tables_expr_destroy(ctx, expr);
2818        kfree(expr);
2819}
2820
2821/*
2822 * Rules
2823 */
2824
2825static struct nft_rule *__nft_rule_lookup(const struct nft_chain *chain,
2826                                          u64 handle)
2827{
2828        struct nft_rule *rule;
2829
2830        // FIXME: this sucks
2831        list_for_each_entry_rcu(rule, &chain->rules, list) {
2832                if (handle == rule->handle)
2833                        return rule;
2834        }
2835
2836        return ERR_PTR(-ENOENT);
2837}
2838
2839static struct nft_rule *nft_rule_lookup(const struct nft_chain *chain,
2840                                        const struct nlattr *nla)
2841{
2842        if (nla == NULL)
2843                return ERR_PTR(-EINVAL);
2844
2845        return __nft_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
2846}
2847
2848static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
2849        [NFTA_RULE_TABLE]       = { .type = NLA_STRING,
2850                                    .len = NFT_TABLE_MAXNAMELEN - 1 },
2851        [NFTA_RULE_CHAIN]       = { .type = NLA_STRING,
2852                                    .len = NFT_CHAIN_MAXNAMELEN - 1 },
2853        [NFTA_RULE_HANDLE]      = { .type = NLA_U64 },
2854        [NFTA_RULE_EXPRESSIONS] = { .type = NLA_NESTED },
2855        [NFTA_RULE_COMPAT]      = { .type = NLA_NESTED },
2856        [NFTA_RULE_POSITION]    = { .type = NLA_U64 },
2857        [NFTA_RULE_USERDATA]    = { .type = NLA_BINARY,
2858                                    .len = NFT_USERDATA_MAXLEN },
2859        [NFTA_RULE_ID]          = { .type = NLA_U32 },
2860        [NFTA_RULE_POSITION_ID] = { .type = NLA_U32 },
2861        [NFTA_RULE_CHAIN_ID]    = { .type = NLA_U32 },
2862};
2863
2864static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
2865                                    u32 portid, u32 seq, int event,
2866                                    u32 flags, int family,
2867                                    const struct nft_table *table,
2868                                    const struct nft_chain *chain,
2869                                    const struct nft_rule *rule,
2870                                    const struct nft_rule *prule)
2871{
2872        struct nlmsghdr *nlh;
2873        const struct nft_expr *expr, *next;
2874        struct nlattr *list;
2875        u16 type = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
2876
2877        nlh = nfnl_msg_put(skb, portid, seq, type, flags, family, NFNETLINK_V0,
2878                           nft_base_seq(net));
2879        if (!nlh)
2880                goto nla_put_failure;
2881
2882        if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
2883                goto nla_put_failure;
2884        if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
2885                goto nla_put_failure;
2886        if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle),
2887                         NFTA_RULE_PAD))
2888                goto nla_put_failure;
2889
2890        if (event != NFT_MSG_DELRULE && prule) {
2891                if (nla_put_be64(skb, NFTA_RULE_POSITION,
2892                                 cpu_to_be64(prule->handle),
2893                                 NFTA_RULE_PAD))
2894                        goto nla_put_failure;
2895        }
2896
2897        if (chain->flags & NFT_CHAIN_HW_OFFLOAD)
2898                nft_flow_rule_stats(chain, rule);
2899
2900        list = nla_nest_start_noflag(skb, NFTA_RULE_EXPRESSIONS);
2901        if (list == NULL)
2902                goto nla_put_failure;
2903        nft_rule_for_each_expr(expr, next, rule) {
2904                if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
2905                        goto nla_put_failure;
2906        }
2907        nla_nest_end(skb, list);
2908
2909        if (rule->udata) {
2910                struct nft_userdata *udata = nft_userdata(rule);
2911                if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
2912                            udata->data) < 0)
2913                        goto nla_put_failure;
2914        }
2915
2916        nlmsg_end(skb, nlh);
2917        return 0;
2918
2919nla_put_failure:
2920        nlmsg_trim(skb, nlh);
2921        return -1;
2922}
2923
2924static void nf_tables_rule_notify(const struct nft_ctx *ctx,
2925                                  const struct nft_rule *rule, int event)
2926{
2927        struct nftables_pernet *nft_net = nft_pernet(ctx->net);
2928        struct sk_buff *skb;
2929        int err;
2930
2931        if (!ctx->report &&
2932            !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
2933                return;
2934
2935        skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
2936        if (skb == NULL)
2937                goto err;
2938
2939        err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
2940                                       event, 0, ctx->family, ctx->table,
2941                                       ctx->chain, rule, NULL);
2942        if (err < 0) {
2943                kfree_skb(skb);
2944                goto err;
2945        }
2946
2947        nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
2948        return;
2949err:
2950        nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
2951}
2952
2953struct nft_rule_dump_ctx {
2954        char *table;
2955        char *chain;
2956};
2957
2958static int __nf_tables_dump_rules(struct sk_buff *skb,
2959                                  unsigned int *idx,
2960                                  struct netlink_callback *cb,
2961                                  const struct nft_table *table,
2962                                  const struct nft_chain *chain)
2963{
2964        struct net *net = sock_net(skb->sk);
2965        const struct nft_rule *rule, *prule;
2966        unsigned int s_idx = cb->args[0];
2967
2968        prule = NULL;
2969        list_for_each_entry_rcu(rule, &chain->rules, list) {
2970                if (!nft_is_active(net, rule))
2971                        goto cont_skip;
2972                if (*idx < s_idx)
2973                        goto cont;
2974                if (*idx > s_idx) {
2975                        memset(&cb->args[1], 0,
2976                                        sizeof(cb->args) - sizeof(cb->args[0]));
2977                }
2978                if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
2979                                        cb->nlh->nlmsg_seq,
2980                                        NFT_MSG_NEWRULE,
2981                                        NLM_F_MULTI | NLM_F_APPEND,
2982                                        table->family,
2983                                        table, chain, rule, prule) < 0)
2984                        return 1;
2985
2986                nl_dump_check_consistent(cb, nlmsg_hdr(skb));
2987cont:
2988                prule = rule;
2989cont_skip:
2990                (*idx)++;
2991        }
2992        return 0;
2993}
2994
2995static int nf_tables_dump_rules(struct sk_buff *skb,
2996                                struct netlink_callback *cb)
2997{
2998        const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2999        const struct nft_rule_dump_ctx *ctx = cb->data;
3000        struct nft_table *table;
3001        const struct nft_chain *chain;
3002        unsigned int idx = 0;
3003        struct net *net = sock_net(skb->sk);
3004        int family = nfmsg->nfgen_family;
3005        struct nftables_pernet *nft_net;
3006
3007        rcu_read_lock();
3008        nft_net = nft_pernet(net);
3009        cb->seq = nft_net->base_seq;
3010
3011        list_for_each_entry_rcu(table, &nft_net->tables, list) {
3012                if (family != NFPROTO_UNSPEC && family != table->family)
3013                        continue;
3014
3015                if (ctx && ctx->table && strcmp(ctx->table, table->name) != 0)
3016                        continue;
3017
3018                if (ctx && ctx->table && ctx->chain) {
3019                        struct rhlist_head *list, *tmp;
3020
3021                        list = rhltable_lookup(&table->chains_ht, ctx->chain,
3022                                               nft_chain_ht_params);
3023                        if (!list)
3024                                goto done;
3025
3026                        rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
3027                                if (!nft_is_active(net, chain))
3028                                        continue;
3029                                __nf_tables_dump_rules(skb, &idx,
3030                                                       cb, table, chain);
3031                                break;
3032                        }
3033                        goto done;
3034                }
3035
3036                list_for_each_entry_rcu(chain, &table->chains, list) {
3037                        if (__nf_tables_dump_rules(skb, &idx, cb, table, chain))
3038                                goto done;
3039                }
3040
3041                if (ctx && ctx->table)
3042                        break;
3043        }
3044done:
3045        rcu_read_unlock();
3046
3047        cb->args[0] = idx;
3048        return skb->len;
3049}
3050
3051static int nf_tables_dump_rules_start(struct netlink_callback *cb)
3052{
3053        const struct nlattr * const *nla = cb->data;
3054        struct nft_rule_dump_ctx *ctx = NULL;
3055
3056        if (nla[NFTA_RULE_TABLE] || nla[NFTA_RULE_CHAIN]) {
3057                ctx = kzalloc(sizeof(*ctx), GFP_ATOMIC);
3058                if (!ctx)
3059                        return -ENOMEM;
3060
3061                if (nla[NFTA_RULE_TABLE]) {
3062                        ctx->table = nla_strdup(nla[NFTA_RULE_TABLE],
3063                                                        GFP_ATOMIC);
3064                        if (!ctx->table) {
3065                                kfree(ctx);
3066                                return -ENOMEM;
3067                        }
3068                }
3069                if (nla[NFTA_RULE_CHAIN]) {
3070                        ctx->chain = nla_strdup(nla[NFTA_RULE_CHAIN],
3071                                                GFP_ATOMIC);
3072                        if (!ctx->chain) {
3073                                kfree(ctx->table);
3074                                kfree(ctx);
3075                                return -ENOMEM;
3076                        }
3077                }
3078        }
3079
3080        cb->data = ctx;
3081        return 0;
3082}
3083
3084static int nf_tables_dump_rules_done(struct netlink_callback *cb)
3085{
3086        struct nft_rule_dump_ctx *ctx = cb->data;
3087
3088        if (ctx) {
3089                kfree(ctx->table);
3090                kfree(ctx->chain);
3091                kfree(ctx);
3092        }
3093        return 0;
3094}
3095
3096/* called with rcu_read_lock held */
3097static int nf_tables_getrule(struct sk_buff *skb, const struct nfnl_info *info,
3098                             const struct nlattr * const nla[])
3099{
3100        struct netlink_ext_ack *extack = info->extack;
3101        u8 genmask = nft_genmask_cur(info->net);
3102        u8 family = info->nfmsg->nfgen_family;
3103        const struct nft_chain *chain;
3104        const struct nft_rule *rule;
3105        struct net *net = info->net;
3106        struct nft_table *table;
3107        struct sk_buff *skb2;
3108        int err;
3109
3110        if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3111                struct netlink_dump_control c = {
3112                        .start= nf_tables_dump_rules_start,
3113                        .dump = nf_tables_dump_rules,
3114                        .done = nf_tables_dump_rules_done,
3115                        .module = THIS_MODULE,
3116                        .data = (void *)nla,
3117                };
3118
3119                return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
3120        }
3121
3122        table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask, 0);
3123        if (IS_ERR(table)) {
3124                NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3125                return PTR_ERR(table);
3126        }
3127
3128        chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
3129        if (IS_ERR(chain)) {
3130                NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3131                return PTR_ERR(chain);
3132        }
3133
3134        rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3135        if (IS_ERR(rule)) {
3136                NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3137                return PTR_ERR(rule);
3138        }
3139
3140        skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
3141        if (!skb2)
3142                return -ENOMEM;
3143
3144        err = nf_tables_fill_rule_info(skb2, net, NETLINK_CB(skb).portid,
3145                                       info->nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
3146                                       family, table, chain, rule, NULL);
3147        if (err < 0)
3148                goto err_fill_rule_info;
3149
3150        return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
3151
3152err_fill_rule_info:
3153        kfree_skb(skb2);
3154        return err;
3155}
3156
3157static void nf_tables_rule_destroy(const struct nft_ctx *ctx,
3158                                   struct nft_rule *rule)
3159{
3160        struct nft_expr *expr, *next;
3161
3162        /*
3163         * Careful: some expressions might not be initialized in case this
3164         * is called on error from nf_tables_newrule().
3165         */
3166        expr = nft_expr_first(rule);
3167        while (nft_expr_more(rule, expr)) {
3168                next = nft_expr_next(expr);
3169                nf_tables_expr_destroy(ctx, expr);
3170                expr = next;
3171        }
3172        kfree(rule);
3173}
3174
3175void nf_tables_rule_release(const struct nft_ctx *ctx, struct nft_rule *rule)
3176{
3177        nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_RELEASE);
3178        nf_tables_rule_destroy(ctx, rule);
3179}
3180
3181int nft_chain_validate(const struct nft_ctx *ctx, const struct nft_chain *chain)
3182{
3183        struct nft_expr *expr, *last;
3184        const struct nft_data *data;
3185        struct nft_rule *rule;
3186        int err;
3187
3188        if (ctx->level == NFT_JUMP_STACK_SIZE)
3189                return -EMLINK;
3190
3191        list_for_each_entry(rule, &chain->rules, list) {
3192                if (!nft_is_active_next(ctx->net, rule))
3193                        continue;
3194
3195                nft_rule_for_each_expr(expr, last, rule) {
3196                        if (!expr->ops->validate)
3197                                continue;
3198
3199                        err = expr->ops->validate(ctx, expr, &data);
3200                        if (err < 0)
3201                                return err;
3202                }
3203        }
3204
3205        return 0;
3206}
3207EXPORT_SYMBOL_GPL(nft_chain_validate);
3208
3209static int nft_table_validate(struct net *net, const struct nft_table *table)
3210{
3211        struct nft_chain *chain;
3212        struct nft_ctx ctx = {
3213                .net    = net,
3214                .family = table->family,
3215        };
3216        int err;
3217
3218        list_for_each_entry(chain, &table->chains, list) {
3219                if (!nft_is_base_chain(chain))
3220                        continue;
3221
3222                ctx.chain = chain;
3223                err = nft_chain_validate(&ctx, chain);
3224                if (err < 0)
3225                        return err;
3226        }
3227
3228        return 0;
3229}
3230
3231static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3232                                             const struct nlattr *nla);
3233
3234#define NFT_RULE_MAXEXPRS       128
3235
3236static int nf_tables_newrule(struct sk_buff *skb, const struct nfnl_info *info,
3237                             const struct nlattr * const nla[])
3238{
3239        struct nftables_pernet *nft_net = nft_pernet(info->net);
3240        struct netlink_ext_ack *extack = info->extack;
3241        unsigned int size, i, n, ulen = 0, usize = 0;
3242        u8 genmask = nft_genmask_next(info->net);
3243        struct nft_rule *rule, *old_rule = NULL;
3244        struct nft_expr_info *expr_info = NULL;
3245        u8 family = info->nfmsg->nfgen_family;
3246        struct nft_flow_rule *flow = NULL;
3247        struct net *net = info->net;
3248        struct nft_userdata *udata;
3249        struct nft_table *table;
3250        struct nft_chain *chain;
3251        struct nft_trans *trans;
3252        u64 handle, pos_handle;
3253        struct nft_expr *expr;
3254        struct nft_ctx ctx;
3255        struct nlattr *tmp;
3256        int err, rem;
3257
3258        lockdep_assert_held(&nft_net->commit_mutex);
3259
3260        table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3261                                 NETLINK_CB(skb).portid);
3262        if (IS_ERR(table)) {
3263                NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3264                return PTR_ERR(table);
3265        }
3266
3267        if (nla[NFTA_RULE_CHAIN]) {
3268                chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3269                                         genmask);
3270                if (IS_ERR(chain)) {
3271                        NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3272                        return PTR_ERR(chain);
3273                }
3274                if (nft_chain_is_bound(chain))
3275                        return -EOPNOTSUPP;
3276
3277        } else if (nla[NFTA_RULE_CHAIN_ID]) {
3278                chain = nft_chain_lookup_byid(net, nla[NFTA_RULE_CHAIN_ID]);
3279                if (IS_ERR(chain)) {
3280                        NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN_ID]);
3281                        return PTR_ERR(chain);
3282                }
3283        } else {
3284                return -EINVAL;
3285        }
3286
3287        if (nla[NFTA_RULE_HANDLE]) {
3288                handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
3289                rule = __nft_rule_lookup(chain, handle);
3290                if (IS_ERR(rule)) {
3291                        NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3292                        return PTR_ERR(rule);
3293                }
3294
3295                if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
3296                        NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3297                        return -EEXIST;
3298                }
3299                if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
3300                        old_rule = rule;
3301                else
3302                        return -EOPNOTSUPP;
3303        } else {
3304                if (!(info->nlh->nlmsg_flags & NLM_F_CREATE) ||
3305                    info->nlh->nlmsg_flags & NLM_F_REPLACE)
3306                        return -EINVAL;
3307                handle = nf_tables_alloc_handle(table);
3308
3309                if (chain->use == UINT_MAX)
3310                        return -EOVERFLOW;
3311
3312                if (nla[NFTA_RULE_POSITION]) {
3313                        pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
3314                        old_rule = __nft_rule_lookup(chain, pos_handle);
3315                        if (IS_ERR(old_rule)) {
3316                                NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION]);
3317                                return PTR_ERR(old_rule);
3318                        }
3319                } else if (nla[NFTA_RULE_POSITION_ID]) {
3320                        old_rule = nft_rule_lookup_byid(net, nla[NFTA_RULE_POSITION_ID]);
3321                        if (IS_ERR(old_rule)) {
3322                                NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION_ID]);
3323                                return PTR_ERR(old_rule);
3324                        }
3325                }
3326        }
3327
3328        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3329
3330        n = 0;
3331        size = 0;
3332        if (nla[NFTA_RULE_EXPRESSIONS]) {
3333                expr_info = kvmalloc_array(NFT_RULE_MAXEXPRS,
3334                                           sizeof(struct nft_expr_info),
3335                                           GFP_KERNEL);
3336                if (!expr_info)
3337                        return -ENOMEM;
3338
3339                nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
3340                        err = -EINVAL;
3341                        if (nla_type(tmp) != NFTA_LIST_ELEM)
3342                                goto err_release_expr;
3343                        if (n == NFT_RULE_MAXEXPRS)
3344                                goto err_release_expr;
3345                        err = nf_tables_expr_parse(&ctx, tmp, &expr_info[n]);
3346                        if (err < 0) {
3347                                NL_SET_BAD_ATTR(extack, tmp);
3348                                goto err_release_expr;
3349                        }
3350                        size += expr_info[n].ops->size;
3351                        n++;
3352                }
3353        }
3354        /* Check for overflow of dlen field */
3355        err = -EFBIG;
3356        if (size >= 1 << 12)
3357                goto err_release_expr;
3358
3359        if (nla[NFTA_RULE_USERDATA]) {
3360                ulen = nla_len(nla[NFTA_RULE_USERDATA]);
3361                if (ulen > 0)
3362                        usize = sizeof(struct nft_userdata) + ulen;
3363        }
3364
3365        err = -ENOMEM;
3366        rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL);
3367        if (rule == NULL)
3368                goto err_release_expr;
3369
3370        nft_activate_next(net, rule);
3371
3372        rule->handle = handle;
3373        rule->dlen   = size;
3374        rule->udata  = ulen ? 1 : 0;
3375
3376        if (ulen) {
3377                udata = nft_userdata(rule);
3378                udata->len = ulen - 1;
3379                nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
3380        }
3381
3382        expr = nft_expr_first(rule);
3383        for (i = 0; i < n; i++) {
3384                err = nf_tables_newexpr(&ctx, &expr_info[i], expr);
3385                if (err < 0) {
3386                        NL_SET_BAD_ATTR(extack, expr_info[i].attr);
3387                        goto err_release_rule;
3388                }
3389
3390                if (expr_info[i].ops->validate)
3391                        nft_validate_state_update(net, NFT_VALIDATE_NEED);
3392
3393                expr_info[i].ops = NULL;
3394                expr = nft_expr_next(expr);
3395        }
3396
3397        if (chain->flags & NFT_CHAIN_HW_OFFLOAD) {
3398                flow = nft_flow_rule_create(net, rule);
3399                if (IS_ERR(flow)) {
3400                        err = PTR_ERR(flow);
3401                        goto err_release_rule;
3402                }
3403        }
3404
3405        if (info->nlh->nlmsg_flags & NLM_F_REPLACE) {
3406                trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3407                if (trans == NULL) {
3408                        err = -ENOMEM;
3409                        goto err_destroy_flow_rule;
3410                }
3411                err = nft_delrule(&ctx, old_rule);
3412                if (err < 0) {
3413                        nft_trans_destroy(trans);
3414                        goto err_destroy_flow_rule;
3415                }
3416
3417                list_add_tail_rcu(&rule->list, &old_rule->list);
3418        } else {
3419                trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3420                if (!trans) {
3421                        err = -ENOMEM;
3422                        goto err_destroy_flow_rule;
3423                }
3424
3425                if (info->nlh->nlmsg_flags & NLM_F_APPEND) {
3426                        if (old_rule)
3427                                list_add_rcu(&rule->list, &old_rule->list);
3428                        else
3429                                list_add_tail_rcu(&rule->list, &chain->rules);
3430                 } else {
3431                        if (old_rule)
3432                                list_add_tail_rcu(&rule->list, &old_rule->list);
3433                        else
3434                                list_add_rcu(&rule->list, &chain->rules);
3435                }
3436        }
3437        kvfree(expr_info);
3438        chain->use++;
3439
3440        if (flow)
3441                nft_trans_flow_rule(trans) = flow;
3442
3443        if (nft_net->validate_state == NFT_VALIDATE_DO)
3444                return nft_table_validate(net, table);
3445
3446        return 0;
3447
3448err_destroy_flow_rule:
3449        if (flow)
3450                nft_flow_rule_destroy(flow);
3451err_release_rule:
3452        nf_tables_rule_release(&ctx, rule);
3453err_release_expr:
3454        for (i = 0; i < n; i++) {
3455                if (expr_info[i].ops) {
3456                        module_put(expr_info[i].ops->type->owner);
3457                        if (expr_info[i].ops->type->release_ops)
3458                                expr_info[i].ops->type->release_ops(expr_info[i].ops);
3459                }
3460        }
3461        kvfree(expr_info);
3462
3463        return err;
3464}
3465
3466static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3467                                             const struct nlattr *nla)
3468{
3469        struct nftables_pernet *nft_net = nft_pernet(net);
3470        u32 id = ntohl(nla_get_be32(nla));
3471        struct nft_trans *trans;
3472
3473        list_for_each_entry(trans, &nft_net->commit_list, list) {
3474                struct nft_rule *rule = nft_trans_rule(trans);
3475
3476                if (trans->msg_type == NFT_MSG_NEWRULE &&
3477                    id == nft_trans_rule_id(trans))
3478                        return rule;
3479        }
3480        return ERR_PTR(-ENOENT);
3481}
3482
3483static int nf_tables_delrule(struct sk_buff *skb, const struct nfnl_info *info,
3484                             const struct nlattr * const nla[])
3485{
3486        struct netlink_ext_ack *extack = info->extack;
3487        u8 genmask = nft_genmask_next(info->net);
3488        u8 family = info->nfmsg->nfgen_family;
3489        struct nft_chain *chain = NULL;
3490        struct net *net = info->net;
3491        struct nft_table *table;
3492        struct nft_rule *rule;
3493        struct nft_ctx ctx;
3494        int err = 0;
3495
3496        table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3497                                 NETLINK_CB(skb).portid);
3498        if (IS_ERR(table)) {
3499                NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3500                return PTR_ERR(table);
3501        }
3502
3503        if (nla[NFTA_RULE_CHAIN]) {
3504                chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3505                                         genmask);
3506                if (IS_ERR(chain)) {
3507                        NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3508                        return PTR_ERR(chain);
3509                }
3510                if (nft_chain_is_bound(chain))
3511                        return -EOPNOTSUPP;
3512        }
3513
3514        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3515
3516        if (chain) {
3517                if (nla[NFTA_RULE_HANDLE]) {
3518                        rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3519                        if (IS_ERR(rule)) {
3520                                NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3521                                return PTR_ERR(rule);
3522                        }
3523
3524                        err = nft_delrule(&ctx, rule);
3525                } else if (nla[NFTA_RULE_ID]) {
3526                        rule = nft_rule_lookup_byid(net, nla[NFTA_RULE_ID]);
3527                        if (IS_ERR(rule)) {
3528                                NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_ID]);
3529                                return PTR_ERR(rule);
3530                        }
3531
3532                        err = nft_delrule(&ctx, rule);
3533                } else {
3534                        err = nft_delrule_by_chain(&ctx);
3535                }
3536        } else {
3537                list_for_each_entry(chain, &table->chains, list) {
3538                        if (!nft_is_active_next(net, chain))
3539                                continue;
3540
3541                        ctx.chain = chain;
3542                        err = nft_delrule_by_chain(&ctx);
3543                        if (err < 0)
3544                                break;
3545                }
3546        }
3547
3548        return err;
3549}
3550
3551/*
3552 * Sets
3553 */
3554static const struct nft_set_type *nft_set_types[] = {
3555        &nft_set_hash_fast_type,
3556        &nft_set_hash_type,
3557        &nft_set_rhash_type,
3558        &nft_set_bitmap_type,
3559        &nft_set_rbtree_type,
3560#if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
3561        &nft_set_pipapo_avx2_type,
3562#endif
3563        &nft_set_pipapo_type,
3564};
3565
3566#define NFT_SET_FEATURES        (NFT_SET_INTERVAL | NFT_SET_MAP | \
3567                                 NFT_SET_TIMEOUT | NFT_SET_OBJECT | \
3568                                 NFT_SET_EVAL)
3569
3570static bool nft_set_ops_candidate(const struct nft_set_type *type, u32 flags)
3571{
3572        return (flags & type->features) == (flags & NFT_SET_FEATURES);
3573}
3574
3575/*
3576 * Select a set implementation based on the data characteristics and the
3577 * given policy. The total memory use might not be known if no size is
3578 * given, in that case the amount of memory per element is used.
3579 */
3580static const struct nft_set_ops *
3581nft_select_set_ops(const struct nft_ctx *ctx,
3582                   const struct nlattr * const nla[],
3583                   const struct nft_set_desc *desc,
3584                   enum nft_set_policies policy)
3585{
3586        struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3587        const struct nft_set_ops *ops, *bops;
3588        struct nft_set_estimate est, best;
3589        const struct nft_set_type *type;
3590        u32 flags = 0;
3591        int i;
3592
3593        lockdep_assert_held(&nft_net->commit_mutex);
3594        lockdep_nfnl_nft_mutex_not_held();
3595
3596        if (nla[NFTA_SET_FLAGS] != NULL)
3597                flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
3598
3599        bops        = NULL;
3600        best.size   = ~0;
3601        best.lookup = ~0;
3602        best.space  = ~0;
3603
3604        for (i = 0; i < ARRAY_SIZE(nft_set_types); i++) {
3605                type = nft_set_types[i];
3606                ops = &type->ops;
3607
3608                if (!nft_set_ops_candidate(type, flags))
3609                        continue;
3610                if (!ops->estimate(desc, flags, &est))
3611                        continue;
3612
3613                switch (policy) {
3614                case NFT_SET_POL_PERFORMANCE:
3615                        if (est.lookup < best.lookup)
3616                                break;
3617                        if (est.lookup == best.lookup &&
3618                            est.space < best.space)
3619                                break;
3620                        continue;
3621                case NFT_SET_POL_MEMORY:
3622                        if (!desc->size) {
3623                                if (est.space < best.space)
3624                                        break;
3625                                if (est.space == best.space &&
3626                                    est.lookup < best.lookup)
3627                                        break;
3628                        } else if (est.size < best.size || !bops) {
3629                                break;
3630                        }
3631                        continue;
3632                default:
3633                        break;
3634                }
3635
3636                bops = ops;
3637                best = est;
3638        }
3639
3640        if (bops != NULL)
3641                return bops;
3642
3643        return ERR_PTR(-EOPNOTSUPP);
3644}
3645
3646static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
3647        [NFTA_SET_TABLE]                = { .type = NLA_STRING,
3648                                            .len = NFT_TABLE_MAXNAMELEN - 1 },
3649        [NFTA_SET_NAME]                 = { .type = NLA_STRING,
3650                                            .len = NFT_SET_MAXNAMELEN - 1 },
3651        [NFTA_SET_FLAGS]                = { .type = NLA_U32 },
3652        [NFTA_SET_KEY_TYPE]             = { .type = NLA_U32 },
3653        [NFTA_SET_KEY_LEN]              = { .type = NLA_U32 },
3654        [NFTA_SET_DATA_TYPE]            = { .type = NLA_U32 },
3655        [NFTA_SET_DATA_LEN]             = { .type = NLA_U32 },
3656        [NFTA_SET_POLICY]               = { .type = NLA_U32 },
3657        [NFTA_SET_DESC]                 = { .type = NLA_NESTED },
3658        [NFTA_SET_ID]                   = { .type = NLA_U32 },
3659        [NFTA_SET_TIMEOUT]              = { .type = NLA_U64 },
3660        [NFTA_SET_GC_INTERVAL]          = { .type = NLA_U32 },
3661        [NFTA_SET_USERDATA]             = { .type = NLA_BINARY,
3662                                            .len  = NFT_USERDATA_MAXLEN },
3663        [NFTA_SET_OBJ_TYPE]             = { .type = NLA_U32 },
3664        [NFTA_SET_HANDLE]               = { .type = NLA_U64 },
3665        [NFTA_SET_EXPR]                 = { .type = NLA_NESTED },
3666        [NFTA_SET_EXPRESSIONS]          = { .type = NLA_NESTED },
3667};
3668
3669static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
3670        [NFTA_SET_DESC_SIZE]            = { .type = NLA_U32 },
3671        [NFTA_SET_DESC_CONCAT]          = { .type = NLA_NESTED },
3672};
3673
3674static struct nft_set *nft_set_lookup(const struct nft_table *table,
3675                                      const struct nlattr *nla, u8 genmask)
3676{
3677        struct nft_set *set;
3678
3679        if (nla == NULL)
3680                return ERR_PTR(-EINVAL);
3681
3682        list_for_each_entry_rcu(set, &table->sets, list) {
3683                if (!nla_strcmp(nla, set->name) &&
3684                    nft_active_genmask(set, genmask))
3685                        return set;
3686        }
3687        return ERR_PTR(-ENOENT);
3688}
3689
3690static struct nft_set *nft_set_lookup_byhandle(const struct nft_table *table,
3691                                               const struct nlattr *nla,
3692                                               u8 genmask)
3693{
3694        struct nft_set *set;
3695
3696        list_for_each_entry(set, &table->sets, list) {
3697                if (be64_to_cpu(nla_get_be64(nla)) == set->handle &&
3698                    nft_active_genmask(set, genmask))
3699                        return set;
3700        }
3701        return ERR_PTR(-ENOENT);
3702}
3703
3704static struct nft_set *nft_set_lookup_byid(const struct net *net,
3705                                           const struct nlattr *nla, u8 genmask)
3706{
3707        struct nftables_pernet *nft_net = nft_pernet(net);
3708        u32 id = ntohl(nla_get_be32(nla));
3709        struct nft_trans *trans;
3710
3711        list_for_each_entry(trans, &nft_net->commit_list, list) {
3712                if (trans->msg_type == NFT_MSG_NEWSET) {
3713                        struct nft_set *set = nft_trans_set(trans);
3714
3715                        if (id == nft_trans_set_id(trans) &&
3716                            nft_active_genmask(set, genmask))
3717                                return set;
3718                }
3719        }
3720        return ERR_PTR(-ENOENT);
3721}
3722
3723struct nft_set *nft_set_lookup_global(const struct net *net,
3724                                      const struct nft_table *table,
3725                                      const struct nlattr *nla_set_name,
3726                                      const struct nlattr *nla_set_id,
3727                                      u8 genmask)
3728{
3729        struct nft_set *set;
3730
3731        set = nft_set_lookup(table, nla_set_name, genmask);
3732        if (IS_ERR(set)) {
3733                if (!nla_set_id)
3734                        return set;
3735
3736                set = nft_set_lookup_byid(net, nla_set_id, genmask);
3737        }
3738        return set;
3739}
3740EXPORT_SYMBOL_GPL(nft_set_lookup_global);
3741
3742static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
3743                                    const char *name)
3744{
3745        const struct nft_set *i;
3746        const char *p;
3747        unsigned long *inuse;
3748        unsigned int n = 0, min = 0;
3749
3750        p = strchr(name, '%');
3751        if (p != NULL) {
3752                if (p[1] != 'd' || strchr(p + 2, '%'))
3753                        return -EINVAL;
3754
3755                inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
3756                if (inuse == NULL)
3757                        return -ENOMEM;
3758cont:
3759                list_for_each_entry(i, &ctx->table->sets, list) {
3760                        int tmp;
3761
3762                        if (!nft_is_active_next(ctx->net, set))
3763                                continue;
3764                        if (!sscanf(i->name, name, &tmp))
3765                                continue;
3766                        if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
3767                                continue;
3768
3769                        set_bit(tmp - min, inuse);
3770                }
3771
3772                n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
3773                if (n >= BITS_PER_BYTE * PAGE_SIZE) {
3774                        min += BITS_PER_BYTE * PAGE_SIZE;
3775                        memset(inuse, 0, PAGE_SIZE);
3776                        goto cont;
3777                }
3778                free_page((unsigned long)inuse);
3779        }
3780
3781        set->name = kasprintf(GFP_KERNEL, name, min + n);
3782        if (!set->name)
3783                return -ENOMEM;
3784
3785        list_for_each_entry(i, &ctx->table->sets, list) {
3786                if (!nft_is_active_next(ctx->net, i))
3787                        continue;
3788                if (!strcmp(set->name, i->name)) {
3789                        kfree(set->name);
3790                        set->name = NULL;
3791                        return -ENFILE;
3792                }
3793        }
3794        return 0;
3795}
3796
3797int nf_msecs_to_jiffies64(const struct nlattr *nla, u64 *result)
3798{
3799        u64 ms = be64_to_cpu(nla_get_be64(nla));
3800        u64 max = (u64)(~((u64)0));
3801
3802        max = div_u64(max, NSEC_PER_MSEC);
3803        if (ms >= max)
3804                return -ERANGE;
3805
3806        ms *= NSEC_PER_MSEC;
3807        *result = nsecs_to_jiffies64(ms);
3808        return 0;
3809}
3810
3811__be64 nf_jiffies64_to_msecs(u64 input)
3812{
3813        return cpu_to_be64(jiffies64_to_msecs(input));
3814}
3815
3816static int nf_tables_fill_set_concat(struct sk_buff *skb,
3817                                     const struct nft_set *set)
3818{
3819        struct nlattr *concat, *field;
3820        int i;
3821
3822        concat = nla_nest_start_noflag(skb, NFTA_SET_DESC_CONCAT);
3823        if (!concat)
3824                return -ENOMEM;
3825
3826        for (i = 0; i < set->field_count; i++) {
3827                field = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
3828                if (!field)
3829                        return -ENOMEM;
3830
3831                if (nla_put_be32(skb, NFTA_SET_FIELD_LEN,
3832                                 htonl(set->field_len[i])))
3833                        return -ENOMEM;
3834
3835                nla_nest_end(skb, field);
3836        }
3837
3838        nla_nest_end(skb, concat);
3839
3840        return 0;
3841}
3842
3843static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
3844                              const struct nft_set *set, u16 event, u16 flags)
3845{
3846        struct nlmsghdr *nlh;
3847        u32 portid = ctx->portid;
3848        struct nlattr *nest;
3849        u32 seq = ctx->seq;
3850        int i;
3851
3852        event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
3853        nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
3854                           NFNETLINK_V0, nft_base_seq(ctx->net));
3855        if (!nlh)
3856                goto nla_put_failure;
3857
3858        if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
3859                goto nla_put_failure;
3860        if (nla_put_string(skb, NFTA_SET_NAME, set->name))
3861                goto nla_put_failure;
3862        if (nla_put_be64(skb, NFTA_SET_HANDLE, cpu_to_be64(set->handle),
3863                         NFTA_SET_PAD))
3864                goto nla_put_failure;
3865        if (set->flags != 0)
3866                if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
3867                        goto nla_put_failure;
3868
3869        if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
3870                goto nla_put_failure;
3871        if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
3872                goto nla_put_failure;
3873        if (set->flags & NFT_SET_MAP) {
3874                if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
3875                        goto nla_put_failure;
3876                if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
3877                        goto nla_put_failure;
3878        }
3879        if (set->flags & NFT_SET_OBJECT &&
3880            nla_put_be32(skb, NFTA_SET_OBJ_TYPE, htonl(set->objtype)))
3881                goto nla_put_failure;
3882
3883        if (set->timeout &&
3884            nla_put_be64(skb, NFTA_SET_TIMEOUT,
3885                         nf_jiffies64_to_msecs(set->timeout),
3886                         NFTA_SET_PAD))
3887                goto nla_put_failure;
3888        if (set->gc_int &&
3889            nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(set->gc_int)))
3890                goto nla_put_failure;
3891
3892        if (set->policy != NFT_SET_POL_PERFORMANCE) {
3893                if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
3894                        goto nla_put_failure;
3895        }
3896
3897        if (set->udata &&
3898            nla_put(skb, NFTA_SET_USERDATA, set->udlen, set->udata))
3899                goto nla_put_failure;
3900
3901        nest = nla_nest_start_noflag(skb, NFTA_SET_DESC);
3902        if (!nest)
3903                goto nla_put_failure;
3904        if (set->size &&
3905            nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
3906                goto nla_put_failure;
3907
3908        if (set->field_count > 1 &&
3909            nf_tables_fill_set_concat(skb, set))
3910                goto nla_put_failure;
3911
3912        nla_nest_end(skb, nest);
3913
3914        if (set->num_exprs == 1) {
3915                nest = nla_nest_start_noflag(skb, NFTA_SET_EXPR);
3916                if (nf_tables_fill_expr_info(skb, set->exprs[0]) < 0)
3917                        goto nla_put_failure;
3918
3919                nla_nest_end(skb, nest);
3920        } else if (set->num_exprs > 1) {
3921                nest = nla_nest_start_noflag(skb, NFTA_SET_EXPRESSIONS);
3922                if (nest == NULL)
3923                        goto nla_put_failure;
3924
3925                for (i = 0; i < set->num_exprs; i++) {
3926                        if (nft_expr_dump(skb, NFTA_LIST_ELEM,
3927                                          set->exprs[i]) < 0)
3928                                goto nla_put_failure;
3929                }
3930                nla_nest_end(skb, nest);
3931        }
3932
3933        nlmsg_end(skb, nlh);
3934        return 0;
3935
3936nla_put_failure:
3937        nlmsg_trim(skb, nlh);
3938        return -1;
3939}
3940
3941static void nf_tables_set_notify(const struct nft_ctx *ctx,
3942                                 const struct nft_set *set, int event,
3943                                 gfp_t gfp_flags)
3944{
3945        struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3946        struct sk_buff *skb;
3947        u32 portid = ctx->portid;
3948        int err;
3949
3950        if (!ctx->report &&
3951            !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
3952                return;
3953
3954        skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
3955        if (skb == NULL)
3956                goto err;
3957
3958        err = nf_tables_fill_set(skb, ctx, set, event, 0);
3959        if (err < 0) {
3960                kfree_skb(skb);
3961                goto err;
3962        }
3963
3964        nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
3965        return;
3966err:
3967        nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
3968}
3969
3970static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
3971{
3972        const struct nft_set *set;
3973        unsigned int idx, s_idx = cb->args[0];
3974        struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
3975        struct net *net = sock_net(skb->sk);
3976        struct nft_ctx *ctx = cb->data, ctx_set;
3977        struct nftables_pernet *nft_net;
3978
3979        if (cb->args[1])
3980                return skb->len;
3981
3982        rcu_read_lock();
3983        nft_net = nft_pernet(net);
3984        cb->seq = nft_net->base_seq;
3985
3986        list_for_each_entry_rcu(table, &nft_net->tables, list) {
3987                if (ctx->family != NFPROTO_UNSPEC &&
3988                    ctx->family != table->family)
3989                        continue;
3990
3991                if (ctx->table && ctx->table != table)
3992                        continue;
3993
3994                if (cur_table) {
3995                        if (cur_table != table)
3996                                continue;
3997
3998                        cur_table = NULL;
3999                }
4000                idx = 0;
4001                list_for_each_entry_rcu(set, &table->sets, list) {
4002                        if (idx < s_idx)
4003                                goto cont;
4004                        if (!nft_is_active(net, set))
4005                                goto cont;
4006
4007                        ctx_set = *ctx;
4008                        ctx_set.table = table;
4009                        ctx_set.family = table->family;
4010
4011                        if (nf_tables_fill_set(skb, &ctx_set, set,
4012                                               NFT_MSG_NEWSET,
4013                                               NLM_F_MULTI) < 0) {
4014                                cb->args[0] = idx;
4015                                cb->args[2] = (unsigned long) table;
4016                                goto done;
4017                        }
4018                        nl_dump_check_consistent(cb, nlmsg_hdr(skb));
4019cont:
4020                        idx++;
4021                }
4022                if (s_idx)
4023                        s_idx = 0;
4024        }
4025        cb->args[1] = 1;
4026done:
4027        rcu_read_unlock();
4028        return skb->len;
4029}
4030
4031static int nf_tables_dump_sets_start(struct netlink_callback *cb)
4032{
4033        struct nft_ctx *ctx_dump = NULL;
4034
4035        ctx_dump = kmemdup(cb->data, sizeof(*ctx_dump), GFP_ATOMIC);
4036        if (ctx_dump == NULL)
4037                return -ENOMEM;
4038
4039        cb->data = ctx_dump;
4040        return 0;
4041}
4042
4043static int nf_tables_dump_sets_done(struct netlink_callback *cb)
4044{
4045        kfree(cb->data);
4046        return 0;
4047}
4048
4049/* called with rcu_read_lock held */
4050static int nf_tables_getset(struct sk_buff *skb, const struct nfnl_info *info,
4051                            const struct nlattr * const nla[])
4052{
4053        struct netlink_ext_ack *extack = info->extack;
4054        u8 genmask = nft_genmask_cur(info->net);
4055        u8 family = info->nfmsg->nfgen_family;
4056        struct nft_table *table = NULL;
4057        struct net *net = info->net;
4058        const struct nft_set *set;
4059        struct sk_buff *skb2;
4060        struct nft_ctx ctx;
4061        int err;
4062
4063        if (nla[NFTA_SET_TABLE]) {
4064                table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4065                                         genmask, 0);
4066                if (IS_ERR(table)) {
4067                        NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4068                        return PTR_ERR(table);
4069                }
4070        }
4071
4072        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4073
4074        if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
4075                struct netlink_dump_control c = {
4076                        .start = nf_tables_dump_sets_start,
4077                        .dump = nf_tables_dump_sets,
4078                        .done = nf_tables_dump_sets_done,
4079                        .data = &ctx,
4080                        .module = THIS_MODULE,
4081                };
4082
4083                return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
4084        }
4085
4086        /* Only accept unspec with dump */
4087        if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4088                return -EAFNOSUPPORT;
4089        if (!nla[NFTA_SET_TABLE])
4090                return -EINVAL;
4091
4092        set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4093        if (IS_ERR(set))
4094                return PTR_ERR(set);
4095
4096        skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
4097        if (skb2 == NULL)
4098                return -ENOMEM;
4099
4100        err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
4101        if (err < 0)
4102                goto err_fill_set_info;
4103
4104        return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
4105
4106err_fill_set_info:
4107        kfree_skb(skb2);
4108        return err;
4109}
4110
4111static const struct nla_policy nft_concat_policy[NFTA_SET_FIELD_MAX + 1] = {
4112        [NFTA_SET_FIELD_LEN]    = { .type = NLA_U32 },
4113};
4114
4115static int nft_set_desc_concat_parse(const struct nlattr *attr,
4116                                     struct nft_set_desc *desc)
4117{
4118        struct nlattr *tb[NFTA_SET_FIELD_MAX + 1];
4119        u32 len;
4120        int err;
4121
4122        err = nla_parse_nested_deprecated(tb, NFTA_SET_FIELD_MAX, attr,
4123                                          nft_concat_policy, NULL);
4124        if (err < 0)
4125                return err;
4126
4127        if (!tb[NFTA_SET_FIELD_LEN])
4128                return -EINVAL;
4129
4130        len = ntohl(nla_get_be32(tb[NFTA_SET_FIELD_LEN]));
4131
4132        if (len * BITS_PER_BYTE / 32 > NFT_REG32_COUNT)
4133                return -E2BIG;
4134
4135        desc->field_len[desc->field_count++] = len;
4136
4137        return 0;
4138}
4139
4140static int nft_set_desc_concat(struct nft_set_desc *desc,
4141                               const struct nlattr *nla)
4142{
4143        struct nlattr *attr;
4144        int rem, err;
4145
4146        nla_for_each_nested(attr, nla, rem) {
4147                if (nla_type(attr) != NFTA_LIST_ELEM)
4148                        return -EINVAL;
4149
4150                err = nft_set_desc_concat_parse(attr, desc);
4151                if (err < 0)
4152                        return err;
4153        }
4154
4155        return 0;
4156}
4157
4158static int nf_tables_set_desc_parse(struct nft_set_desc *desc,
4159                                    const struct nlattr *nla)
4160{
4161        struct nlattr *da[NFTA_SET_DESC_MAX + 1];
4162        int err;
4163
4164        err = nla_parse_nested_deprecated(da, NFTA_SET_DESC_MAX, nla,
4165                                          nft_set_desc_policy, NULL);
4166        if (err < 0)
4167                return err;
4168
4169        if (da[NFTA_SET_DESC_SIZE] != NULL)
4170                desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
4171        if (da[NFTA_SET_DESC_CONCAT])
4172                err = nft_set_desc_concat(desc, da[NFTA_SET_DESC_CONCAT]);
4173
4174        return err;
4175}
4176
4177static int nf_tables_newset(struct sk_buff *skb, const struct nfnl_info *info,
4178                            const struct nlattr * const nla[])
4179{
4180        u32 ktype, dtype, flags, policy, gc_int, objtype;
4181        struct netlink_ext_ack *extack = info->extack;
4182        u8 genmask = nft_genmask_next(info->net);
4183        u8 family = info->nfmsg->nfgen_family;
4184        const struct nft_set_ops *ops;
4185        struct nft_expr *expr = NULL;
4186        struct net *net = info->net;
4187        struct nft_set_desc desc;
4188        struct nft_table *table;
4189        unsigned char *udata;
4190        struct nft_set *set;
4191        struct nft_ctx ctx;
4192        size_t alloc_size;
4193        u64 timeout;
4194        char *name;
4195        int err, i;
4196        u16 udlen;
4197        u64 size;
4198
4199        if (nla[NFTA_SET_TABLE] == NULL ||
4200            nla[NFTA_SET_NAME] == NULL ||
4201            nla[NFTA_SET_KEY_LEN] == NULL ||
4202            nla[NFTA_SET_ID] == NULL)
4203                return -EINVAL;
4204
4205        memset(&desc, 0, sizeof(desc));
4206
4207        ktype = NFT_DATA_VALUE;
4208        if (nla[NFTA_SET_KEY_TYPE] != NULL) {
4209                ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
4210                if ((ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
4211                        return -EINVAL;
4212        }
4213
4214        desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
4215        if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
4216                return -EINVAL;
4217
4218        flags = 0;
4219        if (nla[NFTA_SET_FLAGS] != NULL) {
4220                flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
4221                if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
4222                              NFT_SET_INTERVAL | NFT_SET_TIMEOUT |
4223                              NFT_SET_MAP | NFT_SET_EVAL |
4224                              NFT_SET_OBJECT | NFT_SET_CONCAT | NFT_SET_EXPR))
4225                        return -EOPNOTSUPP;
4226                /* Only one of these operations is supported */
4227                if ((flags & (NFT_SET_MAP | NFT_SET_OBJECT)) ==
4228                             (NFT_SET_MAP | NFT_SET_OBJECT))
4229                        return -EOPNOTSUPP;
4230                if ((flags & (NFT_SET_EVAL | NFT_SET_OBJECT)) ==
4231                             (NFT_SET_EVAL | NFT_SET_OBJECT))
4232                        return -EOPNOTSUPP;
4233        }
4234
4235        dtype = 0;
4236        if (nla[NFTA_SET_DATA_TYPE] != NULL) {
4237                if (!(flags & NFT_SET_MAP))
4238                        return -EINVAL;
4239
4240                dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
4241                if ((dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
4242                    dtype != NFT_DATA_VERDICT)
4243                        return -EINVAL;
4244
4245                if (dtype != NFT_DATA_VERDICT) {
4246                        if (nla[NFTA_SET_DATA_LEN] == NULL)
4247                                return -EINVAL;
4248                        desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
4249                        if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
4250                                return -EINVAL;
4251                } else
4252                        desc.dlen = sizeof(struct nft_verdict);
4253        } else if (flags & NFT_SET_MAP)
4254                return -EINVAL;
4255
4256        if (nla[NFTA_SET_OBJ_TYPE] != NULL) {
4257                if (!(flags & NFT_SET_OBJECT))
4258                        return -EINVAL;
4259
4260                objtype = ntohl(nla_get_be32(nla[NFTA_SET_OBJ_TYPE]));
4261                if (objtype == NFT_OBJECT_UNSPEC ||
4262                    objtype > NFT_OBJECT_MAX)
4263                        return -EOPNOTSUPP;
4264        } else if (flags & NFT_SET_OBJECT)
4265                return -EINVAL;
4266        else
4267                objtype = NFT_OBJECT_UNSPEC;
4268
4269        timeout = 0;
4270        if (nla[NFTA_SET_TIMEOUT] != NULL) {
4271                if (!(flags & NFT_SET_TIMEOUT))
4272                        return -EINVAL;
4273
4274                err = nf_msecs_to_jiffies64(nla[NFTA_SET_TIMEOUT], &timeout);
4275                if (err)
4276                        return err;
4277        }
4278        gc_int = 0;
4279        if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
4280                if (!(flags & NFT_SET_TIMEOUT))
4281                        return -EINVAL;
4282                gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
4283        }
4284
4285        policy = NFT_SET_POL_PERFORMANCE;
4286        if (nla[NFTA_SET_POLICY] != NULL)
4287                policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
4288
4289        if (nla[NFTA_SET_DESC] != NULL) {
4290                err = nf_tables_set_desc_parse(&desc, nla[NFTA_SET_DESC]);
4291                if (err < 0)
4292                        return err;
4293        }
4294
4295        if (nla[NFTA_SET_EXPR] || nla[NFTA_SET_EXPRESSIONS])
4296                desc.expr = true;
4297
4298        table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family, genmask,
4299                                 NETLINK_CB(skb).portid);
4300        if (IS_ERR(table)) {
4301                NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4302                return PTR_ERR(table);
4303        }
4304
4305        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4306
4307        set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4308        if (IS_ERR(set)) {
4309                if (PTR_ERR(set) != -ENOENT) {
4310                        NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4311                        return PTR_ERR(set);
4312                }
4313        } else {
4314                if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
4315                        NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4316                        return -EEXIST;
4317                }
4318                if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
4319                        return -EOPNOTSUPP;
4320
4321                return 0;
4322        }
4323
4324        if (!(info->nlh->nlmsg_flags & NLM_F_CREATE))
4325                return -ENOENT;
4326
4327        ops = nft_select_set_ops(&ctx, nla, &desc, policy);
4328        if (IS_ERR(ops))
4329                return PTR_ERR(ops);
4330
4331        udlen = 0;
4332        if (nla[NFTA_SET_USERDATA])
4333                udlen = nla_len(nla[NFTA_SET_USERDATA]);
4334
4335        size = 0;
4336        if (ops->privsize != NULL)
4337                size = ops->privsize(nla, &desc);
4338        alloc_size = sizeof(*set) + size + udlen;
4339        if (alloc_size < size)
4340                return -ENOMEM;
4341        set = kvzalloc(alloc_size, GFP_KERNEL);
4342        if (!set)
4343                return -ENOMEM;
4344
4345        name = nla_strdup(nla[NFTA_SET_NAME], GFP_KERNEL);
4346        if (!name) {
4347                err = -ENOMEM;
4348                goto err_set_name;
4349        }
4350
4351        err = nf_tables_set_alloc_name(&ctx, set, name);
4352        kfree(name);
4353        if (err < 0)
4354                goto err_set_name;
4355
4356        udata = NULL;
4357        if (udlen) {
4358                udata = set->data + size;
4359                nla_memcpy(udata, nla[NFTA_SET_USERDATA], udlen);
4360        }
4361
4362        INIT_LIST_HEAD(&set->bindings);
4363        INIT_LIST_HEAD(&set->catchall_list);
4364        set->table = table;
4365        write_pnet(&set->net, net);
4366        set->ops = ops;
4367        set->ktype = ktype;
4368        set->klen = desc.klen;
4369        set->dtype = dtype;
4370        set->objtype = objtype;
4371        set->dlen = desc.dlen;
4372        set->flags = flags;
4373        set->size = desc.size;
4374        set->policy = policy;
4375        set->udlen = udlen;
4376        set->udata = udata;
4377        set->timeout = timeout;
4378        set->gc_int = gc_int;
4379
4380        set->field_count = desc.field_count;
4381        for (i = 0; i < desc.field_count; i++)
4382                set->field_len[i] = desc.field_len[i];
4383
4384        err = ops->init(set, &desc, nla);
4385        if (err < 0)
4386                goto err_set_init;
4387
4388        if (nla[NFTA_SET_EXPR]) {
4389                expr = nft_set_elem_expr_alloc(&ctx, set, nla[NFTA_SET_EXPR]);
4390                if (IS_ERR(expr)) {
4391                        err = PTR_ERR(expr);
4392                        goto err_set_expr_alloc;
4393                }
4394                set->exprs[0] = expr;
4395                set->num_exprs++;
4396        } else if (nla[NFTA_SET_EXPRESSIONS]) {
4397                struct nft_expr *expr;
4398                struct nlattr *tmp;
4399                int left;
4400
4401                if (!(flags & NFT_SET_EXPR)) {
4402                        err = -EINVAL;
4403                        goto err_set_expr_alloc;
4404                }
4405                i = 0;
4406                nla_for_each_nested(tmp, nla[NFTA_SET_EXPRESSIONS], left) {
4407                        if (i == NFT_SET_EXPR_MAX) {
4408                                err = -E2BIG;
4409                                goto err_set_expr_alloc;
4410                        }
4411                        if (nla_type(tmp) != NFTA_LIST_ELEM) {
4412                                err = -EINVAL;
4413                                goto err_set_expr_alloc;
4414                        }
4415                        expr = nft_set_elem_expr_alloc(&ctx, set, tmp);
4416                        if (IS_ERR(expr)) {
4417                                err = PTR_ERR(expr);
4418                                goto err_set_expr_alloc;
4419                        }
4420                        set->exprs[i++] = expr;
4421                        set->num_exprs++;
4422                }
4423        }
4424
4425        set->handle = nf_tables_alloc_handle(table);
4426
4427        err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
4428        if (err < 0)
4429                goto err_set_expr_alloc;
4430
4431        list_add_tail_rcu(&set->list, &table->sets);
4432        table->use++;
4433        return 0;
4434
4435err_set_expr_alloc:
4436        for (i = 0; i < set->num_exprs; i++)
4437                nft_expr_destroy(&ctx, set->exprs[i]);
4438
4439        ops->destroy(set);
4440err_set_init:
4441        kfree(set->name);
4442err_set_name:
4443        kvfree(set);
4444        return err;
4445}
4446
4447struct nft_set_elem_catchall {
4448        struct list_head        list;
4449        struct rcu_head         rcu;
4450        void                    *elem;
4451};
4452
4453static void nft_set_catchall_destroy(const struct nft_ctx *ctx,
4454                                     struct nft_set *set)
4455{
4456        struct nft_set_elem_catchall *catchall;
4457
4458        list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4459                list_del_rcu(&catchall->list);
4460                nft_set_elem_destroy(set, catchall->elem, true);
4461                kfree_rcu(catchall);
4462        }
4463}
4464
4465static void nft_set_destroy(const struct nft_ctx *ctx, struct nft_set *set)
4466{
4467        int i;
4468
4469        if (WARN_ON(set->use > 0))
4470                return;
4471
4472        for (i = 0; i < set->num_exprs; i++)
4473                nft_expr_destroy(ctx, set->exprs[i]);
4474
4475        set->ops->destroy(set);
4476        nft_set_catchall_destroy(ctx, set);
4477        kfree(set->name);
4478        kvfree(set);
4479}
4480
4481static int nf_tables_delset(struct sk_buff *skb, const struct nfnl_info *info,
4482                            const struct nlattr * const nla[])
4483{
4484        struct netlink_ext_ack *extack = info->extack;
4485        u8 genmask = nft_genmask_next(info->net);
4486        u8 family = info->nfmsg->nfgen_family;
4487        struct net *net = info->net;
4488        const struct nlattr *attr;
4489        struct nft_table *table;
4490        struct nft_set *set;
4491        struct nft_ctx ctx;
4492
4493        if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4494                return -EAFNOSUPPORT;
4495
4496        table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4497                                 genmask, NETLINK_CB(skb).portid);
4498        if (IS_ERR(table)) {
4499                NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4500                return PTR_ERR(table);
4501        }
4502
4503        if (nla[NFTA_SET_HANDLE]) {
4504                attr = nla[NFTA_SET_HANDLE];
4505                set = nft_set_lookup_byhandle(table, attr, genmask);
4506        } else {
4507                attr = nla[NFTA_SET_NAME];
4508                set = nft_set_lookup(table, attr, genmask);
4509        }
4510
4511        if (IS_ERR(set)) {
4512                NL_SET_BAD_ATTR(extack, attr);
4513                return PTR_ERR(set);
4514        }
4515        if (set->use ||
4516            (info->nlh->nlmsg_flags & NLM_F_NONREC &&
4517             atomic_read(&set->nelems) > 0)) {
4518                NL_SET_BAD_ATTR(extack, attr);
4519                return -EBUSY;
4520        }
4521
4522        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4523
4524        return nft_delset(&ctx, set);
4525}
4526
4527static int nft_validate_register_store(const struct nft_ctx *ctx,
4528                                       enum nft_registers reg,
4529                                       const struct nft_data *data,
4530                                       enum nft_data_types type,
4531                                       unsigned int len);
4532
4533static int nft_setelem_data_validate(const struct nft_ctx *ctx,
4534                                     struct nft_set *set,
4535                                     struct nft_set_elem *elem)
4536{
4537        const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4538        enum nft_registers dreg;
4539
4540        dreg = nft_type_to_reg(set->dtype);
4541        return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
4542                                           set->dtype == NFT_DATA_VERDICT ?
4543                                           NFT_DATA_VERDICT : NFT_DATA_VALUE,
4544                                           set->dlen);
4545}
4546
4547static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
4548                                        struct nft_set *set,
4549                                        const struct nft_set_iter *iter,
4550                                        struct nft_set_elem *elem)
4551{
4552        return nft_setelem_data_validate(ctx, set, elem);
4553}
4554
4555static int nft_set_catchall_bind_check(const struct nft_ctx *ctx,
4556                                       struct nft_set *set)
4557{
4558        u8 genmask = nft_genmask_next(ctx->net);
4559        struct nft_set_elem_catchall *catchall;
4560        struct nft_set_elem elem;
4561        struct nft_set_ext *ext;
4562        int ret = 0;
4563
4564        list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4565                ext = nft_set_elem_ext(set, catchall->elem);
4566                if (!nft_set_elem_active(ext, genmask))
4567                        continue;
4568
4569                elem.priv = catchall->elem;
4570                ret = nft_setelem_data_validate(ctx, set, &elem);
4571                if (ret < 0)
4572                        break;
4573        }
4574
4575        return ret;
4576}
4577
4578int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
4579                       struct nft_set_binding *binding)
4580{
4581        struct nft_set_binding *i;
4582        struct nft_set_iter iter;
4583
4584        if (set->use == UINT_MAX)
4585                return -EOVERFLOW;
4586
4587        if (!list_empty(&set->bindings) && nft_set_is_anonymous(set))
4588                return -EBUSY;
4589
4590        if (binding->flags & NFT_SET_MAP) {
4591                /* If the set is already bound to the same chain all
4592                 * jumps are already validated for that chain.
4593                 */
4594                list_for_each_entry(i, &set->bindings, list) {
4595                        if (i->flags & NFT_SET_MAP &&
4596                            i->chain == binding->chain)
4597                                goto bind;
4598                }
4599
4600                iter.genmask    = nft_genmask_next(ctx->net);
4601                iter.skip       = 0;
4602                iter.count      = 0;
4603                iter.err        = 0;
4604                iter.fn         = nf_tables_bind_check_setelem;
4605
4606                set->ops->walk(ctx, set, &iter);
4607                if (!iter.err)
4608                        iter.err = nft_set_catchall_bind_check(ctx, set);
4609
4610                if (iter.err < 0)
4611                        return iter.err;
4612        }
4613bind:
4614        binding->chain = ctx->chain;
4615        list_add_tail_rcu(&binding->list, &set->bindings);
4616        nft_set_trans_bind(ctx, set);
4617        set->use++;
4618
4619        return 0;
4620}
4621EXPORT_SYMBOL_GPL(nf_tables_bind_set);
4622
4623static void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
4624                                 struct nft_set_binding *binding, bool event)
4625{
4626        list_del_rcu(&binding->list);
4627
4628        if (list_empty(&set->bindings) && nft_set_is_anonymous(set)) {
4629                list_del_rcu(&set->list);
4630                if (event)
4631                        nf_tables_set_notify(ctx, set, NFT_MSG_DELSET,
4632                                             GFP_KERNEL);
4633        }
4634}
4635
4636void nf_tables_deactivate_set(const struct nft_ctx *ctx, struct nft_set *set,
4637                              struct nft_set_binding *binding,
4638                              enum nft_trans_phase phase)
4639{
4640        switch (phase) {
4641        case NFT_TRANS_PREPARE:
4642                set->use--;
4643                return;
4644        case NFT_TRANS_ABORT:
4645        case NFT_TRANS_RELEASE:
4646                set->use--;
4647                fallthrough;
4648        default:
4649                nf_tables_unbind_set(ctx, set, binding,
4650                                     phase == NFT_TRANS_COMMIT);
4651        }
4652}
4653EXPORT_SYMBOL_GPL(nf_tables_deactivate_set);
4654
4655void nf_tables_destroy_set(const struct nft_ctx *ctx, struct nft_set *set)
4656{
4657        if (list_empty(&set->bindings) && nft_set_is_anonymous(set))
4658                nft_set_destroy(ctx, set);
4659}
4660EXPORT_SYMBOL_GPL(nf_tables_destroy_set);
4661
4662const struct nft_set_ext_type nft_set_ext_types[] = {
4663        [NFT_SET_EXT_KEY]               = {
4664                .align  = __alignof__(u32),
4665        },
4666        [NFT_SET_EXT_DATA]              = {
4667                .align  = __alignof__(u32),
4668        },
4669        [NFT_SET_EXT_EXPRESSIONS]       = {
4670                .align  = __alignof__(struct nft_set_elem_expr),
4671        },
4672        [NFT_SET_EXT_OBJREF]            = {
4673                .len    = sizeof(struct nft_object *),
4674                .align  = __alignof__(struct nft_object *),
4675        },
4676        [NFT_SET_EXT_FLAGS]             = {
4677                .len    = sizeof(u8),
4678                .align  = __alignof__(u8),
4679        },
4680        [NFT_SET_EXT_TIMEOUT]           = {
4681                .len    = sizeof(u64),
4682                .align  = __alignof__(u64),
4683        },
4684        [NFT_SET_EXT_EXPIRATION]        = {
4685                .len    = sizeof(u64),
4686                .align  = __alignof__(u64),
4687        },
4688        [NFT_SET_EXT_USERDATA]          = {
4689                .len    = sizeof(struct nft_userdata),
4690                .align  = __alignof__(struct nft_userdata),
4691        },
4692        [NFT_SET_EXT_KEY_END]           = {
4693                .align  = __alignof__(u32),
4694        },
4695};
4696
4697/*
4698 * Set elements
4699 */
4700
4701static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
4702        [NFTA_SET_ELEM_KEY]             = { .type = NLA_NESTED },
4703        [NFTA_SET_ELEM_DATA]            = { .type = NLA_NESTED },
4704        [NFTA_SET_ELEM_FLAGS]           = { .type = NLA_U32 },
4705        [NFTA_SET_ELEM_TIMEOUT]         = { .type = NLA_U64 },
4706        [NFTA_SET_ELEM_EXPIRATION]      = { .type = NLA_U64 },
4707        [NFTA_SET_ELEM_USERDATA]        = { .type = NLA_BINARY,
4708                                            .len = NFT_USERDATA_MAXLEN },
4709        [NFTA_SET_ELEM_EXPR]            = { .type = NLA_NESTED },
4710        [NFTA_SET_ELEM_OBJREF]          = { .type = NLA_STRING,
4711                                            .len = NFT_OBJ_MAXNAMELEN - 1 },
4712        [NFTA_SET_ELEM_KEY_END]         = { .type = NLA_NESTED },
4713        [NFTA_SET_ELEM_EXPRESSIONS]     = { .type = NLA_NESTED },
4714};
4715
4716static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
4717        [NFTA_SET_ELEM_LIST_TABLE]      = { .type = NLA_STRING,
4718                                            .len = NFT_TABLE_MAXNAMELEN - 1 },
4719        [NFTA_SET_ELEM_LIST_SET]        = { .type = NLA_STRING,
4720                                            .len = NFT_SET_MAXNAMELEN - 1 },
4721        [NFTA_SET_ELEM_LIST_ELEMENTS]   = { .type = NLA_NESTED },
4722        [NFTA_SET_ELEM_LIST_SET_ID]     = { .type = NLA_U32 },
4723};
4724
4725static int nft_set_elem_expr_dump(struct sk_buff *skb,
4726                                  const struct nft_set *set,
4727                                  const struct nft_set_ext *ext)
4728{
4729        struct nft_set_elem_expr *elem_expr;
4730        u32 size, num_exprs = 0;
4731        struct nft_expr *expr;
4732        struct nlattr *nest;
4733
4734        elem_expr = nft_set_ext_expr(ext);
4735        nft_setelem_expr_foreach(expr, elem_expr, size)
4736                num_exprs++;
4737
4738        if (num_exprs == 1) {
4739                expr = nft_setelem_expr_at(elem_expr, 0);
4740                if (nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, expr) < 0)
4741                        return -1;
4742
4743                return 0;
4744        } else if (num_exprs > 1) {
4745                nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_EXPRESSIONS);
4746                if (nest == NULL)
4747                        goto nla_put_failure;
4748
4749                nft_setelem_expr_foreach(expr, elem_expr, size) {
4750                        expr = nft_setelem_expr_at(elem_expr, size);
4751                        if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
4752                                goto nla_put_failure;
4753                }
4754                nla_nest_end(skb, nest);
4755        }
4756        return 0;
4757
4758nla_put_failure:
4759        return -1;
4760}
4761
4762static int nf_tables_fill_setelem(struct sk_buff *skb,
4763                                  const struct nft_set *set,
4764                                  const struct nft_set_elem *elem)
4765{
4766        const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4767        unsigned char *b = skb_tail_pointer(skb);
4768        struct nlattr *nest;
4769
4770        nest = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
4771        if (nest == NULL)
4772                goto nla_put_failure;
4773
4774        if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
4775            nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
4776                          NFT_DATA_VALUE, set->klen) < 0)
4777                goto nla_put_failure;
4778
4779        if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
4780            nft_data_dump(skb, NFTA_SET_ELEM_KEY_END, nft_set_ext_key_end(ext),
4781                          NFT_DATA_VALUE, set->klen) < 0)
4782                goto nla_put_failure;
4783
4784        if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
4785            nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
4786                          set->dtype == NFT_DATA_VERDICT ? NFT_DATA_VERDICT : NFT_DATA_VALUE,
4787                          set->dlen) < 0)
4788                goto nla_put_failure;
4789
4790        if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS) &&
4791            nft_set_elem_expr_dump(skb, set, ext))
4792                goto nla_put_failure;
4793
4794        if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
4795            nla_put_string(skb, NFTA_SET_ELEM_OBJREF,
4796                           (*nft_set_ext_obj(ext))->key.name) < 0)
4797                goto nla_put_failure;
4798
4799        if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
4800            nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
4801                         htonl(*nft_set_ext_flags(ext))))
4802                goto nla_put_failure;
4803
4804        if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
4805            nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
4806                         nf_jiffies64_to_msecs(*nft_set_ext_timeout(ext)),
4807                         NFTA_SET_ELEM_PAD))
4808                goto nla_put_failure;
4809
4810        if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
4811                u64 expires, now = get_jiffies_64();
4812
4813                expires = *nft_set_ext_expiration(ext);
4814                if (time_before64(now, expires))
4815                        expires -= now;
4816                else
4817                        expires = 0;
4818
4819                if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
4820                                 nf_jiffies64_to_msecs(expires),
4821                                 NFTA_SET_ELEM_PAD))
4822                        goto nla_put_failure;
4823        }
4824
4825        if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
4826                struct nft_userdata *udata;
4827
4828                udata = nft_set_ext_userdata(ext);
4829                if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
4830                            udata->len + 1, udata->data))
4831                        goto nla_put_failure;
4832        }
4833
4834        nla_nest_end(skb, nest);
4835        return 0;
4836
4837nla_put_failure:
4838        nlmsg_trim(skb, b);
4839        return -EMSGSIZE;
4840}
4841
4842struct nft_set_dump_args {
4843        const struct netlink_callback   *cb;
4844        struct nft_set_iter             iter;
4845        struct sk_buff                  *skb;
4846};
4847
4848static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
4849                                  struct nft_set *set,
4850                                  const struct nft_set_iter *iter,
4851                                  struct nft_set_elem *elem)
4852{
4853        struct nft_set_dump_args *args;
4854
4855        args = container_of(iter, struct nft_set_dump_args, iter);
4856        return nf_tables_fill_setelem(args->skb, set, elem);
4857}
4858
4859struct nft_set_dump_ctx {
4860        const struct nft_set    *set;
4861        struct nft_ctx          ctx;
4862};
4863
4864static int nft_set_catchall_dump(struct net *net, struct sk_buff *skb,
4865                                 const struct nft_set *set)
4866{
4867        struct nft_set_elem_catchall *catchall;
4868        u8 genmask = nft_genmask_cur(net);
4869        struct nft_set_elem elem;
4870        struct nft_set_ext *ext;
4871        int ret = 0;
4872
4873        list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4874                ext = nft_set_elem_ext(set, catchall->elem);
4875                if (!nft_set_elem_active(ext, genmask) ||
4876                    nft_set_elem_expired(ext))
4877                        continue;
4878
4879                elem.priv = catchall->elem;
4880                ret = nf_tables_fill_setelem(skb, set, &elem);
4881                break;
4882        }
4883
4884        return ret;
4885}
4886
4887static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
4888{
4889        struct nft_set_dump_ctx *dump_ctx = cb->data;
4890        struct net *net = sock_net(skb->sk);
4891        struct nftables_pernet *nft_net;
4892        struct nft_table *table;
4893        struct nft_set *set;
4894        struct nft_set_dump_args args;
4895        bool set_found = false;
4896        struct nlmsghdr *nlh;
4897        struct nlattr *nest;
4898        u32 portid, seq;
4899        int event;
4900
4901        rcu_read_lock();
4902        nft_net = nft_pernet(net);
4903        list_for_each_entry_rcu(table, &nft_net->tables, list) {
4904                if (dump_ctx->ctx.family != NFPROTO_UNSPEC &&
4905                    dump_ctx->ctx.family != table->family)
4906                        continue;
4907
4908                if (table != dump_ctx->ctx.table)
4909                        continue;
4910
4911                list_for_each_entry_rcu(set, &table->sets, list) {
4912                        if (set == dump_ctx->set) {
4913                                set_found = true;
4914                                break;
4915                        }
4916                }
4917                break;
4918        }
4919
4920        if (!set_found) {
4921                rcu_read_unlock();
4922                return -ENOENT;
4923        }
4924
4925        event  = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWSETELEM);
4926        portid = NETLINK_CB(cb->skb).portid;
4927        seq    = cb->nlh->nlmsg_seq;
4928
4929        nlh = nfnl_msg_put(skb, portid, seq, event, NLM_F_MULTI,
4930                           table->family, NFNETLINK_V0, nft_base_seq(net));
4931        if (!nlh)
4932                goto nla_put_failure;
4933
4934        if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, table->name))
4935                goto nla_put_failure;
4936        if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
4937                goto nla_put_failure;
4938
4939        nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
4940        if (nest == NULL)
4941                goto nla_put_failure;
4942
4943        args.cb                 = cb;
4944        args.skb                = skb;
4945        args.iter.genmask       = nft_genmask_cur(net);
4946        args.iter.skip          = cb->args[0];
4947        args.iter.count         = 0;
4948        args.iter.err           = 0;
4949        args.iter.fn            = nf_tables_dump_setelem;
4950        set->ops->walk(&dump_ctx->ctx, set, &args.iter);
4951
4952        if (!args.iter.err && args.iter.count == cb->args[0])
4953                args.iter.err = nft_set_catchall_dump(net, skb, set);
4954        rcu_read_unlock();
4955
4956        nla_nest_end(skb, nest);
4957        nlmsg_end(skb, nlh);
4958
4959        if (args.iter.err && args.iter.err != -EMSGSIZE)
4960                return args.iter.err;
4961        if (args.iter.count == cb->args[0])
4962                return 0;
4963
4964        cb->args[0] = args.iter.count;
4965        return skb->len;
4966
4967nla_put_failure:
4968        rcu_read_unlock();
4969        return -ENOSPC;
4970}
4971
4972static int nf_tables_dump_set_start(struct netlink_callback *cb)
4973{
4974        struct nft_set_dump_ctx *dump_ctx = cb->data;
4975
4976        cb->data = kmemdup(dump_ctx, sizeof(*dump_ctx), GFP_ATOMIC);
4977
4978        return cb->data ? 0 : -ENOMEM;
4979}
4980
4981static int nf_tables_dump_set_done(struct netlink_callback *cb)
4982{
4983        kfree(cb->data);
4984        return 0;
4985}
4986
4987static int nf_tables_fill_setelem_info(struct sk_buff *skb,
4988                                       const struct nft_ctx *ctx, u32 seq,
4989                                       u32 portid, int event, u16 flags,
4990                                       const struct nft_set *set,
4991                                       const struct nft_set_elem *elem)
4992{
4993        struct nlmsghdr *nlh;
4994        struct nlattr *nest;
4995        int err;
4996
4997        event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
4998        nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
4999                           NFNETLINK_V0, nft_base_seq(ctx->net));
5000        if (!nlh)
5001                goto nla_put_failure;
5002
5003        if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
5004                goto nla_put_failure;
5005        if (nla_put_string(skb, NFTA_SET_NAME, set->name))
5006                goto nla_put_failure;
5007
5008        nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5009        if (nest == NULL)
5010                goto nla_put_failure;
5011
5012        err = nf_tables_fill_setelem(skb, set, elem);
5013        if (err < 0)
5014                goto nla_put_failure;
5015
5016        nla_nest_end(skb, nest);
5017
5018        nlmsg_end(skb, nlh);
5019        return 0;
5020
5021nla_put_failure:
5022        nlmsg_trim(skb, nlh);
5023        return -1;
5024}
5025
5026static int nft_setelem_parse_flags(const struct nft_set *set,
5027                                   const struct nlattr *attr, u32 *flags)
5028{
5029        if (attr == NULL)
5030                return 0;
5031
5032        *flags = ntohl(nla_get_be32(attr));
5033        if (*flags & ~(NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
5034                return -EOPNOTSUPP;
5035        if (!(set->flags & NFT_SET_INTERVAL) &&
5036            *flags & NFT_SET_ELEM_INTERVAL_END)
5037                return -EINVAL;
5038
5039        return 0;
5040}
5041
5042static int nft_setelem_parse_key(struct nft_ctx *ctx, struct nft_set *set,
5043                                 struct nft_data *key, struct nlattr *attr)
5044{
5045        struct nft_data_desc desc;
5046        int err;
5047
5048        err = nft_data_init(ctx, key, NFT_DATA_VALUE_MAXLEN, &desc, attr);
5049        if (err < 0)
5050                return err;
5051
5052        if (desc.type != NFT_DATA_VALUE || desc.len != set->klen) {
5053                nft_data_release(key, desc.type);
5054                return -EINVAL;
5055        }
5056
5057        return 0;
5058}
5059
5060static int nft_setelem_parse_data(struct nft_ctx *ctx, struct nft_set *set,
5061                                  struct nft_data_desc *desc,
5062                                  struct nft_data *data,
5063                                  struct nlattr *attr)
5064{
5065        int err;
5066
5067        err = nft_data_init(ctx, data, NFT_DATA_VALUE_MAXLEN, desc, attr);
5068        if (err < 0)
5069                return err;
5070
5071        if (desc->type != NFT_DATA_VERDICT && desc->len != set->dlen) {
5072                nft_data_release(data, desc->type);
5073                return -EINVAL;
5074        }
5075
5076        return 0;
5077}
5078
5079static void *nft_setelem_catchall_get(const struct net *net,
5080                                      const struct nft_set *set)
5081{
5082        struct nft_set_elem_catchall *catchall;
5083        u8 genmask = nft_genmask_cur(net);
5084        struct nft_set_ext *ext;
5085        void *priv = NULL;
5086
5087        list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5088                ext = nft_set_elem_ext(set, catchall->elem);
5089                if (!nft_set_elem_active(ext, genmask) ||
5090                    nft_set_elem_expired(ext))
5091                        continue;
5092
5093                priv = catchall->elem;
5094                break;
5095        }
5096
5097        return priv;
5098}
5099
5100static int nft_setelem_get(struct nft_ctx *ctx, struct nft_set *set,
5101                           struct nft_set_elem *elem, u32 flags)
5102{
5103        void *priv;
5104
5105        if (!(flags & NFT_SET_ELEM_CATCHALL)) {
5106                priv = set->ops->get(ctx->net, set, elem, flags);
5107                if (IS_ERR(priv))
5108                        return PTR_ERR(priv);
5109        } else {
5110                priv = nft_setelem_catchall_get(ctx->net, set);
5111                if (!priv)
5112                        return -ENOENT;
5113        }
5114        elem->priv = priv;
5115
5116        return 0;
5117}
5118
5119static int nft_get_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5120                            const struct nlattr *attr)
5121{
5122        struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5123        struct nft_set_elem elem;
5124        struct sk_buff *skb;
5125        uint32_t flags = 0;
5126        int err;
5127
5128        err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5129                                          nft_set_elem_policy, NULL);
5130        if (err < 0)
5131                return err;
5132
5133        err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5134        if (err < 0)
5135                return err;
5136
5137        if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5138                return -EINVAL;
5139
5140        if (nla[NFTA_SET_ELEM_KEY]) {
5141                err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5142                                            nla[NFTA_SET_ELEM_KEY]);
5143                if (err < 0)
5144                        return err;
5145        }
5146
5147        if (nla[NFTA_SET_ELEM_KEY_END]) {
5148                err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5149                                            nla[NFTA_SET_ELEM_KEY_END]);
5150                if (err < 0)
5151                        return err;
5152        }
5153
5154        err = nft_setelem_get(ctx, set, &elem, flags);
5155        if (err < 0)
5156                return err;
5157
5158        err = -ENOMEM;
5159        skb = nlmsg_new(NLMSG_GOODSIZE, GFP_ATOMIC);
5160        if (skb == NULL)
5161                return err;
5162
5163        err = nf_tables_fill_setelem_info(skb, ctx, ctx->seq, ctx->portid,
5164                                          NFT_MSG_NEWSETELEM, 0, set, &elem);
5165        if (err < 0)
5166                goto err_fill_setelem;
5167
5168        return nfnetlink_unicast(skb, ctx->net, ctx->portid);
5169
5170err_fill_setelem:
5171        kfree_skb(skb);
5172        return err;
5173}
5174
5175/* called with rcu_read_lock held */
5176static int nf_tables_getsetelem(struct sk_buff *skb,
5177                                const struct nfnl_info *info,
5178                                const struct nlattr * const nla[])
5179{
5180        struct netlink_ext_ack *extack = info->extack;
5181        u8 genmask = nft_genmask_cur(info->net);
5182        u8 family = info->nfmsg->nfgen_family;
5183        struct net *net = info->net;
5184        struct nft_table *table;
5185        struct nft_set *set;
5186        struct nlattr *attr;
5187        struct nft_ctx ctx;
5188        int rem, err = 0;
5189
5190        table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
5191                                 genmask, NETLINK_CB(skb).portid);
5192        if (IS_ERR(table)) {
5193                NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
5194                return PTR_ERR(table);
5195        }
5196
5197        set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
5198        if (IS_ERR(set))
5199                return PTR_ERR(set);
5200
5201        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5202
5203        if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
5204                struct netlink_dump_control c = {
5205                        .start = nf_tables_dump_set_start,
5206                        .dump = nf_tables_dump_set,
5207                        .done = nf_tables_dump_set_done,
5208                        .module = THIS_MODULE,
5209                };
5210                struct nft_set_dump_ctx dump_ctx = {
5211                        .set = set,
5212                        .ctx = ctx,
5213                };
5214
5215                c.data = &dump_ctx;
5216                return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
5217        }
5218
5219        if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
5220                return -EINVAL;
5221
5222        nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
5223                err = nft_get_set_elem(&ctx, set, attr);
5224                if (err < 0)
5225                        break;
5226        }
5227
5228        return err;
5229}
5230
5231static void nf_tables_setelem_notify(const struct nft_ctx *ctx,
5232                                     const struct nft_set *set,
5233                                     const struct nft_set_elem *elem,
5234                                     int event, u16 flags)
5235{
5236        struct nftables_pernet *nft_net;
5237        struct net *net = ctx->net;
5238        u32 portid = ctx->portid;
5239        struct sk_buff *skb;
5240        int err;
5241
5242        if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
5243                return;
5244
5245        skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
5246        if (skb == NULL)
5247                goto err;
5248
5249        err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
5250                                          set, elem);
5251        if (err < 0) {
5252                kfree_skb(skb);
5253                goto err;
5254        }
5255
5256        nft_net = nft_pernet(net);
5257        nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
5258        return;
5259err:
5260        nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
5261}
5262
5263static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
5264                                              int msg_type,
5265                                              struct nft_set *set)
5266{
5267        struct nft_trans *trans;
5268
5269        trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
5270        if (trans == NULL)
5271                return NULL;
5272
5273        nft_trans_elem_set(trans) = set;
5274        return trans;
5275}
5276
5277struct nft_expr *nft_set_elem_expr_alloc(const struct nft_ctx *ctx,
5278                                         const struct nft_set *set,
5279                                         const struct nlattr *attr)
5280{
5281        struct nft_expr *expr;
5282        int err;
5283
5284        expr = nft_expr_init(ctx, attr);
5285        if (IS_ERR(expr))
5286                return expr;
5287
5288        err = -EOPNOTSUPP;
5289        if (!(expr->ops->type->flags & NFT_EXPR_STATEFUL))
5290                goto err_set_elem_expr;
5291
5292        if (expr->ops->type->flags & NFT_EXPR_GC) {
5293                if (set->flags & NFT_SET_TIMEOUT)
5294                        goto err_set_elem_expr;
5295                if (!set->ops->gc_init)
5296                        goto err_set_elem_expr;
5297                set->ops->gc_init(set);
5298        }
5299
5300        return expr;
5301
5302err_set_elem_expr:
5303        nft_expr_destroy(ctx, expr);
5304        return ERR_PTR(err);
5305}
5306
5307void *nft_set_elem_init(const struct nft_set *set,
5308                        const struct nft_set_ext_tmpl *tmpl,
5309                        const u32 *key, const u32 *key_end,
5310                        const u32 *data, u64 timeout, u64 expiration, gfp_t gfp)
5311{
5312        struct nft_set_ext *ext;
5313        void *elem;
5314
5315        elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
5316        if (elem == NULL)
5317                return NULL;
5318
5319        ext = nft_set_elem_ext(set, elem);
5320        nft_set_ext_init(ext, tmpl);
5321
5322        if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY))
5323                memcpy(nft_set_ext_key(ext), key, set->klen);
5324        if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END))
5325                memcpy(nft_set_ext_key_end(ext), key_end, set->klen);
5326        if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5327                memcpy(nft_set_ext_data(ext), data, set->dlen);
5328        if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5329                *nft_set_ext_expiration(ext) = get_jiffies_64() + expiration;
5330                if (expiration == 0)
5331                        *nft_set_ext_expiration(ext) += timeout;
5332        }
5333        if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
5334                *nft_set_ext_timeout(ext) = timeout;
5335
5336        return elem;
5337}
5338
5339static void __nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5340                                        struct nft_expr *expr)
5341{
5342        if (expr->ops->destroy_clone) {
5343                expr->ops->destroy_clone(ctx, expr);
5344                module_put(expr->ops->type->owner);
5345        } else {
5346                nf_tables_expr_destroy(ctx, expr);
5347        }
5348}
5349
5350static void nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5351                                      struct nft_set_elem_expr *elem_expr)
5352{
5353        struct nft_expr *expr;
5354        u32 size;
5355
5356        nft_setelem_expr_foreach(expr, elem_expr, size)
5357                __nft_set_elem_expr_destroy(ctx, expr);
5358}
5359
5360void nft_set_elem_destroy(const struct nft_set *set, void *elem,
5361                          bool destroy_expr)
5362{
5363        struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5364        struct nft_ctx ctx = {
5365                .net    = read_pnet(&set->net),
5366                .family = set->table->family,
5367        };
5368
5369        nft_data_release(nft_set_ext_key(ext), NFT_DATA_VALUE);
5370        if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5371                nft_data_release(nft_set_ext_data(ext), set->dtype);
5372        if (destroy_expr && nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5373                nft_set_elem_expr_destroy(&ctx, nft_set_ext_expr(ext));
5374
5375        if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
5376                (*nft_set_ext_obj(ext))->use--;
5377        kfree(elem);
5378}
5379EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
5380
5381/* Only called from commit path, nft_setelem_data_deactivate() already deals
5382 * with the refcounting from the preparation phase.
5383 */
5384static void nf_tables_set_elem_destroy(const struct nft_ctx *ctx,
5385                                       const struct nft_set *set, void *elem)
5386{
5387        struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5388
5389        if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5390                nft_set_elem_expr_destroy(ctx, nft_set_ext_expr(ext));
5391
5392        kfree(elem);
5393}
5394
5395int nft_set_elem_expr_clone(const struct nft_ctx *ctx, struct nft_set *set,
5396                            struct nft_expr *expr_array[])
5397{
5398        struct nft_expr *expr;
5399        int err, i, k;
5400
5401        for (i = 0; i < set->num_exprs; i++) {
5402                expr = kzalloc(set->exprs[i]->ops->size, GFP_KERNEL);
5403                if (!expr)
5404                        goto err_expr;
5405
5406                err = nft_expr_clone(expr, set->exprs[i]);
5407                if (err < 0) {
5408                        nft_expr_destroy(ctx, expr);
5409                        goto err_expr;
5410                }
5411                expr_array[i] = expr;
5412        }
5413
5414        return 0;
5415
5416err_expr:
5417        for (k = i - 1; k >= 0; k--)
5418                nft_expr_destroy(ctx, expr_array[k]);
5419
5420        return -ENOMEM;
5421}
5422
5423static int nft_set_elem_expr_setup(struct nft_ctx *ctx,
5424                                   const struct nft_set_ext *ext,
5425                                   struct nft_expr *expr_array[],
5426                                   u32 num_exprs)
5427{
5428        struct nft_set_elem_expr *elem_expr = nft_set_ext_expr(ext);
5429        struct nft_expr *expr;
5430        int i, err;
5431
5432        for (i = 0; i < num_exprs; i++) {
5433                expr = nft_setelem_expr_at(elem_expr, elem_expr->size);
5434                err = nft_expr_clone(expr, expr_array[i]);
5435                if (err < 0)
5436                        goto err_elem_expr_setup;
5437
5438                elem_expr->size += expr_array[i]->ops->size;
5439                nft_expr_destroy(ctx, expr_array[i]);
5440                expr_array[i] = NULL;
5441        }
5442
5443        return 0;
5444
5445err_elem_expr_setup:
5446        for (; i < num_exprs; i++) {
5447                nft_expr_destroy(ctx, expr_array[i]);
5448                expr_array[i] = NULL;
5449        }
5450
5451        return -ENOMEM;
5452}
5453
5454struct nft_set_ext *nft_set_catchall_lookup(const struct net *net,
5455                                            const struct nft_set *set)
5456{
5457        struct nft_set_elem_catchall *catchall;
5458        u8 genmask = nft_genmask_cur(net);
5459        struct nft_set_ext *ext;
5460
5461        list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5462                ext = nft_set_elem_ext(set, catchall->elem);
5463                if (nft_set_elem_active(ext, genmask) &&
5464                    !nft_set_elem_expired(ext))
5465                        return ext;
5466        }
5467
5468        return NULL;
5469}
5470EXPORT_SYMBOL_GPL(nft_set_catchall_lookup);
5471
5472void *nft_set_catchall_gc(const struct nft_set *set)
5473{
5474        struct nft_set_elem_catchall *catchall, *next;
5475        struct nft_set_ext *ext;
5476        void *elem = NULL;
5477
5478        list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5479                ext = nft_set_elem_ext(set, catchall->elem);
5480
5481                if (!nft_set_elem_expired(ext) ||
5482                    nft_set_elem_mark_busy(ext))
5483                        continue;
5484
5485                elem = catchall->elem;
5486                list_del_rcu(&catchall->list);
5487                kfree_rcu(catchall, rcu);
5488                break;
5489        }
5490
5491        return elem;
5492}
5493EXPORT_SYMBOL_GPL(nft_set_catchall_gc);
5494
5495static int nft_setelem_catchall_insert(const struct net *net,
5496                                       struct nft_set *set,
5497                                       const struct nft_set_elem *elem,
5498                                       struct nft_set_ext **pext)
5499{
5500        struct nft_set_elem_catchall *catchall;
5501        u8 genmask = nft_genmask_next(net);
5502        struct nft_set_ext *ext;
5503
5504        list_for_each_entry(catchall, &set->catchall_list, list) {
5505                ext = nft_set_elem_ext(set, catchall->elem);
5506                if (nft_set_elem_active(ext, genmask)) {
5507                        *pext = ext;
5508                        return -EEXIST;
5509                }
5510        }
5511
5512        catchall = kmalloc(sizeof(*catchall), GFP_KERNEL);
5513        if (!catchall)
5514                return -ENOMEM;
5515
5516        catchall->elem = elem->priv;
5517        list_add_tail_rcu(&catchall->list, &set->catchall_list);
5518
5519        return 0;
5520}
5521
5522static int nft_setelem_insert(const struct net *net,
5523                              struct nft_set *set,
5524                              const struct nft_set_elem *elem,
5525                              struct nft_set_ext **ext, unsigned int flags)
5526{
5527        int ret;
5528
5529        if (flags & NFT_SET_ELEM_CATCHALL)
5530                ret = nft_setelem_catchall_insert(net, set, elem, ext);
5531        else
5532                ret = set->ops->insert(net, set, elem, ext);
5533
5534        return ret;
5535}
5536
5537static bool nft_setelem_is_catchall(const struct nft_set *set,
5538                                    const struct nft_set_elem *elem)
5539{
5540        struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5541
5542        if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
5543            *nft_set_ext_flags(ext) & NFT_SET_ELEM_CATCHALL)
5544                return true;
5545
5546        return false;
5547}
5548
5549static void nft_setelem_activate(struct net *net, struct nft_set *set,
5550                                 struct nft_set_elem *elem)
5551{
5552        struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5553
5554        if (nft_setelem_is_catchall(set, elem)) {
5555                nft_set_elem_change_active(net, set, ext);
5556                nft_set_elem_clear_busy(ext);
5557        } else {
5558                set->ops->activate(net, set, elem);
5559        }
5560}
5561
5562static int nft_setelem_catchall_deactivate(const struct net *net,
5563                                           struct nft_set *set,
5564                                           struct nft_set_elem *elem)
5565{
5566        struct nft_set_elem_catchall *catchall;
5567        struct nft_set_ext *ext;
5568
5569        list_for_each_entry(catchall, &set->catchall_list, list) {
5570                ext = nft_set_elem_ext(set, catchall->elem);
5571                if (!nft_is_active(net, ext) ||
5572                    nft_set_elem_mark_busy(ext))
5573                        continue;
5574
5575                kfree(elem->priv);
5576                elem->priv = catchall->elem;
5577                nft_set_elem_change_active(net, set, ext);
5578                return 0;
5579        }
5580
5581        return -ENOENT;
5582}
5583
5584static int __nft_setelem_deactivate(const struct net *net,
5585                                    struct nft_set *set,
5586                                    struct nft_set_elem *elem)
5587{
5588        void *priv;
5589
5590        priv = set->ops->deactivate(net, set, elem);
5591        if (!priv)
5592                return -ENOENT;
5593
5594        kfree(elem->priv);
5595        elem->priv = priv;
5596        set->ndeact++;
5597
5598        return 0;
5599}
5600
5601static int nft_setelem_deactivate(const struct net *net,
5602                                  struct nft_set *set,
5603                                  struct nft_set_elem *elem, u32 flags)
5604{
5605        int ret;
5606
5607        if (flags & NFT_SET_ELEM_CATCHALL)
5608                ret = nft_setelem_catchall_deactivate(net, set, elem);
5609        else
5610                ret = __nft_setelem_deactivate(net, set, elem);
5611
5612        return ret;
5613}
5614
5615static void nft_setelem_catchall_remove(const struct net *net,
5616                                        const struct nft_set *set,
5617                                        const struct nft_set_elem *elem)
5618{
5619        struct nft_set_elem_catchall *catchall, *next;
5620
5621        list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5622                if (catchall->elem == elem->priv) {
5623                        list_del_rcu(&catchall->list);
5624                        kfree_rcu(catchall);
5625                        break;
5626                }
5627        }
5628}
5629
5630static void nft_setelem_remove(const struct net *net,
5631                               const struct nft_set *set,
5632                               const struct nft_set_elem *elem)
5633{
5634        if (nft_setelem_is_catchall(set, elem))
5635                nft_setelem_catchall_remove(net, set, elem);
5636        else
5637                set->ops->remove(net, set, elem);
5638}
5639
5640static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5641                            const struct nlattr *attr, u32 nlmsg_flags)
5642{
5643        struct nft_expr *expr_array[NFT_SET_EXPR_MAX] = {};
5644        struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5645        u8 genmask = nft_genmask_next(ctx->net);
5646        u32 flags = 0, size = 0, num_exprs = 0;
5647        struct nft_set_ext_tmpl tmpl;
5648        struct nft_set_ext *ext, *ext2;
5649        struct nft_set_elem elem;
5650        struct nft_set_binding *binding;
5651        struct nft_object *obj = NULL;
5652        struct nft_userdata *udata;
5653        struct nft_data_desc desc;
5654        enum nft_registers dreg;
5655        struct nft_trans *trans;
5656        u64 timeout;
5657        u64 expiration;
5658        int err, i;
5659        u8 ulen;
5660
5661        err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5662                                          nft_set_elem_policy, NULL);
5663        if (err < 0)
5664                return err;
5665
5666        nft_set_ext_prepare(&tmpl);
5667
5668        err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5669        if (err < 0)
5670                return err;
5671
5672        if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5673                return -EINVAL;
5674
5675        if (flags != 0)
5676                nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
5677
5678        if (set->flags & NFT_SET_MAP) {
5679                if (nla[NFTA_SET_ELEM_DATA] == NULL &&
5680                    !(flags & NFT_SET_ELEM_INTERVAL_END))
5681                        return -EINVAL;
5682        } else {
5683                if (nla[NFTA_SET_ELEM_DATA] != NULL)
5684                        return -EINVAL;
5685        }
5686
5687        if ((flags & NFT_SET_ELEM_INTERVAL_END) &&
5688             (nla[NFTA_SET_ELEM_DATA] ||
5689              nla[NFTA_SET_ELEM_OBJREF] ||
5690              nla[NFTA_SET_ELEM_TIMEOUT] ||
5691              nla[NFTA_SET_ELEM_EXPIRATION] ||
5692              nla[NFTA_SET_ELEM_USERDATA] ||
5693              nla[NFTA_SET_ELEM_EXPR] ||
5694              nla[NFTA_SET_ELEM_EXPRESSIONS]))
5695                return -EINVAL;
5696
5697        timeout = 0;
5698        if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
5699                if (!(set->flags & NFT_SET_TIMEOUT))
5700                        return -EINVAL;
5701                err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_TIMEOUT],
5702                                            &timeout);
5703                if (err)
5704                        return err;
5705        } else if (set->flags & NFT_SET_TIMEOUT) {
5706                timeout = set->timeout;
5707        }
5708
5709        expiration = 0;
5710        if (nla[NFTA_SET_ELEM_EXPIRATION] != NULL) {
5711                if (!(set->flags & NFT_SET_TIMEOUT))
5712                        return -EINVAL;
5713                err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_EXPIRATION],
5714                                            &expiration);
5715                if (err)
5716                        return err;
5717        }
5718
5719        if (nla[NFTA_SET_ELEM_EXPR]) {
5720                struct nft_expr *expr;
5721
5722                if (set->num_exprs && set->num_exprs != 1)
5723                        return -EOPNOTSUPP;
5724
5725                expr = nft_set_elem_expr_alloc(ctx, set,
5726                                               nla[NFTA_SET_ELEM_EXPR]);
5727                if (IS_ERR(expr))
5728                        return PTR_ERR(expr);
5729
5730                expr_array[0] = expr;
5731                num_exprs = 1;
5732
5733                if (set->num_exprs && set->exprs[0]->ops != expr->ops) {
5734                        err = -EOPNOTSUPP;
5735                        goto err_set_elem_expr;
5736                }
5737        } else if (nla[NFTA_SET_ELEM_EXPRESSIONS]) {
5738                struct nft_expr *expr;
5739                struct nlattr *tmp;
5740                int left;
5741
5742                i = 0;
5743                nla_for_each_nested(tmp, nla[NFTA_SET_ELEM_EXPRESSIONS], left) {
5744                        if (i == NFT_SET_EXPR_MAX ||
5745                            (set->num_exprs && set->num_exprs == i)) {
5746                                err = -E2BIG;
5747                                goto err_set_elem_expr;
5748                        }
5749                        if (nla_type(tmp) != NFTA_LIST_ELEM) {
5750                                err = -EINVAL;
5751                                goto err_set_elem_expr;
5752                        }
5753                        expr = nft_set_elem_expr_alloc(ctx, set, tmp);
5754                        if (IS_ERR(expr)) {
5755                                err = PTR_ERR(expr);
5756                                goto err_set_elem_expr;
5757                        }
5758                        expr_array[i] = expr;
5759                        num_exprs++;
5760
5761                        if (set->num_exprs && expr->ops != set->exprs[i]->ops) {
5762                                err = -EOPNOTSUPP;
5763                                goto err_set_elem_expr;
5764                        }
5765                        i++;
5766                }
5767                if (set->num_exprs && set->num_exprs != i) {
5768                        err = -EOPNOTSUPP;
5769                        goto err_set_elem_expr;
5770                }
5771        } else if (set->num_exprs > 0) {
5772                err = nft_set_elem_expr_clone(ctx, set, expr_array);
5773                if (err < 0)
5774                        goto err_set_elem_expr_clone;
5775
5776                num_exprs = set->num_exprs;
5777        }
5778
5779        if (nla[NFTA_SET_ELEM_KEY]) {
5780                err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5781                                            nla[NFTA_SET_ELEM_KEY]);
5782                if (err < 0)
5783                        goto err_set_elem_expr;
5784
5785                nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
5786        }
5787
5788        if (nla[NFTA_SET_ELEM_KEY_END]) {
5789                err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5790                                            nla[NFTA_SET_ELEM_KEY_END]);
5791                if (err < 0)
5792                        goto err_parse_key;
5793
5794                nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
5795        }
5796
5797        if (timeout > 0) {
5798                nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
5799                if (timeout != set->timeout)
5800                        nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
5801        }
5802
5803        if (num_exprs) {
5804                for (i = 0; i < num_exprs; i++)
5805                        size += expr_array[i]->ops->size;
5806
5807                nft_set_ext_add_length(&tmpl, NFT_SET_EXT_EXPRESSIONS,
5808                                       sizeof(struct nft_set_elem_expr) +
5809                                       size);
5810        }
5811
5812        if (nla[NFTA_SET_ELEM_OBJREF] != NULL) {
5813                if (!(set->flags & NFT_SET_OBJECT)) {
5814                        err = -EINVAL;
5815                        goto err_parse_key_end;
5816                }
5817                obj = nft_obj_lookup(ctx->net, ctx->table,
5818                                     nla[NFTA_SET_ELEM_OBJREF],
5819                                     set->objtype, genmask);
5820                if (IS_ERR(obj)) {
5821                        err = PTR_ERR(obj);
5822                        goto err_parse_key_end;
5823                }
5824                nft_set_ext_add(&tmpl, NFT_SET_EXT_OBJREF);
5825        }
5826
5827        if (nla[NFTA_SET_ELEM_DATA] != NULL) {
5828                err = nft_setelem_parse_data(ctx, set, &desc, &elem.data.val,
5829                                             nla[NFTA_SET_ELEM_DATA]);
5830                if (err < 0)
5831                        goto err_parse_key_end;
5832
5833                dreg = nft_type_to_reg(set->dtype);
5834                list_for_each_entry(binding, &set->bindings, list) {
5835                        struct nft_ctx bind_ctx = {
5836                                .net    = ctx->net,
5837                                .family = ctx->family,
5838                                .table  = ctx->table,
5839                                .chain  = (struct nft_chain *)binding->chain,
5840                        };
5841
5842                        if (!(binding->flags & NFT_SET_MAP))
5843                                continue;
5844
5845                        err = nft_validate_register_store(&bind_ctx, dreg,
5846                                                          &elem.data.val,
5847                                                          desc.type, desc.len);
5848                        if (err < 0)
5849                                goto err_parse_data;
5850
5851                        if (desc.type == NFT_DATA_VERDICT &&
5852                            (elem.data.val.verdict.code == NFT_GOTO ||
5853                             elem.data.val.verdict.code == NFT_JUMP))
5854                                nft_validate_state_update(ctx->net,
5855                                                          NFT_VALIDATE_NEED);
5856                }
5857
5858                nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, desc.len);
5859        }
5860
5861        /* The full maximum length of userdata can exceed the maximum
5862         * offset value (U8_MAX) for following extensions, therefor it
5863         * must be the last extension added.
5864         */
5865        ulen = 0;
5866        if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
5867                ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
5868                if (ulen > 0)
5869                        nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
5870                                               ulen);
5871        }
5872
5873        err = -ENOMEM;
5874        elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
5875                                      elem.key_end.val.data, elem.data.val.data,
5876                                      timeout, expiration, GFP_KERNEL);
5877        if (elem.priv == NULL)
5878                goto err_parse_data;
5879
5880        ext = nft_set_elem_ext(set, elem.priv);
5881        if (flags)
5882                *nft_set_ext_flags(ext) = flags;
5883        if (ulen > 0) {
5884                udata = nft_set_ext_userdata(ext);
5885                udata->len = ulen - 1;
5886                nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
5887        }
5888        if (obj) {
5889                *nft_set_ext_obj(ext) = obj;
5890                obj->use++;
5891        }
5892        err = nft_set_elem_expr_setup(ctx, ext, expr_array, num_exprs);
5893        if (err < 0)
5894                goto err_elem_expr;
5895
5896        trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
5897        if (trans == NULL) {
5898                err = -ENOMEM;
5899                goto err_elem_expr;
5900        }
5901
5902        ext->genmask = nft_genmask_cur(ctx->net) | NFT_SET_ELEM_BUSY_MASK;
5903
5904        err = nft_setelem_insert(ctx->net, set, &elem, &ext2, flags);
5905        if (err) {
5906                if (err == -EEXIST) {
5907                        if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
5908                            nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
5909                            nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
5910                            nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF))
5911                                goto err_element_clash;
5912                        if ((nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
5913                             nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) &&
5914                             memcmp(nft_set_ext_data(ext),
5915                                    nft_set_ext_data(ext2), set->dlen) != 0) ||
5916                            (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
5917                             nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
5918                             *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
5919                                goto err_element_clash;
5920                        else if (!(nlmsg_flags & NLM_F_EXCL))
5921                                err = 0;
5922                } else if (err == -ENOTEMPTY) {
5923                        /* ENOTEMPTY reports overlapping between this element
5924                         * and an existing one.
5925                         */
5926                        err = -EEXIST;
5927                }
5928                goto err_element_clash;
5929        }
5930
5931        if (!(flags & NFT_SET_ELEM_CATCHALL) && set->size &&
5932            !atomic_add_unless(&set->nelems, 1, set->size + set->ndeact)) {
5933                err = -ENFILE;
5934                goto err_set_full;
5935        }
5936
5937        nft_trans_elem(trans) = elem;
5938        nft_trans_commit_list_add_tail(ctx->net, trans);
5939        return 0;
5940
5941err_set_full:
5942        nft_setelem_remove(ctx->net, set, &elem);
5943err_element_clash:
5944        kfree(trans);
5945err_elem_expr:
5946        if (obj)
5947                obj->use--;
5948
5949        nf_tables_set_elem_destroy(ctx, set, elem.priv);
5950err_parse_data:
5951        if (nla[NFTA_SET_ELEM_DATA] != NULL)
5952                nft_data_release(&elem.data.val, desc.type);
5953err_parse_key_end:
5954        nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
5955err_parse_key:
5956        nft_data_release(&elem.key.val, NFT_DATA_VALUE);
5957err_set_elem_expr:
5958        for (i = 0; i < num_exprs && expr_array[i]; i++)
5959                nft_expr_destroy(ctx, expr_array[i]);
5960err_set_elem_expr_clone:
5961        return err;
5962}
5963
5964static int nf_tables_newsetelem(struct sk_buff *skb,
5965                                const struct nfnl_info *info,
5966                                const struct nlattr * const nla[])
5967{
5968        struct nftables_pernet *nft_net = nft_pernet(info->net);
5969        struct netlink_ext_ack *extack = info->extack;
5970        u8 genmask = nft_genmask_next(info->net);
5971        u8 family = info->nfmsg->nfgen_family;
5972        struct net *net = info->net;
5973        const struct nlattr *attr;
5974        struct nft_table *table;
5975        struct nft_set *set;
5976        struct nft_ctx ctx;
5977        int rem, err;
5978
5979        if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
5980                return -EINVAL;
5981
5982        table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
5983                                 genmask, NETLINK_CB(skb).portid);
5984        if (IS_ERR(table)) {
5985                NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
5986                return PTR_ERR(table);
5987        }
5988
5989        set = nft_set_lookup_global(net, table, nla[NFTA_SET_ELEM_LIST_SET],
5990                                    nla[NFTA_SET_ELEM_LIST_SET_ID], genmask);
5991        if (IS_ERR(set))
5992                return PTR_ERR(set);
5993
5994        if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
5995                return -EBUSY;
5996
5997        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5998
5999        nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6000                err = nft_add_set_elem(&ctx, set, attr, info->nlh->nlmsg_flags);
6001                if (err < 0)
6002                        return err;
6003        }
6004
6005        if (nft_net->validate_state == NFT_VALIDATE_DO)
6006                return nft_table_validate(net, table);
6007
6008        return 0;
6009}
6010
6011/**
6012 *      nft_data_hold - hold a nft_data item
6013 *
6014 *      @data: struct nft_data to release
6015 *      @type: type of data
6016 *
6017 *      Hold a nft_data item. NFT_DATA_VALUE types can be silently discarded,
6018 *      NFT_DATA_VERDICT bumps the reference to chains in case of NFT_JUMP and
6019 *      NFT_GOTO verdicts. This function must be called on active data objects
6020 *      from the second phase of the commit protocol.
6021 */
6022void nft_data_hold(const struct nft_data *data, enum nft_data_types type)
6023{
6024        struct nft_chain *chain;
6025        struct nft_rule *rule;
6026
6027        if (type == NFT_DATA_VERDICT) {
6028                switch (data->verdict.code) {
6029                case NFT_JUMP:
6030                case NFT_GOTO:
6031                        chain = data->verdict.chain;
6032                        chain->use++;
6033
6034                        if (!nft_chain_is_bound(chain))
6035                                break;
6036
6037                        chain->table->use++;
6038                        list_for_each_entry(rule, &chain->rules, list)
6039                                chain->use++;
6040
6041                        nft_chain_add(chain->table, chain);
6042                        break;
6043                }
6044        }
6045}
6046
6047static void nft_setelem_data_activate(const struct net *net,
6048                                      const struct nft_set *set,
6049                                      struct nft_set_elem *elem)
6050{
6051        const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6052
6053        if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6054                nft_data_hold(nft_set_ext_data(ext), set->dtype);
6055        if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6056                (*nft_set_ext_obj(ext))->use++;
6057}
6058
6059static void nft_setelem_data_deactivate(const struct net *net,
6060                                        const struct nft_set *set,
6061                                        struct nft_set_elem *elem)
6062{
6063        const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6064
6065        if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6066                nft_data_release(nft_set_ext_data(ext), set->dtype);
6067        if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6068                (*nft_set_ext_obj(ext))->use--;
6069}
6070
6071static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
6072                           const struct nlattr *attr)
6073{
6074        struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
6075        struct nft_set_ext_tmpl tmpl;
6076        struct nft_set_elem elem;
6077        struct nft_set_ext *ext;
6078        struct nft_trans *trans;
6079        u32 flags = 0;
6080        int err;
6081
6082        err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
6083                                          nft_set_elem_policy, NULL);
6084        if (err < 0)
6085                return err;
6086
6087        err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
6088        if (err < 0)
6089                return err;
6090
6091        if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
6092                return -EINVAL;
6093
6094        nft_set_ext_prepare(&tmpl);
6095
6096        if (flags != 0)
6097                nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
6098
6099        if (nla[NFTA_SET_ELEM_KEY]) {
6100                err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6101                                            nla[NFTA_SET_ELEM_KEY]);
6102                if (err < 0)
6103                        return err;
6104
6105                nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
6106        }
6107
6108        if (nla[NFTA_SET_ELEM_KEY_END]) {
6109                err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6110                                            nla[NFTA_SET_ELEM_KEY_END]);
6111                if (err < 0)
6112                        return err;
6113
6114                nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
6115        }
6116
6117        err = -ENOMEM;
6118        elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
6119                                      elem.key_end.val.data, NULL, 0, 0,
6120                                      GFP_KERNEL);
6121        if (elem.priv == NULL)
6122                goto fail_elem;
6123
6124        ext = nft_set_elem_ext(set, elem.priv);
6125        if (flags)
6126                *nft_set_ext_flags(ext) = flags;
6127
6128        trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
6129        if (trans == NULL)
6130                goto fail_trans;
6131
6132        err = nft_setelem_deactivate(ctx->net, set, &elem, flags);
6133        if (err < 0)
6134                goto fail_ops;
6135
6136        nft_setelem_data_deactivate(ctx->net, set, &elem);
6137
6138        nft_trans_elem(trans) = elem;
6139        nft_trans_commit_list_add_tail(ctx->net, trans);
6140        return 0;
6141
6142fail_ops:
6143        kfree(trans);
6144fail_trans:
6145        kfree(elem.priv);
6146fail_elem:
6147        nft_data_release(&elem.key.val, NFT_DATA_VALUE);
6148        return err;
6149}
6150
6151static int nft_setelem_flush(const struct nft_ctx *ctx,
6152                             struct nft_set *set,
6153                             const struct nft_set_iter *iter,
6154                             struct nft_set_elem *elem)
6155{
6156        struct nft_trans *trans;
6157        int err;
6158
6159        trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6160                                    sizeof(struct nft_trans_elem), GFP_ATOMIC);
6161        if (!trans)
6162                return -ENOMEM;
6163
6164        if (!set->ops->flush(ctx->net, set, elem->priv)) {
6165                err = -ENOENT;
6166                goto err1;
6167        }
6168        set->ndeact++;
6169
6170        nft_setelem_data_deactivate(ctx->net, set, elem);
6171        nft_trans_elem_set(trans) = set;
6172        nft_trans_elem(trans) = *elem;
6173        nft_trans_commit_list_add_tail(ctx->net, trans);
6174
6175        return 0;
6176err1:
6177        kfree(trans);
6178        return err;
6179}
6180
6181static int __nft_set_catchall_flush(const struct nft_ctx *ctx,
6182                                    struct nft_set *set,
6183                                    struct nft_set_elem *elem)
6184{
6185        struct nft_trans *trans;
6186
6187        trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6188                                    sizeof(struct nft_trans_elem), GFP_KERNEL);
6189        if (!trans)
6190                return -ENOMEM;
6191
6192        nft_setelem_data_deactivate(ctx->net, set, elem);
6193        nft_trans_elem_set(trans) = set;
6194        nft_trans_elem(trans) = *elem;
6195        nft_trans_commit_list_add_tail(ctx->net, trans);
6196
6197        return 0;
6198}
6199
6200static int nft_set_catchall_flush(const struct nft_ctx *ctx,
6201                                  struct nft_set *set)
6202{
6203        u8 genmask = nft_genmask_next(ctx->net);
6204        struct nft_set_elem_catchall *catchall;
6205        struct nft_set_elem elem;
6206        struct nft_set_ext *ext;
6207        int ret = 0;
6208
6209        list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
6210                ext = nft_set_elem_ext(set, catchall->elem);
6211                if (!nft_set_elem_active(ext, genmask) ||
6212                    nft_set_elem_mark_busy(ext))
6213                        continue;
6214
6215                elem.priv = catchall->elem;
6216                ret = __nft_set_catchall_flush(ctx, set, &elem);
6217                if (ret < 0)
6218                        break;
6219        }
6220
6221        return ret;
6222}
6223
6224static int nft_set_flush(struct nft_ctx *ctx, struct nft_set *set, u8 genmask)
6225{
6226        struct nft_set_iter iter = {
6227                .genmask        = genmask,
6228                .fn             = nft_setelem_flush,
6229        };
6230
6231        set->ops->walk(ctx, set, &iter);
6232        if (!iter.err)
6233                iter.err = nft_set_catchall_flush(ctx, set);
6234
6235        return iter.err;
6236}
6237
6238static int nf_tables_delsetelem(struct sk_buff *skb,
6239                                const struct nfnl_info *info,
6240                                const struct nlattr * const nla[])
6241{
6242        struct netlink_ext_ack *extack = info->extack;
6243        u8 genmask = nft_genmask_next(info->net);
6244        u8 family = info->nfmsg->nfgen_family;
6245        struct net *net = info->net;
6246        const struct nlattr *attr;
6247        struct nft_table *table;
6248        struct nft_set *set;
6249        struct nft_ctx ctx;
6250        int rem, err = 0;
6251
6252        table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6253                                 genmask, NETLINK_CB(skb).portid);
6254        if (IS_ERR(table)) {
6255                NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6256                return PTR_ERR(table);
6257        }
6258
6259        set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
6260        if (IS_ERR(set))
6261                return PTR_ERR(set);
6262        if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
6263                return -EBUSY;
6264
6265        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6266
6267        if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
6268                return nft_set_flush(&ctx, set, genmask);
6269
6270        nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6271                err = nft_del_setelem(&ctx, set, attr);
6272                if (err < 0)
6273                        break;
6274        }
6275        return err;
6276}
6277
6278void nft_set_gc_batch_release(struct rcu_head *rcu)
6279{
6280        struct nft_set_gc_batch *gcb;
6281        unsigned int i;
6282
6283        gcb = container_of(rcu, struct nft_set_gc_batch, head.rcu);
6284        for (i = 0; i < gcb->head.cnt; i++)
6285                nft_set_elem_destroy(gcb->head.set, gcb->elems[i], true);
6286        kfree(gcb);
6287}
6288
6289struct nft_set_gc_batch *nft_set_gc_batch_alloc(const struct nft_set *set,
6290                                                gfp_t gfp)
6291{
6292        struct nft_set_gc_batch *gcb;
6293
6294        gcb = kzalloc(sizeof(*gcb), gfp);
6295        if (gcb == NULL)
6296                return gcb;
6297        gcb->head.set = set;
6298        return gcb;
6299}
6300
6301/*
6302 * Stateful objects
6303 */
6304
6305/**
6306 *      nft_register_obj- register nf_tables stateful object type
6307 *      @obj_type: object type
6308 *
6309 *      Registers the object type for use with nf_tables. Returns zero on
6310 *      success or a negative errno code otherwise.
6311 */
6312int nft_register_obj(struct nft_object_type *obj_type)
6313{
6314        if (obj_type->type == NFT_OBJECT_UNSPEC)
6315                return -EINVAL;
6316
6317        nfnl_lock(NFNL_SUBSYS_NFTABLES);
6318        list_add_rcu(&obj_type->list, &nf_tables_objects);
6319        nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6320        return 0;
6321}
6322EXPORT_SYMBOL_GPL(nft_register_obj);
6323
6324/**
6325 *      nft_unregister_obj - unregister nf_tables object type
6326 *      @obj_type: object type
6327 *
6328 *      Unregisters the object type for use with nf_tables.
6329 */
6330void nft_unregister_obj(struct nft_object_type *obj_type)
6331{
6332        nfnl_lock(NFNL_SUBSYS_NFTABLES);
6333        list_del_rcu(&obj_type->list);
6334        nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6335}
6336EXPORT_SYMBOL_GPL(nft_unregister_obj);
6337
6338struct nft_object *nft_obj_lookup(const struct net *net,
6339                                  const struct nft_table *table,
6340                                  const struct nlattr *nla, u32 objtype,
6341                                  u8 genmask)
6342{
6343        struct nft_object_hash_key k = { .table = table };
6344        char search[NFT_OBJ_MAXNAMELEN];
6345        struct rhlist_head *tmp, *list;
6346        struct nft_object *obj;
6347
6348        nla_strscpy(search, nla, sizeof(search));
6349        k.name = search;
6350
6351        WARN_ON_ONCE(!rcu_read_lock_held() &&
6352                     !lockdep_commit_lock_is_held(net));
6353
6354        rcu_read_lock();
6355        list = rhltable_lookup(&nft_objname_ht, &k, nft_objname_ht_params);
6356        if (!list)
6357                goto out;
6358
6359        rhl_for_each_entry_rcu(obj, tmp, list, rhlhead) {
6360                if (objtype == obj->ops->type->type &&
6361                    nft_active_genmask(obj, genmask)) {
6362                        rcu_read_unlock();
6363                        return obj;
6364                }
6365        }
6366out:
6367        rcu_read_unlock();
6368        return ERR_PTR(-ENOENT);
6369}
6370EXPORT_SYMBOL_GPL(nft_obj_lookup);
6371
6372static struct nft_object *nft_obj_lookup_byhandle(const struct nft_table *table,
6373                                                  const struct nlattr *nla,
6374                                                  u32 objtype, u8 genmask)
6375{
6376        struct nft_object *obj;
6377
6378        list_for_each_entry(obj, &table->objects, list) {
6379                if (be64_to_cpu(nla_get_be64(nla)) == obj->handle &&
6380                    objtype == obj->ops->type->type &&
6381                    nft_active_genmask(obj, genmask))
6382                        return obj;
6383        }
6384        return ERR_PTR(-ENOENT);
6385}
6386
6387static const struct nla_policy nft_obj_policy[NFTA_OBJ_MAX + 1] = {
6388        [NFTA_OBJ_TABLE]        = { .type = NLA_STRING,
6389                                    .len = NFT_TABLE_MAXNAMELEN - 1 },
6390        [NFTA_OBJ_NAME]         = { .type = NLA_STRING,
6391                                    .len = NFT_OBJ_MAXNAMELEN - 1 },
6392        [NFTA_OBJ_TYPE]         = { .type = NLA_U32 },
6393        [NFTA_OBJ_DATA]         = { .type = NLA_NESTED },
6394        [NFTA_OBJ_HANDLE]       = { .type = NLA_U64},
6395        [NFTA_OBJ_USERDATA]     = { .type = NLA_BINARY,
6396                                    .len = NFT_USERDATA_MAXLEN },
6397};
6398
6399static struct nft_object *nft_obj_init(const struct nft_ctx *ctx,
6400                                       const struct nft_object_type *type,
6401                                       const struct nlattr *attr)
6402{
6403        struct nlattr **tb;
6404        const struct nft_object_ops *ops;
6405        struct nft_object *obj;
6406        int err = -ENOMEM;
6407
6408        tb = kmalloc_array(type->maxattr + 1, sizeof(*tb), GFP_KERNEL);
6409        if (!tb)
6410                goto err1;
6411
6412        if (attr) {
6413                err = nla_parse_nested_deprecated(tb, type->maxattr, attr,
6414                                                  type->policy, NULL);
6415                if (err < 0)
6416                        goto err2;
6417        } else {
6418                memset(tb, 0, sizeof(tb[0]) * (type->maxattr + 1));
6419        }
6420
6421        if (type->select_ops) {
6422                ops = type->select_ops(ctx, (const struct nlattr * const *)tb);
6423                if (IS_ERR(ops)) {
6424                        err = PTR_ERR(ops);
6425                        goto err2;
6426                }
6427        } else {
6428                ops = type->ops;
6429        }
6430
6431        err = -ENOMEM;
6432        obj = kzalloc(sizeof(*obj) + ops->size, GFP_KERNEL);
6433        if (!obj)
6434                goto err2;
6435
6436        err = ops->init(ctx, (const struct nlattr * const *)tb, obj);
6437        if (err < 0)
6438                goto err3;
6439
6440        obj->ops = ops;
6441
6442        kfree(tb);
6443        return obj;
6444err3:
6445        kfree(obj);
6446err2:
6447        kfree(tb);
6448err1:
6449        return ERR_PTR(err);
6450}
6451
6452static int nft_object_dump(struct sk_buff *skb, unsigned int attr,
6453                           struct nft_object *obj, bool reset)
6454{
6455        struct nlattr *nest;
6456
6457        nest = nla_nest_start_noflag(skb, attr);
6458        if (!nest)
6459                goto nla_put_failure;
6460        if (obj->ops->dump(skb, obj, reset) < 0)
6461                goto nla_put_failure;
6462        nla_nest_end(skb, nest);
6463        return 0;
6464
6465nla_put_failure:
6466        return -1;
6467}
6468
6469static const struct nft_object_type *__nft_obj_type_get(u32 objtype)
6470{
6471        const struct nft_object_type *type;
6472
6473        list_for_each_entry(type, &nf_tables_objects, list) {
6474                if (objtype == type->type)
6475                        return type;
6476        }
6477        return NULL;
6478}
6479
6480static const struct nft_object_type *
6481nft_obj_type_get(struct net *net, u32 objtype)
6482{
6483        const struct nft_object_type *type;
6484
6485        type = __nft_obj_type_get(objtype);
6486        if (type != NULL && try_module_get(type->owner))
6487                return type;
6488
6489        lockdep_nfnl_nft_mutex_not_held();
6490#ifdef CONFIG_MODULES
6491        if (type == NULL) {
6492                if (nft_request_module(net, "nft-obj-%u", objtype) == -EAGAIN)
6493                        return ERR_PTR(-EAGAIN);
6494        }
6495#endif
6496        return ERR_PTR(-ENOENT);
6497}
6498
6499static int nf_tables_updobj(const struct nft_ctx *ctx,
6500                            const struct nft_object_type *type,
6501                            const struct nlattr *attr,
6502                            struct nft_object *obj)
6503{
6504        struct nft_object *newobj;
6505        struct nft_trans *trans;
6506        int err;
6507
6508        trans = nft_trans_alloc(ctx, NFT_MSG_NEWOBJ,
6509                                sizeof(struct nft_trans_obj));
6510        if (!trans)
6511                return -ENOMEM;
6512
6513        newobj = nft_obj_init(ctx, type, attr);
6514        if (IS_ERR(newobj)) {
6515                err = PTR_ERR(newobj);
6516                goto err_free_trans;
6517        }
6518
6519        nft_trans_obj(trans) = obj;
6520        nft_trans_obj_update(trans) = true;
6521        nft_trans_obj_newobj(trans) = newobj;
6522        nft_trans_commit_list_add_tail(ctx->net, trans);
6523
6524        return 0;
6525
6526err_free_trans:
6527        kfree(trans);
6528        return err;
6529}
6530
6531static int nf_tables_newobj(struct sk_buff *skb, const struct nfnl_info *info,
6532                            const struct nlattr * const nla[])
6533{
6534        struct netlink_ext_ack *extack = info->extack;
6535        u8 genmask = nft_genmask_next(info->net);
6536        u8 family = info->nfmsg->nfgen_family;
6537        const struct nft_object_type *type;
6538        struct net *net = info->net;
6539        struct nft_table *table;
6540        struct nft_object *obj;
6541        struct nft_ctx ctx;
6542        u32 objtype;
6543        int err;
6544
6545        if (!nla[NFTA_OBJ_TYPE] ||
6546            !nla[NFTA_OBJ_NAME] ||
6547            !nla[NFTA_OBJ_DATA])
6548                return -EINVAL;
6549
6550        table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
6551                                 NETLINK_CB(skb).portid);
6552        if (IS_ERR(table)) {
6553                NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6554                return PTR_ERR(table);
6555        }
6556
6557        objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6558        obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
6559        if (IS_ERR(obj)) {
6560                err = PTR_ERR(obj);
6561                if (err != -ENOENT) {
6562                        NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6563                        return err;
6564                }
6565        } else {
6566                if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
6567                        NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6568                        return -EEXIST;
6569                }
6570                if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
6571                        return -EOPNOTSUPP;
6572
6573                type = __nft_obj_type_get(objtype);
6574                nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6575
6576                return nf_tables_updobj(&ctx, type, nla[NFTA_OBJ_DATA], obj);
6577        }
6578
6579        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6580
6581        type = nft_obj_type_get(net, objtype);
6582        if (IS_ERR(type))
6583                return PTR_ERR(type);
6584
6585        obj = nft_obj_init(&ctx, type, nla[NFTA_OBJ_DATA]);
6586        if (IS_ERR(obj)) {
6587                err = PTR_ERR(obj);
6588                goto err_init;
6589        }
6590        obj->key.table = table;
6591        obj->handle = nf_tables_alloc_handle(table);
6592
6593        obj->key.name = nla_strdup(nla[NFTA_OBJ_NAME], GFP_KERNEL);
6594        if (!obj->key.name) {
6595                err = -ENOMEM;
6596                goto err_strdup;
6597        }
6598
6599        if (nla[NFTA_OBJ_USERDATA]) {
6600                obj->udata = nla_memdup(nla[NFTA_OBJ_USERDATA], GFP_KERNEL);
6601                if (obj->udata == NULL)
6602                        goto err_userdata;
6603
6604                obj->udlen = nla_len(nla[NFTA_OBJ_USERDATA]);
6605        }
6606
6607        err = nft_trans_obj_add(&ctx, NFT_MSG_NEWOBJ, obj);
6608        if (err < 0)
6609                goto err_trans;
6610
6611        err = rhltable_insert(&nft_objname_ht, &obj->rhlhead,
6612                              nft_objname_ht_params);
6613        if (err < 0)
6614                goto err_obj_ht;
6615
6616        list_add_tail_rcu(&obj->list, &table->objects);
6617        table->use++;
6618        return 0;
6619err_obj_ht:
6620        /* queued in transaction log */
6621        INIT_LIST_HEAD(&obj->list);
6622        return err;
6623err_trans:
6624        kfree(obj->udata);
6625err_userdata:
6626        kfree(obj->key.name);
6627err_strdup:
6628        if (obj->ops->destroy)
6629                obj->ops->destroy(&ctx, obj);
6630        kfree(obj);
6631err_init:
6632        module_put(type->owner);
6633        return err;
6634}
6635
6636static int nf_tables_fill_obj_info(struct sk_buff *skb, struct net *net,
6637                                   u32 portid, u32 seq, int event, u32 flags,
6638                                   int family, const struct nft_table *table,
6639                                   struct nft_object *obj, bool reset)
6640{
6641        struct nlmsghdr *nlh;
6642
6643        event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
6644        nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
6645                           NFNETLINK_V0, nft_base_seq(net));
6646        if (!nlh)
6647                goto nla_put_failure;
6648
6649        if (nla_put_string(skb, NFTA_OBJ_TABLE, table->name) ||
6650            nla_put_string(skb, NFTA_OBJ_NAME, obj->key.name) ||
6651            nla_put_be32(skb, NFTA_OBJ_TYPE, htonl(obj->ops->type->type)) ||
6652            nla_put_be32(skb, NFTA_OBJ_USE, htonl(obj->use)) ||
6653            nft_object_dump(skb, NFTA_OBJ_DATA, obj, reset) ||
6654            nla_put_be64(skb, NFTA_OBJ_HANDLE, cpu_to_be64(obj->handle),
6655                         NFTA_OBJ_PAD))
6656                goto nla_put_failure;
6657
6658        if (obj->udata &&
6659            nla_put(skb, NFTA_OBJ_USERDATA, obj->udlen, obj->udata))
6660                goto nla_put_failure;
6661
6662        nlmsg_end(skb, nlh);
6663        return 0;
6664
6665nla_put_failure:
6666        nlmsg_trim(skb, nlh);
6667        return -1;
6668}
6669
6670struct nft_obj_filter {
6671        char            *table;
6672        u32             type;
6673};
6674
6675static int nf_tables_dump_obj(struct sk_buff *skb, struct netlink_callback *cb)
6676{
6677        const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
6678        const struct nft_table *table;
6679        unsigned int idx = 0, s_idx = cb->args[0];
6680        struct nft_obj_filter *filter = cb->data;
6681        struct net *net = sock_net(skb->sk);
6682        int family = nfmsg->nfgen_family;
6683        struct nftables_pernet *nft_net;
6684        struct nft_object *obj;
6685        bool reset = false;
6686
6687        if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
6688                reset = true;
6689
6690        rcu_read_lock();
6691        nft_net = nft_pernet(net);
6692        cb->seq = nft_net->base_seq;
6693
6694        list_for_each_entry_rcu(table, &nft_net->tables, list) {
6695                if (family != NFPROTO_UNSPEC && family != table->family)
6696                        continue;
6697
6698                list_for_each_entry_rcu(obj, &table->objects, list) {
6699                        if (!nft_is_active(net, obj))
6700                                goto cont;
6701                        if (idx < s_idx)
6702                                goto cont;
6703                        if (idx > s_idx)
6704                                memset(&cb->args[1], 0,
6705                                       sizeof(cb->args) - sizeof(cb->args[0]));
6706                        if (filter && filter->table &&
6707                            strcmp(filter->table, table->name))
6708                                goto cont;
6709                        if (filter &&
6710                            filter->type != NFT_OBJECT_UNSPEC &&
6711                            obj->ops->type->type != filter->type)
6712                                goto cont;
6713                        if (reset) {
6714                                char *buf = kasprintf(GFP_ATOMIC,
6715                                                      "%s:%u",
6716                                                      table->name,
6717                                                      nft_net->base_seq);
6718
6719                                audit_log_nfcfg(buf,
6720                                                family,
6721                                                obj->handle,
6722                                                AUDIT_NFT_OP_OBJ_RESET,
6723                                                GFP_ATOMIC);
6724                                kfree(buf);
6725                        }
6726
6727                        if (nf_tables_fill_obj_info(skb, net, NETLINK_CB(cb->skb).portid,
6728                                                    cb->nlh->nlmsg_seq,
6729                                                    NFT_MSG_NEWOBJ,
6730                                                    NLM_F_MULTI | NLM_F_APPEND,
6731                                                    table->family, table,
6732                                                    obj, reset) < 0)
6733                                goto done;
6734
6735                        nl_dump_check_consistent(cb, nlmsg_hdr(skb));
6736cont:
6737                        idx++;
6738                }
6739        }
6740done:
6741        rcu_read_unlock();
6742
6743        cb->args[0] = idx;
6744        return skb->len;
6745}
6746
6747static int nf_tables_dump_obj_start(struct netlink_callback *cb)
6748{
6749        const struct nlattr * const *nla = cb->data;
6750        struct nft_obj_filter *filter = NULL;
6751
6752        if (nla[NFTA_OBJ_TABLE] || nla[NFTA_OBJ_TYPE]) {
6753                filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
6754                if (!filter)
6755                        return -ENOMEM;
6756
6757                if (nla[NFTA_OBJ_TABLE]) {
6758                        filter->table = nla_strdup(nla[NFTA_OBJ_TABLE], GFP_ATOMIC);
6759                        if (!filter->table) {
6760                                kfree(filter);
6761                                return -ENOMEM;
6762                        }
6763                }
6764
6765                if (nla[NFTA_OBJ_TYPE])
6766                        filter->type = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6767        }
6768
6769        cb->data = filter;
6770        return 0;
6771}
6772
6773static int nf_tables_dump_obj_done(struct netlink_callback *cb)
6774{
6775        struct nft_obj_filter *filter = cb->data;
6776
6777        if (filter) {
6778                kfree(filter->table);
6779                kfree(filter);
6780        }
6781
6782        return 0;
6783}
6784
6785/* called with rcu_read_lock held */
6786static int nf_tables_getobj(struct sk_buff *skb, const struct nfnl_info *info,
6787                            const struct nlattr * const nla[])
6788{
6789        struct netlink_ext_ack *extack = info->extack;
6790        u8 genmask = nft_genmask_cur(info->net);
6791        u8 family = info->nfmsg->nfgen_family;
6792        const struct nft_table *table;
6793        struct net *net = info->net;
6794        struct nft_object *obj;
6795        struct sk_buff *skb2;
6796        bool reset = false;
6797        u32 objtype;
6798        int err;
6799
6800        if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
6801                struct netlink_dump_control c = {
6802                        .start = nf_tables_dump_obj_start,
6803                        .dump = nf_tables_dump_obj,
6804                        .done = nf_tables_dump_obj_done,
6805                        .module = THIS_MODULE,
6806                        .data = (void *)nla,
6807                };
6808
6809                return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
6810        }
6811
6812        if (!nla[NFTA_OBJ_NAME] ||
6813            !nla[NFTA_OBJ_TYPE])
6814                return -EINVAL;
6815
6816        table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask, 0);
6817        if (IS_ERR(table)) {
6818                NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6819                return PTR_ERR(table);
6820        }
6821
6822        objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6823        obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
6824        if (IS_ERR(obj)) {
6825                NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6826                return PTR_ERR(obj);
6827        }
6828
6829        skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
6830        if (!skb2)
6831                return -ENOMEM;
6832
6833        if (NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
6834                reset = true;
6835
6836        if (reset) {
6837                const struct nftables_pernet *nft_net;
6838                char *buf;
6839
6840                nft_net = nft_pernet(net);
6841                buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, nft_net->base_seq);
6842
6843                audit_log_nfcfg(buf,
6844                                family,
6845                                obj->handle,
6846                                AUDIT_NFT_OP_OBJ_RESET,
6847                                GFP_ATOMIC);
6848                kfree(buf);
6849        }
6850
6851        err = nf_tables_fill_obj_info(skb2, net, NETLINK_CB(skb).portid,
6852                                      info->nlh->nlmsg_seq, NFT_MSG_NEWOBJ, 0,
6853                                      family, table, obj, reset);
6854        if (err < 0)
6855                goto err_fill_obj_info;
6856
6857        return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
6858
6859err_fill_obj_info:
6860        kfree_skb(skb2);
6861        return err;
6862}
6863
6864static void nft_obj_destroy(const struct nft_ctx *ctx, struct nft_object *obj)
6865{
6866        if (obj->ops->destroy)
6867                obj->ops->destroy(ctx, obj);
6868
6869        module_put(obj->ops->type->owner);
6870        kfree(obj->key.name);
6871        kfree(obj->udata);
6872        kfree(obj);
6873}
6874
6875static int nf_tables_delobj(struct sk_buff *skb, const struct nfnl_info *info,
6876                            const struct nlattr * const nla[])
6877{
6878        struct netlink_ext_ack *extack = info->extack;
6879        u8 genmask = nft_genmask_next(info->net);
6880        u8 family = info->nfmsg->nfgen_family;
6881        struct net *net = info->net;
6882        const struct nlattr *attr;
6883        struct nft_table *table;
6884        struct nft_object *obj;
6885        struct nft_ctx ctx;
6886        u32 objtype;
6887
6888        if (!nla[NFTA_OBJ_TYPE] ||
6889            (!nla[NFTA_OBJ_NAME] && !nla[NFTA_OBJ_HANDLE]))
6890                return -EINVAL;
6891
6892        table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
6893                                 NETLINK_CB(skb).portid);
6894        if (IS_ERR(table)) {
6895                NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6896                return PTR_ERR(table);
6897        }
6898
6899        objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6900        if (nla[NFTA_OBJ_HANDLE]) {
6901                attr = nla[NFTA_OBJ_HANDLE];
6902                obj = nft_obj_lookup_byhandle(table, attr, objtype, genmask);
6903        } else {
6904                attr = nla[NFTA_OBJ_NAME];
6905                obj = nft_obj_lookup(net, table, attr, objtype, genmask);
6906        }
6907
6908        if (IS_ERR(obj)) {
6909                NL_SET_BAD_ATTR(extack, attr);
6910                return PTR_ERR(obj);
6911        }
6912        if (obj->use > 0) {
6913                NL_SET_BAD_ATTR(extack, attr);
6914                return -EBUSY;
6915        }
6916
6917        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6918
6919        return nft_delobj(&ctx, obj);
6920}
6921
6922void nft_obj_notify(struct net *net, const struct nft_table *table,
6923                    struct nft_object *obj, u32 portid, u32 seq, int event,
6924                    int family, int report, gfp_t gfp)
6925{
6926        struct nftables_pernet *nft_net = nft_pernet(net);
6927        struct sk_buff *skb;
6928        int err;
6929        char *buf = kasprintf(gfp, "%s:%u",
6930                              table->name, nft_net->base_seq);
6931
6932        audit_log_nfcfg(buf,
6933                        family,
6934                        obj->handle,
6935                        event == NFT_MSG_NEWOBJ ?
6936                                 AUDIT_NFT_OP_OBJ_REGISTER :
6937                                 AUDIT_NFT_OP_OBJ_UNREGISTER,
6938                        gfp);
6939        kfree(buf);
6940
6941        if (!report &&
6942            !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
6943                return;
6944
6945        skb = nlmsg_new(NLMSG_GOODSIZE, gfp);
6946        if (skb == NULL)
6947                goto err;
6948
6949        err = nf_tables_fill_obj_info(skb, net, portid, seq, event, 0, family,
6950                                      table, obj, false);
6951        if (err < 0) {
6952                kfree_skb(skb);
6953                goto err;
6954        }
6955
6956        nft_notify_enqueue(skb, report, &nft_net->notify_list);
6957        return;
6958err:
6959        nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
6960}
6961EXPORT_SYMBOL_GPL(nft_obj_notify);
6962
6963static void nf_tables_obj_notify(const struct nft_ctx *ctx,
6964                                 struct nft_object *obj, int event)
6965{
6966        nft_obj_notify(ctx->net, ctx->table, obj, ctx->portid, ctx->seq, event,
6967                       ctx->family, ctx->report, GFP_KERNEL);
6968}
6969
6970/*
6971 * Flow tables
6972 */
6973void nft_register_flowtable_type(struct nf_flowtable_type *type)
6974{
6975        nfnl_lock(NFNL_SUBSYS_NFTABLES);
6976        list_add_tail_rcu(&type->list, &nf_tables_flowtables);
6977        nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6978}
6979EXPORT_SYMBOL_GPL(nft_register_flowtable_type);
6980
6981void nft_unregister_flowtable_type(struct nf_flowtable_type *type)
6982{
6983        nfnl_lock(NFNL_SUBSYS_NFTABLES);
6984        list_del_rcu(&type->list);
6985        nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6986}
6987EXPORT_SYMBOL_GPL(nft_unregister_flowtable_type);
6988
6989static const struct nla_policy nft_flowtable_policy[NFTA_FLOWTABLE_MAX + 1] = {
6990        [NFTA_FLOWTABLE_TABLE]          = { .type = NLA_STRING,
6991                                            .len = NFT_NAME_MAXLEN - 1 },
6992        [NFTA_FLOWTABLE_NAME]           = { .type = NLA_STRING,
6993                                            .len = NFT_NAME_MAXLEN - 1 },
6994        [NFTA_FLOWTABLE_HOOK]           = { .type = NLA_NESTED },
6995        [NFTA_FLOWTABLE_HANDLE]         = { .type = NLA_U64 },
6996        [NFTA_FLOWTABLE_FLAGS]          = { .type = NLA_U32 },
6997};
6998
6999struct nft_flowtable *nft_flowtable_lookup(const struct nft_table *table,
7000                                           const struct nlattr *nla, u8 genmask)
7001{
7002        struct nft_flowtable *flowtable;
7003
7004        list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7005                if (!nla_strcmp(nla, flowtable->name) &&
7006                    nft_active_genmask(flowtable, genmask))
7007                        return flowtable;
7008        }
7009        return ERR_PTR(-ENOENT);
7010}
7011EXPORT_SYMBOL_GPL(nft_flowtable_lookup);
7012
7013void nf_tables_deactivate_flowtable(const struct nft_ctx *ctx,
7014                                    struct nft_flowtable *flowtable,
7015                                    enum nft_trans_phase phase)
7016{
7017        switch (phase) {
7018        case NFT_TRANS_PREPARE:
7019        case NFT_TRANS_ABORT:
7020        case NFT_TRANS_RELEASE:
7021                flowtable->use--;
7022                fallthrough;
7023        default:
7024                return;
7025        }
7026}
7027EXPORT_SYMBOL_GPL(nf_tables_deactivate_flowtable);
7028
7029static struct nft_flowtable *
7030nft_flowtable_lookup_byhandle(const struct nft_table *table,
7031                              const struct nlattr *nla, u8 genmask)
7032{
7033       struct nft_flowtable *flowtable;
7034
7035       list_for_each_entry(flowtable, &table->flowtables, list) {
7036               if (be64_to_cpu(nla_get_be64(nla)) == flowtable->handle &&
7037                   nft_active_genmask(flowtable, genmask))
7038                       return flowtable;
7039       }
7040       return ERR_PTR(-ENOENT);
7041}
7042
7043struct nft_flowtable_hook {
7044        u32                     num;
7045        int                     priority;
7046        struct list_head        list;
7047};
7048
7049static const struct nla_policy nft_flowtable_hook_policy[NFTA_FLOWTABLE_HOOK_MAX + 1] = {
7050        [NFTA_FLOWTABLE_HOOK_NUM]       = { .type = NLA_U32 },
7051        [NFTA_FLOWTABLE_HOOK_PRIORITY]  = { .type = NLA_U32 },
7052        [NFTA_FLOWTABLE_HOOK_DEVS]      = { .type = NLA_NESTED },
7053};
7054
7055static int nft_flowtable_parse_hook(const struct nft_ctx *ctx,
7056                                    const struct nlattr *attr,
7057                                    struct nft_flowtable_hook *flowtable_hook,
7058                                    struct nft_flowtable *flowtable, bool add)
7059{
7060        struct nlattr *tb[NFTA_FLOWTABLE_HOOK_MAX + 1];
7061        struct nft_hook *hook;
7062        int hooknum, priority;
7063        int err;
7064
7065        INIT_LIST_HEAD(&flowtable_hook->list);
7066
7067        err = nla_parse_nested_deprecated(tb, NFTA_FLOWTABLE_HOOK_MAX, attr,
7068                                          nft_flowtable_hook_policy, NULL);
7069        if (err < 0)
7070                return err;
7071
7072        if (add) {
7073                if (!tb[NFTA_FLOWTABLE_HOOK_NUM] ||
7074                    !tb[NFTA_FLOWTABLE_HOOK_PRIORITY])
7075                        return -EINVAL;
7076
7077                hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7078                if (hooknum != NF_NETDEV_INGRESS)
7079                        return -EOPNOTSUPP;
7080
7081                priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7082
7083                flowtable_hook->priority        = priority;
7084                flowtable_hook->num             = hooknum;
7085        } else {
7086                if (tb[NFTA_FLOWTABLE_HOOK_NUM]) {
7087                        hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7088                        if (hooknum != flowtable->hooknum)
7089                                return -EOPNOTSUPP;
7090                }
7091
7092                if (tb[NFTA_FLOWTABLE_HOOK_PRIORITY]) {
7093                        priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7094                        if (priority != flowtable->data.priority)
7095                                return -EOPNOTSUPP;
7096                }
7097
7098                flowtable_hook->priority        = flowtable->data.priority;
7099                flowtable_hook->num             = flowtable->hooknum;
7100        }
7101
7102        if (tb[NFTA_FLOWTABLE_HOOK_DEVS]) {
7103                err = nf_tables_parse_netdev_hooks(ctx->net,
7104                                                   tb[NFTA_FLOWTABLE_HOOK_DEVS],
7105                                                   &flowtable_hook->list);
7106                if (err < 0)
7107                        return err;
7108        }
7109
7110        list_for_each_entry(hook, &flowtable_hook->list, list) {
7111                hook->ops.pf            = NFPROTO_NETDEV;
7112                hook->ops.hooknum       = flowtable_hook->num;
7113                hook->ops.priority      = flowtable_hook->priority;
7114                hook->ops.priv          = &flowtable->data;
7115                hook->ops.hook          = flowtable->data.type->hook;
7116        }
7117
7118        return err;
7119}
7120
7121static const struct nf_flowtable_type *__nft_flowtable_type_get(u8 family)
7122{
7123        const struct nf_flowtable_type *type;
7124
7125        list_for_each_entry(type, &nf_tables_flowtables, list) {
7126                if (family == type->family)
7127                        return type;
7128        }
7129        return NULL;
7130}
7131
7132static const struct nf_flowtable_type *
7133nft_flowtable_type_get(struct net *net, u8 family)
7134{
7135        const struct nf_flowtable_type *type;
7136
7137        type = __nft_flowtable_type_get(family);
7138        if (type != NULL && try_module_get(type->owner))
7139                return type;
7140
7141        lockdep_nfnl_nft_mutex_not_held();
7142#ifdef CONFIG_MODULES
7143        if (type == NULL) {
7144                if (nft_request_module(net, "nf-flowtable-%u", family) == -EAGAIN)
7145                        return ERR_PTR(-EAGAIN);
7146        }
7147#endif
7148        return ERR_PTR(-ENOENT);
7149}
7150
7151/* Only called from error and netdev event paths. */
7152static void nft_unregister_flowtable_hook(struct net *net,
7153                                          struct nft_flowtable *flowtable,
7154                                          struct nft_hook *hook)
7155{
7156        nf_unregister_net_hook(net, &hook->ops);
7157        flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7158                                    FLOW_BLOCK_UNBIND);
7159}
7160
7161static void nft_unregister_flowtable_net_hooks(struct net *net,
7162                                               struct list_head *hook_list)
7163{
7164        struct nft_hook *hook;
7165
7166        list_for_each_entry(hook, hook_list, list)
7167                nf_unregister_net_hook(net, &hook->ops);
7168}
7169
7170static int nft_register_flowtable_net_hooks(struct net *net,
7171                                            struct nft_table *table,
7172                                            struct list_head *hook_list,
7173                                            struct nft_flowtable *flowtable)
7174{
7175        struct nft_hook *hook, *hook2, *next;
7176        struct nft_flowtable *ft;
7177        int err, i = 0;
7178
7179        list_for_each_entry(hook, hook_list, list) {
7180                list_for_each_entry(ft, &table->flowtables, list) {
7181                        if (!nft_is_active_next(net, ft))
7182                                continue;
7183
7184                        list_for_each_entry(hook2, &ft->hook_list, list) {
7185                                if (hook->ops.dev == hook2->ops.dev &&
7186                                    hook->ops.pf == hook2->ops.pf) {
7187                                        err = -EEXIST;
7188                                        goto err_unregister_net_hooks;
7189                                }
7190                        }
7191                }
7192
7193                err = flowtable->data.type->setup(&flowtable->data,
7194                                                  hook->ops.dev,
7195                                                  FLOW_BLOCK_BIND);
7196                if (err < 0)
7197                        goto err_unregister_net_hooks;
7198
7199                err = nf_register_net_hook(net, &hook->ops);
7200                if (err < 0) {
7201                        flowtable->data.type->setup(&flowtable->data,
7202                                                    hook->ops.dev,
7203                                                    FLOW_BLOCK_UNBIND);
7204                        goto err_unregister_net_hooks;
7205                }
7206
7207                i++;
7208        }
7209
7210        return 0;
7211
7212err_unregister_net_hooks:
7213        list_for_each_entry_safe(hook, next, hook_list, list) {
7214                if (i-- <= 0)
7215                        break;
7216
7217                nft_unregister_flowtable_hook(net, flowtable, hook);
7218                list_del_rcu(&hook->list);
7219                kfree_rcu(hook, rcu);
7220        }
7221
7222        return err;
7223}
7224
7225static void nft_flowtable_hooks_destroy(struct list_head *hook_list)
7226{
7227        struct nft_hook *hook, *next;
7228
7229        list_for_each_entry_safe(hook, next, hook_list, list) {
7230                list_del_rcu(&hook->list);
7231                kfree_rcu(hook, rcu);
7232        }
7233}
7234
7235static int nft_flowtable_update(struct nft_ctx *ctx, const struct nlmsghdr *nlh,
7236                                struct nft_flowtable *flowtable)
7237{
7238        const struct nlattr * const *nla = ctx->nla;
7239        struct nft_flowtable_hook flowtable_hook;
7240        struct nft_hook *hook, *next;
7241        struct nft_trans *trans;
7242        bool unregister = false;
7243        u32 flags;
7244        int err;
7245
7246        err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7247                                       &flowtable_hook, flowtable, false);
7248        if (err < 0)
7249                return err;
7250
7251        list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7252                if (nft_hook_list_find(&flowtable->hook_list, hook)) {
7253                        list_del(&hook->list);
7254                        kfree(hook);
7255                }
7256        }
7257
7258        if (nla[NFTA_FLOWTABLE_FLAGS]) {
7259                flags = ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7260                if (flags & ~NFT_FLOWTABLE_MASK)
7261                        return -EOPNOTSUPP;
7262                if ((flowtable->data.flags & NFT_FLOWTABLE_HW_OFFLOAD) ^
7263                    (flags & NFT_FLOWTABLE_HW_OFFLOAD))
7264                        return -EOPNOTSUPP;
7265        } else {
7266                flags = flowtable->data.flags;
7267        }
7268
7269        err = nft_register_flowtable_net_hooks(ctx->net, ctx->table,
7270                                               &flowtable_hook.list, flowtable);
7271        if (err < 0)
7272                goto err_flowtable_update_hook;
7273
7274        trans = nft_trans_alloc(ctx, NFT_MSG_NEWFLOWTABLE,
7275                                sizeof(struct nft_trans_flowtable));
7276        if (!trans) {
7277                unregister = true;
7278                err = -ENOMEM;
7279                goto err_flowtable_update_hook;
7280        }
7281
7282        nft_trans_flowtable_flags(trans) = flags;
7283        nft_trans_flowtable(trans) = flowtable;
7284        nft_trans_flowtable_update(trans) = true;
7285        INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7286        list_splice(&flowtable_hook.list, &nft_trans_flowtable_hooks(trans));
7287
7288        nft_trans_commit_list_add_tail(ctx->net, trans);
7289
7290        return 0;
7291
7292err_flowtable_update_hook:
7293        list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7294                if (unregister)
7295                        nft_unregister_flowtable_hook(ctx->net, flowtable, hook);
7296                list_del_rcu(&hook->list);
7297                kfree_rcu(hook, rcu);
7298        }
7299
7300        return err;
7301
7302}
7303
7304static int nf_tables_newflowtable(struct sk_buff *skb,
7305                                  const struct nfnl_info *info,
7306                                  const struct nlattr * const nla[])
7307{
7308        struct netlink_ext_ack *extack = info->extack;
7309        struct nft_flowtable_hook flowtable_hook;
7310        u8 genmask = nft_genmask_next(info->net);
7311        u8 family = info->nfmsg->nfgen_family;
7312        const struct nf_flowtable_type *type;
7313        struct nft_flowtable *flowtable;
7314        struct nft_hook *hook, *next;
7315        struct net *net = info->net;
7316        struct nft_table *table;
7317        struct nft_ctx ctx;
7318        int err;
7319
7320        if (!nla[NFTA_FLOWTABLE_TABLE] ||
7321            !nla[NFTA_FLOWTABLE_NAME] ||
7322            !nla[NFTA_FLOWTABLE_HOOK])
7323                return -EINVAL;
7324
7325        table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7326                                 genmask, NETLINK_CB(skb).portid);
7327        if (IS_ERR(table)) {
7328                NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7329                return PTR_ERR(table);
7330        }
7331
7332        flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7333                                         genmask);
7334        if (IS_ERR(flowtable)) {
7335                err = PTR_ERR(flowtable);
7336                if (err != -ENOENT) {
7337                        NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7338                        return err;
7339                }
7340        } else {
7341                if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
7342                        NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7343                        return -EEXIST;
7344                }
7345
7346                nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7347
7348                return nft_flowtable_update(&ctx, info->nlh, flowtable);
7349        }
7350
7351        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7352
7353        flowtable = kzalloc(sizeof(*flowtable), GFP_KERNEL);
7354        if (!flowtable)
7355                return -ENOMEM;
7356
7357        flowtable->table = table;
7358        flowtable->handle = nf_tables_alloc_handle(table);
7359        INIT_LIST_HEAD(&flowtable->hook_list);
7360
7361        flowtable->name = nla_strdup(nla[NFTA_FLOWTABLE_NAME], GFP_KERNEL);
7362        if (!flowtable->name) {
7363                err = -ENOMEM;
7364                goto err1;
7365        }
7366
7367        type = nft_flowtable_type_get(net, family);
7368        if (IS_ERR(type)) {
7369                err = PTR_ERR(type);
7370                goto err2;
7371        }
7372
7373        if (nla[NFTA_FLOWTABLE_FLAGS]) {
7374                flowtable->data.flags =
7375                        ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7376                if (flowtable->data.flags & ~NFT_FLOWTABLE_MASK) {
7377                        err = -EOPNOTSUPP;
7378                        goto err3;
7379                }
7380        }
7381
7382        write_pnet(&flowtable->data.net, net);
7383        flowtable->data.type = type;
7384        err = type->init(&flowtable->data);
7385        if (err < 0)
7386                goto err3;
7387
7388        err = nft_flowtable_parse_hook(&ctx, nla[NFTA_FLOWTABLE_HOOK],
7389                                       &flowtable_hook, flowtable, true);
7390        if (err < 0)
7391                goto err4;
7392
7393        list_splice(&flowtable_hook.list, &flowtable->hook_list);
7394        flowtable->data.priority = flowtable_hook.priority;
7395        flowtable->hooknum = flowtable_hook.num;
7396
7397        err = nft_register_flowtable_net_hooks(ctx.net, table,
7398                                               &flowtable->hook_list,
7399                                               flowtable);
7400        if (err < 0) {
7401                nft_flowtable_hooks_destroy(&flowtable->hook_list);
7402                goto err4;
7403        }
7404
7405        err = nft_trans_flowtable_add(&ctx, NFT_MSG_NEWFLOWTABLE, flowtable);
7406        if (err < 0)
7407                goto err5;
7408
7409        list_add_tail_rcu(&flowtable->list, &table->flowtables);
7410        table->use++;
7411
7412        return 0;
7413err5:
7414        list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7415                nft_unregister_flowtable_hook(net, flowtable, hook);
7416                list_del_rcu(&hook->list);
7417                kfree_rcu(hook, rcu);
7418        }
7419err4:
7420        flowtable->data.type->free(&flowtable->data);
7421err3:
7422        module_put(type->owner);
7423err2:
7424        kfree(flowtable->name);
7425err1:
7426        kfree(flowtable);
7427        return err;
7428}
7429
7430static void nft_flowtable_hook_release(struct nft_flowtable_hook *flowtable_hook)
7431{
7432        struct nft_hook *this, *next;
7433
7434        list_for_each_entry_safe(this, next, &flowtable_hook->list, list) {
7435                list_del(&this->list);
7436                kfree(this);
7437        }
7438}
7439
7440static int nft_delflowtable_hook(struct nft_ctx *ctx,
7441                                 struct nft_flowtable *flowtable)
7442{
7443        const struct nlattr * const *nla = ctx->nla;
7444        struct nft_flowtable_hook flowtable_hook;
7445        struct nft_hook *this, *hook;
7446        struct nft_trans *trans;
7447        int err;
7448
7449        err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7450                                       &flowtable_hook, flowtable, false);
7451        if (err < 0)
7452                return err;
7453
7454        list_for_each_entry(this, &flowtable_hook.list, list) {
7455                hook = nft_hook_list_find(&flowtable->hook_list, this);
7456                if (!hook) {
7457                        err = -ENOENT;
7458                        goto err_flowtable_del_hook;
7459                }
7460                hook->inactive = true;
7461        }
7462
7463        trans = nft_trans_alloc(ctx, NFT_MSG_DELFLOWTABLE,
7464                                sizeof(struct nft_trans_flowtable));
7465        if (!trans) {
7466                err = -ENOMEM;
7467                goto err_flowtable_del_hook;
7468        }
7469
7470        nft_trans_flowtable(trans) = flowtable;
7471        nft_trans_flowtable_update(trans) = true;
7472        INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7473        nft_flowtable_hook_release(&flowtable_hook);
7474
7475        nft_trans_commit_list_add_tail(ctx->net, trans);
7476
7477        return 0;
7478
7479err_flowtable_del_hook:
7480        list_for_each_entry(this, &flowtable_hook.list, list) {
7481                hook = nft_hook_list_find(&flowtable->hook_list, this);
7482                if (!hook)
7483                        break;
7484
7485                hook->inactive = false;
7486        }
7487        nft_flowtable_hook_release(&flowtable_hook);
7488
7489        return err;
7490}
7491
7492static int nf_tables_delflowtable(struct sk_buff *skb,
7493                                  const struct nfnl_info *info,
7494                                  const struct nlattr * const nla[])
7495{
7496        struct netlink_ext_ack *extack = info->extack;
7497        u8 genmask = nft_genmask_next(info->net);
7498        u8 family = info->nfmsg->nfgen_family;
7499        struct nft_flowtable *flowtable;
7500        struct net *net = info->net;
7501        const struct nlattr *attr;
7502        struct nft_table *table;
7503        struct nft_ctx ctx;
7504
7505        if (!nla[NFTA_FLOWTABLE_TABLE] ||
7506            (!nla[NFTA_FLOWTABLE_NAME] &&
7507             !nla[NFTA_FLOWTABLE_HANDLE]))
7508                return -EINVAL;
7509
7510        table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7511                                 genmask, NETLINK_CB(skb).portid);
7512        if (IS_ERR(table)) {
7513                NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7514                return PTR_ERR(table);
7515        }
7516
7517        if (nla[NFTA_FLOWTABLE_HANDLE]) {
7518                attr = nla[NFTA_FLOWTABLE_HANDLE];
7519                flowtable = nft_flowtable_lookup_byhandle(table, attr, genmask);
7520        } else {
7521                attr = nla[NFTA_FLOWTABLE_NAME];
7522                flowtable = nft_flowtable_lookup(table, attr, genmask);
7523        }
7524
7525        if (IS_ERR(flowtable)) {
7526                NL_SET_BAD_ATTR(extack, attr);
7527                return PTR_ERR(flowtable);
7528        }
7529
7530        nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7531
7532        if (nla[NFTA_FLOWTABLE_HOOK])
7533                return nft_delflowtable_hook(&ctx, flowtable);
7534
7535        if (flowtable->use > 0) {
7536                NL_SET_BAD_ATTR(extack, attr);
7537                return -EBUSY;
7538        }
7539
7540        return nft_delflowtable(&ctx, flowtable);
7541}
7542
7543static int nf_tables_fill_flowtable_info(struct sk_buff *skb, struct net *net,
7544                                         u32 portid, u32 seq, int event,
7545                                         u32 flags, int family,
7546                                         struct nft_flowtable *flowtable,
7547                                         struct list_head *hook_list)
7548{
7549        struct nlattr *nest, *nest_devs;
7550        struct nft_hook *hook;
7551        struct nlmsghdr *nlh;
7552
7553        event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
7554        nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
7555                           NFNETLINK_V0, nft_base_seq(net));
7556        if (!nlh)
7557                goto nla_put_failure;
7558
7559        if (nla_put_string(skb, NFTA_FLOWTABLE_TABLE, flowtable->table->name) ||
7560            nla_put_string(skb, NFTA_FLOWTABLE_NAME, flowtable->name) ||
7561            nla_put_be32(skb, NFTA_FLOWTABLE_USE, htonl(flowtable->use)) ||
7562            nla_put_be64(skb, NFTA_FLOWTABLE_HANDLE, cpu_to_be64(flowtable->handle),
7563                         NFTA_FLOWTABLE_PAD) ||
7564            nla_put_be32(skb, NFTA_FLOWTABLE_FLAGS, htonl(flowtable->data.flags)))
7565                goto nla_put_failure;
7566
7567        nest = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK);
7568        if (!nest)
7569                goto nla_put_failure;
7570        if (nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_NUM, htonl(flowtable->hooknum)) ||
7571            nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_PRIORITY, htonl(flowtable->data.priority)))
7572                goto nla_put_failure;
7573
7574        nest_devs = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK_DEVS);
7575        if (!nest_devs)
7576                goto nla_put_failure;
7577
7578        list_for_each_entry_rcu(hook, hook_list, list) {
7579                if (nla_put_string(skb, NFTA_DEVICE_NAME, hook->ops.dev->name))
7580                        goto nla_put_failure;
7581        }
7582        nla_nest_end(skb, nest_devs);
7583        nla_nest_end(skb, nest);
7584
7585        nlmsg_end(skb, nlh);
7586        return 0;
7587
7588nla_put_failure:
7589        nlmsg_trim(skb, nlh);
7590        return -1;
7591}
7592
7593struct nft_flowtable_filter {
7594        char            *table;
7595};
7596
7597static int nf_tables_dump_flowtable(struct sk_buff *skb,
7598                                    struct netlink_callback *cb)
7599{
7600        const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
7601        struct nft_flowtable_filter *filter = cb->data;
7602        unsigned int idx = 0, s_idx = cb->args[0];
7603        struct net *net = sock_net(skb->sk);
7604        int family = nfmsg->nfgen_family;
7605        struct nft_flowtable *flowtable;
7606        struct nftables_pernet *nft_net;
7607        const struct nft_table *table;
7608
7609        rcu_read_lock();
7610        nft_net = nft_pernet(net);
7611        cb->seq = nft_net->base_seq;
7612
7613        list_for_each_entry_rcu(table, &nft_net->tables, list) {
7614                if (family != NFPROTO_UNSPEC && family != table->family)
7615                        continue;
7616
7617                list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7618                        if (!nft_is_active(net, flowtable))
7619                                goto cont;
7620                        if (idx < s_idx)
7621                                goto cont;
7622                        if (idx > s_idx)
7623                                memset(&cb->args[1], 0,
7624                                       sizeof(cb->args) - sizeof(cb->args[0]));
7625                        if (filter && filter->table &&
7626                            strcmp(filter->table, table->name))
7627                                goto cont;
7628
7629                        if (nf_tables_fill_flowtable_info(skb, net, NETLINK_CB(cb->skb).portid,
7630                                                          cb->nlh->nlmsg_seq,
7631                                                          NFT_MSG_NEWFLOWTABLE,
7632                                                          NLM_F_MULTI | NLM_F_APPEND,
7633                                                          table->family,
7634                                                          flowtable,
7635                                                          &flowtable->hook_list) < 0)
7636                                goto done;
7637
7638                        nl_dump_check_consistent(cb, nlmsg_hdr(skb));
7639cont:
7640                        idx++;
7641                }
7642        }
7643done:
7644        rcu_read_unlock();
7645
7646        cb->args[0] = idx;
7647        return skb->len;
7648}
7649
7650static int nf_tables_dump_flowtable_start(struct netlink_callback *cb)
7651{
7652        const struct nlattr * const *nla = cb->data;
7653        struct nft_flowtable_filter *filter = NULL;
7654
7655        if (nla[NFTA_FLOWTABLE_TABLE]) {
7656                filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
7657                if (!filter)
7658                        return -ENOMEM;
7659
7660                filter->table = nla_strdup(nla[NFTA_FLOWTABLE_TABLE],
7661                                           GFP_ATOMIC);
7662                if (!filter->table) {
7663                        kfree(filter);
7664                        return -ENOMEM;
7665                }
7666        }
7667
7668        cb->data = filter;
7669        return 0;
7670}
7671
7672static int nf_tables_dump_flowtable_done(struct netlink_callback *cb)
7673{
7674        struct nft_flowtable_filter *filter = cb->data;
7675
7676        if (!filter)
7677                return 0;
7678
7679        kfree(filter->table);
7680        kfree(filter);
7681
7682        return 0;
7683}
7684
7685/* called with rcu_read_lock held */
7686static int nf_tables_getflowtable(struct sk_buff *skb,
7687                                  const struct nfnl_info *info,
7688                                  const struct nlattr * const nla[])
7689{
7690        u8 genmask = nft_genmask_cur(info->net);
7691        u8 family = info->nfmsg->nfgen_family;
7692        struct nft_flowtable *flowtable;
7693        const struct nft_table *table;
7694        struct net *net = info->net;
7695        struct sk_buff *skb2;
7696        int err;
7697
7698        if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
7699                struct netlink_dump_control c = {
7700                        .start = nf_tables_dump_flowtable_start,
7701                        .dump = nf_tables_dump_flowtable,
7702                        .done = nf_tables_dump_flowtable_done,
7703                        .module = THIS_MODULE,
7704                        .data = (void *)nla,
7705                };
7706
7707                return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
7708        }
7709
7710        if (!nla[NFTA_FLOWTABLE_NAME])
7711                return -EINVAL;
7712
7713        table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7714                                 genmask, 0);
7715        if (IS_ERR(table))
7716                return PTR_ERR(table);
7717
7718        flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7719                                         genmask);
7720        if (IS_ERR(flowtable))
7721                return PTR_ERR(flowtable);
7722
7723        skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7724        if (!skb2)
7725                return -ENOMEM;
7726
7727        err = nf_tables_fill_flowtable_info(skb2, net, NETLINK_CB(skb).portid,
7728                                            info->nlh->nlmsg_seq,
7729                                            NFT_MSG_NEWFLOWTABLE, 0, family,
7730                                            flowtable, &flowtable->hook_list);
7731        if (err < 0)
7732                goto err_fill_flowtable_info;
7733
7734        return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
7735
7736err_fill_flowtable_info:
7737        kfree_skb(skb2);
7738        return err;
7739}
7740
7741static void nf_tables_flowtable_notify(struct nft_ctx *ctx,
7742                                       struct nft_flowtable *flowtable,
7743                                       struct list_head *hook_list,
7744                                       int event)
7745{
7746        struct nftables_pernet *nft_net = nft_pernet(ctx->net);
7747        struct sk_buff *skb;
7748        int err;
7749
7750        if (!ctx->report &&
7751            !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
7752                return;
7753
7754        skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
7755        if (skb == NULL)
7756                goto err;
7757
7758        err = nf_tables_fill_flowtable_info(skb, ctx->net, ctx->portid,
7759                                            ctx->seq, event, 0,
7760                                            ctx->family, flowtable, hook_list);
7761        if (err < 0) {
7762                kfree_skb(skb);
7763                goto err;
7764        }
7765
7766        nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
7767        return;
7768err:
7769        nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
7770}
7771
7772static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable)
7773{
7774        struct nft_hook *hook, *next;
7775
7776        flowtable->data.type->free(&flowtable->data);
7777        list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7778                flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7779                                            FLOW_BLOCK_UNBIND);
7780                list_del_rcu(&hook->list);
7781                kfree(hook);
7782        }
7783        kfree(flowtable->name);
7784        module_put(flowtable->data.type->owner);
7785        kfree(flowtable);
7786}
7787
7788static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
7789                                   u32 portid, u32 seq)
7790{
7791        struct nftables_pernet *nft_net = nft_pernet(net);
7792        struct nlmsghdr *nlh;
7793        char buf[TASK_COMM_LEN];
7794        int event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWGEN);
7795
7796        nlh = nfnl_msg_put(skb, portid, seq, event, 0, AF_UNSPEC,
7797                           NFNETLINK_V0, nft_base_seq(net));
7798        if (!nlh)
7799                goto nla_put_failure;
7800
7801        if (nla_put_be32(skb, NFTA_GEN_ID, htonl(nft_net->base_seq)) ||
7802            nla_put_be32(skb, NFTA_GEN_PROC_PID, htonl(task_pid_nr(current))) ||
7803            nla_put_string(skb, NFTA_GEN_PROC_NAME, get_task_comm(buf, current)))
7804                goto nla_put_failure;
7805
7806        nlmsg_end(skb, nlh);
7807        return 0;
7808
7809nla_put_failure:
7810        nlmsg_trim(skb, nlh);
7811        return -EMSGSIZE;
7812}
7813
7814static void nft_flowtable_event(unsigned long event, struct net_device *dev,
7815                                struct nft_flowtable *flowtable)
7816{
7817        struct nft_hook *hook;
7818
7819        list_for_each_entry(hook, &flowtable->hook_list, list) {
7820                if (hook->ops.dev != dev)
7821                        continue;
7822
7823                /* flow_offload_netdev_event() cleans up entries for us. */
7824                nft_unregister_flowtable_hook(dev_net(dev), flowtable, hook);
7825                list_del_rcu(&hook->list);
7826                kfree_rcu(hook, rcu);
7827                break;
7828        }
7829}
7830
7831static int nf_tables_flowtable_event(struct notifier_block *this,
7832                                     unsigned long event, void *ptr)
7833{
7834        struct net_device *dev = netdev_notifier_info_to_dev(ptr);
7835        struct nft_flowtable *flowtable;
7836        struct nftables_pernet *nft_net;
7837        struct nft_table *table;
7838        struct net *net;
7839
7840        if (event != NETDEV_UNREGISTER)
7841                return 0;
7842
7843        net = dev_net(dev);
7844        nft_net = nft_pernet(net);
7845        mutex_lock(&nft_net->commit_mutex);
7846        list_for_each_entry(table, &nft_net->tables, list) {
7847                list_for_each_entry(flowtable, &table->flowtables, list) {
7848                        nft_flowtable_event(event, dev, flowtable);
7849                }
7850        }
7851        mutex_unlock(&nft_net->commit_mutex);
7852
7853        return NOTIFY_DONE;
7854}
7855
7856static struct notifier_block nf_tables_flowtable_notifier = {
7857        .notifier_call  = nf_tables_flowtable_event,
7858};
7859
7860static void nf_tables_gen_notify(struct net *net, struct sk_buff *skb,
7861                                 int event)
7862{
7863        struct nlmsghdr *nlh = nlmsg_hdr(skb);
7864        struct sk_buff *skb2;
7865        int err;
7866
7867        if (!nlmsg_report(nlh) &&
7868            !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
7869                return;
7870
7871        skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
7872        if (skb2 == NULL)
7873                goto err;
7874
7875        err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
7876                                      nlh->nlmsg_seq);
7877        if (err < 0) {
7878                kfree_skb(skb2);
7879                goto err;
7880        }
7881
7882        nfnetlink_send(skb2, net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
7883                       nlmsg_report(nlh), GFP_KERNEL);
7884        return;
7885err:
7886        nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
7887                          -ENOBUFS);
7888}
7889
7890static int nf_tables_getgen(struct sk_buff *skb, const struct nfnl_info *info,
7891                            const struct nlattr * const nla[])
7892{
7893        struct sk_buff *skb2;
7894        int err;
7895
7896        skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7897        if (skb2 == NULL)
7898                return -ENOMEM;
7899
7900        err = nf_tables_fill_gen_info(skb2, info->net, NETLINK_CB(skb).portid,
7901                                      info->nlh->nlmsg_seq);
7902        if (err < 0)
7903                goto err_fill_gen_info;
7904
7905        return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
7906
7907err_fill_gen_info:
7908        kfree_skb(skb2);
7909        return err;
7910}
7911
7912static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
7913        [NFT_MSG_NEWTABLE] = {
7914                .call           = nf_tables_newtable,
7915                .type           = NFNL_CB_BATCH,
7916                .attr_count     = NFTA_TABLE_MAX,
7917                .policy         = nft_table_policy,
7918        },
7919        [NFT_MSG_GETTABLE] = {
7920                .call           = nf_tables_gettable,
7921                .type           = NFNL_CB_RCU,
7922                .attr_count     = NFTA_TABLE_MAX,
7923                .policy         = nft_table_policy,
7924        },
7925        [NFT_MSG_DELTABLE] = {
7926                .call           = nf_tables_deltable,
7927                .type           = NFNL_CB_BATCH,
7928                .attr_count     = NFTA_TABLE_MAX,
7929                .policy         = nft_table_policy,
7930        },
7931        [NFT_MSG_NEWCHAIN] = {
7932                .call           = nf_tables_newchain,
7933                .type           = NFNL_CB_BATCH,
7934                .attr_count     = NFTA_CHAIN_MAX,
7935                .policy         = nft_chain_policy,
7936        },
7937        [NFT_MSG_GETCHAIN] = {
7938                .call           = nf_tables_getchain,
7939                .type           = NFNL_CB_RCU,
7940                .attr_count     = NFTA_CHAIN_MAX,
7941                .policy         = nft_chain_policy,
7942        },
7943        [NFT_MSG_DELCHAIN] = {
7944                .call           = nf_tables_delchain,
7945                .type           = NFNL_CB_BATCH,
7946                .attr_count     = NFTA_CHAIN_MAX,
7947                .policy         = nft_chain_policy,
7948        },
7949        [NFT_MSG_NEWRULE] = {
7950                .call           = nf_tables_newrule,
7951                .type           = NFNL_CB_BATCH,
7952                .attr_count     = NFTA_RULE_MAX,
7953                .policy         = nft_rule_policy,
7954        },
7955        [NFT_MSG_GETRULE] = {
7956                .call           = nf_tables_getrule,
7957                .type           = NFNL_CB_RCU,
7958                .attr_count     = NFTA_RULE_MAX,
7959                .policy         = nft_rule_policy,
7960        },
7961        [NFT_MSG_DELRULE] = {
7962                .call           = nf_tables_delrule,
7963                .type           = NFNL_CB_BATCH,
7964                .attr_count     = NFTA_RULE_MAX,
7965                .policy         = nft_rule_policy,
7966        },
7967        [NFT_MSG_NEWSET] = {
7968                .call           = nf_tables_newset,
7969                .type           = NFNL_CB_BATCH,
7970                .attr_count     = NFTA_SET_MAX,
7971                .policy         = nft_set_policy,
7972        },
7973        [NFT_MSG_GETSET] = {
7974                .call           = nf_tables_getset,
7975                .type           = NFNL_CB_RCU,
7976                .attr_count     = NFTA_SET_MAX,
7977                .policy         = nft_set_policy,
7978        },
7979        [NFT_MSG_DELSET] = {
7980                .call           = nf_tables_delset,
7981                .type           = NFNL_CB_BATCH,
7982                .attr_count     = NFTA_SET_MAX,
7983                .policy         = nft_set_policy,
7984        },
7985        [NFT_MSG_NEWSETELEM] = {
7986                .call           = nf_tables_newsetelem,
7987                .type           = NFNL_CB_BATCH,
7988                .attr_count     = NFTA_SET_ELEM_LIST_MAX,
7989                .policy         = nft_set_elem_list_policy,
7990        },
7991        [NFT_MSG_GETSETELEM] = {
7992                .call           = nf_tables_getsetelem,
7993                .type           = NFNL_CB_RCU,
7994                .attr_count     = NFTA_SET_ELEM_LIST_MAX,
7995                .policy         = nft_set_elem_list_policy,
7996        },
7997        [NFT_MSG_DELSETELEM] = {
7998                .call           = nf_tables_delsetelem,
7999                .type           = NFNL_CB_BATCH,
8000                .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8001                .policy         = nft_set_elem_list_policy,
8002        },
8003        [NFT_MSG_GETGEN] = {
8004                .call           = nf_tables_getgen,
8005                .type           = NFNL_CB_RCU,
8006        },
8007        [NFT_MSG_NEWOBJ] = {
8008                .call           = nf_tables_newobj,
8009                .type           = NFNL_CB_BATCH,
8010                .attr_count     = NFTA_OBJ_MAX,
8011                .policy         = nft_obj_policy,
8012        },
8013        [NFT_MSG_GETOBJ] = {
8014                .call           = nf_tables_getobj,
8015                .type           = NFNL_CB_RCU,
8016                .attr_count     = NFTA_OBJ_MAX,
8017                .policy         = nft_obj_policy,
8018        },
8019        [NFT_MSG_DELOBJ] = {
8020                .call           = nf_tables_delobj,
8021                .type           = NFNL_CB_BATCH,
8022                .attr_count     = NFTA_OBJ_MAX,
8023                .policy         = nft_obj_policy,
8024        },
8025        [NFT_MSG_GETOBJ_RESET] = {
8026                .call           = nf_tables_getobj,
8027                .type           = NFNL_CB_RCU,
8028                .attr_count     = NFTA_OBJ_MAX,
8029                .policy         = nft_obj_policy,
8030        },
8031        [NFT_MSG_NEWFLOWTABLE] = {
8032                .call           = nf_tables_newflowtable,
8033                .type           = NFNL_CB_BATCH,
8034                .attr_count     = NFTA_FLOWTABLE_MAX,
8035                .policy         = nft_flowtable_policy,
8036        },
8037        [NFT_MSG_GETFLOWTABLE] = {
8038                .call           = nf_tables_getflowtable,
8039                .type           = NFNL_CB_RCU,
8040                .attr_count     = NFTA_FLOWTABLE_MAX,
8041                .policy         = nft_flowtable_policy,
8042        },
8043        [NFT_MSG_DELFLOWTABLE] = {
8044                .call           = nf_tables_delflowtable,
8045                .type           = NFNL_CB_BATCH,
8046                .attr_count     = NFTA_FLOWTABLE_MAX,
8047                .policy         = nft_flowtable_policy,
8048        },
8049};
8050
8051static int nf_tables_validate(struct net *net)
8052{
8053        struct nftables_pernet *nft_net = nft_pernet(net);
8054        struct nft_table *table;
8055
8056        switch (nft_net->validate_state) {
8057        case NFT_VALIDATE_SKIP:
8058                break;
8059        case NFT_VALIDATE_NEED:
8060                nft_validate_state_update(net, NFT_VALIDATE_DO);
8061                fallthrough;
8062        case NFT_VALIDATE_DO:
8063                list_for_each_entry(table, &nft_net->tables, list) {
8064                        if (nft_table_validate(net, table) < 0)
8065                                return -EAGAIN;
8066                }
8067                break;
8068        }
8069
8070        return 0;
8071}
8072
8073/* a drop policy has to be deferred until all rules have been activated,
8074 * otherwise a large ruleset that contains a drop-policy base chain will
8075 * cause all packets to get dropped until the full transaction has been
8076 * processed.
8077 *
8078 * We defer the drop policy until the transaction has been finalized.
8079 */
8080static void nft_chain_commit_drop_policy(struct nft_trans *trans)
8081{
8082        struct nft_base_chain *basechain;
8083
8084        if (nft_trans_chain_policy(trans) != NF_DROP)
8085                return;
8086
8087        if (!nft_is_base_chain(trans->ctx.chain))
8088                return;
8089
8090        basechain = nft_base_chain(trans->ctx.chain);
8091        basechain->policy = NF_DROP;
8092}
8093
8094static void nft_chain_commit_update(struct nft_trans *trans)
8095{
8096        struct nft_base_chain *basechain;
8097
8098        if (nft_trans_chain_name(trans)) {
8099                rhltable_remove(&trans->ctx.table->chains_ht,
8100                                &trans->ctx.chain->rhlhead,
8101                                nft_chain_ht_params);
8102                swap(trans->ctx.chain->name, nft_trans_chain_name(trans));
8103                rhltable_insert_key(&trans->ctx.table->chains_ht,
8104                                    trans->ctx.chain->name,
8105                                    &trans->ctx.chain->rhlhead,
8106                                    nft_chain_ht_params);
8107        }
8108
8109        if (!nft_is_base_chain(trans->ctx.chain))
8110                return;
8111
8112        nft_chain_stats_replace(trans);
8113
8114        basechain = nft_base_chain(trans->ctx.chain);
8115
8116        switch (nft_trans_chain_policy(trans)) {
8117        case NF_DROP:
8118        case NF_ACCEPT:
8119                basechain->policy = nft_trans_chain_policy(trans);
8120                break;
8121        }
8122}
8123
8124static void nft_obj_commit_update(struct nft_trans *trans)
8125{
8126        struct nft_object *newobj;
8127        struct nft_object *obj;
8128
8129        obj = nft_trans_obj(trans);
8130        newobj = nft_trans_obj_newobj(trans);
8131
8132        if (obj->ops->update)
8133                obj->ops->update(obj, newobj);
8134
8135        kfree(newobj);
8136}
8137
8138static void nft_commit_release(struct nft_trans *trans)
8139{
8140        switch (trans->msg_type) {
8141        case NFT_MSG_DELTABLE:
8142                nf_tables_table_destroy(&trans->ctx);
8143                break;
8144        case NFT_MSG_NEWCHAIN:
8145                free_percpu(nft_trans_chain_stats(trans));
8146                kfree(nft_trans_chain_name(trans));
8147                break;
8148        case NFT_MSG_DELCHAIN:
8149                nf_tables_chain_destroy(&trans->ctx);
8150                break;
8151        case NFT_MSG_DELRULE:
8152                nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8153                break;
8154        case NFT_MSG_DELSET:
8155                nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8156                break;
8157        case NFT_MSG_DELSETELEM:
8158                nf_tables_set_elem_destroy(&trans->ctx,
8159                                           nft_trans_elem_set(trans),
8160                                           nft_trans_elem(trans).priv);
8161                break;
8162        case NFT_MSG_DELOBJ:
8163                nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8164                break;
8165        case NFT_MSG_DELFLOWTABLE:
8166                if (nft_trans_flowtable_update(trans))
8167                        nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8168                else
8169                        nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8170                break;
8171        }
8172
8173        if (trans->put_net)
8174                put_net(trans->ctx.net);
8175
8176        kfree(trans);
8177}
8178
8179static void nf_tables_trans_destroy_work(struct work_struct *w)
8180{
8181        struct nft_trans *trans, *next;
8182        LIST_HEAD(head);
8183
8184        spin_lock(&nf_tables_destroy_list_lock);
8185        list_splice_init(&nf_tables_destroy_list, &head);
8186        spin_unlock(&nf_tables_destroy_list_lock);
8187
8188        if (list_empty(&head))
8189                return;
8190
8191        synchronize_rcu();
8192
8193        list_for_each_entry_safe(trans, next, &head, list) {
8194                list_del(&trans->list);
8195                nft_commit_release(trans);
8196        }
8197}
8198
8199void nf_tables_trans_destroy_flush_work(void)
8200{
8201        flush_work(&trans_destroy_work);
8202}
8203EXPORT_SYMBOL_GPL(nf_tables_trans_destroy_flush_work);
8204
8205static int nf_tables_commit_chain_prepare(struct net *net, struct nft_chain *chain)
8206{
8207        struct nft_rule *rule;
8208        unsigned int alloc = 0;
8209        int i;
8210
8211        /* already handled or inactive chain? */
8212        if (chain->rules_next || !nft_is_active_next(net, chain))
8213                return 0;
8214
8215        rule = list_entry(&chain->rules, struct nft_rule, list);
8216        i = 0;
8217
8218        list_for_each_entry_continue(rule, &chain->rules, list) {
8219                if (nft_is_active_next(net, rule))
8220                        alloc++;
8221        }
8222
8223        chain->rules_next = nf_tables_chain_alloc_rules(chain, alloc);
8224        if (!chain->rules_next)
8225                return -ENOMEM;
8226
8227        list_for_each_entry_continue(rule, &chain->rules, list) {
8228                if (nft_is_active_next(net, rule))
8229                        chain->rules_next[i++] = rule;
8230        }
8231
8232        chain->rules_next[i] = NULL;
8233        return 0;
8234}
8235
8236static void nf_tables_commit_chain_prepare_cancel(struct net *net)
8237{
8238        struct nftables_pernet *nft_net = nft_pernet(net);
8239        struct nft_trans *trans, *next;
8240
8241        list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8242                struct nft_chain *chain = trans->ctx.chain;
8243
8244                if (trans->msg_type == NFT_MSG_NEWRULE ||
8245                    trans->msg_type == NFT_MSG_DELRULE) {
8246                        kvfree(chain->rules_next);
8247                        chain->rules_next = NULL;
8248                }
8249        }
8250}
8251
8252static void __nf_tables_commit_chain_free_rules_old(struct rcu_head *h)
8253{
8254        struct nft_rules_old *o = container_of(h, struct nft_rules_old, h);
8255
8256        kvfree(o->start);
8257}
8258
8259static void nf_tables_commit_chain_free_rules_old(struct nft_rule **rules)
8260{
8261        struct nft_rule **r = rules;
8262        struct nft_rules_old *old;
8263
8264        while (*r)
8265                r++;
8266
8267        r++;    /* rcu_head is after end marker */
8268        old = (void *) r;
8269        old->start = rules;
8270
8271        call_rcu(&old->h, __nf_tables_commit_chain_free_rules_old);
8272}
8273
8274static void nf_tables_commit_chain(struct net *net, struct nft_chain *chain)
8275{
8276        struct nft_rule **g0, **g1;
8277        bool next_genbit;
8278
8279        next_genbit = nft_gencursor_next(net);
8280
8281        g0 = rcu_dereference_protected(chain->rules_gen_0,
8282                                       lockdep_commit_lock_is_held(net));
8283        g1 = rcu_dereference_protected(chain->rules_gen_1,
8284                                       lockdep_commit_lock_is_held(net));
8285
8286        /* No changes to this chain? */
8287        if (chain->rules_next == NULL) {
8288                /* chain had no change in last or next generation */
8289                if (g0 == g1)
8290                        return;
8291                /*
8292                 * chain had no change in this generation; make sure next
8293                 * one uses same rules as current generation.
8294                 */
8295                if (next_genbit) {
8296                        rcu_assign_pointer(chain->rules_gen_1, g0);
8297                        nf_tables_commit_chain_free_rules_old(g1);
8298                } else {
8299                        rcu_assign_pointer(chain->rules_gen_0, g1);
8300                        nf_tables_commit_chain_free_rules_old(g0);
8301                }
8302
8303                return;
8304        }
8305
8306        if (next_genbit)
8307                rcu_assign_pointer(chain->rules_gen_1, chain->rules_next);
8308        else
8309                rcu_assign_pointer(chain->rules_gen_0, chain->rules_next);
8310
8311        chain->rules_next = NULL;
8312
8313        if (g0 == g1)
8314                return;
8315
8316        if (next_genbit)
8317                nf_tables_commit_chain_free_rules_old(g1);
8318        else
8319                nf_tables_commit_chain_free_rules_old(g0);
8320}
8321
8322static void nft_obj_del(struct nft_object *obj)
8323{
8324        rhltable_remove(&nft_objname_ht, &obj->rhlhead, nft_objname_ht_params);
8325        list_del_rcu(&obj->list);
8326}
8327
8328void nft_chain_del(struct nft_chain *chain)
8329{
8330        struct nft_table *table = chain->table;
8331
8332        WARN_ON_ONCE(rhltable_remove(&table->chains_ht, &chain->rhlhead,
8333                                     nft_chain_ht_params));
8334        list_del_rcu(&chain->list);
8335}
8336
8337static void nft_flowtable_hooks_del(struct nft_flowtable *flowtable,
8338                                    struct list_head *hook_list)
8339{
8340        struct nft_hook *hook, *next;
8341
8342        list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
8343                if (hook->inactive)
8344                        list_move(&hook->list, hook_list);
8345        }
8346}
8347
8348static void nf_tables_module_autoload_cleanup(struct net *net)
8349{
8350        struct nftables_pernet *nft_net = nft_pernet(net);
8351        struct nft_module_request *req, *next;
8352
8353        WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
8354        list_for_each_entry_safe(req, next, &nft_net->module_list, list) {
8355                WARN_ON_ONCE(!req->done);
8356                list_del(&req->list);
8357                kfree(req);
8358        }
8359}
8360
8361static void nf_tables_commit_release(struct net *net)
8362{
8363        struct nftables_pernet *nft_net = nft_pernet(net);
8364        struct nft_trans *trans;
8365
8366        /* all side effects have to be made visible.
8367         * For example, if a chain named 'foo' has been deleted, a
8368         * new transaction must not find it anymore.
8369         *
8370         * Memory reclaim happens asynchronously from work queue
8371         * to prevent expensive synchronize_rcu() in commit phase.
8372         */
8373        if (list_empty(&nft_net->commit_list)) {
8374                nf_tables_module_autoload_cleanup(net);
8375                mutex_unlock(&nft_net->commit_mutex);
8376                return;
8377        }
8378
8379        trans = list_last_entry(&nft_net->commit_list,
8380                                struct nft_trans, list);
8381        get_net(trans->ctx.net);
8382        WARN_ON_ONCE(trans->put_net);
8383
8384        trans->put_net = true;
8385        spin_lock(&nf_tables_destroy_list_lock);
8386        list_splice_tail_init(&nft_net->commit_list, &nf_tables_destroy_list);
8387        spin_unlock(&nf_tables_destroy_list_lock);
8388
8389        nf_tables_module_autoload_cleanup(net);
8390        schedule_work(&trans_destroy_work);
8391
8392        mutex_unlock(&nft_net->commit_mutex);
8393}
8394
8395static void nft_commit_notify(struct net *net, u32 portid)
8396{
8397        struct nftables_pernet *nft_net = nft_pernet(net);
8398        struct sk_buff *batch_skb = NULL, *nskb, *skb;
8399        unsigned char *data;
8400        int len;
8401
8402        list_for_each_entry_safe(skb, nskb, &nft_net->notify_list, list) {
8403                if (!batch_skb) {
8404new_batch:
8405                        batch_skb = skb;
8406                        len = NLMSG_GOODSIZE - skb->len;
8407                        list_del(&skb->list);
8408                        continue;
8409                }
8410                len -= skb->len;
8411                if (len > 0 && NFT_CB(skb).report == NFT_CB(batch_skb).report) {
8412                        data = skb_put(batch_skb, skb->len);
8413                        memcpy(data, skb->data, skb->len);
8414                        list_del(&skb->list);
8415                        kfree_skb(skb);
8416                        continue;
8417                }
8418                nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8419                               NFT_CB(batch_skb).report, GFP_KERNEL);
8420                goto new_batch;
8421        }
8422
8423        if (batch_skb) {
8424                nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8425                               NFT_CB(batch_skb).report, GFP_KERNEL);
8426        }
8427
8428        WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
8429}
8430
8431static int nf_tables_commit_audit_alloc(struct list_head *adl,
8432                                        struct nft_table *table)
8433{
8434        struct nft_audit_data *adp;
8435
8436        list_for_each_entry(adp, adl, list) {
8437                if (adp->table == table)
8438                        return 0;
8439        }
8440        adp = kzalloc(sizeof(*adp), GFP_KERNEL);
8441        if (!adp)
8442                return -ENOMEM;
8443        adp->table = table;
8444        list_add(&adp->list, adl);
8445        return 0;
8446}
8447
8448static void nf_tables_commit_audit_free(struct list_head *adl)
8449{
8450        struct nft_audit_data *adp, *adn;
8451
8452        list_for_each_entry_safe(adp, adn, adl, list) {
8453                list_del(&adp->list);
8454                kfree(adp);
8455        }
8456}
8457
8458static void nf_tables_commit_audit_collect(struct list_head *adl,
8459                                           struct nft_table *table, u32 op)
8460{
8461        struct nft_audit_data *adp;
8462
8463        list_for_each_entry(adp, adl, list) {
8464                if (adp->table == table)
8465                        goto found;
8466        }
8467        WARN_ONCE(1, "table=%s not expected in commit list", table->name);
8468        return;
8469found:
8470        adp->entries++;
8471        if (!adp->op || adp->op > op)
8472                adp->op = op;
8473}
8474
8475#define AUNFTABLENAMELEN (NFT_TABLE_MAXNAMELEN + 22)
8476
8477static void nf_tables_commit_audit_log(struct list_head *adl, u32 generation)
8478{
8479        struct nft_audit_data *adp, *adn;
8480        char aubuf[AUNFTABLENAMELEN];
8481
8482        list_for_each_entry_safe(adp, adn, adl, list) {
8483                snprintf(aubuf, AUNFTABLENAMELEN, "%s:%u", adp->table->name,
8484                         generation);
8485                audit_log_nfcfg(aubuf, adp->table->family, adp->entries,
8486                                nft2audit_op[adp->op], GFP_KERNEL);
8487                list_del(&adp->list);
8488                kfree(adp);
8489        }
8490}
8491
8492static int nf_tables_commit(struct net *net, struct sk_buff *skb)
8493{
8494        struct nftables_pernet *nft_net = nft_pernet(net);
8495        struct nft_trans *trans, *next;
8496        struct nft_trans_elem *te;
8497        struct nft_chain *chain;
8498        struct nft_table *table;
8499        LIST_HEAD(adl);
8500        int err;
8501
8502        if (list_empty(&nft_net->commit_list)) {
8503                mutex_unlock(&nft_net->commit_mutex);
8504                return 0;
8505        }
8506
8507        /* 0. Validate ruleset, otherwise roll back for error reporting. */
8508        if (nf_tables_validate(net) < 0)
8509                return -EAGAIN;
8510
8511        err = nft_flow_rule_offload_commit(net);
8512        if (err < 0)
8513                return err;
8514
8515        /* 1.  Allocate space for next generation rules_gen_X[] */
8516        list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8517                int ret;
8518
8519                ret = nf_tables_commit_audit_alloc(&adl, trans->ctx.table);
8520                if (ret) {
8521                        nf_tables_commit_chain_prepare_cancel(net);
8522                        nf_tables_commit_audit_free(&adl);
8523                        return ret;
8524                }
8525                if (trans->msg_type == NFT_MSG_NEWRULE ||
8526                    trans->msg_type == NFT_MSG_DELRULE) {
8527                        chain = trans->ctx.chain;
8528
8529                        ret = nf_tables_commit_chain_prepare(net, chain);
8530                        if (ret < 0) {
8531                                nf_tables_commit_chain_prepare_cancel(net);
8532                                nf_tables_commit_audit_free(&adl);
8533                                return ret;
8534                        }
8535                }
8536        }
8537
8538        /* step 2.  Make rules_gen_X visible to packet path */
8539        list_for_each_entry(table, &nft_net->tables, list) {
8540                list_for_each_entry(chain, &table->chains, list)
8541                        nf_tables_commit_chain(net, chain);
8542        }
8543
8544        /*
8545         * Bump generation counter, invalidate any dump in progress.
8546         * Cannot fail after this point.
8547         */
8548        while (++nft_net->base_seq == 0)
8549                ;
8550
8551        /* step 3. Start new generation, rules_gen_X now in use. */
8552        net->nft.gencursor = nft_gencursor_next(net);
8553
8554        list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8555                nf_tables_commit_audit_collect(&adl, trans->ctx.table,
8556                                               trans->msg_type);
8557                switch (trans->msg_type) {
8558                case NFT_MSG_NEWTABLE:
8559                        if (nft_trans_table_update(trans)) {
8560                                if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
8561                                        nft_trans_destroy(trans);
8562                                        break;
8563                                }
8564                                if (trans->ctx.table->flags & NFT_TABLE_F_DORMANT)
8565                                        nf_tables_table_disable(net, trans->ctx.table);
8566
8567                                trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
8568                        } else {
8569                                nft_clear(net, trans->ctx.table);
8570                        }
8571                        nf_tables_table_notify(&trans->ctx, NFT_MSG_NEWTABLE);
8572                        nft_trans_destroy(trans);
8573                        break;
8574                case NFT_MSG_DELTABLE:
8575                        list_del_rcu(&trans->ctx.table->list);
8576                        nf_tables_table_notify(&trans->ctx, NFT_MSG_DELTABLE);
8577                        break;
8578                case NFT_MSG_NEWCHAIN:
8579                        if (nft_trans_chain_update(trans)) {
8580                                nft_chain_commit_update(trans);
8581                                nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
8582                                /* trans destroyed after rcu grace period */
8583                        } else {
8584                                nft_chain_commit_drop_policy(trans);
8585                                nft_clear(net, trans->ctx.chain);
8586                                nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
8587                                nft_trans_destroy(trans);
8588                        }
8589                        break;
8590                case NFT_MSG_DELCHAIN:
8591                        nft_chain_del(trans->ctx.chain);
8592                        nf_tables_chain_notify(&trans->ctx, NFT_MSG_DELCHAIN);
8593                        nf_tables_unregister_hook(trans->ctx.net,
8594                                                  trans->ctx.table,
8595                                                  trans->ctx.chain);
8596                        break;
8597                case NFT_MSG_NEWRULE:
8598                        nft_clear(trans->ctx.net, nft_trans_rule(trans));
8599                        nf_tables_rule_notify(&trans->ctx,
8600                                              nft_trans_rule(trans),
8601                                              NFT_MSG_NEWRULE);
8602                        nft_trans_destroy(trans);
8603                        break;
8604                case NFT_MSG_DELRULE:
8605                        list_del_rcu(&nft_trans_rule(trans)->list);
8606                        nf_tables_rule_notify(&trans->ctx,
8607                                              nft_trans_rule(trans),
8608                                              NFT_MSG_DELRULE);
8609                        nft_rule_expr_deactivate(&trans->ctx,
8610                                                 nft_trans_rule(trans),
8611                                                 NFT_TRANS_COMMIT);
8612                        break;
8613                case NFT_MSG_NEWSET:
8614                        nft_clear(net, nft_trans_set(trans));
8615                        /* This avoids hitting -EBUSY when deleting the table
8616                         * from the transaction.
8617                         */
8618                        if (nft_set_is_anonymous(nft_trans_set(trans)) &&
8619                            !list_empty(&nft_trans_set(trans)->bindings))
8620                                trans->ctx.table->use--;
8621
8622                        nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
8623                                             NFT_MSG_NEWSET, GFP_KERNEL);
8624                        nft_trans_destroy(trans);
8625                        break;
8626                case NFT_MSG_DELSET:
8627                        list_del_rcu(&nft_trans_set(trans)->list);
8628                        nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
8629                                             NFT_MSG_DELSET, GFP_KERNEL);
8630                        break;
8631                case NFT_MSG_NEWSETELEM:
8632                        te = (struct nft_trans_elem *)trans->data;
8633
8634                        nft_setelem_activate(net, te->set, &te->elem);
8635                        nf_tables_setelem_notify(&trans->ctx, te->set,
8636                                                 &te->elem,
8637                                                 NFT_MSG_NEWSETELEM, 0);
8638                        nft_trans_destroy(trans);
8639                        break;
8640                case NFT_MSG_DELSETELEM:
8641                        te = (struct nft_trans_elem *)trans->data;
8642
8643                        nf_tables_setelem_notify(&trans->ctx, te->set,
8644                                                 &te->elem,
8645                                                 NFT_MSG_DELSETELEM, 0);
8646                        nft_setelem_remove(net, te->set, &te->elem);
8647                        if (!nft_setelem_is_catchall(te->set, &te->elem)) {
8648                                atomic_dec(&te->set->nelems);
8649                                te->set->ndeact--;
8650                        }
8651                        break;
8652                case NFT_MSG_NEWOBJ:
8653                        if (nft_trans_obj_update(trans)) {
8654                                nft_obj_commit_update(trans);
8655                                nf_tables_obj_notify(&trans->ctx,
8656                                                     nft_trans_obj(trans),
8657                                                     NFT_MSG_NEWOBJ);
8658                        } else {
8659                                nft_clear(net, nft_trans_obj(trans));
8660                                nf_tables_obj_notify(&trans->ctx,
8661                                                     nft_trans_obj(trans),
8662                                                     NFT_MSG_NEWOBJ);
8663                                nft_trans_destroy(trans);
8664                        }
8665                        break;
8666                case NFT_MSG_DELOBJ:
8667                        nft_obj_del(nft_trans_obj(trans));
8668                        nf_tables_obj_notify(&trans->ctx, nft_trans_obj(trans),
8669                                             NFT_MSG_DELOBJ);
8670                        break;
8671                case NFT_MSG_NEWFLOWTABLE:
8672                        if (nft_trans_flowtable_update(trans)) {
8673                                nft_trans_flowtable(trans)->data.flags =
8674                                        nft_trans_flowtable_flags(trans);
8675                                nf_tables_flowtable_notify(&trans->ctx,
8676                                                           nft_trans_flowtable(trans),
8677                                                           &nft_trans_flowtable_hooks(trans),
8678                                                           NFT_MSG_NEWFLOWTABLE);
8679                                list_splice(&nft_trans_flowtable_hooks(trans),
8680                                            &nft_trans_flowtable(trans)->hook_list);
8681                        } else {
8682                                nft_clear(net, nft_trans_flowtable(trans));
8683                                nf_tables_flowtable_notify(&trans->ctx,
8684                                                           nft_trans_flowtable(trans),
8685                                                           &nft_trans_flowtable(trans)->hook_list,
8686                                                           NFT_MSG_NEWFLOWTABLE);
8687                        }
8688                        nft_trans_destroy(trans);
8689                        break;
8690                case NFT_MSG_DELFLOWTABLE:
8691                        if (nft_trans_flowtable_update(trans)) {
8692                                nft_flowtable_hooks_del(nft_trans_flowtable(trans),
8693                                                        &nft_trans_flowtable_hooks(trans));
8694                                nf_tables_flowtable_notify(&trans->ctx,
8695                                                           nft_trans_flowtable(trans),
8696                                                           &nft_trans_flowtable_hooks(trans),
8697                                                           NFT_MSG_DELFLOWTABLE);
8698                                nft_unregister_flowtable_net_hooks(net,
8699                                                                   &nft_trans_flowtable_hooks(trans));
8700                        } else {
8701                                list_del_rcu(&nft_trans_flowtable(trans)->list);
8702                                nf_tables_flowtable_notify(&trans->ctx,
8703                                                           nft_trans_flowtable(trans),
8704                                                           &nft_trans_flowtable(trans)->hook_list,
8705                                                           NFT_MSG_DELFLOWTABLE);
8706                                nft_unregister_flowtable_net_hooks(net,
8707                                                &nft_trans_flowtable(trans)->hook_list);
8708                        }
8709                        break;
8710                }
8711        }
8712
8713        nft_commit_notify(net, NETLINK_CB(skb).portid);
8714        nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
8715        nf_tables_commit_audit_log(&adl, nft_net->base_seq);
8716        nf_tables_commit_release(net);
8717
8718        return 0;
8719}
8720
8721static void nf_tables_module_autoload(struct net *net)
8722{
8723        struct nftables_pernet *nft_net = nft_pernet(net);
8724        struct nft_module_request *req, *next;
8725        LIST_HEAD(module_list);
8726
8727        list_splice_init(&nft_net->module_list, &module_list);
8728        mutex_unlock(&nft_net->commit_mutex);
8729        list_for_each_entry_safe(req, next, &module_list, list) {
8730                request_module("%s", req->module);
8731                req->done = true;
8732        }
8733        mutex_lock(&nft_net->commit_mutex);
8734        list_splice(&module_list, &nft_net->module_list);
8735}
8736
8737static void nf_tables_abort_release(struct nft_trans *trans)
8738{
8739        switch (trans->msg_type) {
8740        case NFT_MSG_NEWTABLE:
8741                nf_tables_table_destroy(&trans->ctx);
8742                break;
8743        case NFT_MSG_NEWCHAIN:
8744                nf_tables_chain_destroy(&trans->ctx);
8745                break;
8746        case NFT_MSG_NEWRULE:
8747                nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8748                break;
8749        case NFT_MSG_NEWSET:
8750                nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8751                break;
8752        case NFT_MSG_NEWSETELEM:
8753                nft_set_elem_destroy(nft_trans_elem_set(trans),
8754                                     nft_trans_elem(trans).priv, true);
8755                break;
8756        case NFT_MSG_NEWOBJ:
8757                nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8758                break;
8759        case NFT_MSG_NEWFLOWTABLE:
8760                if (nft_trans_flowtable_update(trans))
8761                        nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8762                else
8763                        nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8764                break;
8765        }
8766        kfree(trans);
8767}
8768
8769static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action)
8770{
8771        struct nftables_pernet *nft_net = nft_pernet(net);
8772        struct nft_trans *trans, *next;
8773        struct nft_trans_elem *te;
8774        struct nft_hook *hook;
8775
8776        if (action == NFNL_ABORT_VALIDATE &&
8777            nf_tables_validate(net) < 0)
8778                return -EAGAIN;
8779
8780        list_for_each_entry_safe_reverse(trans, next, &nft_net->commit_list,
8781                                         list) {
8782                switch (trans->msg_type) {
8783                case NFT_MSG_NEWTABLE:
8784                        if (nft_trans_table_update(trans)) {
8785                                if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
8786                                        nft_trans_destroy(trans);
8787                                        break;
8788                                }
8789                                if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_DORMANT) {
8790                                        nf_tables_table_disable(net, trans->ctx.table);
8791                                        trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
8792                                } else if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_AWAKEN) {
8793                                        trans->ctx.table->flags &= ~NFT_TABLE_F_DORMANT;
8794                                }
8795                                trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
8796                                nft_trans_destroy(trans);
8797                        } else {
8798                                list_del_rcu(&trans->ctx.table->list);
8799                        }
8800                        break;
8801                case NFT_MSG_DELTABLE:
8802                        nft_clear(trans->ctx.net, trans->ctx.table);
8803                        nft_trans_destroy(trans);
8804                        break;
8805                case NFT_MSG_NEWCHAIN:
8806                        if (nft_trans_chain_update(trans)) {
8807                                free_percpu(nft_trans_chain_stats(trans));
8808                                kfree(nft_trans_chain_name(trans));
8809                                nft_trans_destroy(trans);
8810                        } else {
8811                                if (nft_chain_is_bound(trans->ctx.chain)) {
8812                                        nft_trans_destroy(trans);
8813                                        break;
8814                                }
8815                                trans->ctx.table->use--;
8816                                nft_chain_del(trans->ctx.chain);
8817                                nf_tables_unregister_hook(trans->ctx.net,
8818                                                          trans->ctx.table,
8819                                                          trans->ctx.chain);
8820                        }
8821                        break;
8822                case NFT_MSG_DELCHAIN:
8823                        trans->ctx.table->use++;
8824                        nft_clear(trans->ctx.net, trans->ctx.chain);
8825                        nft_trans_destroy(trans);
8826                        break;
8827                case NFT_MSG_NEWRULE:
8828                        trans->ctx.chain->use--;
8829                        list_del_rcu(&nft_trans_rule(trans)->list);
8830                        nft_rule_expr_deactivate(&trans->ctx,
8831                                                 nft_trans_rule(trans),
8832                                                 NFT_TRANS_ABORT);
8833                        if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
8834                                nft_flow_rule_destroy(nft_trans_flow_rule(trans));
8835                        break;
8836                case NFT_MSG_DELRULE:
8837                        trans->ctx.chain->use++;
8838                        nft_clear(trans->ctx.net, nft_trans_rule(trans));
8839                        nft_rule_expr_activate(&trans->ctx, nft_trans_rule(trans));
8840                        if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
8841                                nft_flow_rule_destroy(nft_trans_flow_rule(trans));
8842
8843                        nft_trans_destroy(trans);
8844                        break;
8845                case NFT_MSG_NEWSET:
8846                        trans->ctx.table->use--;
8847                        if (nft_trans_set_bound(trans)) {
8848                                nft_trans_destroy(trans);
8849                                break;
8850                        }
8851                        list_del_rcu(&nft_trans_set(trans)->list);
8852                        break;
8853                case NFT_MSG_DELSET:
8854                        trans->ctx.table->use++;
8855                        nft_clear(trans->ctx.net, nft_trans_set(trans));
8856                        nft_trans_destroy(trans);
8857                        break;
8858                case NFT_MSG_NEWSETELEM:
8859                        if (nft_trans_elem_set_bound(trans)) {
8860                                nft_trans_destroy(trans);
8861                                break;
8862                        }
8863                        te = (struct nft_trans_elem *)trans->data;
8864                        nft_setelem_remove(net, te->set, &te->elem);
8865                        if (!nft_setelem_is_catchall(te->set, &te->elem))
8866                                atomic_dec(&te->set->nelems);
8867                        break;
8868                case NFT_MSG_DELSETELEM:
8869                        te = (struct nft_trans_elem *)trans->data;
8870
8871                        nft_setelem_data_activate(net, te->set, &te->elem);
8872                        nft_setelem_activate(net, te->set, &te->elem);
8873                        if (!nft_setelem_is_catchall(te->set, &te->elem))
8874                                te->set->ndeact--;
8875
8876                        nft_trans_destroy(trans);
8877                        break;
8878                case NFT_MSG_NEWOBJ:
8879                        if (nft_trans_obj_update(trans)) {
8880                                kfree(nft_trans_obj_newobj(trans));
8881                                nft_trans_destroy(trans);
8882                        } else {
8883                                trans->ctx.table->use--;
8884                                nft_obj_del(nft_trans_obj(trans));
8885                        }
8886                        break;
8887                case NFT_MSG_DELOBJ:
8888                        trans->ctx.table->use++;
8889                        nft_clear(trans->ctx.net, nft_trans_obj(trans));
8890                        nft_trans_destroy(trans);
8891                        break;
8892                case NFT_MSG_NEWFLOWTABLE:
8893                        if (nft_trans_flowtable_update(trans)) {
8894                                nft_unregister_flowtable_net_hooks(net,
8895                                                &nft_trans_flowtable_hooks(trans));
8896                        } else {
8897                                trans->ctx.table->use--;
8898                                list_del_rcu(&nft_trans_flowtable(trans)->list);
8899                                nft_unregister_flowtable_net_hooks(net,
8900                                                &nft_trans_flowtable(trans)->hook_list);
8901                        }
8902                        break;
8903                case NFT_MSG_DELFLOWTABLE:
8904                        if (nft_trans_flowtable_update(trans)) {
8905                                list_for_each_entry(hook, &nft_trans_flowtable(trans)->hook_list, list)
8906                                        hook->inactive = false;
8907                        } else {
8908                                trans->ctx.table->use++;
8909                                nft_clear(trans->ctx.net, nft_trans_flowtable(trans));
8910                        }
8911                        nft_trans_destroy(trans);
8912                        break;
8913                }
8914        }
8915
8916        synchronize_rcu();
8917
8918        list_for_each_entry_safe_reverse(trans, next,
8919                                         &nft_net->commit_list, list) {
8920                list_del(&trans->list);
8921                nf_tables_abort_release(trans);
8922        }
8923
8924        if (action == NFNL_ABORT_AUTOLOAD)
8925                nf_tables_module_autoload(net);
8926        else
8927                nf_tables_module_autoload_cleanup(net);
8928
8929        return 0;
8930}
8931
8932static void nf_tables_cleanup(struct net *net)
8933{
8934        nft_validate_state_update(net, NFT_VALIDATE_SKIP);
8935}
8936
8937static int nf_tables_abort(struct net *net, struct sk_buff *skb,
8938                           enum nfnl_abort_action action)
8939{
8940        struct nftables_pernet *nft_net = nft_pernet(net);
8941        int ret = __nf_tables_abort(net, action);
8942
8943        mutex_unlock(&nft_net->commit_mutex);
8944
8945        return ret;
8946}
8947
8948static bool nf_tables_valid_genid(struct net *net, u32 genid)
8949{
8950        struct nftables_pernet *nft_net = nft_pernet(net);
8951        bool genid_ok;
8952
8953        mutex_lock(&nft_net->commit_mutex);
8954
8955        genid_ok = genid == 0 || nft_net->base_seq == genid;
8956        if (!genid_ok)
8957                mutex_unlock(&nft_net->commit_mutex);
8958
8959        /* else, commit mutex has to be released by commit or abort function */
8960        return genid_ok;
8961}
8962
8963static const struct nfnetlink_subsystem nf_tables_subsys = {
8964        .name           = "nf_tables",
8965        .subsys_id      = NFNL_SUBSYS_NFTABLES,
8966        .cb_count       = NFT_MSG_MAX,
8967        .cb             = nf_tables_cb,
8968        .commit         = nf_tables_commit,
8969        .abort          = nf_tables_abort,
8970        .cleanup        = nf_tables_cleanup,
8971        .valid_genid    = nf_tables_valid_genid,
8972        .owner          = THIS_MODULE,
8973};
8974
8975int nft_chain_validate_dependency(const struct nft_chain *chain,
8976                                  enum nft_chain_types type)
8977{
8978        const struct nft_base_chain *basechain;
8979
8980        if (nft_is_base_chain(chain)) {
8981                basechain = nft_base_chain(chain);
8982                if (basechain->type->type != type)
8983                        return -EOPNOTSUPP;
8984        }
8985        return 0;
8986}
8987EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
8988
8989int nft_chain_validate_hooks(const struct nft_chain *chain,
8990                             unsigned int hook_flags)
8991{
8992        struct nft_base_chain *basechain;
8993
8994        if (nft_is_base_chain(chain)) {
8995                basechain = nft_base_chain(chain);
8996
8997                if ((1 << basechain->ops.hooknum) & hook_flags)
8998                        return 0;
8999
9000                return -EOPNOTSUPP;
9001        }
9002
9003        return 0;
9004}
9005EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
9006
9007/*
9008 * Loop detection - walk through the ruleset beginning at the destination chain
9009 * of a new jump until either the source chain is reached (loop) or all
9010 * reachable chains have been traversed.
9011 *
9012 * The loop check is performed whenever a new jump verdict is added to an
9013 * expression or verdict map or a verdict map is bound to a new chain.
9014 */
9015
9016static int nf_tables_check_loops(const struct nft_ctx *ctx,
9017                                 const struct nft_chain *chain);
9018
9019static int nft_check_loops(const struct nft_ctx *ctx,
9020                           const struct nft_set_ext *ext)
9021{
9022        const struct nft_data *data;
9023        int ret;
9024
9025        data = nft_set_ext_data(ext);
9026        switch (data->verdict.code) {
9027        case NFT_JUMP:
9028        case NFT_GOTO:
9029                ret = nf_tables_check_loops(ctx, data->verdict.chain);
9030                break;
9031        default:
9032                ret = 0;
9033                break;
9034        }
9035
9036        return ret;
9037}
9038
9039static int nf_tables_loop_check_setelem(const struct nft_ctx *ctx,
9040                                        struct nft_set *set,
9041                                        const struct nft_set_iter *iter,
9042                                        struct nft_set_elem *elem)
9043{
9044        const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
9045
9046        if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
9047            *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
9048                return 0;
9049
9050        return nft_check_loops(ctx, ext);
9051}
9052
9053static int nft_set_catchall_loops(const struct nft_ctx *ctx,
9054                                  struct nft_set *set)
9055{
9056        u8 genmask = nft_genmask_next(ctx->net);
9057        struct nft_set_elem_catchall *catchall;
9058        struct nft_set_ext *ext;
9059        int ret = 0;
9060
9061        list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
9062                ext = nft_set_elem_ext(set, catchall->elem);
9063                if (!nft_set_elem_active(ext, genmask))
9064                        continue;
9065
9066                ret = nft_check_loops(ctx, ext);
9067                if (ret < 0)
9068                        return ret;
9069        }
9070
9071        return ret;
9072}
9073
9074static int nf_tables_check_loops(const struct nft_ctx *ctx,
9075                                 const struct nft_chain *chain)
9076{
9077        const struct nft_rule *rule;
9078        const struct nft_expr *expr, *last;
9079        struct nft_set *set;
9080        struct nft_set_binding *binding;
9081        struct nft_set_iter iter;
9082
9083        if (ctx->chain == chain)
9084                return -ELOOP;
9085
9086        list_for_each_entry(rule, &chain->rules, list) {
9087                nft_rule_for_each_expr(expr, last, rule) {
9088                        struct nft_immediate_expr *priv;
9089                        const struct nft_data *data;
9090                        int err;
9091
9092                        if (strcmp(expr->ops->type->name, "immediate"))
9093                                continue;
9094
9095                        priv = nft_expr_priv(expr);
9096                        if (priv->dreg != NFT_REG_VERDICT)
9097                                continue;
9098
9099                        data = &priv->data;
9100                        switch (data->verdict.code) {
9101                        case NFT_JUMP:
9102                        case NFT_GOTO:
9103                                err = nf_tables_check_loops(ctx,
9104                                                        data->verdict.chain);
9105                                if (err < 0)
9106                                        return err;
9107                                break;
9108                        default:
9109                                break;
9110                        }
9111                }
9112        }
9113
9114        list_for_each_entry(set, &ctx->table->sets, list) {
9115                if (!nft_is_active_next(ctx->net, set))
9116                        continue;
9117                if (!(set->flags & NFT_SET_MAP) ||
9118                    set->dtype != NFT_DATA_VERDICT)
9119                        continue;
9120
9121                list_for_each_entry(binding, &set->bindings, list) {
9122                        if (!(binding->flags & NFT_SET_MAP) ||
9123                            binding->chain != chain)
9124                                continue;
9125
9126                        iter.genmask    = nft_genmask_next(ctx->net);
9127                        iter.skip       = 0;
9128                        iter.count      = 0;
9129                        iter.err        = 0;
9130                        iter.fn         = nf_tables_loop_check_setelem;
9131
9132                        set->ops->walk(ctx, set, &iter);
9133                        if (!iter.err)
9134                                iter.err = nft_set_catchall_loops(ctx, set);
9135
9136                        if (iter.err < 0)
9137                                return iter.err;
9138                }
9139        }
9140
9141        return 0;
9142}
9143
9144/**
9145 *      nft_parse_u32_check - fetch u32 attribute and check for maximum value
9146 *
9147 *      @attr: netlink attribute to fetch value from
9148 *      @max: maximum value to be stored in dest
9149 *      @dest: pointer to the variable
9150 *
9151 *      Parse, check and store a given u32 netlink attribute into variable.
9152 *      This function returns -ERANGE if the value goes over maximum value.
9153 *      Otherwise a 0 is returned and the attribute value is stored in the
9154 *      destination variable.
9155 */
9156int nft_parse_u32_check(const struct nlattr *attr, int max, u32 *dest)
9157{
9158        u32 val;
9159
9160        val = ntohl(nla_get_be32(attr));
9161        if (val > max)
9162                return -ERANGE;
9163
9164        *dest = val;
9165        return 0;
9166}
9167EXPORT_SYMBOL_GPL(nft_parse_u32_check);
9168
9169static unsigned int nft_parse_register(const struct nlattr *attr)
9170{
9171        unsigned int reg;
9172
9173        reg = ntohl(nla_get_be32(attr));
9174        switch (reg) {
9175        case NFT_REG_VERDICT...NFT_REG_4:
9176                return reg * NFT_REG_SIZE / NFT_REG32_SIZE;
9177        default:
9178                return reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
9179        }
9180}
9181
9182/**
9183 *      nft_dump_register - dump a register value to a netlink attribute
9184 *
9185 *      @skb: socket buffer
9186 *      @attr: attribute number
9187 *      @reg: register number
9188 *
9189 *      Construct a netlink attribute containing the register number. For
9190 *      compatibility reasons, register numbers being a multiple of 4 are
9191 *      translated to the corresponding 128 bit register numbers.
9192 */
9193int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
9194{
9195        if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
9196                reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
9197        else
9198                reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
9199
9200        return nla_put_be32(skb, attr, htonl(reg));
9201}
9202EXPORT_SYMBOL_GPL(nft_dump_register);
9203
9204static int nft_validate_register_load(enum nft_registers reg, unsigned int len)
9205{
9206        if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9207                return -EINVAL;
9208        if (len == 0)
9209                return -EINVAL;
9210        if (reg * NFT_REG32_SIZE + len > sizeof_field(struct nft_regs, data))
9211                return -ERANGE;
9212
9213        return 0;
9214}
9215
9216int nft_parse_register_load(const struct nlattr *attr, u8 *sreg, u32 len)
9217{
9218        u32 reg;
9219        int err;
9220
9221        reg = nft_parse_register(attr);
9222        err = nft_validate_register_load(reg, len);
9223        if (err < 0)
9224                return err;
9225
9226        *sreg = reg;
9227        return 0;
9228}
9229EXPORT_SYMBOL_GPL(nft_parse_register_load);
9230
9231static int nft_validate_register_store(const struct nft_ctx *ctx,
9232                                       enum nft_registers reg,
9233                                       const struct nft_data *data,
9234                                       enum nft_data_types type,
9235                                       unsigned int len)
9236{
9237        int err;
9238
9239        switch (reg) {
9240        case NFT_REG_VERDICT:
9241                if (type != NFT_DATA_VERDICT)
9242                        return -EINVAL;
9243
9244                if (data != NULL &&
9245                    (data->verdict.code == NFT_GOTO ||
9246                     data->verdict.code == NFT_JUMP)) {
9247                        err = nf_tables_check_loops(ctx, data->verdict.chain);
9248                        if (err < 0)
9249                                return err;
9250                }
9251
9252                return 0;
9253        default:
9254                if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9255                        return -EINVAL;
9256                if (len == 0)
9257                        return -EINVAL;
9258                if (reg * NFT_REG32_SIZE + len >
9259                    sizeof_field(struct nft_regs, data))
9260                        return -ERANGE;
9261
9262                if (data != NULL && type != NFT_DATA_VALUE)
9263                        return -EINVAL;
9264                return 0;
9265        }
9266}
9267
9268int nft_parse_register_store(const struct nft_ctx *ctx,
9269                             const struct nlattr *attr, u8 *dreg,
9270                             const struct nft_data *data,
9271                             enum nft_data_types type, unsigned int len)
9272{
9273        int err;
9274        u32 reg;
9275
9276        reg = nft_parse_register(attr);
9277        err = nft_validate_register_store(ctx, reg, data, type, len);
9278        if (err < 0)
9279                return err;
9280
9281        *dreg = reg;
9282        return 0;
9283}
9284EXPORT_SYMBOL_GPL(nft_parse_register_store);
9285
9286static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
9287        [NFTA_VERDICT_CODE]     = { .type = NLA_U32 },
9288        [NFTA_VERDICT_CHAIN]    = { .type = NLA_STRING,
9289                                    .len = NFT_CHAIN_MAXNAMELEN - 1 },
9290        [NFTA_VERDICT_CHAIN_ID] = { .type = NLA_U32 },
9291};
9292
9293static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
9294                            struct nft_data_desc *desc, const struct nlattr *nla)
9295{
9296        u8 genmask = nft_genmask_next(ctx->net);
9297        struct nlattr *tb[NFTA_VERDICT_MAX + 1];
9298        struct nft_chain *chain;
9299        int err;
9300
9301        err = nla_parse_nested_deprecated(tb, NFTA_VERDICT_MAX, nla,
9302                                          nft_verdict_policy, NULL);
9303        if (err < 0)
9304                return err;
9305
9306        if (!tb[NFTA_VERDICT_CODE])
9307                return -EINVAL;
9308        data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
9309
9310        switch (data->verdict.code) {
9311        default:
9312                switch (data->verdict.code & NF_VERDICT_MASK) {
9313                case NF_ACCEPT:
9314                case NF_DROP:
9315                case NF_QUEUE:
9316                        break;
9317                default:
9318                        return -EINVAL;
9319                }
9320                fallthrough;
9321        case NFT_CONTINUE:
9322        case NFT_BREAK:
9323        case NFT_RETURN:
9324                break;
9325        case NFT_JUMP:
9326        case NFT_GOTO:
9327                if (tb[NFTA_VERDICT_CHAIN]) {
9328                        chain = nft_chain_lookup(ctx->net, ctx->table,
9329                                                 tb[NFTA_VERDICT_CHAIN],
9330                                                 genmask);
9331                } else if (tb[NFTA_VERDICT_CHAIN_ID]) {
9332                        chain = nft_chain_lookup_byid(ctx->net,
9333                                                      tb[NFTA_VERDICT_CHAIN_ID]);
9334                        if (IS_ERR(chain))
9335                                return PTR_ERR(chain);
9336                } else {
9337                        return -EINVAL;
9338                }
9339
9340                if (IS_ERR(chain))
9341                        return PTR_ERR(chain);
9342                if (nft_is_base_chain(chain))
9343                        return -EOPNOTSUPP;
9344
9345                chain->use++;
9346                data->verdict.chain = chain;
9347                break;
9348        }
9349
9350        desc->len = sizeof(data->verdict);
9351        desc->type = NFT_DATA_VERDICT;
9352        return 0;
9353}
9354
9355static void nft_verdict_uninit(const struct nft_data *data)
9356{
9357        struct nft_chain *chain;
9358        struct nft_rule *rule;
9359
9360        switch (data->verdict.code) {
9361        case NFT_JUMP:
9362        case NFT_GOTO:
9363                chain = data->verdict.chain;
9364                chain->use--;
9365
9366                if (!nft_chain_is_bound(chain))
9367                        break;
9368
9369                chain->table->use--;
9370                list_for_each_entry(rule, &chain->rules, list)
9371                        chain->use--;
9372
9373                nft_chain_del(chain);
9374                break;
9375        }
9376}
9377
9378int nft_verdict_dump(struct sk_buff *skb, int type, const struct nft_verdict *v)
9379{
9380        struct nlattr *nest;
9381
9382        nest = nla_nest_start_noflag(skb, type);
9383        if (!nest)
9384                goto nla_put_failure;
9385
9386        if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(v->code)))
9387                goto nla_put_failure;
9388
9389        switch (v->code) {
9390        case NFT_JUMP:
9391        case NFT_GOTO:
9392                if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
9393                                   v->chain->name))
9394                        goto nla_put_failure;
9395        }
9396        nla_nest_end(skb, nest);
9397        return 0;
9398
9399nla_put_failure:
9400        return -1;
9401}
9402
9403static int nft_value_init(const struct nft_ctx *ctx,
9404                          struct nft_data *data, unsigned int size,
9405                          struct nft_data_desc *desc, const struct nlattr *nla)
9406{
9407        unsigned int len;
9408
9409        len = nla_len(nla);
9410        if (len == 0)
9411                return -EINVAL;
9412        if (len > size)
9413                return -EOVERFLOW;
9414
9415        nla_memcpy(data->data, nla, len);
9416        desc->type = NFT_DATA_VALUE;
9417        desc->len  = len;
9418        return 0;
9419}
9420
9421static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
9422                          unsigned int len)
9423{
9424        return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
9425}
9426
9427static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
9428        [NFTA_DATA_VALUE]       = { .type = NLA_BINARY },
9429        [NFTA_DATA_VERDICT]     = { .type = NLA_NESTED },
9430};
9431
9432/**
9433 *      nft_data_init - parse nf_tables data netlink attributes
9434 *
9435 *      @ctx: context of the expression using the data
9436 *      @data: destination struct nft_data
9437 *      @size: maximum data length
9438 *      @desc: data description
9439 *      @nla: netlink attribute containing data
9440 *
9441 *      Parse the netlink data attributes and initialize a struct nft_data.
9442 *      The type and length of data are returned in the data description.
9443 *
9444 *      The caller can indicate that it only wants to accept data of type
9445 *      NFT_DATA_VALUE by passing NULL for the ctx argument.
9446 */
9447int nft_data_init(const struct nft_ctx *ctx,
9448                  struct nft_data *data, unsigned int size,
9449                  struct nft_data_desc *desc, const struct nlattr *nla)
9450{
9451        struct nlattr *tb[NFTA_DATA_MAX + 1];
9452        int err;
9453
9454        err = nla_parse_nested_deprecated(tb, NFTA_DATA_MAX, nla,
9455                                          nft_data_policy, NULL);
9456        if (err < 0)
9457                return err;
9458
9459        if (tb[NFTA_DATA_VALUE])
9460                return nft_value_init(ctx, data, size, desc,
9461                                      tb[NFTA_DATA_VALUE]);
9462        if (tb[NFTA_DATA_VERDICT] && ctx != NULL)
9463                return nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
9464        return -EINVAL;
9465}
9466EXPORT_SYMBOL_GPL(nft_data_init);
9467
9468/**
9469 *      nft_data_release - release a nft_data item
9470 *
9471 *      @data: struct nft_data to release
9472 *      @type: type of data
9473 *
9474 *      Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
9475 *      all others need to be released by calling this function.
9476 */
9477void nft_data_release(const struct nft_data *data, enum nft_data_types type)
9478{
9479        if (type < NFT_DATA_VERDICT)
9480                return;
9481        switch (type) {
9482        case NFT_DATA_VERDICT:
9483                return nft_verdict_uninit(data);
9484        default:
9485                WARN_ON(1);
9486        }
9487}
9488EXPORT_SYMBOL_GPL(nft_data_release);
9489
9490int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
9491                  enum nft_data_types type, unsigned int len)
9492{
9493        struct nlattr *nest;
9494        int err;
9495
9496        nest = nla_nest_start_noflag(skb, attr);
9497        if (nest == NULL)
9498                return -1;
9499
9500        switch (type) {
9501        case NFT_DATA_VALUE:
9502                err = nft_value_dump(skb, data, len);
9503                break;
9504        case NFT_DATA_VERDICT:
9505                err = nft_verdict_dump(skb, NFTA_DATA_VERDICT, &data->verdict);
9506                break;
9507        default:
9508                err = -EINVAL;
9509                WARN_ON(1);
9510        }
9511
9512        nla_nest_end(skb, nest);
9513        return err;
9514}
9515EXPORT_SYMBOL_GPL(nft_data_dump);
9516
9517int __nft_release_basechain(struct nft_ctx *ctx)
9518{
9519        struct nft_rule *rule, *nr;
9520
9521        if (WARN_ON(!nft_is_base_chain(ctx->chain)))
9522                return 0;
9523
9524        nf_tables_unregister_hook(ctx->net, ctx->chain->table, ctx->chain);
9525        list_for_each_entry_safe(rule, nr, &ctx->chain->rules, list) {
9526                list_del(&rule->list);
9527                ctx->chain->use--;
9528                nf_tables_rule_release(ctx, rule);
9529        }
9530        nft_chain_del(ctx->chain);
9531        ctx->table->use--;
9532        nf_tables_chain_destroy(ctx);
9533
9534        return 0;
9535}
9536EXPORT_SYMBOL_GPL(__nft_release_basechain);
9537
9538static void __nft_release_hook(struct net *net, struct nft_table *table)
9539{
9540        struct nft_chain *chain;
9541
9542        list_for_each_entry(chain, &table->chains, list)
9543                nf_tables_unregister_hook(net, table, chain);
9544}
9545
9546static void __nft_release_hooks(struct net *net)
9547{
9548        struct nftables_pernet *nft_net = nft_pernet(net);
9549        struct nft_table *table;
9550
9551        list_for_each_entry(table, &nft_net->tables, list) {
9552                if (nft_table_has_owner(table))
9553                        continue;
9554
9555                __nft_release_hook(net, table);
9556        }
9557}
9558
9559static void __nft_release_table(struct net *net, struct nft_table *table)
9560{
9561        struct nft_flowtable *flowtable, *nf;
9562        struct nft_chain *chain, *nc;
9563        struct nft_object *obj, *ne;
9564        struct nft_rule *rule, *nr;
9565        struct nft_set *set, *ns;
9566        struct nft_ctx ctx = {
9567                .net    = net,
9568                .family = NFPROTO_NETDEV,
9569        };
9570
9571        ctx.family = table->family;
9572        ctx.table = table;
9573        list_for_each_entry(chain, &table->chains, list) {
9574                ctx.chain = chain;
9575                list_for_each_entry_safe(rule, nr, &chain->rules, list) {
9576                        list_del(&rule->list);
9577                        chain->use--;
9578                        nf_tables_rule_release(&ctx, rule);
9579                }
9580        }
9581        list_for_each_entry_safe(flowtable, nf, &table->flowtables, list) {
9582                list_del(&flowtable->list);
9583                table->use--;
9584                nf_tables_flowtable_destroy(flowtable);
9585        }
9586        list_for_each_entry_safe(set, ns, &table->sets, list) {
9587                list_del(&set->list);
9588                table->use--;
9589                nft_set_destroy(&ctx, set);
9590        }
9591        list_for_each_entry_safe(obj, ne, &table->objects, list) {
9592                nft_obj_del(obj);
9593                table->use--;
9594                nft_obj_destroy(&ctx, obj);
9595        }
9596        list_for_each_entry_safe(chain, nc, &table->chains, list) {
9597                ctx.chain = chain;
9598                nft_chain_del(chain);
9599                table->use--;
9600                nf_tables_chain_destroy(&ctx);
9601        }
9602        list_del(&table->list);
9603        nf_tables_table_destroy(&ctx);
9604}
9605
9606static void __nft_release_tables(struct net *net)
9607{
9608        struct nftables_pernet *nft_net = nft_pernet(net);
9609        struct nft_table *table, *nt;
9610
9611        list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
9612                if (nft_table_has_owner(table))
9613                        continue;
9614
9615                __nft_release_table(net, table);
9616        }
9617}
9618
9619static int nft_rcv_nl_event(struct notifier_block *this, unsigned long event,
9620                            void *ptr)
9621{
9622        struct nftables_pernet *nft_net;
9623        struct netlink_notify *n = ptr;
9624        struct nft_table *table, *nt;
9625        struct net *net = n->net;
9626        bool release = false;
9627
9628        if (event != NETLINK_URELEASE || n->protocol != NETLINK_NETFILTER)
9629                return NOTIFY_DONE;
9630
9631        nft_net = nft_pernet(net);
9632        mutex_lock(&nft_net->commit_mutex);
9633        list_for_each_entry(table, &nft_net->tables, list) {
9634                if (nft_table_has_owner(table) &&
9635                    n->portid == table->nlpid) {
9636                        __nft_release_hook(net, table);
9637                        release = true;
9638                }
9639        }
9640        if (release) {
9641                synchronize_rcu();
9642                list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
9643                        if (nft_table_has_owner(table) &&
9644                            n->portid == table->nlpid)
9645                                __nft_release_table(net, table);
9646                }
9647        }
9648        mutex_unlock(&nft_net->commit_mutex);
9649
9650        return NOTIFY_DONE;
9651}
9652
9653static struct notifier_block nft_nl_notifier = {
9654        .notifier_call  = nft_rcv_nl_event,
9655};
9656
9657static int __net_init nf_tables_init_net(struct net *net)
9658{
9659        struct nftables_pernet *nft_net = nft_pernet(net);
9660
9661        INIT_LIST_HEAD(&nft_net->tables);
9662        INIT_LIST_HEAD(&nft_net->commit_list);
9663        INIT_LIST_HEAD(&nft_net->module_list);
9664        INIT_LIST_HEAD(&nft_net->notify_list);
9665        mutex_init(&nft_net->commit_mutex);
9666        nft_net->base_seq = 1;
9667        nft_net->validate_state = NFT_VALIDATE_SKIP;
9668
9669        return 0;
9670}
9671
9672static void __net_exit nf_tables_pre_exit_net(struct net *net)
9673{
9674        __nft_release_hooks(net);
9675}
9676
9677static void __net_exit nf_tables_exit_net(struct net *net)
9678{
9679        struct nftables_pernet *nft_net = nft_pernet(net);
9680
9681        mutex_lock(&nft_net->commit_mutex);
9682        if (!list_empty(&nft_net->commit_list))
9683                __nf_tables_abort(net, NFNL_ABORT_NONE);
9684        __nft_release_tables(net);
9685        mutex_unlock(&nft_net->commit_mutex);
9686        WARN_ON_ONCE(!list_empty(&nft_net->tables));
9687        WARN_ON_ONCE(!list_empty(&nft_net->module_list));
9688        WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
9689}
9690
9691static struct pernet_operations nf_tables_net_ops = {
9692        .init           = nf_tables_init_net,
9693        .pre_exit       = nf_tables_pre_exit_net,
9694        .exit           = nf_tables_exit_net,
9695        .id             = &nf_tables_net_id,
9696        .size           = sizeof(struct nftables_pernet),
9697};
9698
9699static int __init nf_tables_module_init(void)
9700{
9701        int err;
9702
9703        err = register_pernet_subsys(&nf_tables_net_ops);
9704        if (err < 0)
9705                return err;
9706
9707        err = nft_chain_filter_init();
9708        if (err < 0)
9709                goto err_chain_filter;
9710
9711        err = nf_tables_core_module_init();
9712        if (err < 0)
9713                goto err_core_module;
9714
9715        err = register_netdevice_notifier(&nf_tables_flowtable_notifier);
9716        if (err < 0)
9717                goto err_netdev_notifier;
9718
9719        err = rhltable_init(&nft_objname_ht, &nft_objname_ht_params);
9720        if (err < 0)
9721                goto err_rht_objname;
9722
9723        err = nft_offload_init();
9724        if (err < 0)
9725                goto err_offload;
9726
9727        err = netlink_register_notifier(&nft_nl_notifier);
9728        if (err < 0)
9729                goto err_netlink_notifier;
9730
9731        /* must be last */
9732        err = nfnetlink_subsys_register(&nf_tables_subsys);
9733        if (err < 0)
9734                goto err_nfnl_subsys;
9735
9736        nft_chain_route_init();
9737
9738        return err;
9739
9740err_nfnl_subsys:
9741        netlink_unregister_notifier(&nft_nl_notifier);
9742err_netlink_notifier:
9743        nft_offload_exit();
9744err_offload:
9745        rhltable_destroy(&nft_objname_ht);
9746err_rht_objname:
9747        unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
9748err_netdev_notifier:
9749        nf_tables_core_module_exit();
9750err_core_module:
9751        nft_chain_filter_fini();
9752err_chain_filter:
9753        unregister_pernet_subsys(&nf_tables_net_ops);
9754        return err;
9755}
9756
9757static void __exit nf_tables_module_exit(void)
9758{
9759        nfnetlink_subsys_unregister(&nf_tables_subsys);
9760        netlink_unregister_notifier(&nft_nl_notifier);
9761        nft_offload_exit();
9762        unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
9763        nft_chain_filter_fini();
9764        nft_chain_route_fini();
9765        unregister_pernet_subsys(&nf_tables_net_ops);
9766        cancel_work_sync(&trans_destroy_work);
9767        rcu_barrier();
9768        rhltable_destroy(&nft_objname_ht);
9769        nf_tables_core_module_exit();
9770}
9771
9772module_init(nf_tables_module_init);
9773module_exit(nf_tables_module_exit);
9774
9775MODULE_LICENSE("GPL");
9776MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
9777MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);
9778